1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/convertnode.hpp"
30 #include "opto/memnode.hpp"
31 #include "opto/mulnode.hpp"
32 #include "opto/phaseX.hpp"
33 #include "opto/subnode.hpp"
34 #include "utilities/powerOfTwo.hpp"
35
36 // Portions of code courtesy of Clifford Click
37
38
39 //=============================================================================
40 //------------------------------hash-------------------------------------------
41 // Hash function over MulNodes. Needs to be commutative; i.e., I swap
42 // (commute) inputs to MulNodes willy-nilly so the hash function must return
43 // the same value in the presence of edge swapping.
44 uint MulNode::hash() const {
45 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode();
46 }
47
48 //------------------------------Identity---------------------------------------
49 // Multiplying a one preserves the other argument
50 Node* MulNode::Identity(PhaseGVN* phase) {
51 const Type *one = mul_id(); // The multiplicative identity
52 if( phase->type( in(1) )->higher_equal( one ) ) return in(2);
53 if( phase->type( in(2) )->higher_equal( one ) ) return in(1);
54
55 return this;
56 }
57
58 //------------------------------Ideal------------------------------------------
59 // We also canonicalize the Node, moving constants to the right input,
60 // and flatten expressions (so that 1+x+2 becomes x+3).
61 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) {
62 Node* in1 = in(1);
63 Node* in2 = in(2);
64 Node* progress = nullptr; // Progress flag
65
66 // This code is used by And nodes too, but some conversions are
67 // only valid for the actual Mul nodes.
68 uint op = Opcode();
69 bool real_mul = (op == Op_MulI) || (op == Op_MulL) ||
70 (op == Op_MulF) || (op == Op_MulD);
71
72 // Convert "(-a)*(-b)" into "a*b".
73 if (real_mul && in1->is_Sub() && in2->is_Sub()) {
74 if (phase->type(in1->in(1))->is_zero_type() &&
75 phase->type(in2->in(1))->is_zero_type()) {
76 set_req_X(1, in1->in(2), phase);
77 set_req_X(2, in2->in(2), phase);
78 in1 = in(1);
79 in2 = in(2);
80 progress = this;
81 }
82 }
83
84 // convert "max(a,b) * min(a,b)" into "a*b".
85 if ((in(1)->Opcode() == max_opcode() && in(2)->Opcode() == min_opcode())
86 || (in(1)->Opcode() == min_opcode() && in(2)->Opcode() == max_opcode())) {
87 Node *in11 = in(1)->in(1);
88 Node *in12 = in(1)->in(2);
89
90 Node *in21 = in(2)->in(1);
91 Node *in22 = in(2)->in(2);
92
93 if ((in11 == in21 && in12 == in22) ||
94 (in11 == in22 && in12 == in21)) {
95 set_req_X(1, in11, phase);
96 set_req_X(2, in12, phase);
97 in1 = in(1);
98 in2 = in(2);
99 progress = this;
100 }
101 }
102
103 const Type* t1 = phase->type(in1);
104 const Type* t2 = phase->type(in2);
105
106 // We are OK if right is a constant, or right is a load and
107 // left is a non-constant.
108 if( !(t2->singleton() ||
109 (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
110 if( t1->singleton() || // Left input is a constant?
111 // Otherwise, sort inputs (commutativity) to help value numbering.
112 (in(1)->_idx > in(2)->_idx) ) {
113 swap_edges(1, 2);
114 const Type *t = t1;
115 t1 = t2;
116 t2 = t;
117 progress = this; // Made progress
118 }
119 }
120
121 // If the right input is a constant, and the left input is a product of a
122 // constant, flatten the expression tree.
123 if( t2->singleton() && // Right input is a constant?
124 op != Op_MulF && // Float & double cannot reassociate
125 op != Op_MulD ) {
126 if( t2 == Type::TOP ) return nullptr;
127 Node *mul1 = in(1);
128 #ifdef ASSERT
129 // Check for dead loop
130 int op1 = mul1->Opcode();
131 if ((mul1 == this) || (in(2) == this) ||
132 ((op1 == mul_opcode() || op1 == add_opcode()) &&
133 ((mul1->in(1) == this) || (mul1->in(2) == this) ||
134 (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) {
135 assert(false, "dead loop in MulNode::Ideal");
136 }
137 #endif
138
139 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply?
140 // Mul of a constant?
141 const Type *t12 = phase->type( mul1->in(2) );
142 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
143 // Compute new constant; check for overflow
144 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
145 if( tcon01->singleton() ) {
146 // The Mul of the flattened expression
147 set_req_X(1, mul1->in(1), phase);
148 set_req_X(2, phase->makecon(tcon01), phase);
149 t2 = tcon01;
150 progress = this; // Made progress
151 }
152 }
153 }
154 // If the right input is a constant, and the left input is an add of a
155 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
156 const Node *add1 = in(1);
157 if( add1->Opcode() == add_opcode() ) { // Left input is an add?
158 // Add of a constant?
159 const Type *t12 = phase->type( add1->in(2) );
160 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
161 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
162 // Compute new constant; check for overflow
163 const Type *tcon01 = mul_ring(t2,t12);
164 if( tcon01->singleton() ) {
165
166 // Convert (X+con1)*con0 into X*con0
167 Node *mul = clone(); // mul = ()*con0
168 mul->set_req(1,add1->in(1)); // mul = X*con0
169 mul = phase->transform(mul);
170
171 Node *add2 = add1->clone();
172 add2->set_req(1, mul); // X*con0 + con0*con1
173 add2->set_req(2, phase->makecon(tcon01) );
174 progress = add2;
175 }
176 }
177 } // End of is left input an add
178 } // End of is right input a Mul
179
180 return progress;
181 }
182
183 //------------------------------Value-----------------------------------------
184 const Type* MulNode::Value(PhaseGVN* phase) const {
185 const Type *t1 = phase->type( in(1) );
186 const Type *t2 = phase->type( in(2) );
187 // Either input is TOP ==> the result is TOP
188 if( t1 == Type::TOP ) return Type::TOP;
189 if( t2 == Type::TOP ) return Type::TOP;
190
191 // Either input is ZERO ==> the result is ZERO.
192 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
193 int op = Opcode();
194 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
195 const Type *zero = add_id(); // The multiplicative zero
196 if( t1->higher_equal( zero ) ) return zero;
197 if( t2->higher_equal( zero ) ) return zero;
198 }
199
200 // Code pattern on return from a call that returns an __Value. Can
201 // be optimized away if the return value turns out to be an oop.
202 if (op == Op_AndX &&
203 in(1) != nullptr &&
204 in(1)->Opcode() == Op_CastP2X &&
205 in(1)->in(1) != nullptr &&
206 phase->type(in(1)->in(1))->isa_oopptr() &&
207 t2->isa_intptr_t()->_lo >= 0 &&
208 t2->isa_intptr_t()->_hi <= MinObjAlignmentInBytesMask) {
209 return add_id();
210 }
211
212 // Either input is BOTTOM ==> the result is the local BOTTOM
213 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
214 return bottom_type();
215
216 #if defined(IA32)
217 // Can't trust native compilers to properly fold strict double
218 // multiplication with round-to-zero on this platform.
219 if (op == Op_MulD) {
220 return TypeD::DOUBLE;
221 }
222 #endif
223
224 return mul_ring(t1,t2); // Local flavor of type multiplication
225 }
226
227 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
228 switch (bt) {
229 case T_INT:
230 return new MulINode(in1, in2);
231 case T_LONG:
232 return new MulLNode(in1, in2);
233 default:
234 fatal("Not implemented for %s", type2name(bt));
235 }
236 return nullptr;
237 }
238
239
240 //=============================================================================
241 //------------------------------Ideal------------------------------------------
242 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
243 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
244 const jint con = in(2)->find_int_con(0);
245 if (con == 0) {
246 // If in(2) is not a constant, call Ideal() of the parent class to
247 // try to move constant to the right side.
248 return MulNode::Ideal(phase, can_reshape);
249 }
250
251 // Now we have a constant Node on the right and the constant in con.
252 if (con == 1) {
253 // By one is handled by Identity call
254 return nullptr;
255 }
256
257 // Check for negative constant; if so negate the final result
258 bool sign_flip = false;
259
260 unsigned int abs_con = uabs(con);
261 if (abs_con != (unsigned int)con) {
262 sign_flip = true;
263 }
264
265 // Get low bit; check for being the only bit
266 Node *res = nullptr;
267 unsigned int bit1 = submultiple_power_of_2(abs_con);
268 if (bit1 == abs_con) { // Found a power of 2?
269 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
270 } else {
271 // Check for constant with 2 bits set
272 unsigned int bit2 = abs_con - bit1;
273 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
274 if (bit2 + bit1 == abs_con) { // Found all bits in con?
275 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
276 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
277 res = new AddINode(n2, n1);
278 } else if (is_power_of_2(abs_con + 1)) {
279 // Sleezy: power-of-2 - 1. Next time be generic.
280 unsigned int temp = abs_con + 1;
281 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
282 res = new SubINode(n1, in(1));
283 } else {
284 return MulNode::Ideal(phase, can_reshape);
285 }
286 }
287
288 if (sign_flip) { // Need to negate result?
289 res = phase->transform(res);// Transform, before making the zero con
290 res = new SubINode(phase->intcon(0),res);
291 }
292
293 return res; // Return final result
294 }
295
296 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
297 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
298 //
299 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
300 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
301 // - no overflow (i.e. product fits into x bits)
302 // - overflow (i.e. product does not fit into x bits)
303 //
304 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
305 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
306 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
307 // The cross product consists of the following four multiplications with 2x-bit results:
308 // (1) left_lo * right_lo
309 // (2) left_lo * right_hi
310 // (3) left_hi * right_lo
311 // (4) left_hi * right_hi
312 //
313 // Let's define the following two functions:
314 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
315 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
316 //
317 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
318 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
319 // Ux(p) = 0
320 // Lx(p) = p
321 //
322 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
323 // with x bits:
324 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
325 //
326 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
327 // (i.e. all values in the x-bit space could be possible):
328 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
329 //
330 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
331 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
332 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
333 // the lower x bits the same number of times:
334 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
335 //
336 // If all upper x bits are equal, we can conclude:
337 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
338 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
339 //
340 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
341 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
342 //
343 //
344 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
345 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
346 // upper x bits of the 2x-bit result are all set to ones which is minus one in two's complement. If there is an underflow,
347 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
348 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
349 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
350 // that there is no underflow.
351 //
352 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
353 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
354 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
355 // can compare the upper x bits in a unified way with N() being the normalization function:
356 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
357 template<typename NativeType>
358 class IntegerTypeMultiplication {
359
360 NativeType _lo_left;
361 NativeType _lo_right;
362 NativeType _hi_left;
363 NativeType _hi_right;
364 short _widen_left;
365 short _widen_right;
366
367 static const Type* overflow_type();
368 static NativeType multiply_high(NativeType x, NativeType y);
369 const Type* create_type(NativeType lo, NativeType hi) const;
370
371 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
372 return normalize_overflow_value(x, y, multiply_high(x, y));
373 }
374
375 bool cross_product_not_same_overflow_value() const {
376 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
377 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
378 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
379 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
380 return lo_lo_high_product != lo_hi_high_product ||
381 lo_hi_high_product != hi_lo_high_product ||
382 hi_lo_high_product != hi_hi_high_product;
383 }
384
385 bool does_product_overflow(NativeType x, NativeType y) const {
386 return multiply_high_signed_overflow_value(x, y) != 0;
387 }
388
389 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
390 return java_multiply(x, y) < 0 ? result + 1 : result;
391 }
392
393 public:
394 template<class IntegerType>
395 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
396 : _lo_left(left->_lo), _lo_right(right->_lo),
397 _hi_left(left->_hi), _hi_right(right->_hi),
398 _widen_left(left->_widen), _widen_right(right->_widen) {}
399
400 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
401 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
402 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
403 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
404 const Type* compute() const {
405 if (cross_product_not_same_overflow_value()) {
406 return overflow_type();
407 }
408
409 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
410 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
411 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
412 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
413 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
414 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
415 return create_type(min, max);
416 }
417
418 bool does_overflow() const {
419 return does_product_overflow(_lo_left, _lo_right) ||
420 does_product_overflow(_lo_left, _hi_right) ||
421 does_product_overflow(_hi_left, _lo_right) ||
422 does_product_overflow(_hi_left, _hi_right);
423 }
424 };
425
426 template <>
427 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
428 return TypeInt::INT;
429 }
430
431 template <>
432 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
433 const jlong x_64 = x;
434 const jlong y_64 = y;
435 const jlong product = x_64 * y_64;
436 return (jint)((uint64_t)product >> 32u);
437 }
438
439 template <>
440 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
441 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
442 }
443
444 template <>
445 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
446 return TypeLong::LONG;
447 }
448
449 template <>
450 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
451 return multiply_high_signed(x, y);
452 }
453
454 template <>
455 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
456 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
457 }
458
459 // Compute the product type of two integer ranges into this node.
460 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
461 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
462 return integer_multiplication.compute();
463 }
464
465 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
466 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
467 return integer_multiplication.does_overflow();
468 }
469
470 // Compute the product type of two long ranges into this node.
471 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
472 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
473 return integer_multiplication.compute();
474 }
475
476 //=============================================================================
477 //------------------------------Ideal------------------------------------------
478 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
479 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
480 const jlong con = in(2)->find_long_con(0);
481 if (con == 0) {
482 // If in(2) is not a constant, call Ideal() of the parent class to
483 // try to move constant to the right side.
484 return MulNode::Ideal(phase, can_reshape);
485 }
486
487 // Now we have a constant Node on the right and the constant in con.
488 if (con == 1) {
489 // By one is handled by Identity call
490 return nullptr;
491 }
492
493 // Check for negative constant; if so negate the final result
494 bool sign_flip = false;
495 julong abs_con = uabs(con);
496 if (abs_con != (julong)con) {
497 sign_flip = true;
498 }
499
500 // Get low bit; check for being the only bit
501 Node *res = nullptr;
502 julong bit1 = submultiple_power_of_2(abs_con);
503 if (bit1 == abs_con) { // Found a power of 2?
504 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
505 } else {
506
507 // Check for constant with 2 bits set
508 julong bit2 = abs_con-bit1;
509 bit2 = bit2 & (0-bit2); // Extract 2nd bit
510 if (bit2 + bit1 == abs_con) { // Found all bits in con?
511 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
512 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
513 res = new AddLNode(n2, n1);
514
515 } else if (is_power_of_2(abs_con+1)) {
516 // Sleezy: power-of-2 -1. Next time be generic.
517 julong temp = abs_con + 1;
518 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
519 res = new SubLNode(n1, in(1));
520 } else {
521 return MulNode::Ideal(phase, can_reshape);
522 }
523 }
524
525 if (sign_flip) { // Need to negate result?
526 res = phase->transform(res);// Transform, before making the zero con
527 res = new SubLNode(phase->longcon(0),res);
528 }
529
530 return res; // Return final result
531 }
532
533 //=============================================================================
534 //------------------------------mul_ring---------------------------------------
535 // Compute the product type of two double ranges into this node.
536 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
537 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
538 return TypeF::make( t0->getf() * t1->getf() );
539 }
540
541 //------------------------------Ideal---------------------------------------
542 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
543 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
544 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
545
546 // x * 2 -> x + x
547 if (t2 != nullptr && t2->getf() == 2) {
548 Node* base = in(1);
549 return new AddFNode(base, base);
550 }
551
552 return MulNode::Ideal(phase, can_reshape);
553 }
554
555 //=============================================================================
556 //------------------------------mul_ring---------------------------------------
557 // Compute the product type of two double ranges into this node.
558 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
559 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
560 // We must be multiplying 2 double constants.
561 return TypeD::make( t0->getd() * t1->getd() );
562 }
563
564 //------------------------------Ideal---------------------------------------
565 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
566 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
567 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
568
569 // x * 2 -> x + x
570 if (t2 != nullptr && t2->getd() == 2) {
571 Node* base = in(1);
572 return new AddDNode(base, base);
573 }
574
575 return MulNode::Ideal(phase, can_reshape);
576 }
577
578 //=============================================================================
579 //------------------------------Value------------------------------------------
580 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
581 const Type *t1 = phase->type( in(1) );
582 const Type *t2 = phase->type( in(2) );
583 const Type *bot = bottom_type();
584 return MulHiValue(t1, t2, bot);
585 }
586
587 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
588 const Type *t1 = phase->type( in(1) );
589 const Type *t2 = phase->type( in(2) );
590 const Type *bot = bottom_type();
591 return MulHiValue(t1, t2, bot);
592 }
593
594 // A common routine used by UMulHiLNode and MulHiLNode
595 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
596 // Either input is TOP ==> the result is TOP
597 if( t1 == Type::TOP ) return Type::TOP;
598 if( t2 == Type::TOP ) return Type::TOP;
599
600 // Either input is BOTTOM ==> the result is the local BOTTOM
601 if( (t1 == bot) || (t2 == bot) ||
602 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
603 return bot;
604
605 // It is not worth trying to constant fold this stuff!
606 return TypeLong::LONG;
607 }
608
609 template<typename IntegerType>
610 static const IntegerType* and_value(const IntegerType* r0, const IntegerType* r1) {
611 typedef typename IntegerType::NativeType NativeType;
612 static_assert(std::is_signed<NativeType>::value, "Native type of IntegerType must be signed!");
613
614 int widen = MAX2(r0->_widen, r1->_widen);
615
616 // If both types are constants, we can calculate a constant result.
617 if (r0->is_con() && r1->is_con()) {
618 return IntegerType::make(r0->get_con() & r1->get_con());
619 }
620
621 // If both ranges are positive, the result will range from 0 up to the hi value of the smaller range. The minimum
622 // of the two constrains the upper bound because any higher value in the other range will see all zeroes, so it will be masked out.
623 if (r0->_lo >= 0 && r1->_lo >= 0) {
624 return IntegerType::make(0, MIN2(r0->_hi, r1->_hi), widen);
625 }
626
627 // If only one range is positive, the result will range from 0 up to that range's maximum value.
628 // For the operation 'x & C' where C is a positive constant, the result will be in the range [0..C]. With that observation,
629 // we can say that for any integer c such that 0 <= c <= C will also be in the range [0..C]. Therefore, 'x & [c..C]'
630 // where c >= 0 will be in the range [0..C].
631 if (r0->_lo >= 0) {
632 return IntegerType::make(0, r0->_hi, widen);
633 }
634
635 if (r1->_lo >= 0) {
636 return IntegerType::make(0, r1->_hi, widen);
637 }
638
639 // At this point, all positive ranges will have already been handled, so the only remaining cases will be negative ranges
640 // and constants.
641
642 assert(r0->_lo < 0 && r1->_lo < 0, "positive ranges should already be handled!");
643
644 // As two's complement means that both numbers will start with leading 1s, the lower bound of both ranges will contain
645 // the common leading 1s of both minimum values. In order to count them with count_leading_zeros, the bits are inverted.
646 NativeType sel_val = ~MIN2(r0->_lo, r1->_lo);
647
648 NativeType min;
649 if (sel_val == 0) {
650 // Since count_leading_zeros is undefined at 0, we short-circuit the condition where both ranges have a minimum of -1.
651 min = -1;
652 } else {
653 // To get the number of bits to shift, we count the leading 0-bits and then subtract one, as the sign bit is already set.
654 int shift_bits = count_leading_zeros(sel_val) - 1;
655 min = std::numeric_limits<NativeType>::min() >> shift_bits;
656 }
657
658 NativeType max;
659 if (r0->_hi < 0 && r1->_hi < 0) {
660 // If both ranges are negative, then the same optimization as both positive ranges will apply, and the smaller hi
661 // value will mask off any bits set by higher values.
662 max = MIN2(r0->_hi, r1->_hi);
663 } else {
664 // In the case of ranges that cross zero, negative values can cause the higher order bits to be set, so the maximum
665 // positive value can be as high as the larger hi value.
666 max = MAX2(r0->_hi, r1->_hi);
667 }
668
669 return IntegerType::make(min, max, widen);
670 }
671
672 //=============================================================================
673 //------------------------------mul_ring---------------------------------------
674 // Supplied function returns the product of the inputs IN THE CURRENT RING.
675 // For the logical operations the ring's MUL is really a logical AND function.
676 // This also type-checks the inputs for sanity. Guaranteed never to
677 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
678 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
679 const TypeInt* r0 = t0->is_int();
680 const TypeInt* r1 = t1->is_int();
681
682 return and_value<TypeInt>(r0, r1);
683 }
684
685 const Type* AndINode::Value(PhaseGVN* phase) const {
686 // patterns similar to (v << 2) & 3
687 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_INT, true)) {
688 return TypeInt::ZERO;
689 }
690
691 return MulNode::Value(phase);
692 }
693
694 //------------------------------Identity---------------------------------------
695 // Masking off the high bits of an unsigned load is not required
696 Node* AndINode::Identity(PhaseGVN* phase) {
697
698 // x & x => x
699 if (in(1) == in(2)) {
700 return in(1);
701 }
702
703 Node* in1 = in(1);
704 uint op = in1->Opcode();
705 const TypeInt* t2 = phase->type(in(2))->isa_int();
706 if (t2 && t2->is_con()) {
707 int con = t2->get_con();
708 // Masking off high bits which are always zero is useless.
709 const TypeInt* t1 = phase->type(in(1))->isa_int();
710 if (t1 != nullptr && t1->_lo >= 0) {
711 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
712 if ((t1_support & con) == t1_support)
713 return in1;
714 }
715 // Masking off the high bits of a unsigned-shift-right is not
716 // needed either.
717 if (op == Op_URShiftI) {
718 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
719 if (t12 && t12->is_con()) { // Shift is by a constant
720 int shift = t12->get_con();
721 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
722 int mask = max_juint >> shift;
723 if ((mask & con) == mask) // If AND is useless, skip it
724 return in1;
725 }
726 }
727 }
728 return MulNode::Identity(phase);
729 }
730
731 //------------------------------Ideal------------------------------------------
732 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
733 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3
734 Node* progress = AndIL_add_shift_and_mask(phase, T_INT);
735 if (progress != nullptr) {
736 return progress;
737 }
738
739 // Convert "(~a) & (~b)" into "~(a | b)"
740 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
741 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
742 Node* tn = phase->transform(or_a_b);
743 return AddNode::make_not(phase, tn, T_INT);
744 }
745
746 // Special case constant AND mask
747 const TypeInt *t2 = phase->type( in(2) )->isa_int();
748 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
749 const int mask = t2->get_con();
750 Node *load = in(1);
751 uint lop = load->Opcode();
752
753 // Masking bits off of a Character? Hi bits are already zero.
754 if( lop == Op_LoadUS &&
755 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
756 return new AndINode(load,phase->intcon(mask&0xFFFF));
757
758 // Masking bits off of a Short? Loading a Character does some masking
759 if (can_reshape &&
760 load->outcnt() == 1 && load->unique_out() == this) {
761 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
762 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
763 ldus = phase->transform(ldus);
764 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
765 }
766
767 // Masking sign bits off of a Byte? Do an unsigned byte load plus
768 // an and.
769 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
770 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
771 ldub = phase->transform(ldub);
772 return new AndINode(ldub, phase->intcon(mask));
773 }
774 }
775
776 // Masking off sign bits? Dont make them!
777 if( lop == Op_RShiftI ) {
778 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
779 if( t12 && t12->is_con() ) { // Shift is by a constant
780 int shift = t12->get_con();
781 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
782 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
783 // If the AND'ing of the 2 masks has no bits, then only original shifted
784 // bits survive. NO sign-extension bits survive the maskings.
785 if( (sign_bits_mask & mask) == 0 ) {
786 // Use zero-fill shift instead
787 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
788 return new AndINode( zshift, in(2) );
789 }
790 }
791 }
792
793 // Check for 'negate/and-1', a pattern emitted when someone asks for
794 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
795 // plus 1) and the mask is of the low order bit. Skip the negate.
796 if( lop == Op_SubI && mask == 1 && load->in(1) &&
797 phase->type(load->in(1)) == TypeInt::ZERO )
798 return new AndINode( load->in(2), in(2) );
799
800 return MulNode::Ideal(phase, can_reshape);
801 }
802
803 //=============================================================================
804 //------------------------------mul_ring---------------------------------------
805 // Supplied function returns the product of the inputs IN THE CURRENT RING.
806 // For the logical operations the ring's MUL is really a logical AND function.
807 // This also type-checks the inputs for sanity. Guaranteed never to
808 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
809 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
810 const TypeLong* r0 = t0->is_long();
811 const TypeLong* r1 = t1->is_long();
812
813 return and_value<TypeLong>(r0, r1);
814 }
815
816 const Type* AndLNode::Value(PhaseGVN* phase) const {
817 // patterns similar to (v << 2) & 3
818 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_LONG, true)) {
819 return TypeLong::ZERO;
820 }
821
822 return MulNode::Value(phase);
823 }
824
825 //------------------------------Identity---------------------------------------
826 // Masking off the high bits of an unsigned load is not required
827 Node* AndLNode::Identity(PhaseGVN* phase) {
828
829 // x & x => x
830 if (in(1) == in(2)) {
831 return in(1);
832 }
833
834 Node *usr = in(1);
835 const TypeLong *t2 = phase->type( in(2) )->isa_long();
836 if( t2 && t2->is_con() ) {
837 jlong con = t2->get_con();
838 // Masking off high bits which are always zero is useless.
839 const TypeLong* t1 = phase->type( in(1) )->isa_long();
840 if (t1 != nullptr && t1->_lo >= 0) {
841 int bit_count = log2i_graceful(t1->_hi) + 1;
842 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
843 if ((t1_support & con) == t1_support)
844 return usr;
845 }
846 uint lop = usr->Opcode();
847 // Masking off the high bits of a unsigned-shift-right is not
848 // needed either.
849 if( lop == Op_URShiftL ) {
850 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
851 if( t12 && t12->is_con() ) { // Shift is by a constant
852 int shift = t12->get_con();
853 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
854 jlong mask = max_julong >> shift;
855 if( (mask&con) == mask ) // If AND is useless, skip it
856 return usr;
857 }
858 }
859 }
860 return MulNode::Identity(phase);
861 }
862
863 //------------------------------Ideal------------------------------------------
864 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
865 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3
866 Node* progress = AndIL_add_shift_and_mask(phase, T_LONG);
867 if (progress != nullptr) {
868 return progress;
869 }
870
871 // Convert "(~a) & (~b)" into "~(a | b)"
872 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
873 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
874 Node* tn = phase->transform(or_a_b);
875 return AddNode::make_not(phase, tn, T_LONG);
876 }
877
878 // Special case constant AND mask
879 const TypeLong *t2 = phase->type( in(2) )->isa_long();
880 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
881 const jlong mask = t2->get_con();
882
883 Node* in1 = in(1);
884 int op = in1->Opcode();
885
886 // Are we masking a long that was converted from an int with a mask
887 // that fits in 32-bits? Commute them and use an AndINode. Don't
888 // convert masks which would cause a sign extension of the integer
889 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
890 // would be optimized away later in Identity.
891 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
892 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
893 andi = phase->transform(andi);
894 return new ConvI2LNode(andi);
895 }
896
897 // Masking off sign bits? Dont make them!
898 if (op == Op_RShiftL) {
899 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
900 if( t12 && t12->is_con() ) { // Shift is by a constant
901 int shift = t12->get_con();
902 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
903 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
904 // If the AND'ing of the 2 masks has no bits, then only original shifted
905 // bits survive. NO sign-extension bits survive the maskings.
906 if( (sign_bits_mask & mask) == 0 ) {
907 // Use zero-fill shift instead
908 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
909 return new AndLNode(zshift, in(2));
910 }
911 }
912 }
913
914 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
915 Node* load1 = in(1);
916 Node* load2 = nullptr;
917 if (load1->is_Phi() && phase->type(load1)->isa_long()) {
918 load1 = in(1)->in(1);
919 load2 = in(1)->in(2);
920 }
921 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
922 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
923 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
924 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
925 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
926 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
927 if (mask == markWord::inline_type_pattern) {
928 if (adr_t1->is_inlinetypeptr()) {
929 set_req_X(1, in(2), phase);
930 return this;
931 } else if (!adr_t1->can_be_inline_type()) {
932 set_req_X(1, phase->longcon(0), phase);
933 return this;
934 }
935 } else if (mask == markWord::null_free_array_bit_in_place) {
936 if (adr_t1->is_null_free()) {
937 set_req_X(1, in(2), phase);
938 return this;
939 } else if (adr_t1->is_not_null_free()) {
940 set_req_X(1, phase->longcon(0), phase);
941 return this;
942 }
943 } else if (mask == markWord::flat_array_bit_in_place) {
944 if (adr_t1->is_flat()) {
945 set_req_X(1, in(2), phase);
946 return this;
947 } else if (adr_t1->is_not_flat()) {
948 set_req_X(1, phase->longcon(0), phase);
949 return this;
950 }
951 }
952 }
953 }
954
955 return MulNode::Ideal(phase, can_reshape);
956 }
957
958 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
959 switch (bt) {
960 case T_INT:
961 return new LShiftINode(in1, in2);
962 case T_LONG:
963 return new LShiftLNode(in1, in2);
964 default:
965 fatal("Not implemented for %s", type2name(bt));
966 }
967 return nullptr;
968 }
969
970 //=============================================================================
971
972 static bool const_shift_count(PhaseGVN* phase, Node* shiftNode, int* count) {
973 const TypeInt* tcount = phase->type(shiftNode->in(2))->isa_int();
974 if (tcount != nullptr && tcount->is_con()) {
975 *count = tcount->get_con();
976 return true;
977 }
978 return false;
979 }
980
981 static int maskShiftAmount(PhaseGVN* phase, Node* shiftNode, int nBits) {
982 int count = 0;
983 if (const_shift_count(phase, shiftNode, &count)) {
984 int maskedShift = count & (nBits - 1);
985 if (maskedShift == 0) {
986 // Let Identity() handle 0 shift count.
987 return 0;
988 }
989
990 if (count != maskedShift) {
991 shiftNode->set_req(2, phase->intcon(maskedShift)); // Replace shift count with masked value.
992 PhaseIterGVN* igvn = phase->is_IterGVN();
993 if (igvn) {
994 igvn->rehash_node_delayed(shiftNode);
995 }
996 }
997 return maskedShift;
998 }
999 return 0;
1000 }
1001
1002 //------------------------------Identity---------------------------------------
1003 Node* LShiftINode::Identity(PhaseGVN* phase) {
1004 int count = 0;
1005 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1006 // Shift by a multiple of 32 does nothing
1007 return in(1);
1008 }
1009 return this;
1010 }
1011
1012 //------------------------------Ideal------------------------------------------
1013 // If the right input is a constant, and the left input is an add of a
1014 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1015 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1016 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
1017 if (con == 0) {
1018 return nullptr;
1019 }
1020
1021 // Left input is an add?
1022 Node *add1 = in(1);
1023 int add1_op = add1->Opcode();
1024 if( add1_op == Op_AddI ) { // Left input is an add?
1025 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" );
1026
1027 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1028 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1029 if( con < 16 ) {
1030 // Left input is an add of the same number?
1031 if (add1->in(1) == add1->in(2)) {
1032 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1033 // In general, this optimization cannot be applied for c0 == 31 since
1034 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1035 return new LShiftINode(add1->in(1), phase->intcon(con + 1));
1036 }
1037
1038 // Left input is an add of a constant?
1039 const TypeInt *t12 = phase->type(add1->in(2))->isa_int();
1040 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1041 // Compute X << con0
1042 Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) );
1043 // Compute X<<con0 + (con1<<con0)
1044 return new AddINode( lsh, phase->intcon(t12->get_con() << con));
1045 }
1046 }
1047 }
1048
1049 // Check for "(x >> C1) << C2"
1050 if (add1_op == Op_RShiftI || add1_op == Op_URShiftI) {
1051 int add1Con = 0;
1052 const_shift_count(phase, add1, &add1Con);
1053
1054 // Special case C1 == C2, which just masks off low bits
1055 if (add1Con > 0 && con == add1Con) {
1056 // Convert to "(x & -(1 << C2))"
1057 return new AndINode(add1->in(1), phase->intcon(java_negate(jint(1 << con))));
1058 } else {
1059 // Wait until the right shift has been sharpened to the correct count
1060 if (add1Con > 0 && add1Con < BitsPerJavaInteger) {
1061 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1062 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1063 if (phase->is_IterGVN()) {
1064 if (con > add1Con) {
1065 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1066 Node* lshift = phase->transform(new LShiftINode(add1->in(1), phase->intcon(con - add1Con)));
1067 return new AndINode(lshift, phase->intcon(java_negate(jint(1 << con))));
1068 } else {
1069 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1070 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1071
1072 // Handle logical and arithmetic shifts
1073 Node* rshift;
1074 if (add1_op == Op_RShiftI) {
1075 rshift = phase->transform(new RShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1076 } else {
1077 rshift = phase->transform(new URShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1078 }
1079
1080 return new AndINode(rshift, phase->intcon(java_negate(jint(1 << con))));
1081 }
1082 } else {
1083 phase->record_for_igvn(this);
1084 }
1085 }
1086 }
1087 }
1088
1089 // Check for "((x >> C1) & Y) << C2"
1090 if (add1_op == Op_AndI) {
1091 Node *add2 = add1->in(1);
1092 int add2_op = add2->Opcode();
1093 if (add2_op == Op_RShiftI || add2_op == Op_URShiftI) {
1094 // Special case C1 == C2, which just masks off low bits
1095 if (add2->in(2) == in(2)) {
1096 // Convert to "(x & (Y << C2))"
1097 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1098 return new AndINode(add2->in(1), y_sh);
1099 }
1100
1101 int add2Con = 0;
1102 const_shift_count(phase, add2, &add2Con);
1103 if (add2Con > 0 && add2Con < BitsPerJavaInteger) {
1104 if (phase->is_IterGVN()) {
1105 // Convert to "((x >> C1) << C2) & (Y << C2)"
1106
1107 // Make "(x >> C1) << C2", which will get folded away by the rule above
1108 Node* x_sh = phase->transform(new LShiftINode(add2, phase->intcon(con)));
1109 // Make "Y << C2", which will simplify when Y is a constant
1110 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1111
1112 return new AndINode(x_sh, y_sh);
1113 } else {
1114 phase->record_for_igvn(this);
1115 }
1116 }
1117 }
1118 }
1119
1120 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1121 // before shifting them away.
1122 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con);
1123 if( add1_op == Op_AndI &&
1124 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) )
1125 return new LShiftINode( add1->in(1), in(2) );
1126
1127 return nullptr;
1128 }
1129
1130 //------------------------------Value------------------------------------------
1131 // A LShiftINode shifts its input2 left by input1 amount.
1132 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1133 const Type *t1 = phase->type( in(1) );
1134 const Type *t2 = phase->type( in(2) );
1135 // Either input is TOP ==> the result is TOP
1136 if( t1 == Type::TOP ) return Type::TOP;
1137 if( t2 == Type::TOP ) return Type::TOP;
1138
1139 // Left input is ZERO ==> the result is ZERO.
1140 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1141 // Shift by zero does nothing
1142 if( t2 == TypeInt::ZERO ) return t1;
1143
1144 // Either input is BOTTOM ==> the result is BOTTOM
1145 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) ||
1146 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1147 return TypeInt::INT;
1148
1149 const TypeInt *r1 = t1->is_int(); // Handy access
1150 const TypeInt *r2 = t2->is_int(); // Handy access
1151
1152 if (!r2->is_con())
1153 return TypeInt::INT;
1154
1155 uint shift = r2->get_con();
1156 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1157 // Shift by a multiple of 32 does nothing:
1158 if (shift == 0) return t1;
1159
1160 // If the shift is a constant, shift the bounds of the type,
1161 // unless this could lead to an overflow.
1162 if (!r1->is_con()) {
1163 jint lo = r1->_lo, hi = r1->_hi;
1164 if (((lo << shift) >> shift) == lo &&
1165 ((hi << shift) >> shift) == hi) {
1166 // No overflow. The range shifts up cleanly.
1167 return TypeInt::make((jint)lo << (jint)shift,
1168 (jint)hi << (jint)shift,
1169 MAX2(r1->_widen,r2->_widen));
1170 }
1171 return TypeInt::INT;
1172 }
1173
1174 return TypeInt::make( (jint)r1->get_con() << (jint)shift );
1175 }
1176
1177 //=============================================================================
1178 //------------------------------Identity---------------------------------------
1179 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1180 int count = 0;
1181 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1182 // Shift by a multiple of 64 does nothing
1183 return in(1);
1184 }
1185 return this;
1186 }
1187
1188 //------------------------------Ideal------------------------------------------
1189 // If the right input is a constant, and the left input is an add of a
1190 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1191 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1192 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1193 if (con == 0) {
1194 return nullptr;
1195 }
1196
1197 // Left input is an add?
1198 Node *add1 = in(1);
1199 int add1_op = add1->Opcode();
1200 if( add1_op == Op_AddL ) { // Left input is an add?
1201 // Avoid dead data cycles from dead loops
1202 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" );
1203
1204 // Left input is an add of the same number?
1205 if (con != (BitsPerJavaLong - 1) && add1->in(1) == add1->in(2)) {
1206 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1207 // Can only be applied if c0 != 63 because:
1208 // (x + x) << 63 = 2x << 63, while
1209 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1210 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1211 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1212 return new LShiftLNode(add1->in(1), phase->intcon(con + 1));
1213 }
1214
1215 // Left input is an add of a constant?
1216 const TypeLong *t12 = phase->type(add1->in(2))->isa_long();
1217 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1218 // Compute X << con0
1219 Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) );
1220 // Compute X<<con0 + (con1<<con0)
1221 return new AddLNode( lsh, phase->longcon(t12->get_con() << con));
1222 }
1223 }
1224
1225 // Check for "(x >> C1) << C2"
1226 if (add1_op == Op_RShiftL || add1_op == Op_URShiftL) {
1227 int add1Con = 0;
1228 const_shift_count(phase, add1, &add1Con);
1229
1230 // Special case C1 == C2, which just masks off low bits
1231 if (add1Con > 0 && con == add1Con) {
1232 // Convert to "(x & -(1 << C2))"
1233 return new AndLNode(add1->in(1), phase->longcon(java_negate(jlong(CONST64(1) << con))));
1234 } else {
1235 // Wait until the right shift has been sharpened to the correct count
1236 if (add1Con > 0 && add1Con < BitsPerJavaLong) {
1237 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1238 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1239 if (phase->is_IterGVN()) {
1240 if (con > add1Con) {
1241 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1242 Node* lshift = phase->transform(new LShiftLNode(add1->in(1), phase->intcon(con - add1Con)));
1243 return new AndLNode(lshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1244 } else {
1245 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1246 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1247
1248 // Handle logical and arithmetic shifts
1249 Node* rshift;
1250 if (add1_op == Op_RShiftL) {
1251 rshift = phase->transform(new RShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1252 } else {
1253 rshift = phase->transform(new URShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1254 }
1255
1256 return new AndLNode(rshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1257 }
1258 } else {
1259 phase->record_for_igvn(this);
1260 }
1261 }
1262 }
1263 }
1264
1265 // Check for "((x >> C1) & Y) << C2"
1266 if (add1_op == Op_AndL) {
1267 Node* add2 = add1->in(1);
1268 int add2_op = add2->Opcode();
1269 if (add2_op == Op_RShiftL || add2_op == Op_URShiftL) {
1270 // Special case C1 == C2, which just masks off low bits
1271 if (add2->in(2) == in(2)) {
1272 // Convert to "(x & (Y << C2))"
1273 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1274 return new AndLNode(add2->in(1), y_sh);
1275 }
1276
1277 int add2Con = 0;
1278 const_shift_count(phase, add2, &add2Con);
1279 if (add2Con > 0 && add2Con < BitsPerJavaLong) {
1280 if (phase->is_IterGVN()) {
1281 // Convert to "((x >> C1) << C2) & (Y << C2)"
1282
1283 // Make "(x >> C1) << C2", which will get folded away by the rule above
1284 Node* x_sh = phase->transform(new LShiftLNode(add2, phase->intcon(con)));
1285 // Make "Y << C2", which will simplify when Y is a constant
1286 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1287
1288 return new AndLNode(x_sh, y_sh);
1289 } else {
1290 phase->record_for_igvn(this);
1291 }
1292 }
1293 }
1294 }
1295
1296 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits
1297 // before shifting them away.
1298 const jlong bits_mask = jlong(max_julong >> con);
1299 if( add1_op == Op_AndL &&
1300 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) )
1301 return new LShiftLNode( add1->in(1), in(2) );
1302
1303 return nullptr;
1304 }
1305
1306 //------------------------------Value------------------------------------------
1307 // A LShiftLNode shifts its input2 left by input1 amount.
1308 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1309 const Type *t1 = phase->type( in(1) );
1310 const Type *t2 = phase->type( in(2) );
1311 // Either input is TOP ==> the result is TOP
1312 if( t1 == Type::TOP ) return Type::TOP;
1313 if( t2 == Type::TOP ) return Type::TOP;
1314
1315 // Left input is ZERO ==> the result is ZERO.
1316 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1317 // Shift by zero does nothing
1318 if( t2 == TypeInt::ZERO ) return t1;
1319
1320 // Either input is BOTTOM ==> the result is BOTTOM
1321 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) ||
1322 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1323 return TypeLong::LONG;
1324
1325 const TypeLong *r1 = t1->is_long(); // Handy access
1326 const TypeInt *r2 = t2->is_int(); // Handy access
1327
1328 if (!r2->is_con())
1329 return TypeLong::LONG;
1330
1331 uint shift = r2->get_con();
1332 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1333 // Shift by a multiple of 64 does nothing:
1334 if (shift == 0) return t1;
1335
1336 // If the shift is a constant, shift the bounds of the type,
1337 // unless this could lead to an overflow.
1338 if (!r1->is_con()) {
1339 jlong lo = r1->_lo, hi = r1->_hi;
1340 if (((lo << shift) >> shift) == lo &&
1341 ((hi << shift) >> shift) == hi) {
1342 // No overflow. The range shifts up cleanly.
1343 return TypeLong::make((jlong)lo << (jint)shift,
1344 (jlong)hi << (jint)shift,
1345 MAX2(r1->_widen,r2->_widen));
1346 }
1347 return TypeLong::LONG;
1348 }
1349
1350 return TypeLong::make( (jlong)r1->get_con() << (jint)shift );
1351 }
1352
1353 //=============================================================================
1354 //------------------------------Identity---------------------------------------
1355 Node* RShiftINode::Identity(PhaseGVN* phase) {
1356 int count = 0;
1357 if (const_shift_count(phase, this, &count)) {
1358 if ((count & (BitsPerJavaInteger - 1)) == 0) {
1359 // Shift by a multiple of 32 does nothing
1360 return in(1);
1361 }
1362 // Check for useless sign-masking
1363 if (in(1)->Opcode() == Op_LShiftI &&
1364 in(1)->req() == 3 &&
1365 in(1)->in(2) == in(2)) {
1366 count &= BitsPerJavaInteger-1; // semantics of Java shifts
1367 // Compute masks for which this shifting doesn't change
1368 int lo = (-1 << (BitsPerJavaInteger - ((uint)count)-1)); // FFFF8000
1369 int hi = ~lo; // 00007FFF
1370 const TypeInt* t11 = phase->type(in(1)->in(1))->isa_int();
1371 if (t11 == nullptr) {
1372 return this;
1373 }
1374 // Does actual value fit inside of mask?
1375 if (lo <= t11->_lo && t11->_hi <= hi) {
1376 return in(1)->in(1); // Then shifting is a nop
1377 }
1378 }
1379 }
1380 return this;
1381 }
1382
1383 //------------------------------Ideal------------------------------------------
1384 Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1385 // Inputs may be TOP if they are dead.
1386 const TypeInt *t1 = phase->type(in(1))->isa_int();
1387 if (!t1) return nullptr; // Left input is an integer
1388 const TypeInt *t3; // type of in(1).in(2)
1389 int shift = maskShiftAmount(phase, this, BitsPerJavaInteger);
1390 if (shift == 0) {
1391 return nullptr;
1392 }
1393
1394 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1395 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1396 const Node *mask = in(1);
1397 if( mask->Opcode() == Op_AndI &&
1398 (t3 = phase->type(mask->in(2))->isa_int()) &&
1399 t3->is_con() ) {
1400 Node *x = mask->in(1);
1401 jint maskbits = t3->get_con();
1402 // Convert to "(x >> shift) & (mask >> shift)"
1403 Node *shr_nomask = phase->transform( new RShiftINode(mask->in(1), in(2)) );
1404 return new AndINode(shr_nomask, phase->intcon( maskbits >> shift));
1405 }
1406
1407 // Check for "(short[i] <<16)>>16" which simply sign-extends
1408 const Node *shl = in(1);
1409 if( shl->Opcode() != Op_LShiftI ) return nullptr;
1410
1411 if( shift == 16 &&
1412 (t3 = phase->type(shl->in(2))->isa_int()) &&
1413 t3->is_con(16) ) {
1414 Node *ld = shl->in(1);
1415 if( ld->Opcode() == Op_LoadS ) {
1416 // Sign extension is just useless here. Return a RShiftI of zero instead
1417 // returning 'ld' directly. We cannot return an old Node directly as
1418 // that is the job of 'Identity' calls and Identity calls only work on
1419 // direct inputs ('ld' is an extra Node removed from 'this'). The
1420 // combined optimization requires Identity only return direct inputs.
1421 set_req_X(1, ld, phase);
1422 set_req_X(2, phase->intcon(0), phase);
1423 return this;
1424 }
1425 else if (can_reshape &&
1426 ld->Opcode() == Op_LoadUS &&
1427 ld->outcnt() == 1 && ld->unique_out() == shl)
1428 // Replace zero-extension-load with sign-extension-load
1429 return ld->as_Load()->convert_to_signed_load(*phase);
1430 }
1431
1432 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1433 if( shift == 24 &&
1434 (t3 = phase->type(shl->in(2))->isa_int()) &&
1435 t3->is_con(24) ) {
1436 Node *ld = shl->in(1);
1437 if (ld->Opcode() == Op_LoadB) {
1438 // Sign extension is just useless here
1439 set_req_X(1, ld, phase);
1440 set_req_X(2, phase->intcon(0), phase);
1441 return this;
1442 }
1443 }
1444
1445 return nullptr;
1446 }
1447
1448 //------------------------------Value------------------------------------------
1449 // A RShiftINode shifts its input2 right by input1 amount.
1450 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1451 const Type *t1 = phase->type( in(1) );
1452 const Type *t2 = phase->type( in(2) );
1453 // Either input is TOP ==> the result is TOP
1454 if( t1 == Type::TOP ) return Type::TOP;
1455 if( t2 == Type::TOP ) return Type::TOP;
1456
1457 // Left input is ZERO ==> the result is ZERO.
1458 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1459 // Shift by zero does nothing
1460 if( t2 == TypeInt::ZERO ) return t1;
1461
1462 // Either input is BOTTOM ==> the result is BOTTOM
1463 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1464 return TypeInt::INT;
1465
1466 const TypeInt *r1 = t1->is_int(); // Handy access
1467 const TypeInt *r2 = t2->is_int(); // Handy access
1468
1469 // If the shift is a constant, just shift the bounds of the type.
1470 // For example, if the shift is 31, we just propagate sign bits.
1471 if (!r1->is_con() && r2->is_con()) {
1472 uint shift = r2->get_con();
1473 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1474 // Shift by a multiple of 32 does nothing:
1475 if (shift == 0) return t1;
1476 // Calculate reasonably aggressive bounds for the result.
1477 // This is necessary if we are to correctly type things
1478 // like (x<<24>>24) == ((byte)x).
1479 jint lo = (jint)r1->_lo >> (jint)shift;
1480 jint hi = (jint)r1->_hi >> (jint)shift;
1481 assert(lo <= hi, "must have valid bounds");
1482 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1483 #ifdef ASSERT
1484 // Make sure we get the sign-capture idiom correct.
1485 if (shift == BitsPerJavaInteger-1) {
1486 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0");
1487 if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1");
1488 }
1489 #endif
1490 return ti;
1491 }
1492
1493 if (!r1->is_con() || !r2->is_con()) {
1494 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1495 if (r1->_lo >= 0) {
1496 return TypeInt::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen));
1497 }
1498
1499 // Conversely, if the left input is negative then the result must be negative.
1500 if (r1->_hi <= -1) {
1501 return TypeInt::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen));
1502 }
1503
1504 return TypeInt::INT;
1505 }
1506
1507 // Signed shift right
1508 return TypeInt::make(r1->get_con() >> (r2->get_con() & 31));
1509 }
1510
1511 //=============================================================================
1512 //------------------------------Identity---------------------------------------
1513 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1514 const TypeInt *ti = phase->type(in(2))->isa_int(); // Shift count is an int.
1515 return (ti && ti->is_con() && (ti->get_con() & (BitsPerJavaLong - 1)) == 0) ? in(1) : this;
1516 }
1517
1518 //------------------------------Value------------------------------------------
1519 // A RShiftLNode shifts its input2 right by input1 amount.
1520 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1521 const Type *t1 = phase->type( in(1) );
1522 const Type *t2 = phase->type( in(2) );
1523 // Either input is TOP ==> the result is TOP
1524 if( t1 == Type::TOP ) return Type::TOP;
1525 if( t2 == Type::TOP ) return Type::TOP;
1526
1527 // Left input is ZERO ==> the result is ZERO.
1528 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1529 // Shift by zero does nothing
1530 if( t2 == TypeInt::ZERO ) return t1;
1531
1532 // Either input is BOTTOM ==> the result is BOTTOM
1533 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1534 return TypeLong::LONG;
1535
1536 const TypeLong *r1 = t1->is_long(); // Handy access
1537 const TypeInt *r2 = t2->is_int (); // Handy access
1538
1539 // If the shift is a constant, just shift the bounds of the type.
1540 // For example, if the shift is 63, we just propagate sign bits.
1541 if (!r1->is_con() && r2->is_con()) {
1542 uint shift = r2->get_con();
1543 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts
1544 // Shift by a multiple of 64 does nothing:
1545 if (shift == 0) return t1;
1546 // Calculate reasonably aggressive bounds for the result.
1547 // This is necessary if we are to correctly type things
1548 // like (x<<24>>24) == ((byte)x).
1549 jlong lo = (jlong)r1->_lo >> (jlong)shift;
1550 jlong hi = (jlong)r1->_hi >> (jlong)shift;
1551 assert(lo <= hi, "must have valid bounds");
1552 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1553 #ifdef ASSERT
1554 // Make sure we get the sign-capture idiom correct.
1555 if (shift == (2*BitsPerJavaInteger)-1) {
1556 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0");
1557 if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1");
1558 }
1559 #endif
1560 return tl;
1561 }
1562
1563 if (!r1->is_con() || !r2->is_con()) {
1564 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1565 if (r1->_lo >= 0) {
1566 return TypeLong::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen));
1567 }
1568
1569 // Conversely, if the left input is negative then the result must be negative.
1570 if (r1->_hi <= -1) {
1571 return TypeLong::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen));
1572 }
1573
1574 return TypeLong::LONG;
1575 }
1576
1577 return TypeLong::make(r1->get_con() >> (r2->get_con() & 63));
1578 }
1579
1580 //=============================================================================
1581 //------------------------------Identity---------------------------------------
1582 Node* URShiftINode::Identity(PhaseGVN* phase) {
1583 int count = 0;
1584 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1585 // Shift by a multiple of 32 does nothing
1586 return in(1);
1587 }
1588
1589 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1590 // Happens during new-array length computation.
1591 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1592 Node *add = in(1);
1593 if (add->Opcode() == Op_AddI) {
1594 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1595 if (t2 && t2->is_con(wordSize - 1) &&
1596 add->in(1)->Opcode() == Op_LShiftI) {
1597 // Check that shift_counts are LogBytesPerWord.
1598 Node *lshift_count = add->in(1)->in(2);
1599 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1600 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1601 t_lshift_count == phase->type(in(2))) {
1602 Node *x = add->in(1)->in(1);
1603 const TypeInt *t_x = phase->type(x)->isa_int();
1604 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1605 return x;
1606 }
1607 }
1608 }
1609 }
1610
1611 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1612 }
1613
1614 //------------------------------Ideal------------------------------------------
1615 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1616 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
1617 if (con == 0) {
1618 return nullptr;
1619 }
1620
1621 // We'll be wanting the right-shift amount as a mask of that many bits
1622 const int mask = right_n_bits(BitsPerJavaInteger - con);
1623
1624 int in1_op = in(1)->Opcode();
1625
1626 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1627 if( in1_op == Op_URShiftI ) {
1628 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1629 if( t12 && t12->is_con() ) { // Right input is a constant
1630 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1631 const int con2 = t12->get_con() & 31; // Shift count is always masked
1632 const int con3 = con+con2;
1633 if( con3 < 32 ) // Only merge shifts if total is < 32
1634 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1635 }
1636 }
1637
1638 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1639 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1640 // If Q is "X << z" the rounding is useless. Look for patterns like
1641 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1642 Node *add = in(1);
1643 const TypeInt *t2 = phase->type(in(2))->isa_int();
1644 if (in1_op == Op_AddI) {
1645 Node *lshl = add->in(1);
1646 if( lshl->Opcode() == Op_LShiftI &&
1647 phase->type(lshl->in(2)) == t2 ) {
1648 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1649 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1650 return new AndINode( sum, phase->intcon(mask) );
1651 }
1652 }
1653
1654 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1655 // This shortens the mask. Also, if we are extracting a high byte and
1656 // storing it to a buffer, the mask will be removed completely.
1657 Node *andi = in(1);
1658 if( in1_op == Op_AndI ) {
1659 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1660 if( t3 && t3->is_con() ) { // Right input is a constant
1661 jint mask2 = t3->get_con();
1662 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1663 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1664 return new AndINode(newshr, phase->intcon(mask2));
1665 // The negative values are easier to materialize than positive ones.
1666 // A typical case from address arithmetic is ((x & ~15) >> 4).
1667 // It's better to change that to ((x >> 4) & ~0) versus
1668 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1669 }
1670 }
1671
1672 // Check for "(X << z ) >>> z" which simply zero-extends
1673 Node *shl = in(1);
1674 if( in1_op == Op_LShiftI &&
1675 phase->type(shl->in(2)) == t2 )
1676 return new AndINode( shl->in(1), phase->intcon(mask) );
1677
1678 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1679 Node *shr = in(1);
1680 if ( in1_op == Op_RShiftI ) {
1681 Node *in11 = shr->in(1);
1682 Node *in12 = shr->in(2);
1683 const TypeInt *t11 = phase->type(in11)->isa_int();
1684 const TypeInt *t12 = phase->type(in12)->isa_int();
1685 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1686 return new URShiftINode(in11, phase->intcon(31));
1687 }
1688 }
1689
1690 return nullptr;
1691 }
1692
1693 //------------------------------Value------------------------------------------
1694 // A URShiftINode shifts its input2 right by input1 amount.
1695 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1696 // (This is a near clone of RShiftINode::Value.)
1697 const Type *t1 = phase->type( in(1) );
1698 const Type *t2 = phase->type( in(2) );
1699 // Either input is TOP ==> the result is TOP
1700 if( t1 == Type::TOP ) return Type::TOP;
1701 if( t2 == Type::TOP ) return Type::TOP;
1702
1703 // Left input is ZERO ==> the result is ZERO.
1704 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1705 // Shift by zero does nothing
1706 if( t2 == TypeInt::ZERO ) return t1;
1707
1708 // Either input is BOTTOM ==> the result is BOTTOM
1709 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1710 return TypeInt::INT;
1711
1712 if (t2 == TypeInt::INT)
1713 return TypeInt::INT;
1714
1715 const TypeInt *r1 = t1->is_int(); // Handy access
1716 const TypeInt *r2 = t2->is_int(); // Handy access
1717
1718 if (r2->is_con()) {
1719 uint shift = r2->get_con();
1720 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1721 // Shift by a multiple of 32 does nothing:
1722 if (shift == 0) return t1;
1723 // Calculate reasonably aggressive bounds for the result.
1724 jint lo = (juint)r1->_lo >> (juint)shift;
1725 jint hi = (juint)r1->_hi >> (juint)shift;
1726 if (r1->_hi >= 0 && r1->_lo < 0) {
1727 // If the type has both negative and positive values,
1728 // there are two separate sub-domains to worry about:
1729 // The positive half and the negative half.
1730 jint neg_lo = lo;
1731 jint neg_hi = (juint)-1 >> (juint)shift;
1732 jint pos_lo = (juint) 0 >> (juint)shift;
1733 jint pos_hi = hi;
1734 lo = MIN2(neg_lo, pos_lo); // == 0
1735 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1736 }
1737 assert(lo <= hi, "must have valid bounds");
1738 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1739 #ifdef ASSERT
1740 // Make sure we get the sign-capture idiom correct.
1741 if (shift == BitsPerJavaInteger-1) {
1742 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1743 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1744 }
1745 #endif
1746 return ti;
1747 }
1748
1749 //
1750 // Do not support shifted oops in info for GC
1751 //
1752 // else if( t1->base() == Type::InstPtr ) {
1753 //
1754 // const TypeInstPtr *o = t1->is_instptr();
1755 // if( t1->singleton() )
1756 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1757 // }
1758 // else if( t1->base() == Type::KlassPtr ) {
1759 // const TypeKlassPtr *o = t1->is_klassptr();
1760 // if( t1->singleton() )
1761 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1762 // }
1763
1764 return TypeInt::INT;
1765 }
1766
1767 //=============================================================================
1768 //------------------------------Identity---------------------------------------
1769 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1770 int count = 0;
1771 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1772 // Shift by a multiple of 64 does nothing
1773 return in(1);
1774 }
1775 return this;
1776 }
1777
1778 //------------------------------Ideal------------------------------------------
1779 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1780 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1781 if (con == 0) {
1782 return nullptr;
1783 }
1784
1785 // We'll be wanting the right-shift amount as a mask of that many bits
1786 const jlong mask = jlong(max_julong >> con);
1787
1788 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1789 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1790 // If Q is "X << z" the rounding is useless. Look for patterns like
1791 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1792 Node *add = in(1);
1793 const TypeInt *t2 = phase->type(in(2))->isa_int();
1794 if (add->Opcode() == Op_AddL) {
1795 Node *lshl = add->in(1);
1796 if( lshl->Opcode() == Op_LShiftL &&
1797 phase->type(lshl->in(2)) == t2 ) {
1798 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1799 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1800 return new AndLNode( sum, phase->longcon(mask) );
1801 }
1802 }
1803
1804 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1805 // This shortens the mask. Also, if we are extracting a high byte and
1806 // storing it to a buffer, the mask will be removed completely.
1807 Node *andi = in(1);
1808 if( andi->Opcode() == Op_AndL ) {
1809 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1810 if( t3 && t3->is_con() ) { // Right input is a constant
1811 jlong mask2 = t3->get_con();
1812 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1813 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1814 return new AndLNode(newshr, phase->longcon(mask2));
1815 }
1816 }
1817
1818 // Check for "(X << z ) >>> z" which simply zero-extends
1819 Node *shl = in(1);
1820 if( shl->Opcode() == Op_LShiftL &&
1821 phase->type(shl->in(2)) == t2 )
1822 return new AndLNode( shl->in(1), phase->longcon(mask) );
1823
1824 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1825 Node *shr = in(1);
1826 if ( shr->Opcode() == Op_RShiftL ) {
1827 Node *in11 = shr->in(1);
1828 Node *in12 = shr->in(2);
1829 const TypeLong *t11 = phase->type(in11)->isa_long();
1830 const TypeInt *t12 = phase->type(in12)->isa_int();
1831 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1832 return new URShiftLNode(in11, phase->intcon(63));
1833 }
1834 }
1835 return nullptr;
1836 }
1837
1838 //------------------------------Value------------------------------------------
1839 // A URShiftINode shifts its input2 right by input1 amount.
1840 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1841 // (This is a near clone of RShiftLNode::Value.)
1842 const Type *t1 = phase->type( in(1) );
1843 const Type *t2 = phase->type( in(2) );
1844 // Either input is TOP ==> the result is TOP
1845 if( t1 == Type::TOP ) return Type::TOP;
1846 if( t2 == Type::TOP ) return Type::TOP;
1847
1848 // Left input is ZERO ==> the result is ZERO.
1849 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1850 // Shift by zero does nothing
1851 if( t2 == TypeInt::ZERO ) return t1;
1852
1853 // Either input is BOTTOM ==> the result is BOTTOM
1854 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1855 return TypeLong::LONG;
1856
1857 if (t2 == TypeInt::INT)
1858 return TypeLong::LONG;
1859
1860 const TypeLong *r1 = t1->is_long(); // Handy access
1861 const TypeInt *r2 = t2->is_int (); // Handy access
1862
1863 if (r2->is_con()) {
1864 uint shift = r2->get_con();
1865 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1866 // Shift by a multiple of 64 does nothing:
1867 if (shift == 0) return t1;
1868 // Calculate reasonably aggressive bounds for the result.
1869 jlong lo = (julong)r1->_lo >> (juint)shift;
1870 jlong hi = (julong)r1->_hi >> (juint)shift;
1871 if (r1->_hi >= 0 && r1->_lo < 0) {
1872 // If the type has both negative and positive values,
1873 // there are two separate sub-domains to worry about:
1874 // The positive half and the negative half.
1875 jlong neg_lo = lo;
1876 jlong neg_hi = (julong)-1 >> (juint)shift;
1877 jlong pos_lo = (julong) 0 >> (juint)shift;
1878 jlong pos_hi = hi;
1879 //lo = MIN2(neg_lo, pos_lo); // == 0
1880 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1881 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1882 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1883 }
1884 assert(lo <= hi, "must have valid bounds");
1885 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1886 #ifdef ASSERT
1887 // Make sure we get the sign-capture idiom correct.
1888 if (shift == BitsPerJavaLong - 1) {
1889 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1890 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1891 }
1892 #endif
1893 return tl;
1894 }
1895
1896 return TypeLong::LONG; // Give up
1897 }
1898
1899 //=============================================================================
1900 //------------------------------Ideal------------------------------------------
1901 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1902 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1903 // This reduces the number of rules in the matcher, as we only need to check
1904 // for negations on the second argument, and not the symmetric case where
1905 // the first argument is negated.
1906 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1907 swap_edges(1, 2);
1908 return this;
1909 }
1910 return nullptr;
1911 }
1912
1913 //=============================================================================
1914 //------------------------------Value------------------------------------------
1915 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1916 const Type *t1 = phase->type(in(1));
1917 if (t1 == Type::TOP) return Type::TOP;
1918 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1919 const Type *t2 = phase->type(in(2));
1920 if (t2 == Type::TOP) return Type::TOP;
1921 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1922 const Type *t3 = phase->type(in(3));
1923 if (t3 == Type::TOP) return Type::TOP;
1924 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1925 #ifndef __STDC_IEC_559__
1926 return Type::DOUBLE;
1927 #else
1928 double d1 = t1->getd();
1929 double d2 = t2->getd();
1930 double d3 = t3->getd();
1931 return TypeD::make(fma(d1, d2, d3));
1932 #endif
1933 }
1934
1935 //=============================================================================
1936 //------------------------------Value------------------------------------------
1937 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1938 const Type *t1 = phase->type(in(1));
1939 if (t1 == Type::TOP) return Type::TOP;
1940 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1941 const Type *t2 = phase->type(in(2));
1942 if (t2 == Type::TOP) return Type::TOP;
1943 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1944 const Type *t3 = phase->type(in(3));
1945 if (t3 == Type::TOP) return Type::TOP;
1946 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1947 #ifndef __STDC_IEC_559__
1948 return Type::FLOAT;
1949 #else
1950 float f1 = t1->getf();
1951 float f2 = t2->getf();
1952 float f3 = t3->getf();
1953 return TypeF::make(fma(f1, f2, f3));
1954 #endif
1955 }
1956
1957 //=============================================================================
1958 //------------------------------hash-------------------------------------------
1959 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1960 // The hash function must return the same value when edge swapping is performed.
1961 uint MulAddS2INode::hash() const {
1962 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1963 }
1964
1965 //------------------------------Rotate Operations ------------------------------
1966
1967 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1968 const Type* t1 = phase->type(in(1));
1969 if (t1 == Type::TOP) {
1970 return this;
1971 }
1972 int count = 0;
1973 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1974 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1975 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1976 // Rotate by a multiple of 32/64 does nothing
1977 return in(1);
1978 }
1979 return this;
1980 }
1981
1982 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1983 const Type* t1 = phase->type(in(1));
1984 const Type* t2 = phase->type(in(2));
1985 // Either input is TOP ==> the result is TOP
1986 if (t1 == Type::TOP || t2 == Type::TOP) {
1987 return Type::TOP;
1988 }
1989
1990 if (t1->isa_int()) {
1991 const TypeInt* r1 = t1->is_int();
1992 const TypeInt* r2 = t2->is_int();
1993
1994 // Left input is ZERO ==> the result is ZERO.
1995 if (r1 == TypeInt::ZERO) {
1996 return TypeInt::ZERO;
1997 }
1998 // Rotate by zero does nothing
1999 if (r2 == TypeInt::ZERO) {
2000 return r1;
2001 }
2002 if (r1->is_con() && r2->is_con()) {
2003 juint r1_con = (juint)r1->get_con();
2004 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2005 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2006 }
2007 return TypeInt::INT;
2008 } else {
2009 assert(t1->isa_long(), "Type must be a long");
2010 const TypeLong* r1 = t1->is_long();
2011 const TypeInt* r2 = t2->is_int();
2012
2013 // Left input is ZERO ==> the result is ZERO.
2014 if (r1 == TypeLong::ZERO) {
2015 return TypeLong::ZERO;
2016 }
2017 // Rotate by zero does nothing
2018 if (r2 == TypeInt::ZERO) {
2019 return r1;
2020 }
2021 if (r1->is_con() && r2->is_con()) {
2022 julong r1_con = (julong)r1->get_con();
2023 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2024 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2025 }
2026 return TypeLong::LONG;
2027 }
2028 }
2029
2030 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2031 const Type* t1 = phase->type(in(1));
2032 const Type* t2 = phase->type(in(2));
2033 if (t2->isa_int() && t2->is_int()->is_con()) {
2034 if (t1->isa_int()) {
2035 int lshift = t2->is_int()->get_con() & 31;
2036 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2037 } else if (t1 != Type::TOP) {
2038 assert(t1->isa_long(), "Type must be a long");
2039 int lshift = t2->is_int()->get_con() & 63;
2040 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2041 }
2042 }
2043 return nullptr;
2044 }
2045
2046 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2047 const Type* t1 = phase->type(in(1));
2048 if (t1 == Type::TOP) {
2049 return this;
2050 }
2051 int count = 0;
2052 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2053 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2054 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2055 // Rotate by a multiple of 32/64 does nothing
2056 return in(1);
2057 }
2058 return this;
2059 }
2060
2061 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2062 const Type* t1 = phase->type(in(1));
2063 const Type* t2 = phase->type(in(2));
2064 // Either input is TOP ==> the result is TOP
2065 if (t1 == Type::TOP || t2 == Type::TOP) {
2066 return Type::TOP;
2067 }
2068
2069 if (t1->isa_int()) {
2070 const TypeInt* r1 = t1->is_int();
2071 const TypeInt* r2 = t2->is_int();
2072
2073 // Left input is ZERO ==> the result is ZERO.
2074 if (r1 == TypeInt::ZERO) {
2075 return TypeInt::ZERO;
2076 }
2077 // Rotate by zero does nothing
2078 if (r2 == TypeInt::ZERO) {
2079 return r1;
2080 }
2081 if (r1->is_con() && r2->is_con()) {
2082 juint r1_con = (juint)r1->get_con();
2083 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2084 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2085 }
2086 return TypeInt::INT;
2087 } else {
2088 assert(t1->isa_long(), "Type must be a long");
2089 const TypeLong* r1 = t1->is_long();
2090 const TypeInt* r2 = t2->is_int();
2091 // Left input is ZERO ==> the result is ZERO.
2092 if (r1 == TypeLong::ZERO) {
2093 return TypeLong::ZERO;
2094 }
2095 // Rotate by zero does nothing
2096 if (r2 == TypeInt::ZERO) {
2097 return r1;
2098 }
2099 if (r1->is_con() && r2->is_con()) {
2100 julong r1_con = (julong)r1->get_con();
2101 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2102 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2103 }
2104 return TypeLong::LONG;
2105 }
2106 }
2107
2108 // Given an expression (AndX shift mask) or (AndX mask shift),
2109 // determine if the AndX must always produce zero, because the
2110 // the shift (x<<N) is bitwise disjoint from the mask #M.
2111 // The X in AndX must be I or L, depending on bt.
2112 // Specifically, the following cases fold to zero,
2113 // when the shift value N is large enough to zero out
2114 // all the set positions of the and-mask M.
2115 // (AndI (LShiftI _ #N) #M) => #0
2116 // (AndL (LShiftL _ #N) #M) => #0
2117 // (AndL (ConvI2L (LShiftI _ #N)) #M) => #0
2118 // The M and N values must satisfy ((-1 << N) & M) == 0.
2119 // Because the optimization might work for a non-constant
2120 // mask M, we check the AndX for both operand orders.
2121 bool MulNode::AndIL_shift_and_mask_is_always_zero(PhaseGVN* phase, Node* shift, Node* mask, BasicType bt, bool check_reverse) {
2122 if (mask == nullptr || shift == nullptr) {
2123 return false;
2124 }
2125 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2126 if (mask_t == nullptr || phase->type(shift)->isa_integer(bt) == nullptr) {
2127 return false;
2128 }
2129 shift = shift->uncast();
2130 if (shift == nullptr) {
2131 return false;
2132 }
2133 if (phase->type(shift)->isa_integer(bt) == nullptr) {
2134 return false;
2135 }
2136 BasicType shift_bt = bt;
2137 if (bt == T_LONG && shift->Opcode() == Op_ConvI2L) {
2138 bt = T_INT;
2139 Node* val = shift->in(1);
2140 if (val == nullptr) {
2141 return false;
2142 }
2143 val = val->uncast();
2144 if (val == nullptr) {
2145 return false;
2146 }
2147 if (val->Opcode() == Op_LShiftI) {
2148 shift_bt = T_INT;
2149 shift = val;
2150 if (phase->type(shift)->isa_integer(bt) == nullptr) {
2151 return false;
2152 }
2153 }
2154 }
2155 if (shift->Opcode() != Op_LShift(shift_bt)) {
2156 if (check_reverse &&
2157 (mask->Opcode() == Op_LShift(bt) ||
2158 (bt == T_LONG && mask->Opcode() == Op_ConvI2L))) {
2159 // try it the other way around
2160 return AndIL_shift_and_mask_is_always_zero(phase, mask, shift, bt, false);
2161 }
2162 return false;
2163 }
2164 Node* shift2 = shift->in(2);
2165 if (shift2 == nullptr) {
2166 return false;
2167 }
2168 const Type* shift2_t = phase->type(shift2);
2169 if (!shift2_t->isa_int() || !shift2_t->is_int()->is_con()) {
2170 return false;
2171 }
2172
2173 jint shift_con = shift2_t->is_int()->get_con() & ((shift_bt == T_INT ? BitsPerJavaInteger : BitsPerJavaLong) - 1);
2174 if ((((jlong)1) << shift_con) > mask_t->hi_as_long() && mask_t->lo_as_long() >= 0) {
2175 return true;
2176 }
2177
2178 return false;
2179 }
2180
2181 // Given an expression (AndX (AddX v1 (LShiftX v2 #N)) #M)
2182 // determine if the AndX must always produce (AndX v1 #M),
2183 // because the shift (v2<<N) is bitwise disjoint from the mask #M.
2184 // The X in AndX will be I or L, depending on bt.
2185 // Specifically, the following cases fold,
2186 // when the shift value N is large enough to zero out
2187 // all the set positions of the and-mask M.
2188 // (AndI (AddI v1 (LShiftI _ #N)) #M) => (AndI v1 #M)
2189 // (AndL (AddI v1 (LShiftL _ #N)) #M) => (AndL v1 #M)
2190 // (AndL (AddL v1 (ConvI2L (LShiftI _ #N))) #M) => (AndL v1 #M)
2191 // The M and N values must satisfy ((-1 << N) & M) == 0.
2192 // Because the optimization might work for a non-constant
2193 // mask M, and because the AddX operands can come in either
2194 // order, we check for every operand order.
2195 Node* MulNode::AndIL_add_shift_and_mask(PhaseGVN* phase, BasicType bt) {
2196 Node* add = in(1);
2197 Node* mask = in(2);
2198 if (add == nullptr || mask == nullptr) {
2199 return nullptr;
2200 }
2201 int addidx = 0;
2202 if (add->Opcode() == Op_Add(bt)) {
2203 addidx = 1;
2204 } else if (mask->Opcode() == Op_Add(bt)) {
2205 mask = add;
2206 addidx = 2;
2207 add = in(addidx);
2208 }
2209 if (addidx > 0) {
2210 Node* add1 = add->in(1);
2211 Node* add2 = add->in(2);
2212 if (add1 != nullptr && add2 != nullptr) {
2213 if (AndIL_shift_and_mask_is_always_zero(phase, add1, mask, bt, false)) {
2214 set_req_X(addidx, add2, phase);
2215 return this;
2216 } else if (AndIL_shift_and_mask_is_always_zero(phase, add2, mask, bt, false)) {
2217 set_req_X(addidx, add1, phase);
2218 return this;
2219 }
2220 }
2221 }
2222 return nullptr;
2223 }