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 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
203 // Code pattern on return from a call that returns an __Value. Can
204 // be optimized away if the return value turns out to be an oop.
205 if (op == Op_AndX &&
206 in(1) != nullptr &&
207 in(1)->Opcode() == Op_CastP2X &&
208 in(1)->in(1) != nullptr &&
209 phase->type(in(1)->in(1))->isa_oopptr() &&
210 t2->isa_intptr_t()->_lo >= 0 &&
211 t2->isa_intptr_t()->_hi <= MinObjAlignmentInBytesMask) {
212 return add_id();
213 }
214
215 // Either input is BOTTOM ==> the result is the local BOTTOM
216 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
217 return bottom_type();
218
219 return mul_ring(t1,t2); // Local flavor of type multiplication
220 }
221
222 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
223 switch (bt) {
224 case T_INT:
225 return new MulINode(in1, in2);
226 case T_LONG:
227 return new MulLNode(in1, in2);
228 default:
229 fatal("Not implemented for %s", type2name(bt));
230 }
231 return nullptr;
232 }
233
234 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
235 switch (bt) {
236 case T_INT:
237 return new AndINode(in1, in2);
238 case T_LONG:
239 return new AndLNode(in1, in2);
240 default:
241 fatal("Not implemented for %s", type2name(bt));
242 }
243 return nullptr;
244 }
245
246
247 //=============================================================================
248 //------------------------------Ideal------------------------------------------
249 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
250 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
251 const jint con = in(2)->find_int_con(0);
252 if (con == 0) {
253 // If in(2) is not a constant, call Ideal() of the parent class to
254 // try to move constant to the right side.
255 return MulNode::Ideal(phase, can_reshape);
256 }
257
258 // Now we have a constant Node on the right and the constant in con.
259 if (con == 1) {
260 // By one is handled by Identity call
261 return nullptr;
262 }
263
264 // Check for negative constant; if so negate the final result
265 bool sign_flip = false;
266
267 unsigned int abs_con = g_uabs(con);
268 if (abs_con != (unsigned int)con) {
269 sign_flip = true;
270 }
271
272 // Get low bit; check for being the only bit
273 Node *res = nullptr;
274 unsigned int bit1 = submultiple_power_of_2(abs_con);
275 if (bit1 == abs_con) { // Found a power of 2?
276 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
277 } else {
278 // Check for constant with 2 bits set
279 unsigned int bit2 = abs_con - bit1;
280 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
281 if (bit2 + bit1 == abs_con) { // Found all bits in con?
282 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
283 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
284 res = new AddINode(n2, n1);
285 } else if (is_power_of_2(abs_con + 1)) {
286 // Sleezy: power-of-2 - 1. Next time be generic.
287 unsigned int temp = abs_con + 1;
288 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
289 res = new SubINode(n1, in(1));
290 } else {
291 return MulNode::Ideal(phase, can_reshape);
292 }
293 }
294
295 if (sign_flip) { // Need to negate result?
296 res = phase->transform(res);// Transform, before making the zero con
297 res = new SubINode(phase->intcon(0),res);
298 }
299
300 return res; // Return final result
301 }
302
303 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
304 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
305 //
306 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
307 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
308 // - no overflow (i.e. product fits into x bits)
309 // - overflow (i.e. product does not fit into x bits)
310 //
311 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
312 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
313 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
314 // The cross product consists of the following four multiplications with 2x-bit results:
315 // (1) left_lo * right_lo
316 // (2) left_lo * right_hi
317 // (3) left_hi * right_lo
318 // (4) left_hi * right_hi
319 //
320 // Let's define the following two functions:
321 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
322 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
323 //
324 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
325 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
326 // Ux(p) = 0
327 // Lx(p) = p
328 //
329 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
330 // with x bits:
331 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
332 //
333 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
334 // (i.e. all values in the x-bit space could be possible):
335 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
336 //
337 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
338 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
339 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
340 // the lower x bits the same number of times:
341 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
342 //
343 // If all upper x bits are equal, we can conclude:
344 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
345 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
346 //
347 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
348 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
349 //
350 //
351 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
352 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
353 // 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,
354 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
355 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
356 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
357 // that there is no underflow.
358 //
359 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
360 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
361 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
362 // can compare the upper x bits in a unified way with N() being the normalization function:
363 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
364 template<typename NativeType>
365 class IntegerTypeMultiplication {
366
367 NativeType _lo_left;
368 NativeType _lo_right;
369 NativeType _hi_left;
370 NativeType _hi_right;
371 short _widen_left;
372 short _widen_right;
373
374 static const Type* overflow_type();
375 static NativeType multiply_high(NativeType x, NativeType y);
376 const Type* create_type(NativeType lo, NativeType hi) const;
377
378 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
379 return normalize_overflow_value(x, y, multiply_high(x, y));
380 }
381
382 bool cross_product_not_same_overflow_value() const {
383 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
384 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
385 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
386 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
387 return lo_lo_high_product != lo_hi_high_product ||
388 lo_hi_high_product != hi_lo_high_product ||
389 hi_lo_high_product != hi_hi_high_product;
390 }
391
392 bool does_product_overflow(NativeType x, NativeType y) const {
393 return multiply_high_signed_overflow_value(x, y) != 0;
394 }
395
396 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
397 return java_multiply(x, y) < 0 ? result + 1 : result;
398 }
399
400 public:
401 template<class IntegerType>
402 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
403 : _lo_left(left->_lo), _lo_right(right->_lo),
404 _hi_left(left->_hi), _hi_right(right->_hi),
405 _widen_left(left->_widen), _widen_right(right->_widen) {}
406
407 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
408 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
409 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
410 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
411 const Type* compute() const {
412 if (cross_product_not_same_overflow_value()) {
413 return overflow_type();
414 }
415
416 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
417 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
418 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
419 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
420 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
421 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
422 return create_type(min, max);
423 }
424
425 bool does_overflow() const {
426 return does_product_overflow(_lo_left, _lo_right) ||
427 does_product_overflow(_lo_left, _hi_right) ||
428 does_product_overflow(_hi_left, _lo_right) ||
429 does_product_overflow(_hi_left, _hi_right);
430 }
431 };
432
433 template <>
434 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
435 return TypeInt::INT;
436 }
437
438 template <>
439 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
440 const jlong x_64 = x;
441 const jlong y_64 = y;
442 const jlong product = x_64 * y_64;
443 return (jint)((uint64_t)product >> 32u);
444 }
445
446 template <>
447 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
448 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
449 }
450
451 template <>
452 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
453 return TypeLong::LONG;
454 }
455
456 template <>
457 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
458 return multiply_high_signed(x, y);
459 }
460
461 template <>
462 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
463 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
464 }
465
466 // Compute the product type of two integer ranges into this node.
467 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
468 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
469 return integer_multiplication.compute();
470 }
471
472 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
473 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
474 return integer_multiplication.does_overflow();
475 }
476
477 // Compute the product type of two long ranges into this node.
478 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
479 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
480 return integer_multiplication.compute();
481 }
482
483 //=============================================================================
484 //------------------------------Ideal------------------------------------------
485 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
486 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
487 const jlong con = in(2)->find_long_con(0);
488 if (con == 0) {
489 // If in(2) is not a constant, call Ideal() of the parent class to
490 // try to move constant to the right side.
491 return MulNode::Ideal(phase, can_reshape);
492 }
493
494 // Now we have a constant Node on the right and the constant in con.
495 if (con == 1) {
496 // By one is handled by Identity call
497 return nullptr;
498 }
499
500 // Check for negative constant; if so negate the final result
501 bool sign_flip = false;
502 julong abs_con = g_uabs(con);
503 if (abs_con != (julong)con) {
504 sign_flip = true;
505 }
506
507 // Get low bit; check for being the only bit
508 Node *res = nullptr;
509 julong bit1 = submultiple_power_of_2(abs_con);
510 if (bit1 == abs_con) { // Found a power of 2?
511 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
512 } else {
513
514 // Check for constant with 2 bits set
515 julong bit2 = abs_con-bit1;
516 bit2 = bit2 & (0-bit2); // Extract 2nd bit
517 if (bit2 + bit1 == abs_con) { // Found all bits in con?
518 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
519 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
520 res = new AddLNode(n2, n1);
521
522 } else if (is_power_of_2(abs_con+1)) {
523 // Sleezy: power-of-2 -1. Next time be generic.
524 julong temp = abs_con + 1;
525 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
526 res = new SubLNode(n1, in(1));
527 } else {
528 return MulNode::Ideal(phase, can_reshape);
529 }
530 }
531
532 if (sign_flip) { // Need to negate result?
533 res = phase->transform(res);// Transform, before making the zero con
534 res = new SubLNode(phase->longcon(0),res);
535 }
536
537 return res; // Return final result
538 }
539
540 //=============================================================================
541 //------------------------------mul_ring---------------------------------------
542 // Compute the product type of two double ranges into this node.
543 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
544 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
545 return TypeF::make( t0->getf() * t1->getf() );
546 }
547
548 //------------------------------Ideal---------------------------------------
549 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
550 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
551 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
552
553 // x * 2 -> x + x
554 if (t2 != nullptr && t2->getf() == 2) {
555 Node* base = in(1);
556 return new AddFNode(base, base);
557 }
558 return MulNode::Ideal(phase, can_reshape);
559 }
560
561 //=============================================================================
562 //------------------------------Ideal------------------------------------------
563 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
564 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
565 const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
566
567 // x * 2 -> x + x
568 if (t2 != nullptr && t2->getf() == 2) {
569 Node* base = in(1);
570 return new AddHFNode(base, base);
571 }
572 return MulNode::Ideal(phase, can_reshape);
573 }
574
575 // Compute the product type of two half float ranges into this node.
576 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
577 if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
578 return Type::HALF_FLOAT;
579 }
580 return TypeH::make(t0->getf() * t1->getf());
581 }
582
583 //=============================================================================
584 //------------------------------mul_ring---------------------------------------
585 // Compute the product type of two double ranges into this node.
586 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
587 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
588 // We must be multiplying 2 double constants.
589 return TypeD::make( t0->getd() * t1->getd() );
590 }
591
592 //------------------------------Ideal---------------------------------------
593 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
594 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
595 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
596
597 // x * 2 -> x + x
598 if (t2 != nullptr && t2->getd() == 2) {
599 Node* base = in(1);
600 return new AddDNode(base, base);
601 }
602
603 return MulNode::Ideal(phase, can_reshape);
604 }
605
606 //=============================================================================
607 //------------------------------Value------------------------------------------
608 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
609 const Type *t1 = phase->type( in(1) );
610 const Type *t2 = phase->type( in(2) );
611 const Type *bot = bottom_type();
612 return MulHiValue(t1, t2, bot);
613 }
614
615 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
616 const Type *t1 = phase->type( in(1) );
617 const Type *t2 = phase->type( in(2) );
618 const Type *bot = bottom_type();
619 return MulHiValue(t1, t2, bot);
620 }
621
622 // A common routine used by UMulHiLNode and MulHiLNode
623 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
624 // Either input is TOP ==> the result is TOP
625 if( t1 == Type::TOP ) return Type::TOP;
626 if( t2 == Type::TOP ) return Type::TOP;
627
628 // Either input is BOTTOM ==> the result is the local BOTTOM
629 if( (t1 == bot) || (t2 == bot) ||
630 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
631 return bot;
632
633 // It is not worth trying to constant fold this stuff!
634 return TypeLong::LONG;
635 }
636
637 //=============================================================================
638 //------------------------------mul_ring---------------------------------------
639 // Supplied function returns the product of the inputs IN THE CURRENT RING.
640 // For the logical operations the ring's MUL is really a logical AND function.
641 // This also type-checks the inputs for sanity. Guaranteed never to
642 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
643 const Type* AndINode::mul_ring(const Type* t1, const Type* t2) const {
644 return RangeInference::infer_and(t1->is_int(), t2->is_int());
645 }
646
647 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
648
649 const Type* AndINode::Value(PhaseGVN* phase) const {
650 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
651 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
652 return TypeInt::ZERO;
653 }
654
655 return MulNode::Value(phase);
656 }
657
658 //------------------------------Identity---------------------------------------
659 // Masking off the high bits of an unsigned load is not required
660 Node* AndINode::Identity(PhaseGVN* phase) {
661
662 // x & x => x
663 if (in(1) == in(2)) {
664 return in(1);
665 }
666
667 const TypeInt* t1 = phase->type(in(1))->is_int();
668 const TypeInt* t2 = phase->type(in(2))->is_int();
669
670 if ((~t1->_bits._ones & ~t2->_bits._zeros) == 0) {
671 // All bits that might be 0 in in1 are known to be 0 in in2
672 return in(2);
673 }
674
675 if ((~t2->_bits._ones & ~t1->_bits._zeros) == 0) {
676 // All bits that might be 0 in in2 are known to be 0 in in1
677 return in(1);
678 }
679
680 return MulNode::Identity(phase);
681 }
682
683 //------------------------------Ideal------------------------------------------
684 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
685 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
686 Node* progress = AndIL_sum_and_mask(phase, T_INT);
687 if (progress != nullptr) {
688 return progress;
689 }
690
691 // Convert "(~a) & (~b)" into "~(a | b)"
692 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
693 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
694 Node* tn = phase->transform(or_a_b);
695 return AddNode::make_not(phase, tn, T_INT);
696 }
697
698 // Special case constant AND mask
699 const TypeInt *t2 = phase->type( in(2) )->isa_int();
700 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
701 const int mask = t2->get_con();
702 Node *load = in(1);
703 uint lop = load->Opcode();
704
705 // Masking bits off of a Character? Hi bits are already zero.
706 if( lop == Op_LoadUS &&
707 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
708 return new AndINode(load,phase->intcon(mask&0xFFFF));
709
710 // Masking bits off of a Short? Loading a Character does some masking
711 if (can_reshape &&
712 load->outcnt() == 1 && load->unique_out() == this) {
713 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
714 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
715 ldus = phase->transform(ldus);
716 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
717 }
718
719 // Masking sign bits off of a Byte? Do an unsigned byte load plus
720 // an and.
721 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
722 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
723 ldub = phase->transform(ldub);
724 return new AndINode(ldub, phase->intcon(mask));
725 }
726 }
727
728 // Masking off sign bits? Dont make them!
729 if( lop == Op_RShiftI ) {
730 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
731 if( t12 && t12->is_con() ) { // Shift is by a constant
732 int shift = t12->get_con();
733 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
734 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
735 // If the AND'ing of the 2 masks has no bits, then only original shifted
736 // bits survive. NO sign-extension bits survive the maskings.
737 if( (sign_bits_mask & mask) == 0 ) {
738 // Use zero-fill shift instead
739 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
740 return new AndINode( zshift, in(2) );
741 }
742 }
743 }
744
745 // Check for 'negate/and-1', a pattern emitted when someone asks for
746 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
747 // plus 1) and the mask is of the low order bit. Skip the negate.
748 if( lop == Op_SubI && mask == 1 && load->in(1) &&
749 phase->type(load->in(1)) == TypeInt::ZERO )
750 return new AndINode( load->in(2), in(2) );
751
752 return MulNode::Ideal(phase, can_reshape);
753 }
754
755 //=============================================================================
756 //------------------------------mul_ring---------------------------------------
757 // Supplied function returns the product of the inputs IN THE CURRENT RING.
758 // For the logical operations the ring's MUL is really a logical AND function.
759 // This also type-checks the inputs for sanity. Guaranteed never to
760 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
761 const Type* AndLNode::mul_ring(const Type* t1, const Type* t2) const {
762 return RangeInference::infer_and(t1->is_long(), t2->is_long());
763 }
764
765 const Type* AndLNode::Value(PhaseGVN* phase) const {
766 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
767 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
768 return TypeLong::ZERO;
769 }
770
771 return MulNode::Value(phase);
772 }
773
774 //------------------------------Identity---------------------------------------
775 // Masking off the high bits of an unsigned load is not required
776 Node* AndLNode::Identity(PhaseGVN* phase) {
777
778 // x & x => x
779 if (in(1) == in(2)) {
780 return in(1);
781 }
782
783 const TypeLong* t1 = phase->type(in(1))->is_long();
784 const TypeLong* t2 = phase->type(in(2))->is_long();
785
786 if ((~t1->_bits._ones & ~t2->_bits._zeros) == 0) {
787 // All bits that might be 0 in in1 are known to be 0 in in2
788 return in(2);
789 }
790
791 if ((~t2->_bits._ones & ~t1->_bits._zeros) == 0) {
792 // All bits that might be 0 in in2 are known to be 0 in in1
793 return in(1);
794 }
795
796 return MulNode::Identity(phase);
797 }
798
799 //------------------------------Ideal------------------------------------------
800 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
801 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
802 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
803 if (progress != nullptr) {
804 return progress;
805 }
806
807 // Convert "(~a) & (~b)" into "~(a | b)"
808 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
809 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
810 Node* tn = phase->transform(or_a_b);
811 return AddNode::make_not(phase, tn, T_LONG);
812 }
813
814 // Special case constant AND mask
815 const TypeLong *t2 = phase->type( in(2) )->isa_long();
816 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
817 const jlong mask = t2->get_con();
818
819 Node* in1 = in(1);
820 int op = in1->Opcode();
821
822 // Are we masking a long that was converted from an int with a mask
823 // that fits in 32-bits? Commute them and use an AndINode. Don't
824 // convert masks which would cause a sign extension of the integer
825 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
826 // would be optimized away later in Identity.
827 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
828 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
829 andi = phase->transform(andi);
830 return new ConvI2LNode(andi);
831 }
832
833 // Masking off sign bits? Dont make them!
834 if (op == Op_RShiftL) {
835 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
836 if( t12 && t12->is_con() ) { // Shift is by a constant
837 int shift = t12->get_con();
838 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
839 if (shift != 0) {
840 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
841 // If the AND'ing of the 2 masks has no bits, then only original shifted
842 // bits survive. NO sign-extension bits survive the maskings.
843 if( (sign_bits_mask & mask) == 0 ) {
844 // Use zero-fill shift instead
845 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
846 return new AndLNode(zshift, in(2));
847 }
848 }
849 }
850 }
851
852 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
853 Node* load1 = in(1);
854 Node* load2 = nullptr;
855 if (load1->is_Phi() && phase->type(load1)->isa_long()) {
856 load1 = in(1)->in(1);
857 load2 = in(1)->in(2);
858 }
859 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
860 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
861 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
862 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
863 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
864 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
865 if (mask == markWord::inline_type_pattern) {
866 if (adr_t1->is_inlinetypeptr()) {
867 set_req_X(1, in(2), phase);
868 return this;
869 } else if (!adr_t1->can_be_inline_type()) {
870 set_req_X(1, phase->longcon(0), phase);
871 return this;
872 }
873 } else if (mask == markWord::null_free_array_bit_in_place) {
874 if (adr_t1->is_null_free()) {
875 set_req_X(1, in(2), phase);
876 return this;
877 } else if (adr_t1->is_not_null_free()) {
878 set_req_X(1, phase->longcon(0), phase);
879 return this;
880 }
881 } else if (mask == markWord::flat_array_bit_in_place) {
882 if (adr_t1->is_flat()) {
883 set_req_X(1, in(2), phase);
884 return this;
885 } else if (adr_t1->is_not_flat()) {
886 set_req_X(1, phase->longcon(0), phase);
887 return this;
888 }
889 }
890 }
891 }
892
893 return MulNode::Ideal(phase, can_reshape);
894 }
895
896 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
897 switch (bt) {
898 case T_INT:
899 return new LShiftINode(in1, in2);
900 case T_LONG:
901 return new LShiftLNode(in1, in2);
902 default:
903 fatal("Not implemented for %s", type2name(bt));
904 }
905 return nullptr;
906 }
907
908 // Returns whether the shift amount is constant. If so, sets count.
909 static bool const_shift_count(PhaseGVN* phase, const Node* shift_node, int* count) {
910 const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
911 if (tcount != nullptr && tcount->is_con()) {
912 *count = tcount->get_con();
913 return true;
914 }
915 return false;
916 }
917
918 // Returns whether the shift amount is constant. If so, sets real_shift and masked_shift.
919 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, int& real_shift, uint& masked_shift) {
920 if (const_shift_count(phase, shift_node, &real_shift)) {
921 masked_shift = real_shift & (nBits - 1);
922 return true;
923 }
924 return false;
925 }
926
927 // Convenience for when we don't care about the real amount
928 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, uint& masked_shift) {
929 int real_shift;
930 return mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift);
931 }
932
933 // Use this in ::Ideal only with shiftNode == this!
934 // Sets masked_shift to the masked shift amount if constant or 0 if not constant.
935 // Returns shift_node if the shift amount input node was modified, nullptr otherwise.
936 static Node* mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits, uint& masked_shift) {
937 int real_shift;
938 if (mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift)) {
939 if (masked_shift == 0) {
940 // Let Identity() handle 0 shift count.
941 return nullptr;
942 }
943
944 if (real_shift != (int)masked_shift) {
945 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
946
947 // We need to notify the caller that the graph was reshaped, as Ideal needs
948 // to return the root of the reshaped graph if any change was made.
949 return shift_node;
950 }
951 } else {
952 // Not a shift by a constant.
953 masked_shift = 0;
954 }
955 return nullptr;
956 }
957
958 // Called with
959 // outer_shift = (_ << rhs_outer)
960 // We are looking for the pattern:
961 // outer_shift = ((X << rhs_inner) << rhs_outer)
962 // where rhs_outer and rhs_inner are constant
963 // we denote inner_shift the nested expression (X << rhs_inner)
964 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
965 // where nbits is the number of bits of the shifts
966 //
967 // There are 2 cases:
968 // if con_outer + con_inner >= nbits => 0
969 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
970 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
971 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
972 const Node* inner_shift = outer_shift->in(1);
973 if (inner_shift->Opcode() != Op_LShift(bt)) {
974 return nullptr;
975 }
976
977 uint nbits = bits_per_java_integer(bt);
978 uint con_inner;
979 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
980 return nullptr;
981 }
982
983 if (con_inner == 0) {
984 // We let the Identity() of the inner shift do its job.
985 return nullptr;
986 }
987
988 if (con_outer + con_inner >= nbits) {
989 // While it might be tempting to use
990 // phase->zerocon(bt);
991 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
992 // to return a new node, this or nullptr, but not an old (cached) node.
993 return ConNode::make(TypeInteger::zero(bt));
994 }
995
996 // con0 + con1 < nbits ==> actual shift happens now
997 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
998 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
999 }
1000
1001 //------------------------------Identity---------------------------------------
1002 Node* LShiftINode::Identity(PhaseGVN* phase) {
1003 return IdentityIL(phase, T_INT);
1004 }
1005
1006 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1007 uint con;
1008 Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), con);
1009 if (con == 0) {
1010 return nullptr;
1011 }
1012
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* add1 = in(1);
1016 int add1_op = add1->Opcode();
1017 if (add1_op == Op_Add(bt)) { // Left input is an add?
1018 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
1019
1020 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1021 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1022 if (bt != T_INT || con < 16) {
1023 // Left input is an add of the same number?
1024 if (con != (bits_per_java_integer(bt) - 1) && add1->in(1) == add1->in(2)) {
1025 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1026 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
1027 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1028 // or c0 != 63 (for LShiftL) because:
1029 // (x + x) << 63 = 2x << 63, while
1030 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1031 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1032 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1033 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1034 }
1035
1036 // Left input is an add of a constant?
1037 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1038 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1039 // Compute X << con0
1040 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1041 // Compute X<<con0 + (con1<<con0)
1042 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1043 }
1044 }
1045 }
1046 // Check for "(con0 - X) << con1"
1047 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1048 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1049 if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) { // Left input is a sub?
1050 // Left input is a sub from a constant?
1051 const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
1052 if (t11 != nullptr && t11->is_con()) {
1053 // Compute X << con0
1054 Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
1055 // Compute (con1<<con0) - (X<<con0)
1056 return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
1057 }
1058 }
1059
1060 // Check for "(x >> C1) << C2"
1061 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1062 int add1Con = 0;
1063 const_shift_count(phase, add1, &add1Con);
1064
1065 // Special case C1 == C2, which just masks off low bits
1066 if (add1Con > 0 && con == (uint)add1Con) {
1067 // Convert to "(x & -(1 << C2))"
1068 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1069 } else {
1070 // Wait until the right shift has been sharpened to the correct count
1071 if (add1Con > 0 && (uint)add1Con < bits_per_java_integer(bt)) {
1072 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1073 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1074 if (phase->is_IterGVN()) {
1075 if (con > (uint)add1Con) {
1076 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1077 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1078 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1079 } else {
1080 assert(con < (uint)add1Con, "must be (%d < %d)", con, add1Con);
1081 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1082
1083 // Handle logical and arithmetic shifts
1084 Node* rshift;
1085 if (add1_op == Op_RShift(bt)) {
1086 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1087 } else {
1088 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1089 }
1090
1091 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1092 }
1093 } else {
1094 phase->record_for_igvn(this);
1095 }
1096 }
1097 }
1098 }
1099
1100 // Check for "((x >> C1) & Y) << C2"
1101 if (add1_op == Op_And(bt)) {
1102 Node* add2 = add1->in(1);
1103 int add2_op = add2->Opcode();
1104 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1105 // Special case C1 == C2, which just masks off low bits
1106 if (add2->in(2) == in(2)) {
1107 // Convert to "(x & (Y << C2))"
1108 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1109 return MulNode::make_and(add2->in(1), y_sh, bt);
1110 }
1111
1112 int add2Con = 0;
1113 const_shift_count(phase, add2, &add2Con);
1114 if (add2Con > 0 && (uint)add2Con < bits_per_java_integer(bt)) {
1115 if (phase->is_IterGVN()) {
1116 // Convert to "((x >> C1) << C2) & (Y << C2)"
1117
1118 // Make "(x >> C1) << C2", which will get folded away by the rule above
1119 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1120 // Make "Y << C2", which will simplify when Y is a constant
1121 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1122
1123 return MulNode::make_and(x_sh, y_sh, bt);
1124 } else {
1125 phase->record_for_igvn(this);
1126 }
1127 }
1128 }
1129 }
1130
1131 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1132 // before shifting them away.
1133 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1134 assert(bt != T_INT || bits_mask == right_n_bits(bits_per_java_integer(bt)-con), "inconsistent");
1135 if (add1_op == Op_And(bt) &&
1136 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1137 return LShiftNode::make(add1->in(1), in(2), bt);
1138 }
1139
1140 // Collapse nested left-shifts with constant rhs:
1141 // (X << con1) << con2 ==> X << (con1 + con2)
1142 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1143 if (doubleShift != nullptr) {
1144 return doubleShift;
1145 }
1146
1147 return progress;
1148 }
1149
1150 //------------------------------Ideal------------------------------------------
1151 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1152 return IdealIL(phase, can_reshape, T_INT);
1153 }
1154
1155 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1156 const Type* t1 = phase->type(in(1));
1157 const Type* t2 = phase->type(in(2));
1158 // Either input is TOP ==> the result is TOP
1159 if (t1 == Type::TOP) {
1160 return Type::TOP;
1161 }
1162 if (t2 == Type::TOP) {
1163 return Type::TOP;
1164 }
1165
1166 // Left input is ZERO ==> the result is ZERO.
1167 if (t1 == TypeInteger::zero(bt)) {
1168 return TypeInteger::zero(bt);
1169 }
1170 // Shift by zero does nothing
1171 if (t2 == TypeInt::ZERO) {
1172 return t1;
1173 }
1174
1175 // If nothing is known about the shift amount then the result is BOTTOM
1176 if (t2 == TypeInt::INT) {
1177 return TypeInteger::bottom(bt);
1178 }
1179
1180 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1181 // Since the shift semantics in Java take into account only the bottom five
1182 // bits for ints and the bottom six bits for longs, we can further constrain
1183 // the range of values of the shift amount by ANDing with the right mask based
1184 // on whether the type is int or long.
1185 const TypeInt* mask = TypeInt::make(bits_per_java_integer(bt) - 1);
1186 const TypeInt* r2 = RangeInference::infer_and(t2->is_int(), mask);
1187
1188 if (!r2->is_con()) {
1189 return TypeInteger::bottom(bt);
1190 }
1191
1192 uint shift = r2->get_con();
1193 // Shift by a multiple of 32/64 does nothing:
1194 if (shift == 0) {
1195 return t1;
1196 }
1197
1198 // If the shift is a constant, shift the bounds of the type,
1199 // unless this could lead to an overflow.
1200 if (!r1->is_con()) {
1201 #ifdef ASSERT
1202 if (bt == T_INT) {
1203 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1204 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1205 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1206 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1207 }
1208 #endif
1209
1210 if (bt == T_INT) {
1211 return RangeInference::infer_lshift(r1->is_int(), shift);
1212 }
1213
1214 return RangeInference::infer_lshift(r1->is_long(), shift);
1215 }
1216
1217 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1218 }
1219
1220 //------------------------------Value------------------------------------------
1221 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1222 return ValueIL(phase, T_INT);
1223 }
1224
1225 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1226 int count = 0;
1227 if (const_shift_count(phase, this, &count) && (count & (bits_per_java_integer(bt) - 1)) == 0) {
1228 // Shift by a multiple of 32/64 does nothing
1229 return in(1);
1230 }
1231 return this;
1232 }
1233
1234 //=============================================================================
1235 //------------------------------Identity---------------------------------------
1236 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1237 return IdentityIL(phase, T_LONG);
1238 }
1239
1240 //------------------------------Ideal------------------------------------------
1241 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1242 return IdealIL(phase, can_reshape, T_LONG);
1243 }
1244
1245 //------------------------------Value------------------------------------------
1246 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1247 return ValueIL(phase, T_LONG);
1248 }
1249
1250 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1251 switch (bt) {
1252 case T_INT:
1253 return new RShiftINode(in1, in2);
1254 case T_LONG:
1255 return new RShiftLNode(in1, in2);
1256 default:
1257 fatal("Not implemented for %s", type2name(bt));
1258 }
1259 return nullptr;
1260 }
1261
1262
1263 //=============================================================================
1264 //------------------------------Identity---------------------------------------
1265 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1266 int count = 0;
1267 if (const_shift_count(phase, this, &count)) {
1268 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1269 // Shift by a multiple of 32/64 does nothing
1270 return in(1);
1271 }
1272 // Check for useless sign-masking
1273 int lshift_count = 0;
1274 if (in(1)->Opcode() == Op_LShift(bt) &&
1275 in(1)->req() == 3 &&
1276 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1277 // negative constant (e.g. -1 vs 31)
1278 const_shift_count(phase, in(1), &lshift_count)) {
1279 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1280 lshift_count &= bits_per_java_integer(bt) - 1;
1281 if (count == lshift_count) {
1282 // Compute masks for which this shifting doesn't change
1283 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1284 jlong hi = ~lo; // 00007FFF
1285 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1286 if (t11 == nullptr) {
1287 return this;
1288 }
1289 // Does actual value fit inside of mask?
1290 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1291 return in(1)->in(1); // Then shifting is a nop
1292 }
1293 }
1294 }
1295 }
1296 return this;
1297 }
1298
1299 Node* RShiftINode::Identity(PhaseGVN* phase) {
1300 return IdentityIL(phase, T_INT);
1301 }
1302
1303 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1304 // Inputs may be TOP if they are dead.
1305 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1306 if (t1 == nullptr) {
1307 return NodeSentinel; // Left input is an integer
1308 }
1309
1310 uint shift;
1311 Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), shift);
1312 if (shift == 0) {
1313 return NodeSentinel;
1314 }
1315
1316 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1317 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1318 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1319 const Node* and_node = in(1);
1320 if (and_node->Opcode() != Op_And(bt)) {
1321 return progress;
1322 }
1323 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1324 if (mask_t != nullptr && mask_t->is_con()) {
1325 jlong maskbits = mask_t->get_con_as_long(bt);
1326 // Convert to "(x >> shift) & (mask >> shift)"
1327 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1328 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1329 }
1330
1331 return progress;
1332 }
1333
1334 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1335 Node* progress = IdealIL(phase, can_reshape, T_INT);
1336 if (progress == NodeSentinel) {
1337 return nullptr;
1338 }
1339 if (progress != nullptr) {
1340 return progress;
1341 }
1342 uint shift;
1343 progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, shift);
1344 assert(shift != 0, "handled by IdealIL");
1345
1346 // Check for "(short[i] <<16)>>16" which simply sign-extends
1347 const Node *shl = in(1);
1348 if (shl->Opcode() != Op_LShiftI) {
1349 return progress;
1350 }
1351
1352 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1353 if (left_shift_t == nullptr) {
1354 return progress;
1355 }
1356 if (shift == 16 && left_shift_t->is_con(16)) {
1357 Node *ld = shl->in(1);
1358 if (ld->Opcode() == Op_LoadS) {
1359 // Sign extension is just useless here. Return a RShiftI of zero instead
1360 // returning 'ld' directly. We cannot return an old Node directly as
1361 // that is the job of 'Identity' calls and Identity calls only work on
1362 // direct inputs ('ld' is an extra Node removed from 'this'). The
1363 // combined optimization requires Identity only return direct inputs.
1364 set_req_X(1, ld, phase);
1365 set_req_X(2, phase->intcon(0), phase);
1366 return this;
1367 }
1368 else if (can_reshape &&
1369 ld->Opcode() == Op_LoadUS &&
1370 ld->outcnt() == 1 && ld->unique_out() == shl)
1371 // Replace zero-extension-load with sign-extension-load
1372 return ld->as_Load()->convert_to_signed_load(*phase);
1373 }
1374
1375 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1376 if (shift == 24 && left_shift_t->is_con(24)) {
1377 Node *ld = shl->in(1);
1378 if (ld->Opcode() == Op_LoadB) {
1379 // Sign extension is just useless here
1380 set_req_X(1, ld, phase);
1381 set_req_X(2, phase->intcon(0), phase);
1382 return this;
1383 }
1384 }
1385
1386 return progress;
1387 }
1388
1389 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1390 const Type* t1 = phase->type(in(1));
1391 const Type* t2 = phase->type(in(2));
1392 // Either input is TOP ==> the result is TOP
1393 if (t1 == Type::TOP) {
1394 return Type::TOP;
1395 }
1396 if (t2 == Type::TOP) {
1397 return Type::TOP;
1398 }
1399
1400 // Left input is ZERO ==> the result is ZERO.
1401 if (t1 == TypeInteger::zero(bt)) {
1402 return TypeInteger::zero(bt);
1403 }
1404 // Shift by zero does nothing
1405 if (t2 == TypeInt::ZERO) {
1406 return t1;
1407 }
1408
1409 // Either input is BOTTOM ==> the result is BOTTOM
1410 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1411 return TypeInteger::bottom(bt);
1412 }
1413
1414 const TypeInteger* r1 = t1->isa_integer(bt);
1415 const TypeInt* r2 = t2->isa_int();
1416
1417 // If the shift is a constant, just shift the bounds of the type.
1418 // For example, if the shift is 31/63, we just propagate sign bits.
1419 if (!r1->is_con() && r2->is_con()) {
1420 uint shift = r2->get_con();
1421 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1422 // Shift by a multiple of 32/64 does nothing:
1423 if (shift == 0) {
1424 return t1;
1425 }
1426 // Calculate reasonably aggressive bounds for the result.
1427 // This is necessary if we are to correctly type things
1428 // like (x<<24>>24) == ((byte)x).
1429 jlong lo = r1->lo_as_long() >> (jint)shift;
1430 jlong hi = r1->hi_as_long() >> (jint)shift;
1431 assert(lo <= hi, "must have valid bounds");
1432 #ifdef ASSERT
1433 if (bt == T_INT) {
1434 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1435 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1436 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1437 }
1438 #endif
1439 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1440 #ifdef ASSERT
1441 // Make sure we get the sign-capture idiom correct.
1442 if (shift == bits_per_java_integer(bt) - 1) {
1443 if (r1->lo_as_long() >= 0) {
1444 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1445 }
1446 if (r1->hi_as_long() < 0) {
1447 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1448 }
1449 }
1450 #endif
1451 return ti;
1452 }
1453
1454 if (!r1->is_con() || !r2->is_con()) {
1455 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1456 if (r1->lo_as_long() >= 0) {
1457 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1458 }
1459
1460 // Conversely, if the left input is negative then the result must be negative.
1461 if (r1->hi_as_long() <= -1) {
1462 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1463 }
1464
1465 return TypeInteger::bottom(bt);
1466 }
1467
1468 // Signed shift right
1469 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1470 }
1471
1472 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1473 return ValueIL(phase, T_INT);
1474 }
1475
1476 //=============================================================================
1477 //------------------------------Identity---------------------------------------
1478 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1479 return IdentityIL(phase, T_LONG);
1480 }
1481
1482 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1483 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1484 if (progress == NodeSentinel) {
1485 return nullptr;
1486 }
1487 return progress;
1488 }
1489
1490 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1491 return ValueIL(phase, T_LONG);
1492 }
1493
1494 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1495 switch (bt) {
1496 case T_INT:
1497 return new URShiftINode(in1, in2);
1498 case T_LONG:
1499 return new URShiftLNode(in1, in2);
1500 default:
1501 fatal("Not implemented for %s", type2name(bt));
1502 }
1503 return nullptr;
1504 }
1505
1506 //=============================================================================
1507 //------------------------------Identity---------------------------------------
1508 Node* URShiftINode::Identity(PhaseGVN* phase) {
1509 int count = 0;
1510 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1511 // Shift by a multiple of 32 does nothing
1512 return in(1);
1513 }
1514
1515 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1516 // Happens during new-array length computation.
1517 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1518 Node *add = in(1);
1519 if (add->Opcode() == Op_AddI) {
1520 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1521 if (t2 && t2->is_con(wordSize - 1) &&
1522 add->in(1)->Opcode() == Op_LShiftI) {
1523 // Check that shift_counts are LogBytesPerWord.
1524 Node *lshift_count = add->in(1)->in(2);
1525 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1526 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1527 t_lshift_count == phase->type(in(2))) {
1528 Node *x = add->in(1)->in(1);
1529 const TypeInt *t_x = phase->type(x)->isa_int();
1530 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1531 return x;
1532 }
1533 }
1534 }
1535 }
1536
1537 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1538 }
1539
1540 //------------------------------Ideal------------------------------------------
1541 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1542 uint con;
1543 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, con);
1544 if (con == 0) {
1545 return nullptr;
1546 }
1547
1548 // We'll be wanting the right-shift amount as a mask of that many bits
1549 const int mask = right_n_bits(BitsPerJavaInteger - con);
1550
1551 int in1_op = in(1)->Opcode();
1552
1553 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1554 if( in1_op == Op_URShiftI ) {
1555 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1556 if( t12 && t12->is_con() ) { // Right input is a constant
1557 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1558 const int con2 = t12->get_con() & 31; // Shift count is always masked
1559 const int con3 = con+con2;
1560 if( con3 < 32 ) // Only merge shifts if total is < 32
1561 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1562 }
1563 }
1564
1565 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1566 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1567 // If Q is "X << z" the rounding is useless. Look for patterns like
1568 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1569 Node *add = in(1);
1570 if (in1_op == Op_AddI) {
1571 Node *lshl = add->in(1);
1572 Node *y = add->in(2);
1573 if (lshl->Opcode() != Op_LShiftI) {
1574 lshl = add->in(2);
1575 y = add->in(1);
1576 }
1577 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1578 // negative constant (e.g. -1 vs 31)
1579 int lshl_con = 0;
1580 if (lshl->Opcode() == Op_LShiftI &&
1581 const_shift_count(phase, lshl, &lshl_con) &&
1582 (lshl_con & (BitsPerJavaInteger - 1)) == con) {
1583 Node *y_z = phase->transform(new URShiftINode(y, in(2)));
1584 Node *sum = phase->transform(new AddINode(lshl->in(1), y_z));
1585 return new AndINode(sum, phase->intcon(mask));
1586 }
1587 }
1588
1589 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1590 // This shortens the mask. Also, if we are extracting a high byte and
1591 // storing it to a buffer, the mask will be removed completely.
1592 Node *andi = in(1);
1593 if( in1_op == Op_AndI ) {
1594 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1595 if( t3 && t3->is_con() ) { // Right input is a constant
1596 jint mask2 = t3->get_con();
1597 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1598 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1599 return new AndINode(newshr, phase->intcon(mask2));
1600 // The negative values are easier to materialize than positive ones.
1601 // A typical case from address arithmetic is ((x & ~15) >> 4).
1602 // It's better to change that to ((x >> 4) & ~0) versus
1603 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1604 }
1605 }
1606
1607 // Check for "(X << z ) >>> z" which simply zero-extends
1608 Node *shl = in(1);
1609 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1610 // negative constant (e.g. -1 vs 31)
1611 int shl_con = 0;
1612 if (in1_op == Op_LShiftI &&
1613 const_shift_count(phase, shl, &shl_con) &&
1614 (shl_con & (BitsPerJavaInteger - 1)) == con)
1615 return new AndINode(shl->in(1), phase->intcon(mask));
1616
1617 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1618 const TypeInt* t2 = phase->type(in(2))->isa_int();
1619 Node *shr = in(1);
1620 if ( in1_op == Op_RShiftI ) {
1621 Node *in11 = shr->in(1);
1622 Node *in12 = shr->in(2);
1623 const TypeInt *t11 = phase->type(in11)->isa_int();
1624 const TypeInt *t12 = phase->type(in12)->isa_int();
1625 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1626 return new URShiftINode(in11, phase->intcon(31));
1627 }
1628 }
1629
1630 return progress;
1631 }
1632
1633 //------------------------------Value------------------------------------------
1634 // A URShiftINode shifts its input2 right by input1 amount.
1635 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1636 // (This is a near clone of RShiftINode::Value.)
1637 const Type *t1 = phase->type( in(1) );
1638 const Type *t2 = phase->type( in(2) );
1639 // Either input is TOP ==> the result is TOP
1640 if( t1 == Type::TOP ) return Type::TOP;
1641 if( t2 == Type::TOP ) return Type::TOP;
1642
1643 // Left input is ZERO ==> the result is ZERO.
1644 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1645 // Shift by zero does nothing
1646 if( t2 == TypeInt::ZERO ) return t1;
1647
1648 // Either input is BOTTOM ==> the result is BOTTOM
1649 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1650 return TypeInt::INT;
1651
1652 if (t2 == TypeInt::INT)
1653 return TypeInt::INT;
1654
1655 const TypeInt *r1 = t1->is_int(); // Handy access
1656 const TypeInt *r2 = t2->is_int(); // Handy access
1657
1658 if (r2->is_con()) {
1659 uint shift = r2->get_con();
1660 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1661 // Shift by a multiple of 32 does nothing:
1662 if (shift == 0) return t1;
1663 // Calculate reasonably aggressive bounds for the result.
1664 jint lo = (juint)r1->_lo >> (juint)shift;
1665 jint hi = (juint)r1->_hi >> (juint)shift;
1666 if (r1->_hi >= 0 && r1->_lo < 0) {
1667 // If the type has both negative and positive values,
1668 // there are two separate sub-domains to worry about:
1669 // The positive half and the negative half.
1670 jint neg_lo = lo;
1671 jint neg_hi = (juint)-1 >> (juint)shift;
1672 jint pos_lo = (juint) 0 >> (juint)shift;
1673 jint pos_hi = hi;
1674 lo = MIN2(neg_lo, pos_lo); // == 0
1675 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1676 }
1677 assert(lo <= hi, "must have valid bounds");
1678 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1679 #ifdef ASSERT
1680 // Make sure we get the sign-capture idiom correct.
1681 if (shift == BitsPerJavaInteger-1) {
1682 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1683 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1684 }
1685 #endif
1686 return ti;
1687 }
1688
1689 //
1690 // Do not support shifted oops in info for GC
1691 //
1692 // else if( t1->base() == Type::InstPtr ) {
1693 //
1694 // const TypeInstPtr *o = t1->is_instptr();
1695 // if( t1->singleton() )
1696 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1697 // }
1698 // else if( t1->base() == Type::KlassPtr ) {
1699 // const TypeKlassPtr *o = t1->is_klassptr();
1700 // if( t1->singleton() )
1701 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1702 // }
1703
1704 return TypeInt::INT;
1705 }
1706
1707 //=============================================================================
1708 //------------------------------Identity---------------------------------------
1709 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1710 int count = 0;
1711 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1712 // Shift by a multiple of 64 does nothing
1713 return in(1);
1714 }
1715 return this;
1716 }
1717
1718 //------------------------------Ideal------------------------------------------
1719 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1720 uint con;
1721 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong, con);
1722 if (con == 0) {
1723 return nullptr;
1724 }
1725
1726 // We'll be wanting the right-shift amount as a mask of that many bits
1727 const jlong mask = jlong(max_julong >> con);
1728
1729 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1730 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1731 // If Q is "X << z" the rounding is useless. Look for patterns like
1732 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1733 Node *add = in(1);
1734 const TypeInt *t2 = phase->type(in(2))->isa_int();
1735 if (add->Opcode() == Op_AddL) {
1736 Node *lshl = add->in(1);
1737 Node *y = add->in(2);
1738 if (lshl->Opcode() != Op_LShiftL) {
1739 lshl = add->in(2);
1740 y = add->in(1);
1741 }
1742 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1743 // negative constant (e.g. -1 vs 63)
1744 int lshl_con = 0;
1745 if (lshl->Opcode() == Op_LShiftL &&
1746 const_shift_count(phase, lshl, &lshl_con) &&
1747 (lshl_con & (BitsPerJavaLong - 1)) == con) {
1748 Node* y_z = phase->transform(new URShiftLNode(y, in(2)));
1749 Node* sum = phase->transform(new AddLNode(lshl->in(1), y_z));
1750 return new AndLNode(sum, phase->longcon(mask));
1751 }
1752 }
1753
1754 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1755 // This shortens the mask. Also, if we are extracting a high byte and
1756 // storing it to a buffer, the mask will be removed completely.
1757 Node *andi = in(1);
1758 if( andi->Opcode() == Op_AndL ) {
1759 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1760 if( t3 && t3->is_con() ) { // Right input is a constant
1761 jlong mask2 = t3->get_con();
1762 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1763 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1764 return new AndLNode(newshr, phase->longcon(mask2));
1765 }
1766 }
1767
1768 // Check for "(X << z ) >>> z" which simply zero-extends
1769 Node *shl = in(1);
1770 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1771 // negative constant (e.g. -1 vs 63)
1772 int shl_con = 0;
1773 if (shl->Opcode() == Op_LShiftL &&
1774 const_shift_count(phase, shl, &shl_con) &&
1775 (shl_con & (BitsPerJavaLong - 1)) == con) {
1776 return new AndLNode(shl->in(1), phase->longcon(mask));
1777 }
1778
1779 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1780 Node *shr = in(1);
1781 if ( shr->Opcode() == Op_RShiftL ) {
1782 Node *in11 = shr->in(1);
1783 Node *in12 = shr->in(2);
1784 const TypeLong *t11 = phase->type(in11)->isa_long();
1785 const TypeInt *t12 = phase->type(in12)->isa_int();
1786 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1787 return new URShiftLNode(in11, phase->intcon(63));
1788 }
1789 }
1790
1791 return progress;
1792 }
1793
1794 //------------------------------Value------------------------------------------
1795 // A URShiftINode shifts its input2 right by input1 amount.
1796 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1797 // (This is a near clone of RShiftLNode::Value.)
1798 const Type *t1 = phase->type( in(1) );
1799 const Type *t2 = phase->type( in(2) );
1800 // Either input is TOP ==> the result is TOP
1801 if( t1 == Type::TOP ) return Type::TOP;
1802 if( t2 == Type::TOP ) return Type::TOP;
1803
1804 // Left input is ZERO ==> the result is ZERO.
1805 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1806 // Shift by zero does nothing
1807 if( t2 == TypeInt::ZERO ) return t1;
1808
1809 // Either input is BOTTOM ==> the result is BOTTOM
1810 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1811 return TypeLong::LONG;
1812
1813 if (t2 == TypeInt::INT)
1814 return TypeLong::LONG;
1815
1816 const TypeLong *r1 = t1->is_long(); // Handy access
1817 const TypeInt *r2 = t2->is_int (); // Handy access
1818
1819 if (r2->is_con()) {
1820 uint shift = r2->get_con();
1821 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1822 // Shift by a multiple of 64 does nothing:
1823 if (shift == 0) return t1;
1824 // Calculate reasonably aggressive bounds for the result.
1825 jlong lo = (julong)r1->_lo >> (juint)shift;
1826 jlong hi = (julong)r1->_hi >> (juint)shift;
1827 if (r1->_hi >= 0 && r1->_lo < 0) {
1828 // If the type has both negative and positive values,
1829 // there are two separate sub-domains to worry about:
1830 // The positive half and the negative half.
1831 jlong neg_lo = lo;
1832 jlong neg_hi = (julong)-1 >> (juint)shift;
1833 jlong pos_lo = (julong) 0 >> (juint)shift;
1834 jlong pos_hi = hi;
1835 //lo = MIN2(neg_lo, pos_lo); // == 0
1836 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1837 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1838 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1839 }
1840 assert(lo <= hi, "must have valid bounds");
1841 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1842 #ifdef ASSERT
1843 // Make sure we get the sign-capture idiom correct.
1844 if (shift == BitsPerJavaLong - 1) {
1845 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1846 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1847 }
1848 #endif
1849 return tl;
1850 }
1851
1852 return TypeLong::LONG; // Give up
1853 }
1854
1855 //=============================================================================
1856 //------------------------------Ideal------------------------------------------
1857 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1858 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1859 // This reduces the number of rules in the matcher, as we only need to check
1860 // for negations on the second argument, and not the symmetric case where
1861 // the first argument is negated.
1862 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1863 swap_edges(1, 2);
1864 return this;
1865 }
1866 return nullptr;
1867 }
1868
1869 //=============================================================================
1870 //------------------------------Value------------------------------------------
1871 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1872 const Type *t1 = phase->type(in(1));
1873 if (t1 == Type::TOP) return Type::TOP;
1874 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1875 const Type *t2 = phase->type(in(2));
1876 if (t2 == Type::TOP) return Type::TOP;
1877 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1878 const Type *t3 = phase->type(in(3));
1879 if (t3 == Type::TOP) return Type::TOP;
1880 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1881 #ifndef __STDC_IEC_559__
1882 return Type::DOUBLE;
1883 #else
1884 double d1 = t1->getd();
1885 double d2 = t2->getd();
1886 double d3 = t3->getd();
1887 return TypeD::make(fma(d1, d2, d3));
1888 #endif
1889 }
1890
1891 //=============================================================================
1892 //------------------------------Value------------------------------------------
1893 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1894 const Type *t1 = phase->type(in(1));
1895 if (t1 == Type::TOP) return Type::TOP;
1896 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1897 const Type *t2 = phase->type(in(2));
1898 if (t2 == Type::TOP) return Type::TOP;
1899 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1900 const Type *t3 = phase->type(in(3));
1901 if (t3 == Type::TOP) return Type::TOP;
1902 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1903 #ifndef __STDC_IEC_559__
1904 return Type::FLOAT;
1905 #else
1906 float f1 = t1->getf();
1907 float f2 = t2->getf();
1908 float f3 = t3->getf();
1909 return TypeF::make(fma(f1, f2, f3));
1910 #endif
1911 }
1912
1913 //=============================================================================
1914 //------------------------------Value------------------------------------------
1915 const Type* FmaHFNode::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::HalfFloatCon) { return Type::HALF_FLOAT; }
1919 const Type* t2 = phase->type(in(2));
1920 if (t2 == Type::TOP) { return Type::TOP; }
1921 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1922 const Type* t3 = phase->type(in(3));
1923 if (t3 == Type::TOP) { return Type::TOP; }
1924 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1925 #ifndef __STDC_IEC_559__
1926 return Type::HALF_FLOAT;
1927 #else
1928 float f1 = t1->getf();
1929 float f2 = t2->getf();
1930 float f3 = t3->getf();
1931 return TypeH::make(fma(f1, f2, f3));
1932 #endif
1933 }
1934
1935 //=============================================================================
1936 //------------------------------hash-------------------------------------------
1937 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1938 // The hash function must return the same value when edge swapping is performed.
1939 uint MulAddS2INode::hash() const {
1940 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1941 }
1942
1943 //------------------------------Rotate Operations ------------------------------
1944
1945 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1946 const Type* t1 = phase->type(in(1));
1947 if (t1 == Type::TOP) {
1948 return this;
1949 }
1950 int count = 0;
1951 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1952 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1953 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1954 // Rotate by a multiple of 32/64 does nothing
1955 return in(1);
1956 }
1957 return this;
1958 }
1959
1960 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1961 const Type* t1 = phase->type(in(1));
1962 const Type* t2 = phase->type(in(2));
1963 // Either input is TOP ==> the result is TOP
1964 if (t1 == Type::TOP || t2 == Type::TOP) {
1965 return Type::TOP;
1966 }
1967
1968 if (t1->isa_int()) {
1969 const TypeInt* r1 = t1->is_int();
1970 const TypeInt* r2 = t2->is_int();
1971
1972 // Left input is ZERO ==> the result is ZERO.
1973 if (r1 == TypeInt::ZERO) {
1974 return TypeInt::ZERO;
1975 }
1976 // Rotate by zero does nothing
1977 if (r2 == TypeInt::ZERO) {
1978 return r1;
1979 }
1980 if (r1->is_con() && r2->is_con()) {
1981 juint r1_con = (juint)r1->get_con();
1982 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1983 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
1984 }
1985 return TypeInt::INT;
1986 } else {
1987 assert(t1->isa_long(), "Type must be a long");
1988 const TypeLong* r1 = t1->is_long();
1989 const TypeInt* r2 = t2->is_int();
1990
1991 // Left input is ZERO ==> the result is ZERO.
1992 if (r1 == TypeLong::ZERO) {
1993 return TypeLong::ZERO;
1994 }
1995 // Rotate by zero does nothing
1996 if (r2 == TypeInt::ZERO) {
1997 return r1;
1998 }
1999 if (r1->is_con() && r2->is_con()) {
2000 julong r1_con = (julong)r1->get_con();
2001 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2002 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2003 }
2004 return TypeLong::LONG;
2005 }
2006 }
2007
2008 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2009 const Type* t1 = phase->type(in(1));
2010 const Type* t2 = phase->type(in(2));
2011 if (t2->isa_int() && t2->is_int()->is_con()) {
2012 if (t1->isa_int()) {
2013 int lshift = t2->is_int()->get_con() & 31;
2014 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2015 } else if (t1 != Type::TOP) {
2016 assert(t1->isa_long(), "Type must be a long");
2017 int lshift = t2->is_int()->get_con() & 63;
2018 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2019 }
2020 }
2021 return nullptr;
2022 }
2023
2024 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2025 const Type* t1 = phase->type(in(1));
2026 if (t1 == Type::TOP) {
2027 return this;
2028 }
2029 int count = 0;
2030 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2031 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2032 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2033 // Rotate by a multiple of 32/64 does nothing
2034 return in(1);
2035 }
2036 return this;
2037 }
2038
2039 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2040 const Type* t1 = phase->type(in(1));
2041 const Type* t2 = phase->type(in(2));
2042 // Either input is TOP ==> the result is TOP
2043 if (t1 == Type::TOP || t2 == Type::TOP) {
2044 return Type::TOP;
2045 }
2046
2047 if (t1->isa_int()) {
2048 const TypeInt* r1 = t1->is_int();
2049 const TypeInt* r2 = t2->is_int();
2050
2051 // Left input is ZERO ==> the result is ZERO.
2052 if (r1 == TypeInt::ZERO) {
2053 return TypeInt::ZERO;
2054 }
2055 // Rotate by zero does nothing
2056 if (r2 == TypeInt::ZERO) {
2057 return r1;
2058 }
2059 if (r1->is_con() && r2->is_con()) {
2060 juint r1_con = (juint)r1->get_con();
2061 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2062 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2063 }
2064 return TypeInt::INT;
2065 } else {
2066 assert(t1->isa_long(), "Type must be a long");
2067 const TypeLong* r1 = t1->is_long();
2068 const TypeInt* r2 = t2->is_int();
2069 // Left input is ZERO ==> the result is ZERO.
2070 if (r1 == TypeLong::ZERO) {
2071 return TypeLong::ZERO;
2072 }
2073 // Rotate by zero does nothing
2074 if (r2 == TypeInt::ZERO) {
2075 return r1;
2076 }
2077 if (r1->is_con() && r2->is_con()) {
2078 julong r1_con = (julong)r1->get_con();
2079 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2080 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2081 }
2082 return TypeLong::LONG;
2083 }
2084 }
2085
2086 //------------------------------ Sum & Mask ------------------------------
2087
2088 // Returns a lower bound on the number of trailing zeros in expr.
2089 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2090 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2091 if (type == nullptr) {
2092 return 0;
2093 }
2094
2095 expr = expr->uncast();
2096 type = phase->type(expr)->isa_integer(bt);
2097 if (type == nullptr) {
2098 return 0;
2099 }
2100
2101 if (type->is_con()) {
2102 jlong con = type->get_con_as_long(bt);
2103 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2104 }
2105
2106 if (expr->Opcode() == Op_ConvI2L) {
2107 expr = expr->in(1)->uncast();
2108 bt = T_INT;
2109 type = phase->type(expr)->isa_int();
2110 }
2111
2112 // Pattern: expr = (x << shift)
2113 if (expr->Opcode() == Op_LShift(bt)) {
2114 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2115 if (shift_t == nullptr || !shift_t->is_con()) {
2116 return 0;
2117 }
2118 // We need to truncate the shift, as it may not have been canonicalized yet.
2119 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2120 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2121 // (JLS: "Shift Operators")
2122 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2123 return shift_t->get_con() & shift_mask;
2124 }
2125
2126 return 0;
2127 }
2128
2129 // Checks whether expr is neutral additive element (zero) under mask,
2130 // i.e. whether an expression of the form:
2131 // (AndX (AddX (expr addend) mask)
2132 // (expr + addend) & mask
2133 // is equivalent to
2134 // (AndX addend mask)
2135 // addend & mask
2136 // for any addend.
2137 // (The X in AndX must be I or L, depending on bt).
2138 //
2139 // We check for the sufficient condition when the lowest set bit in expr is higher than
2140 // the highest set bit in mask, i.e.:
2141 // expr: eeeeee0000000000000
2142 // mask: 000000mmmmmmmmmmmmm
2143 // <--w bits--->
2144 // We do not test for other cases.
2145 //
2146 // Correctness:
2147 // Given "expr" with at least "w" trailing zeros,
2148 // let "mod = 2^w", "suffix_mask = mod - 1"
2149 //
2150 // Since "mask" only has bits set where "suffix_mask" does, we have:
2151 // mask = suffix_mask & mask (SUFFIX_MASK)
2152 //
2153 // And since expr only has bits set above w, and suffix_mask only below:
2154 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2155 //
2156 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2157 // a power of 2, and we are computing in a ring of powers of 2, we know that
2158 // (x + y) % mod = (x % mod + y) % mod
2159 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2160 //
2161 // We can now prove the equality:
2162 // (expr + addend) & mask
2163 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2164 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2165 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2166 // = addend & mask (SUFFIX_MASK)
2167 //
2168 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2169 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2170 // When the mask is negative, it has the most significant bit set.
2171 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2172 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2173 return false;
2174 }
2175
2176 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2177 if (mask_t->hi_as_long() == 0) {
2178 assert(mask_t->lo_as_long() == 0, "checked earlier");
2179 return false;
2180 }
2181
2182 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2183 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2184 return expr_trailing_zeros >= mask_bit_width;
2185 }
2186
2187 // Reduces the pattern:
2188 // (AndX (AddX add1 add2) mask)
2189 // to
2190 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2191 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2192 Node* add = in(1);
2193 Node* mask = in(2);
2194 int addidx = 0;
2195 if (add->Opcode() == Op_Add(bt)) {
2196 addidx = 1;
2197 } else if (mask->Opcode() == Op_Add(bt)) {
2198 mask = add;
2199 addidx = 2;
2200 add = in(addidx);
2201 }
2202 if (addidx > 0) {
2203 Node* add1 = add->in(1);
2204 Node* add2 = add->in(2);
2205 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2206 set_req_X(addidx, add2, phase);
2207 return this;
2208 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2209 set_req_X(addidx, add1, phase);
2210 return this;
2211 }
2212 }
2213 return nullptr;
2214 }