1 /*
2 * Copyright (c) 1997, 2025, 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/subnode.hpp"
33 #include "utilities/powerOfTwo.hpp"
34
35 // Portions of code courtesy of Clifford Click
36
37
38 //=============================================================================
39 //------------------------------hash-------------------------------------------
40 // Hash function over MulNodes. Needs to be commutative; i.e., I swap
41 // (commute) inputs to MulNodes willy-nilly so the hash function must return
42 // the same value in the presence of edge swapping.
43 uint MulNode::hash() const {
44 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode();
45 }
46
47 //------------------------------Identity---------------------------------------
48 // Multiplying a one preserves the other argument
49 Node* MulNode::Identity(PhaseGVN* phase) {
50 const Type *one = mul_id(); // The multiplicative identity
51 if( phase->type( in(1) )->higher_equal( one ) ) return in(2);
52 if( phase->type( in(2) )->higher_equal( one ) ) return in(1);
53
54 return this;
55 }
56
57 //------------------------------Ideal------------------------------------------
58 // We also canonicalize the Node, moving constants to the right input,
59 // and flatten expressions (so that 1+x+2 becomes x+3).
60 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) {
61 Node* in1 = in(1);
62 Node* in2 = in(2);
63 Node* progress = nullptr; // Progress flag
64
65 // This code is used by And nodes too, but some conversions are
66 // only valid for the actual Mul nodes.
67 uint op = Opcode();
68 bool real_mul = (op == Op_MulI) || (op == Op_MulL) ||
69 (op == Op_MulF) || (op == Op_MulD) ||
70 (op == Op_MulHF);
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 op != Op_MulHF) {
127 if( t2 == Type::TOP ) return nullptr;
128 Node *mul1 = in(1);
129 #ifdef ASSERT
130 // Check for dead loop
131 int op1 = mul1->Opcode();
132 if ((mul1 == this) || (in(2) == this) ||
133 ((op1 == mul_opcode() || op1 == add_opcode()) &&
134 ((mul1->in(1) == this) || (mul1->in(2) == this) ||
135 (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) {
136 assert(false, "dead loop in MulNode::Ideal");
137 }
138 #endif
139
140 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply?
141 // Mul of a constant?
142 const Type *t12 = phase->type( mul1->in(2) );
143 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
144 // Compute new constant; check for overflow
145 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
146 if( tcon01->singleton() ) {
147 // The Mul of the flattened expression
148 set_req_X(1, mul1->in(1), phase);
149 set_req_X(2, phase->makecon(tcon01), phase);
150 t2 = tcon01;
151 progress = this; // Made progress
152 }
153 }
154 }
155 // If the right input is a constant, and the left input is an add of a
156 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
157 const Node *add1 = in(1);
158 if( add1->Opcode() == add_opcode() ) { // Left input is an add?
159 // Add of a constant?
160 const Type *t12 = phase->type( add1->in(2) );
161 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
162 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
163 // Compute new constant; check for overflow
164 const Type *tcon01 = mul_ring(t2,t12);
165 if( tcon01->singleton() ) {
166
167 // Convert (X+con1)*con0 into X*con0
168 Node *mul = clone(); // mul = ()*con0
169 mul->set_req(1,add1->in(1)); // mul = X*con0
170 mul = phase->transform(mul);
171
172 Node *add2 = add1->clone();
173 add2->set_req(1, mul); // X*con0 + con0*con1
174 add2->set_req(2, phase->makecon(tcon01) );
175 progress = add2;
176 }
177 }
178 } // End of is left input an add
179 } // End of is right input a Mul
180
181 return progress;
182 }
183
184 //------------------------------Value-----------------------------------------
185 const Type* MulNode::Value(PhaseGVN* phase) const {
186 const Type *t1 = phase->type( in(1) );
187 const Type *t2 = phase->type( in(2) );
188 // Either input is TOP ==> the result is TOP
189 if( t1 == Type::TOP ) return Type::TOP;
190 if( t2 == Type::TOP ) return Type::TOP;
191
192 // Either input is ZERO ==> the result is ZERO.
193 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
194 int op = Opcode();
195 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
196 const Type *zero = add_id(); // The multiplicative zero
197 if( t1->higher_equal( zero ) ) return zero;
198 if( t2->higher_equal( zero ) ) return zero;
199 }
200
201 // Either input is BOTTOM ==> the result is the local BOTTOM
202 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
203 return bottom_type();
204
205 #if defined(IA32)
206 // Can't trust native compilers to properly fold strict double
207 // multiplication with round-to-zero on this platform.
208 if (op == Op_MulD) {
209 return TypeD::DOUBLE;
210 }
211 #endif
212
213 return mul_ring(t1,t2); // Local flavor of type multiplication
214 }
215
216 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
217 switch (bt) {
218 case T_INT:
219 return new MulINode(in1, in2);
220 case T_LONG:
221 return new MulLNode(in1, in2);
222 default:
223 fatal("Not implemented for %s", type2name(bt));
224 }
225 return nullptr;
226 }
227
228 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
229 switch (bt) {
230 case T_INT:
231 return new AndINode(in1, in2);
232 case T_LONG:
233 return new AndLNode(in1, in2);
234 default:
235 fatal("Not implemented for %s", type2name(bt));
236 }
237 return nullptr;
238 }
239
240
241 //=============================================================================
242 //------------------------------Ideal------------------------------------------
243 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
244 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
245 const jint con = in(2)->find_int_con(0);
246 if (con == 0) {
247 // If in(2) is not a constant, call Ideal() of the parent class to
248 // try to move constant to the right side.
249 return MulNode::Ideal(phase, can_reshape);
250 }
251
252 // Now we have a constant Node on the right and the constant in con.
253 if (con == 1) {
254 // By one is handled by Identity call
255 return nullptr;
256 }
257
258 // Check for negative constant; if so negate the final result
259 bool sign_flip = false;
260
261 unsigned int abs_con = uabs(con);
262 if (abs_con != (unsigned int)con) {
263 sign_flip = true;
264 }
265
266 // Get low bit; check for being the only bit
267 Node *res = nullptr;
268 unsigned int bit1 = submultiple_power_of_2(abs_con);
269 if (bit1 == abs_con) { // Found a power of 2?
270 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
271 } else {
272 // Check for constant with 2 bits set
273 unsigned int bit2 = abs_con - bit1;
274 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
275 if (bit2 + bit1 == abs_con) { // Found all bits in con?
276 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
277 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
278 res = new AddINode(n2, n1);
279 } else if (is_power_of_2(abs_con + 1)) {
280 // Sleezy: power-of-2 - 1. Next time be generic.
281 unsigned int temp = abs_con + 1;
282 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
283 res = new SubINode(n1, in(1));
284 } else {
285 return MulNode::Ideal(phase, can_reshape);
286 }
287 }
288
289 if (sign_flip) { // Need to negate result?
290 res = phase->transform(res);// Transform, before making the zero con
291 res = new SubINode(phase->intcon(0),res);
292 }
293
294 return res; // Return final result
295 }
296
297 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
298 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
299 //
300 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
301 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
302 // - no overflow (i.e. product fits into x bits)
303 // - overflow (i.e. product does not fit into x bits)
304 //
305 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
306 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
307 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
308 // The cross product consists of the following four multiplications with 2x-bit results:
309 // (1) left_lo * right_lo
310 // (2) left_lo * right_hi
311 // (3) left_hi * right_lo
312 // (4) left_hi * right_hi
313 //
314 // Let's define the following two functions:
315 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
316 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
317 //
318 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
319 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
320 // Ux(p) = 0
321 // Lx(p) = p
322 //
323 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
324 // with x bits:
325 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
326 //
327 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
328 // (i.e. all values in the x-bit space could be possible):
329 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
330 //
331 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
332 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
333 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
334 // the lower x bits the same number of times:
335 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
336 //
337 // If all upper x bits are equal, we can conclude:
338 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
339 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
340 //
341 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
342 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
343 //
344 //
345 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
346 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
347 // 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,
348 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
349 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
350 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
351 // that there is no underflow.
352 //
353 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
354 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
355 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
356 // can compare the upper x bits in a unified way with N() being the normalization function:
357 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
358 template<typename NativeType>
359 class IntegerTypeMultiplication {
360
361 NativeType _lo_left;
362 NativeType _lo_right;
363 NativeType _hi_left;
364 NativeType _hi_right;
365 short _widen_left;
366 short _widen_right;
367
368 static const Type* overflow_type();
369 static NativeType multiply_high(NativeType x, NativeType y);
370 const Type* create_type(NativeType lo, NativeType hi) const;
371
372 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
373 return normalize_overflow_value(x, y, multiply_high(x, y));
374 }
375
376 bool cross_product_not_same_overflow_value() const {
377 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
378 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
379 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
380 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
381 return lo_lo_high_product != lo_hi_high_product ||
382 lo_hi_high_product != hi_lo_high_product ||
383 hi_lo_high_product != hi_hi_high_product;
384 }
385
386 bool does_product_overflow(NativeType x, NativeType y) const {
387 return multiply_high_signed_overflow_value(x, y) != 0;
388 }
389
390 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
391 return java_multiply(x, y) < 0 ? result + 1 : result;
392 }
393
394 public:
395 template<class IntegerType>
396 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
397 : _lo_left(left->_lo), _lo_right(right->_lo),
398 _hi_left(left->_hi), _hi_right(right->_hi),
399 _widen_left(left->_widen), _widen_right(right->_widen) {}
400
401 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
402 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
403 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
404 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
405 const Type* compute() const {
406 if (cross_product_not_same_overflow_value()) {
407 return overflow_type();
408 }
409
410 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
411 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
412 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
413 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
414 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
415 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
416 return create_type(min, max);
417 }
418
419 bool does_overflow() const {
420 return does_product_overflow(_lo_left, _lo_right) ||
421 does_product_overflow(_lo_left, _hi_right) ||
422 does_product_overflow(_hi_left, _lo_right) ||
423 does_product_overflow(_hi_left, _hi_right);
424 }
425 };
426
427 template <>
428 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
429 return TypeInt::INT;
430 }
431
432 template <>
433 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
434 const jlong x_64 = x;
435 const jlong y_64 = y;
436 const jlong product = x_64 * y_64;
437 return (jint)((uint64_t)product >> 32u);
438 }
439
440 template <>
441 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
442 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
443 }
444
445 template <>
446 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
447 return TypeLong::LONG;
448 }
449
450 template <>
451 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
452 return multiply_high_signed(x, y);
453 }
454
455 template <>
456 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
457 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
458 }
459
460 // Compute the product type of two integer ranges into this node.
461 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
462 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
463 return integer_multiplication.compute();
464 }
465
466 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
467 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
468 return integer_multiplication.does_overflow();
469 }
470
471 // Compute the product type of two long ranges into this node.
472 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
473 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
474 return integer_multiplication.compute();
475 }
476
477 //=============================================================================
478 //------------------------------Ideal------------------------------------------
479 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
480 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
481 const jlong con = in(2)->find_long_con(0);
482 if (con == 0) {
483 // If in(2) is not a constant, call Ideal() of the parent class to
484 // try to move constant to the right side.
485 return MulNode::Ideal(phase, can_reshape);
486 }
487
488 // Now we have a constant Node on the right and the constant in con.
489 if (con == 1) {
490 // By one is handled by Identity call
491 return nullptr;
492 }
493
494 // Check for negative constant; if so negate the final result
495 bool sign_flip = false;
496 julong abs_con = uabs(con);
497 if (abs_con != (julong)con) {
498 sign_flip = true;
499 }
500
501 // Get low bit; check for being the only bit
502 Node *res = nullptr;
503 julong bit1 = submultiple_power_of_2(abs_con);
504 if (bit1 == abs_con) { // Found a power of 2?
505 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
506 } else {
507
508 // Check for constant with 2 bits set
509 julong bit2 = abs_con-bit1;
510 bit2 = bit2 & (0-bit2); // Extract 2nd bit
511 if (bit2 + bit1 == abs_con) { // Found all bits in con?
512 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
513 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
514 res = new AddLNode(n2, n1);
515
516 } else if (is_power_of_2(abs_con+1)) {
517 // Sleezy: power-of-2 -1. Next time be generic.
518 julong temp = abs_con + 1;
519 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
520 res = new SubLNode(n1, in(1));
521 } else {
522 return MulNode::Ideal(phase, can_reshape);
523 }
524 }
525
526 if (sign_flip) { // Need to negate result?
527 res = phase->transform(res);// Transform, before making the zero con
528 res = new SubLNode(phase->longcon(0),res);
529 }
530
531 return res; // Return final result
532 }
533
534 //=============================================================================
535 //------------------------------mul_ring---------------------------------------
536 // Compute the product type of two double ranges into this node.
537 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
538 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
539 return TypeF::make( t0->getf() * t1->getf() );
540 }
541
542 //------------------------------Ideal---------------------------------------
543 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
544 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
545 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
546
547 // x * 2 -> x + x
548 if (t2 != nullptr && t2->getf() == 2) {
549 Node* base = in(1);
550 return new AddFNode(base, base);
551 }
552 return MulNode::Ideal(phase, can_reshape);
553 }
554
555 //=============================================================================
556 //------------------------------Ideal------------------------------------------
557 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
558 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
559 const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
560
561 // x * 2 -> x + x
562 if (t2 != nullptr && t2->getf() == 2) {
563 Node* base = in(1);
564 return new AddHFNode(base, base);
565 }
566 return MulNode::Ideal(phase, can_reshape);
567 }
568
569 // Compute the product type of two half float ranges into this node.
570 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
571 if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
572 return Type::HALF_FLOAT;
573 }
574 return TypeH::make(t0->getf() * t1->getf());
575 }
576
577 //=============================================================================
578 //------------------------------mul_ring---------------------------------------
579 // Compute the product type of two double ranges into this node.
580 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
581 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
582 // We must be multiplying 2 double constants.
583 return TypeD::make( t0->getd() * t1->getd() );
584 }
585
586 //------------------------------Ideal---------------------------------------
587 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
588 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
589 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
590
591 // x * 2 -> x + x
592 if (t2 != nullptr && t2->getd() == 2) {
593 Node* base = in(1);
594 return new AddDNode(base, base);
595 }
596
597 return MulNode::Ideal(phase, can_reshape);
598 }
599
600 //=============================================================================
601 //------------------------------Value------------------------------------------
602 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
603 const Type *t1 = phase->type( in(1) );
604 const Type *t2 = phase->type( in(2) );
605 const Type *bot = bottom_type();
606 return MulHiValue(t1, t2, bot);
607 }
608
609 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
610 const Type *t1 = phase->type( in(1) );
611 const Type *t2 = phase->type( in(2) );
612 const Type *bot = bottom_type();
613 return MulHiValue(t1, t2, bot);
614 }
615
616 // A common routine used by UMulHiLNode and MulHiLNode
617 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
618 // Either input is TOP ==> the result is TOP
619 if( t1 == Type::TOP ) return Type::TOP;
620 if( t2 == Type::TOP ) return Type::TOP;
621
622 // Either input is BOTTOM ==> the result is the local BOTTOM
623 if( (t1 == bot) || (t2 == bot) ||
624 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
625 return bot;
626
627 // It is not worth trying to constant fold this stuff!
628 return TypeLong::LONG;
629 }
630
631 template<typename IntegerType>
632 static const IntegerType* and_value(const IntegerType* r0, const IntegerType* r1) {
633 typedef typename IntegerType::NativeType NativeType;
634 static_assert(std::is_signed<NativeType>::value, "Native type of IntegerType must be signed!");
635
636 int widen = MAX2(r0->_widen, r1->_widen);
637
638 // If both types are constants, we can calculate a constant result.
639 if (r0->is_con() && r1->is_con()) {
640 return IntegerType::make(r0->get_con() & r1->get_con());
641 }
642
643 // If both ranges are positive, the result will range from 0 up to the hi value of the smaller range. The minimum
644 // 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.
645 if (r0->_lo >= 0 && r1->_lo >= 0) {
646 return IntegerType::make(0, MIN2(r0->_hi, r1->_hi), widen);
647 }
648
649 // If only one range is positive, the result will range from 0 up to that range's maximum value.
650 // For the operation 'x & C' where C is a positive constant, the result will be in the range [0..C]. With that observation,
651 // 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]'
652 // where c >= 0 will be in the range [0..C].
653 if (r0->_lo >= 0) {
654 return IntegerType::make(0, r0->_hi, widen);
655 }
656
657 if (r1->_lo >= 0) {
658 return IntegerType::make(0, r1->_hi, widen);
659 }
660
661 // At this point, all positive ranges will have already been handled, so the only remaining cases will be negative ranges
662 // and constants.
663
664 assert(r0->_lo < 0 && r1->_lo < 0, "positive ranges should already be handled!");
665
666 // As two's complement means that both numbers will start with leading 1s, the lower bound of both ranges will contain
667 // the common leading 1s of both minimum values. In order to count them with count_leading_zeros, the bits are inverted.
668 NativeType sel_val = ~MIN2(r0->_lo, r1->_lo);
669
670 NativeType min;
671 if (sel_val == 0) {
672 // Since count_leading_zeros is undefined at 0, we short-circuit the condition where both ranges have a minimum of -1.
673 min = -1;
674 } else {
675 // 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.
676 int shift_bits = count_leading_zeros(sel_val) - 1;
677 min = std::numeric_limits<NativeType>::min() >> shift_bits;
678 }
679
680 NativeType max;
681 if (r0->_hi < 0 && r1->_hi < 0) {
682 // If both ranges are negative, then the same optimization as both positive ranges will apply, and the smaller hi
683 // value will mask off any bits set by higher values.
684 max = MIN2(r0->_hi, r1->_hi);
685 } else {
686 // In the case of ranges that cross zero, negative values can cause the higher order bits to be set, so the maximum
687 // positive value can be as high as the larger hi value.
688 max = MAX2(r0->_hi, r1->_hi);
689 }
690
691 return IntegerType::make(min, max, widen);
692 }
693
694 //=============================================================================
695 //------------------------------mul_ring---------------------------------------
696 // Supplied function returns the product of the inputs IN THE CURRENT RING.
697 // For the logical operations the ring's MUL is really a logical AND function.
698 // This also type-checks the inputs for sanity. Guaranteed never to
699 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
700 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
701 const TypeInt* r0 = t0->is_int();
702 const TypeInt* r1 = t1->is_int();
703
704 return and_value<TypeInt>(r0, r1);
705 }
706
707 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
708
709 const Type* AndINode::Value(PhaseGVN* phase) const {
710 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
711 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
712 return TypeInt::ZERO;
713 }
714
715 return MulNode::Value(phase);
716 }
717
718 //------------------------------Identity---------------------------------------
719 // Masking off the high bits of an unsigned load is not required
720 Node* AndINode::Identity(PhaseGVN* phase) {
721
722 // x & x => x
723 if (in(1) == in(2)) {
724 return in(1);
725 }
726
727 Node* in1 = in(1);
728 uint op = in1->Opcode();
729 const TypeInt* t2 = phase->type(in(2))->isa_int();
730 if (t2 && t2->is_con()) {
731 int con = t2->get_con();
732 // Masking off high bits which are always zero is useless.
733 const TypeInt* t1 = phase->type(in(1))->isa_int();
734 if (t1 != nullptr && t1->_lo >= 0) {
735 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
736 if ((t1_support & con) == t1_support)
737 return in1;
738 }
739 // Masking off the high bits of a unsigned-shift-right is not
740 // needed either.
741 if (op == Op_URShiftI) {
742 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
743 if (t12 && t12->is_con()) { // Shift is by a constant
744 int shift = t12->get_con();
745 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
746 int mask = max_juint >> shift;
747 if ((mask & con) == mask) // If AND is useless, skip it
748 return in1;
749 }
750 }
751 }
752 return MulNode::Identity(phase);
753 }
754
755 //------------------------------Ideal------------------------------------------
756 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
757 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
758 Node* progress = AndIL_sum_and_mask(phase, T_INT);
759 if (progress != nullptr) {
760 return progress;
761 }
762
763 // Convert "(~a) & (~b)" into "~(a | b)"
764 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
765 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
766 Node* tn = phase->transform(or_a_b);
767 return AddNode::make_not(phase, tn, T_INT);
768 }
769
770 // Special case constant AND mask
771 const TypeInt *t2 = phase->type( in(2) )->isa_int();
772 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
773 const int mask = t2->get_con();
774 Node *load = in(1);
775 uint lop = load->Opcode();
776
777 // Masking bits off of a Character? Hi bits are already zero.
778 if( lop == Op_LoadUS &&
779 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
780 return new AndINode(load,phase->intcon(mask&0xFFFF));
781
782 // Masking bits off of a Short? Loading a Character does some masking
783 if (can_reshape &&
784 load->outcnt() == 1 && load->unique_out() == this) {
785 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
786 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
787 ldus = phase->transform(ldus);
788 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
789 }
790
791 // Masking sign bits off of a Byte? Do an unsigned byte load plus
792 // an and.
793 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
794 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
795 ldub = phase->transform(ldub);
796 return new AndINode(ldub, phase->intcon(mask));
797 }
798 }
799
800 // Masking off sign bits? Dont make them!
801 if( lop == Op_RShiftI ) {
802 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
803 if( t12 && t12->is_con() ) { // Shift is by a constant
804 int shift = t12->get_con();
805 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
806 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
807 // If the AND'ing of the 2 masks has no bits, then only original shifted
808 // bits survive. NO sign-extension bits survive the maskings.
809 if( (sign_bits_mask & mask) == 0 ) {
810 // Use zero-fill shift instead
811 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
812 return new AndINode( zshift, in(2) );
813 }
814 }
815 }
816
817 // Check for 'negate/and-1', a pattern emitted when someone asks for
818 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
819 // plus 1) and the mask is of the low order bit. Skip the negate.
820 if( lop == Op_SubI && mask == 1 && load->in(1) &&
821 phase->type(load->in(1)) == TypeInt::ZERO )
822 return new AndINode( load->in(2), in(2) );
823
824 return MulNode::Ideal(phase, can_reshape);
825 }
826
827 //=============================================================================
828 //------------------------------mul_ring---------------------------------------
829 // Supplied function returns the product of the inputs IN THE CURRENT RING.
830 // For the logical operations the ring's MUL is really a logical AND function.
831 // This also type-checks the inputs for sanity. Guaranteed never to
832 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
833 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
834 const TypeLong* r0 = t0->is_long();
835 const TypeLong* r1 = t1->is_long();
836
837 return and_value<TypeLong>(r0, r1);
838 }
839
840 const Type* AndLNode::Value(PhaseGVN* phase) const {
841 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
842 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
843 return TypeLong::ZERO;
844 }
845
846 return MulNode::Value(phase);
847 }
848
849 //------------------------------Identity---------------------------------------
850 // Masking off the high bits of an unsigned load is not required
851 Node* AndLNode::Identity(PhaseGVN* phase) {
852
853 // x & x => x
854 if (in(1) == in(2)) {
855 return in(1);
856 }
857
858 Node *usr = in(1);
859 const TypeLong *t2 = phase->type( in(2) )->isa_long();
860 if( t2 && t2->is_con() ) {
861 jlong con = t2->get_con();
862 // Masking off high bits which are always zero is useless.
863 const TypeLong* t1 = phase->type( in(1) )->isa_long();
864 if (t1 != nullptr && t1->_lo >= 0) {
865 int bit_count = log2i_graceful(t1->_hi) + 1;
866 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
867 if ((t1_support & con) == t1_support)
868 return usr;
869 }
870 uint lop = usr->Opcode();
871 // Masking off the high bits of a unsigned-shift-right is not
872 // needed either.
873 if( lop == Op_URShiftL ) {
874 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
875 if( t12 && t12->is_con() ) { // Shift is by a constant
876 int shift = t12->get_con();
877 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
878 jlong mask = max_julong >> shift;
879 if( (mask&con) == mask ) // If AND is useless, skip it
880 return usr;
881 }
882 }
883 }
884 return MulNode::Identity(phase);
885 }
886
887 //------------------------------Ideal------------------------------------------
888 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
889 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
890 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
891 if (progress != nullptr) {
892 return progress;
893 }
894
895 // Convert "(~a) & (~b)" into "~(a | b)"
896 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
897 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
898 Node* tn = phase->transform(or_a_b);
899 return AddNode::make_not(phase, tn, T_LONG);
900 }
901
902 // Special case constant AND mask
903 const TypeLong *t2 = phase->type( in(2) )->isa_long();
904 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
905 const jlong mask = t2->get_con();
906
907 Node* in1 = in(1);
908 int op = in1->Opcode();
909
910 // Are we masking a long that was converted from an int with a mask
911 // that fits in 32-bits? Commute them and use an AndINode. Don't
912 // convert masks which would cause a sign extension of the integer
913 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
914 // would be optimized away later in Identity.
915 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
916 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
917 andi = phase->transform(andi);
918 return new ConvI2LNode(andi);
919 }
920
921 // Masking off sign bits? Dont make them!
922 if (op == Op_RShiftL) {
923 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
924 if( t12 && t12->is_con() ) { // Shift is by a constant
925 int shift = t12->get_con();
926 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
927 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
928 // If the AND'ing of the 2 masks has no bits, then only original shifted
929 // bits survive. NO sign-extension bits survive the maskings.
930 if( (sign_bits_mask & mask) == 0 ) {
931 // Use zero-fill shift instead
932 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
933 return new AndLNode(zshift, in(2));
934 }
935 }
936 }
937
938 return MulNode::Ideal(phase, can_reshape);
939 }
940
941 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
942 switch (bt) {
943 case T_INT:
944 return new LShiftINode(in1, in2);
945 case T_LONG:
946 return new LShiftLNode(in1, in2);
947 default:
948 fatal("Not implemented for %s", type2name(bt));
949 }
950 return nullptr;
951 }
952
953 //=============================================================================
954
955 static bool const_shift_count(PhaseGVN* phase, Node* shiftNode, int* count) {
956 const TypeInt* tcount = phase->type(shiftNode->in(2))->isa_int();
957 if (tcount != nullptr && tcount->is_con()) {
958 *count = tcount->get_con();
959 return true;
960 }
961 return false;
962 }
963
964 static int maskShiftAmount(PhaseGVN* phase, Node* shiftNode, uint nBits) {
965 int count = 0;
966 if (const_shift_count(phase, shiftNode, &count)) {
967 int maskedShift = count & (nBits - 1);
968 if (maskedShift == 0) {
969 // Let Identity() handle 0 shift count.
970 return 0;
971 }
972
973 if (count != maskedShift) {
974 shiftNode->set_req(2, phase->intcon(maskedShift)); // Replace shift count with masked value.
975 PhaseIterGVN* igvn = phase->is_IterGVN();
976 if (igvn) {
977 igvn->rehash_node_delayed(shiftNode);
978 }
979 }
980 return maskedShift;
981 }
982 return 0;
983 }
984
985 // Called with
986 // outer_shift = (_ << con0)
987 // We are looking for the pattern:
988 // outer_shift = ((X << con1) << con0)
989 // we denote inner_shift the nested expression (X << con1)
990 //
991 // con0 and con1 are both in [0..nbits), as they are computed by maskShiftAmount.
992 //
993 // There are 2 cases:
994 // if con0 + con1 >= nbits => 0
995 // if con0 + con1 < nbits => X << (con1 + con0)
996 static Node* collapse_nested_shift_left(PhaseGVN* phase, Node* outer_shift, int con0, BasicType bt) {
997 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
998 int nbits = static_cast<int>(bits_per_java_integer(bt));
999 Node* inner_shift = outer_shift->in(1);
1000 if (inner_shift->Opcode() != Op_LShift(bt)) {
1001 return nullptr;
1002 }
1003
1004 int con1 = maskShiftAmount(phase, inner_shift, nbits);
1005 if (con1 == 0) { // Either non-const, or actually 0 (up to mask) and then delegated to Identity()
1006 return nullptr;
1007 }
1008
1009 if (con0 + con1 >= nbits) {
1010 // While it might be tempting to use
1011 // phase->zerocon(bt);
1012 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
1013 // to return a new node, this or nullptr, but not an old (cached) node.
1014 return ConNode::make(TypeInteger::zero(bt));
1015 }
1016
1017 // con0 + con1 < nbits ==> actual shift happens now
1018 Node* con0_plus_con1 = phase->intcon(con0 + con1);
1019 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
1020 }
1021
1022 //------------------------------Identity---------------------------------------
1023 Node* LShiftINode::Identity(PhaseGVN* phase) {
1024 int count = 0;
1025 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1026 // Shift by a multiple of 32 does nothing
1027 return in(1);
1028 }
1029 return this;
1030 }
1031
1032 //------------------------------Ideal------------------------------------------
1033 // If the right input is a constant, and the left input is an add of a
1034 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1035 //
1036 // Also collapse nested left-shifts with constant rhs:
1037 // (X << con1) << con2 ==> X << (con1 + con2)
1038 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1039 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
1040 if (con == 0) {
1041 return nullptr;
1042 }
1043
1044 // Left input is an add?
1045 Node *add1 = in(1);
1046 int add1_op = add1->Opcode();
1047 if( add1_op == Op_AddI ) { // Left input is an add?
1048 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" );
1049
1050 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1051 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1052 if( con < 16 ) {
1053 // Left input is an add of the same number?
1054 if (add1->in(1) == add1->in(2)) {
1055 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1056 // In general, this optimization cannot be applied for c0 == 31 since
1057 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1058 return new LShiftINode(add1->in(1), phase->intcon(con + 1));
1059 }
1060
1061 // Left input is an add of a constant?
1062 const TypeInt *t12 = phase->type(add1->in(2))->isa_int();
1063 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1064 // Compute X << con0
1065 Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) );
1066 // Compute X<<con0 + (con1<<con0)
1067 return new AddINode( lsh, phase->intcon(t12->get_con() << con));
1068 }
1069 }
1070 }
1071
1072 // Check for "(x >> C1) << C2"
1073 if (add1_op == Op_RShiftI || add1_op == Op_URShiftI) {
1074 int add1Con = 0;
1075 const_shift_count(phase, add1, &add1Con);
1076
1077 // Special case C1 == C2, which just masks off low bits
1078 if (add1Con > 0 && con == add1Con) {
1079 // Convert to "(x & -(1 << C2))"
1080 return new AndINode(add1->in(1), phase->intcon(java_negate(jint(1 << con))));
1081 } else {
1082 // Wait until the right shift has been sharpened to the correct count
1083 if (add1Con > 0 && add1Con < BitsPerJavaInteger) {
1084 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1085 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1086 if (phase->is_IterGVN()) {
1087 if (con > add1Con) {
1088 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1089 Node* lshift = phase->transform(new LShiftINode(add1->in(1), phase->intcon(con - add1Con)));
1090 return new AndINode(lshift, phase->intcon(java_negate(jint(1 << con))));
1091 } else {
1092 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1093 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1094
1095 // Handle logical and arithmetic shifts
1096 Node* rshift;
1097 if (add1_op == Op_RShiftI) {
1098 rshift = phase->transform(new RShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1099 } else {
1100 rshift = phase->transform(new URShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1101 }
1102
1103 return new AndINode(rshift, phase->intcon(java_negate(jint(1 << con))));
1104 }
1105 } else {
1106 phase->record_for_igvn(this);
1107 }
1108 }
1109 }
1110 }
1111
1112 // Check for "((x >> C1) & Y) << C2"
1113 if (add1_op == Op_AndI) {
1114 Node *add2 = add1->in(1);
1115 int add2_op = add2->Opcode();
1116 if (add2_op == Op_RShiftI || add2_op == Op_URShiftI) {
1117 // Special case C1 == C2, which just masks off low bits
1118 if (add2->in(2) == in(2)) {
1119 // Convert to "(x & (Y << C2))"
1120 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1121 return new AndINode(add2->in(1), y_sh);
1122 }
1123
1124 int add2Con = 0;
1125 const_shift_count(phase, add2, &add2Con);
1126 if (add2Con > 0 && add2Con < BitsPerJavaInteger) {
1127 if (phase->is_IterGVN()) {
1128 // Convert to "((x >> C1) << C2) & (Y << C2)"
1129
1130 // Make "(x >> C1) << C2", which will get folded away by the rule above
1131 Node* x_sh = phase->transform(new LShiftINode(add2, phase->intcon(con)));
1132 // Make "Y << C2", which will simplify when Y is a constant
1133 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1134
1135 return new AndINode(x_sh, y_sh);
1136 } else {
1137 phase->record_for_igvn(this);
1138 }
1139 }
1140 }
1141 }
1142
1143 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1144 // before shifting them away.
1145 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con);
1146 if( add1_op == Op_AndI &&
1147 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) )
1148 return new LShiftINode( add1->in(1), in(2) );
1149
1150 // Performs:
1151 // (X << con1) << con2 ==> X << (con1 + con2)
1152 Node* doubleShift = collapse_nested_shift_left(phase, this, con, T_INT);
1153 if (doubleShift != nullptr) {
1154 return doubleShift;
1155 }
1156
1157 return nullptr;
1158 }
1159
1160 //------------------------------Value------------------------------------------
1161 // A LShiftINode shifts its input2 left by input1 amount.
1162 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1163 const Type *t1 = phase->type( in(1) );
1164 const Type *t2 = phase->type( in(2) );
1165 // Either input is TOP ==> the result is TOP
1166 if( t1 == Type::TOP ) return Type::TOP;
1167 if( t2 == Type::TOP ) return Type::TOP;
1168
1169 // Left input is ZERO ==> the result is ZERO.
1170 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1171 // Shift by zero does nothing
1172 if( t2 == TypeInt::ZERO ) return t1;
1173
1174 // Either input is BOTTOM ==> the result is BOTTOM
1175 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) ||
1176 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1177 return TypeInt::INT;
1178
1179 const TypeInt *r1 = t1->is_int(); // Handy access
1180 const TypeInt *r2 = t2->is_int(); // Handy access
1181
1182 if (!r2->is_con())
1183 return TypeInt::INT;
1184
1185 uint shift = r2->get_con();
1186 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1187 // Shift by a multiple of 32 does nothing:
1188 if (shift == 0) return t1;
1189
1190 // If the shift is a constant, shift the bounds of the type,
1191 // unless this could lead to an overflow.
1192 if (!r1->is_con()) {
1193 jint lo = r1->_lo, hi = r1->_hi;
1194 if (((lo << shift) >> shift) == lo &&
1195 ((hi << shift) >> shift) == hi) {
1196 // No overflow. The range shifts up cleanly.
1197 return TypeInt::make((jint)lo << (jint)shift,
1198 (jint)hi << (jint)shift,
1199 MAX2(r1->_widen,r2->_widen));
1200 }
1201 return TypeInt::INT;
1202 }
1203
1204 return TypeInt::make( (jint)r1->get_con() << (jint)shift );
1205 }
1206
1207 //=============================================================================
1208 //------------------------------Identity---------------------------------------
1209 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1210 int count = 0;
1211 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1212 // Shift by a multiple of 64 does nothing
1213 return in(1);
1214 }
1215 return this;
1216 }
1217
1218 //------------------------------Ideal------------------------------------------
1219 // If the right input is a constant, and the left input is an add of a
1220 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1221 //
1222 // Also collapse nested left-shifts with constant rhs:
1223 // (X << con1) << con2 ==> X << (con1 + con2)
1224 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1225 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1226 if (con == 0) {
1227 return nullptr;
1228 }
1229
1230 // Left input is an add?
1231 Node *add1 = in(1);
1232 int add1_op = add1->Opcode();
1233 if( add1_op == Op_AddL ) { // Left input is an add?
1234 // Avoid dead data cycles from dead loops
1235 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" );
1236
1237 // Left input is an add of the same number?
1238 if (con != (BitsPerJavaLong - 1) && add1->in(1) == add1->in(2)) {
1239 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1240 // Can only be applied if c0 != 63 because:
1241 // (x + x) << 63 = 2x << 63, while
1242 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1243 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1244 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1245 return new LShiftLNode(add1->in(1), phase->intcon(con + 1));
1246 }
1247
1248 // Left input is an add of a constant?
1249 const TypeLong *t12 = phase->type(add1->in(2))->isa_long();
1250 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1251 // Compute X << con0
1252 Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) );
1253 // Compute X<<con0 + (con1<<con0)
1254 return new AddLNode( lsh, phase->longcon(t12->get_con() << con));
1255 }
1256 }
1257
1258 // Check for "(x >> C1) << C2"
1259 if (add1_op == Op_RShiftL || add1_op == Op_URShiftL) {
1260 int add1Con = 0;
1261 const_shift_count(phase, add1, &add1Con);
1262
1263 // Special case C1 == C2, which just masks off low bits
1264 if (add1Con > 0 && con == add1Con) {
1265 // Convert to "(x & -(1 << C2))"
1266 return new AndLNode(add1->in(1), phase->longcon(java_negate(jlong(CONST64(1) << con))));
1267 } else {
1268 // Wait until the right shift has been sharpened to the correct count
1269 if (add1Con > 0 && add1Con < BitsPerJavaLong) {
1270 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1271 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1272 if (phase->is_IterGVN()) {
1273 if (con > add1Con) {
1274 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1275 Node* lshift = phase->transform(new LShiftLNode(add1->in(1), phase->intcon(con - add1Con)));
1276 return new AndLNode(lshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1277 } else {
1278 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1279 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1280
1281 // Handle logical and arithmetic shifts
1282 Node* rshift;
1283 if (add1_op == Op_RShiftL) {
1284 rshift = phase->transform(new RShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1285 } else {
1286 rshift = phase->transform(new URShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1287 }
1288
1289 return new AndLNode(rshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1290 }
1291 } else {
1292 phase->record_for_igvn(this);
1293 }
1294 }
1295 }
1296 }
1297
1298 // Check for "((x >> C1) & Y) << C2"
1299 if (add1_op == Op_AndL) {
1300 Node* add2 = add1->in(1);
1301 int add2_op = add2->Opcode();
1302 if (add2_op == Op_RShiftL || add2_op == Op_URShiftL) {
1303 // Special case C1 == C2, which just masks off low bits
1304 if (add2->in(2) == in(2)) {
1305 // Convert to "(x & (Y << C2))"
1306 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1307 return new AndLNode(add2->in(1), y_sh);
1308 }
1309
1310 int add2Con = 0;
1311 const_shift_count(phase, add2, &add2Con);
1312 if (add2Con > 0 && add2Con < BitsPerJavaLong) {
1313 if (phase->is_IterGVN()) {
1314 // Convert to "((x >> C1) << C2) & (Y << C2)"
1315
1316 // Make "(x >> C1) << C2", which will get folded away by the rule above
1317 Node* x_sh = phase->transform(new LShiftLNode(add2, phase->intcon(con)));
1318 // Make "Y << C2", which will simplify when Y is a constant
1319 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1320
1321 return new AndLNode(x_sh, y_sh);
1322 } else {
1323 phase->record_for_igvn(this);
1324 }
1325 }
1326 }
1327 }
1328
1329 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits
1330 // before shifting them away.
1331 const jlong bits_mask = jlong(max_julong >> con);
1332 if( add1_op == Op_AndL &&
1333 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) )
1334 return new LShiftLNode( add1->in(1), in(2) );
1335
1336 // Performs:
1337 // (X << con1) << con2 ==> X << (con1 + con2)
1338 Node* doubleShift = collapse_nested_shift_left(phase, this, con, T_LONG);
1339 if (doubleShift != nullptr) {
1340 return doubleShift;
1341 }
1342
1343 return nullptr;
1344 }
1345
1346 //------------------------------Value------------------------------------------
1347 // A LShiftLNode shifts its input2 left by input1 amount.
1348 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1349 const Type *t1 = phase->type( in(1) );
1350 const Type *t2 = phase->type( in(2) );
1351 // Either input is TOP ==> the result is TOP
1352 if( t1 == Type::TOP ) return Type::TOP;
1353 if( t2 == Type::TOP ) return Type::TOP;
1354
1355 // Left input is ZERO ==> the result is ZERO.
1356 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1357 // Shift by zero does nothing
1358 if( t2 == TypeInt::ZERO ) return t1;
1359
1360 // Either input is BOTTOM ==> the result is BOTTOM
1361 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) ||
1362 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1363 return TypeLong::LONG;
1364
1365 const TypeLong *r1 = t1->is_long(); // Handy access
1366 const TypeInt *r2 = t2->is_int(); // Handy access
1367
1368 if (!r2->is_con())
1369 return TypeLong::LONG;
1370
1371 uint shift = r2->get_con();
1372 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1373 // Shift by a multiple of 64 does nothing:
1374 if (shift == 0) return t1;
1375
1376 // If the shift is a constant, shift the bounds of the type,
1377 // unless this could lead to an overflow.
1378 if (!r1->is_con()) {
1379 jlong lo = r1->_lo, hi = r1->_hi;
1380 if (((lo << shift) >> shift) == lo &&
1381 ((hi << shift) >> shift) == hi) {
1382 // No overflow. The range shifts up cleanly.
1383 return TypeLong::make((jlong)lo << (jint)shift,
1384 (jlong)hi << (jint)shift,
1385 MAX2(r1->_widen,r2->_widen));
1386 }
1387 return TypeLong::LONG;
1388 }
1389
1390 return TypeLong::make( (jlong)r1->get_con() << (jint)shift );
1391 }
1392
1393 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1394 switch (bt) {
1395 case T_INT:
1396 return new RShiftINode(in1, in2);
1397 case T_LONG:
1398 return new RShiftLNode(in1, in2);
1399 default:
1400 fatal("Not implemented for %s", type2name(bt));
1401 }
1402 return nullptr;
1403 }
1404
1405
1406 //=============================================================================
1407 //------------------------------Identity---------------------------------------
1408 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1409 int count = 0;
1410 if (const_shift_count(phase, this, &count)) {
1411 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1412 // Shift by a multiple of 32/64 does nothing
1413 return in(1);
1414 }
1415 // Check for useless sign-masking
1416 if (in(1)->Opcode() == Op_LShift(bt) &&
1417 in(1)->req() == 3 &&
1418 in(1)->in(2) == in(2)) {
1419 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1420 // Compute masks for which this shifting doesn't change
1421 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1422 jlong hi = ~lo; // 00007FFF
1423 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1424 if (t11 == nullptr) {
1425 return this;
1426 }
1427 // Does actual value fit inside of mask?
1428 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1429 return in(1)->in(1); // Then shifting is a nop
1430 }
1431 }
1432 }
1433 return this;
1434 }
1435
1436 Node* RShiftINode::Identity(PhaseGVN* phase) {
1437 return IdentityIL(phase, T_INT);
1438 }
1439
1440 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1441 // Inputs may be TOP if they are dead.
1442 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1443 if (t1 == nullptr) {
1444 return NodeSentinel; // Left input is an integer
1445 }
1446 int shift = maskShiftAmount(phase, this, bits_per_java_integer(bt));
1447 if (shift == 0) {
1448 return NodeSentinel;
1449 }
1450
1451 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1452 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1453 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1454 const Node* and_node = in(1);
1455 if (and_node->Opcode() != Op_And(bt)) {
1456 return nullptr;
1457 }
1458 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1459 if (mask_t != nullptr && mask_t->is_con()) {
1460 jlong maskbits = mask_t->get_con_as_long(bt);
1461 // Convert to "(x >> shift) & (mask >> shift)"
1462 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1463 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1464 }
1465 return nullptr;
1466 }
1467
1468 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1469 Node* progress = IdealIL(phase, can_reshape, T_INT);
1470 if (progress == NodeSentinel) {
1471 return nullptr;
1472 }
1473 if (progress != nullptr) {
1474 return progress;
1475 }
1476 int shift = maskShiftAmount(phase, this, BitsPerJavaInteger);
1477 assert(shift != 0, "handled by IdealIL");
1478
1479 // Check for "(short[i] <<16)>>16" which simply sign-extends
1480 const Node *shl = in(1);
1481 if (shl->Opcode() != Op_LShiftI) {
1482 return nullptr;
1483 }
1484
1485 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1486 if (left_shift_t == nullptr) {
1487 return nullptr;
1488 }
1489 if (shift == 16 && left_shift_t->is_con(16)) {
1490 Node *ld = shl->in(1);
1491 if (ld->Opcode() == Op_LoadS) {
1492 // Sign extension is just useless here. Return a RShiftI of zero instead
1493 // returning 'ld' directly. We cannot return an old Node directly as
1494 // that is the job of 'Identity' calls and Identity calls only work on
1495 // direct inputs ('ld' is an extra Node removed from 'this'). The
1496 // combined optimization requires Identity only return direct inputs.
1497 set_req_X(1, ld, phase);
1498 set_req_X(2, phase->intcon(0), phase);
1499 return this;
1500 }
1501 else if (can_reshape &&
1502 ld->Opcode() == Op_LoadUS &&
1503 ld->outcnt() == 1 && ld->unique_out() == shl)
1504 // Replace zero-extension-load with sign-extension-load
1505 return ld->as_Load()->convert_to_signed_load(*phase);
1506 }
1507
1508 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1509 if (shift == 24 && left_shift_t->is_con(24)) {
1510 Node *ld = shl->in(1);
1511 if (ld->Opcode() == Op_LoadB) {
1512 // Sign extension is just useless here
1513 set_req_X(1, ld, phase);
1514 set_req_X(2, phase->intcon(0), phase);
1515 return this;
1516 }
1517 }
1518
1519 return nullptr;
1520 }
1521
1522 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1523 const Type* t1 = phase->type(in(1));
1524 const Type* t2 = phase->type(in(2));
1525 // Either input is TOP ==> the result is TOP
1526 if (t1 == Type::TOP) {
1527 return Type::TOP;
1528 }
1529 if (t2 == Type::TOP) {
1530 return Type::TOP;
1531 }
1532
1533 // Left input is ZERO ==> the result is ZERO.
1534 if (t1 == TypeInteger::zero(bt)) {
1535 return TypeInteger::zero(bt);
1536 }
1537 // Shift by zero does nothing
1538 if (t2 == TypeInt::ZERO) {
1539 return t1;
1540 }
1541
1542 // Either input is BOTTOM ==> the result is BOTTOM
1543 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1544 return TypeInteger::bottom(bt);
1545 }
1546
1547 const TypeInteger* r1 = t1->isa_integer(bt);
1548 const TypeInt* r2 = t2->isa_int();
1549
1550 // If the shift is a constant, just shift the bounds of the type.
1551 // For example, if the shift is 31/63, we just propagate sign bits.
1552 if (!r1->is_con() && r2->is_con()) {
1553 uint shift = r2->get_con();
1554 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1555 // Shift by a multiple of 32/64 does nothing:
1556 if (shift == 0) {
1557 return t1;
1558 }
1559 // Calculate reasonably aggressive bounds for the result.
1560 // This is necessary if we are to correctly type things
1561 // like (x<<24>>24) == ((byte)x).
1562 jlong lo = r1->lo_as_long() >> (jint)shift;
1563 jlong hi = r1->hi_as_long() >> (jint)shift;
1564 assert(lo <= hi, "must have valid bounds");
1565 #ifdef ASSERT
1566 if (bt == T_INT) {
1567 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1568 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1569 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1570 }
1571 #endif
1572 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1573 #ifdef ASSERT
1574 // Make sure we get the sign-capture idiom correct.
1575 if (shift == bits_per_java_integer(bt) - 1) {
1576 if (r1->lo_as_long() >= 0) {
1577 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1578 }
1579 if (r1->hi_as_long() < 0) {
1580 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1581 }
1582 }
1583 #endif
1584 return ti;
1585 }
1586
1587 if (!r1->is_con() || !r2->is_con()) {
1588 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1589 if (r1->lo_as_long() >= 0) {
1590 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1591 }
1592
1593 // Conversely, if the left input is negative then the result must be negative.
1594 if (r1->hi_as_long() <= -1) {
1595 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1596 }
1597
1598 return TypeInteger::bottom(bt);
1599 }
1600
1601 // Signed shift right
1602 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1603 }
1604
1605 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1606 return ValueIL(phase, T_INT);
1607 }
1608
1609 //=============================================================================
1610 //------------------------------Identity---------------------------------------
1611 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1612 return IdentityIL(phase, T_LONG);
1613 }
1614
1615 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1616 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1617 if (progress == NodeSentinel) {
1618 return nullptr;
1619 }
1620 return progress;
1621 }
1622
1623 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1624 return ValueIL(phase, T_LONG);
1625 }
1626
1627 //=============================================================================
1628 //------------------------------Identity---------------------------------------
1629 Node* URShiftINode::Identity(PhaseGVN* phase) {
1630 int count = 0;
1631 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1632 // Shift by a multiple of 32 does nothing
1633 return in(1);
1634 }
1635
1636 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1637 // Happens during new-array length computation.
1638 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1639 Node *add = in(1);
1640 if (add->Opcode() == Op_AddI) {
1641 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1642 if (t2 && t2->is_con(wordSize - 1) &&
1643 add->in(1)->Opcode() == Op_LShiftI) {
1644 // Check that shift_counts are LogBytesPerWord.
1645 Node *lshift_count = add->in(1)->in(2);
1646 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1647 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1648 t_lshift_count == phase->type(in(2))) {
1649 Node *x = add->in(1)->in(1);
1650 const TypeInt *t_x = phase->type(x)->isa_int();
1651 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1652 return x;
1653 }
1654 }
1655 }
1656 }
1657
1658 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1659 }
1660
1661 //------------------------------Ideal------------------------------------------
1662 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1663 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
1664 if (con == 0) {
1665 return nullptr;
1666 }
1667
1668 // We'll be wanting the right-shift amount as a mask of that many bits
1669 const int mask = right_n_bits(BitsPerJavaInteger - con);
1670
1671 int in1_op = in(1)->Opcode();
1672
1673 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1674 if( in1_op == Op_URShiftI ) {
1675 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1676 if( t12 && t12->is_con() ) { // Right input is a constant
1677 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1678 const int con2 = t12->get_con() & 31; // Shift count is always masked
1679 const int con3 = con+con2;
1680 if( con3 < 32 ) // Only merge shifts if total is < 32
1681 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1682 }
1683 }
1684
1685 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1686 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1687 // If Q is "X << z" the rounding is useless. Look for patterns like
1688 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1689 Node *add = in(1);
1690 const TypeInt *t2 = phase->type(in(2))->isa_int();
1691 if (in1_op == Op_AddI) {
1692 Node *lshl = add->in(1);
1693 if( lshl->Opcode() == Op_LShiftI &&
1694 phase->type(lshl->in(2)) == t2 ) {
1695 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1696 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1697 return new AndINode( sum, phase->intcon(mask) );
1698 }
1699 }
1700
1701 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1702 // This shortens the mask. Also, if we are extracting a high byte and
1703 // storing it to a buffer, the mask will be removed completely.
1704 Node *andi = in(1);
1705 if( in1_op == Op_AndI ) {
1706 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1707 if( t3 && t3->is_con() ) { // Right input is a constant
1708 jint mask2 = t3->get_con();
1709 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1710 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1711 return new AndINode(newshr, phase->intcon(mask2));
1712 // The negative values are easier to materialize than positive ones.
1713 // A typical case from address arithmetic is ((x & ~15) >> 4).
1714 // It's better to change that to ((x >> 4) & ~0) versus
1715 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1716 }
1717 }
1718
1719 // Check for "(X << z ) >>> z" which simply zero-extends
1720 Node *shl = in(1);
1721 if( in1_op == Op_LShiftI &&
1722 phase->type(shl->in(2)) == t2 )
1723 return new AndINode( shl->in(1), phase->intcon(mask) );
1724
1725 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1726 Node *shr = in(1);
1727 if ( in1_op == Op_RShiftI ) {
1728 Node *in11 = shr->in(1);
1729 Node *in12 = shr->in(2);
1730 const TypeInt *t11 = phase->type(in11)->isa_int();
1731 const TypeInt *t12 = phase->type(in12)->isa_int();
1732 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1733 return new URShiftINode(in11, phase->intcon(31));
1734 }
1735 }
1736
1737 return nullptr;
1738 }
1739
1740 //------------------------------Value------------------------------------------
1741 // A URShiftINode shifts its input2 right by input1 amount.
1742 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1743 // (This is a near clone of RShiftINode::Value.)
1744 const Type *t1 = phase->type( in(1) );
1745 const Type *t2 = phase->type( in(2) );
1746 // Either input is TOP ==> the result is TOP
1747 if( t1 == Type::TOP ) return Type::TOP;
1748 if( t2 == Type::TOP ) return Type::TOP;
1749
1750 // Left input is ZERO ==> the result is ZERO.
1751 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1752 // Shift by zero does nothing
1753 if( t2 == TypeInt::ZERO ) return t1;
1754
1755 // Either input is BOTTOM ==> the result is BOTTOM
1756 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1757 return TypeInt::INT;
1758
1759 if (t2 == TypeInt::INT)
1760 return TypeInt::INT;
1761
1762 const TypeInt *r1 = t1->is_int(); // Handy access
1763 const TypeInt *r2 = t2->is_int(); // Handy access
1764
1765 if (r2->is_con()) {
1766 uint shift = r2->get_con();
1767 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1768 // Shift by a multiple of 32 does nothing:
1769 if (shift == 0) return t1;
1770 // Calculate reasonably aggressive bounds for the result.
1771 jint lo = (juint)r1->_lo >> (juint)shift;
1772 jint hi = (juint)r1->_hi >> (juint)shift;
1773 if (r1->_hi >= 0 && r1->_lo < 0) {
1774 // If the type has both negative and positive values,
1775 // there are two separate sub-domains to worry about:
1776 // The positive half and the negative half.
1777 jint neg_lo = lo;
1778 jint neg_hi = (juint)-1 >> (juint)shift;
1779 jint pos_lo = (juint) 0 >> (juint)shift;
1780 jint pos_hi = hi;
1781 lo = MIN2(neg_lo, pos_lo); // == 0
1782 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1783 }
1784 assert(lo <= hi, "must have valid bounds");
1785 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1786 #ifdef ASSERT
1787 // Make sure we get the sign-capture idiom correct.
1788 if (shift == BitsPerJavaInteger-1) {
1789 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1790 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1791 }
1792 #endif
1793 return ti;
1794 }
1795
1796 //
1797 // Do not support shifted oops in info for GC
1798 //
1799 // else if( t1->base() == Type::InstPtr ) {
1800 //
1801 // const TypeInstPtr *o = t1->is_instptr();
1802 // if( t1->singleton() )
1803 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1804 // }
1805 // else if( t1->base() == Type::KlassPtr ) {
1806 // const TypeKlassPtr *o = t1->is_klassptr();
1807 // if( t1->singleton() )
1808 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1809 // }
1810
1811 return TypeInt::INT;
1812 }
1813
1814 //=============================================================================
1815 //------------------------------Identity---------------------------------------
1816 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1817 int count = 0;
1818 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1819 // Shift by a multiple of 64 does nothing
1820 return in(1);
1821 }
1822 return this;
1823 }
1824
1825 //------------------------------Ideal------------------------------------------
1826 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1827 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1828 if (con == 0) {
1829 return nullptr;
1830 }
1831
1832 // We'll be wanting the right-shift amount as a mask of that many bits
1833 const jlong mask = jlong(max_julong >> con);
1834
1835 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1836 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1837 // If Q is "X << z" the rounding is useless. Look for patterns like
1838 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1839 Node *add = in(1);
1840 const TypeInt *t2 = phase->type(in(2))->isa_int();
1841 if (add->Opcode() == Op_AddL) {
1842 Node *lshl = add->in(1);
1843 if( lshl->Opcode() == Op_LShiftL &&
1844 phase->type(lshl->in(2)) == t2 ) {
1845 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1846 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1847 return new AndLNode( sum, phase->longcon(mask) );
1848 }
1849 }
1850
1851 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1852 // This shortens the mask. Also, if we are extracting a high byte and
1853 // storing it to a buffer, the mask will be removed completely.
1854 Node *andi = in(1);
1855 if( andi->Opcode() == Op_AndL ) {
1856 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1857 if( t3 && t3->is_con() ) { // Right input is a constant
1858 jlong mask2 = t3->get_con();
1859 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1860 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1861 return new AndLNode(newshr, phase->longcon(mask2));
1862 }
1863 }
1864
1865 // Check for "(X << z ) >>> z" which simply zero-extends
1866 Node *shl = in(1);
1867 if( shl->Opcode() == Op_LShiftL &&
1868 phase->type(shl->in(2)) == t2 )
1869 return new AndLNode( shl->in(1), phase->longcon(mask) );
1870
1871 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1872 Node *shr = in(1);
1873 if ( shr->Opcode() == Op_RShiftL ) {
1874 Node *in11 = shr->in(1);
1875 Node *in12 = shr->in(2);
1876 const TypeLong *t11 = phase->type(in11)->isa_long();
1877 const TypeInt *t12 = phase->type(in12)->isa_int();
1878 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1879 return new URShiftLNode(in11, phase->intcon(63));
1880 }
1881 }
1882 return nullptr;
1883 }
1884
1885 //------------------------------Value------------------------------------------
1886 // A URShiftINode shifts its input2 right by input1 amount.
1887 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1888 // (This is a near clone of RShiftLNode::Value.)
1889 const Type *t1 = phase->type( in(1) );
1890 const Type *t2 = phase->type( in(2) );
1891 // Either input is TOP ==> the result is TOP
1892 if( t1 == Type::TOP ) return Type::TOP;
1893 if( t2 == Type::TOP ) return Type::TOP;
1894
1895 // Left input is ZERO ==> the result is ZERO.
1896 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1897 // Shift by zero does nothing
1898 if( t2 == TypeInt::ZERO ) return t1;
1899
1900 // Either input is BOTTOM ==> the result is BOTTOM
1901 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1902 return TypeLong::LONG;
1903
1904 if (t2 == TypeInt::INT)
1905 return TypeLong::LONG;
1906
1907 const TypeLong *r1 = t1->is_long(); // Handy access
1908 const TypeInt *r2 = t2->is_int (); // Handy access
1909
1910 if (r2->is_con()) {
1911 uint shift = r2->get_con();
1912 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1913 // Shift by a multiple of 64 does nothing:
1914 if (shift == 0) return t1;
1915 // Calculate reasonably aggressive bounds for the result.
1916 jlong lo = (julong)r1->_lo >> (juint)shift;
1917 jlong hi = (julong)r1->_hi >> (juint)shift;
1918 if (r1->_hi >= 0 && r1->_lo < 0) {
1919 // If the type has both negative and positive values,
1920 // there are two separate sub-domains to worry about:
1921 // The positive half and the negative half.
1922 jlong neg_lo = lo;
1923 jlong neg_hi = (julong)-1 >> (juint)shift;
1924 jlong pos_lo = (julong) 0 >> (juint)shift;
1925 jlong pos_hi = hi;
1926 //lo = MIN2(neg_lo, pos_lo); // == 0
1927 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1928 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1929 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1930 }
1931 assert(lo <= hi, "must have valid bounds");
1932 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1933 #ifdef ASSERT
1934 // Make sure we get the sign-capture idiom correct.
1935 if (shift == BitsPerJavaLong - 1) {
1936 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1937 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1938 }
1939 #endif
1940 return tl;
1941 }
1942
1943 return TypeLong::LONG; // Give up
1944 }
1945
1946 //=============================================================================
1947 //------------------------------Ideal------------------------------------------
1948 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1949 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1950 // This reduces the number of rules in the matcher, as we only need to check
1951 // for negations on the second argument, and not the symmetric case where
1952 // the first argument is negated.
1953 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1954 swap_edges(1, 2);
1955 return this;
1956 }
1957 return nullptr;
1958 }
1959
1960 //=============================================================================
1961 //------------------------------Value------------------------------------------
1962 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1963 const Type *t1 = phase->type(in(1));
1964 if (t1 == Type::TOP) return Type::TOP;
1965 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1966 const Type *t2 = phase->type(in(2));
1967 if (t2 == Type::TOP) return Type::TOP;
1968 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1969 const Type *t3 = phase->type(in(3));
1970 if (t3 == Type::TOP) return Type::TOP;
1971 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1972 #ifndef __STDC_IEC_559__
1973 return Type::DOUBLE;
1974 #else
1975 double d1 = t1->getd();
1976 double d2 = t2->getd();
1977 double d3 = t3->getd();
1978 return TypeD::make(fma(d1, d2, d3));
1979 #endif
1980 }
1981
1982 //=============================================================================
1983 //------------------------------Value------------------------------------------
1984 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1985 const Type *t1 = phase->type(in(1));
1986 if (t1 == Type::TOP) return Type::TOP;
1987 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1988 const Type *t2 = phase->type(in(2));
1989 if (t2 == Type::TOP) return Type::TOP;
1990 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1991 const Type *t3 = phase->type(in(3));
1992 if (t3 == Type::TOP) return Type::TOP;
1993 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1994 #ifndef __STDC_IEC_559__
1995 return Type::FLOAT;
1996 #else
1997 float f1 = t1->getf();
1998 float f2 = t2->getf();
1999 float f3 = t3->getf();
2000 return TypeF::make(fma(f1, f2, f3));
2001 #endif
2002 }
2003
2004 //=============================================================================
2005 //------------------------------Value------------------------------------------
2006 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
2007 const Type* t1 = phase->type(in(1));
2008 if (t1 == Type::TOP) { return Type::TOP; }
2009 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2010 const Type* t2 = phase->type(in(2));
2011 if (t2 == Type::TOP) { return Type::TOP; }
2012 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2013 const Type* t3 = phase->type(in(3));
2014 if (t3 == Type::TOP) { return Type::TOP; }
2015 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2016 #ifndef __STDC_IEC_559__
2017 return Type::HALF_FLOAT;
2018 #else
2019 float f1 = t1->getf();
2020 float f2 = t2->getf();
2021 float f3 = t3->getf();
2022 return TypeH::make(fma(f1, f2, f3));
2023 #endif
2024 }
2025
2026 //=============================================================================
2027 //------------------------------hash-------------------------------------------
2028 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
2029 // The hash function must return the same value when edge swapping is performed.
2030 uint MulAddS2INode::hash() const {
2031 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
2032 }
2033
2034 //------------------------------Rotate Operations ------------------------------
2035
2036 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
2037 const Type* t1 = phase->type(in(1));
2038 if (t1 == Type::TOP) {
2039 return this;
2040 }
2041 int count = 0;
2042 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2043 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2044 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2045 // Rotate by a multiple of 32/64 does nothing
2046 return in(1);
2047 }
2048 return this;
2049 }
2050
2051 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
2052 const Type* t1 = phase->type(in(1));
2053 const Type* t2 = phase->type(in(2));
2054 // Either input is TOP ==> the result is TOP
2055 if (t1 == Type::TOP || t2 == Type::TOP) {
2056 return Type::TOP;
2057 }
2058
2059 if (t1->isa_int()) {
2060 const TypeInt* r1 = t1->is_int();
2061 const TypeInt* r2 = t2->is_int();
2062
2063 // Left input is ZERO ==> the result is ZERO.
2064 if (r1 == TypeInt::ZERO) {
2065 return TypeInt::ZERO;
2066 }
2067 // Rotate by zero does nothing
2068 if (r2 == TypeInt::ZERO) {
2069 return r1;
2070 }
2071 if (r1->is_con() && r2->is_con()) {
2072 juint r1_con = (juint)r1->get_con();
2073 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2074 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2075 }
2076 return TypeInt::INT;
2077 } else {
2078 assert(t1->isa_long(), "Type must be a long");
2079 const TypeLong* r1 = t1->is_long();
2080 const TypeInt* r2 = t2->is_int();
2081
2082 // Left input is ZERO ==> the result is ZERO.
2083 if (r1 == TypeLong::ZERO) {
2084 return TypeLong::ZERO;
2085 }
2086 // Rotate by zero does nothing
2087 if (r2 == TypeInt::ZERO) {
2088 return r1;
2089 }
2090 if (r1->is_con() && r2->is_con()) {
2091 julong r1_con = (julong)r1->get_con();
2092 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2093 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2094 }
2095 return TypeLong::LONG;
2096 }
2097 }
2098
2099 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2100 const Type* t1 = phase->type(in(1));
2101 const Type* t2 = phase->type(in(2));
2102 if (t2->isa_int() && t2->is_int()->is_con()) {
2103 if (t1->isa_int()) {
2104 int lshift = t2->is_int()->get_con() & 31;
2105 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2106 } else if (t1 != Type::TOP) {
2107 assert(t1->isa_long(), "Type must be a long");
2108 int lshift = t2->is_int()->get_con() & 63;
2109 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2110 }
2111 }
2112 return nullptr;
2113 }
2114
2115 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2116 const Type* t1 = phase->type(in(1));
2117 if (t1 == Type::TOP) {
2118 return this;
2119 }
2120 int count = 0;
2121 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2122 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2123 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2124 // Rotate by a multiple of 32/64 does nothing
2125 return in(1);
2126 }
2127 return this;
2128 }
2129
2130 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2131 const Type* t1 = phase->type(in(1));
2132 const Type* t2 = phase->type(in(2));
2133 // Either input is TOP ==> the result is TOP
2134 if (t1 == Type::TOP || t2 == Type::TOP) {
2135 return Type::TOP;
2136 }
2137
2138 if (t1->isa_int()) {
2139 const TypeInt* r1 = t1->is_int();
2140 const TypeInt* r2 = t2->is_int();
2141
2142 // Left input is ZERO ==> the result is ZERO.
2143 if (r1 == TypeInt::ZERO) {
2144 return TypeInt::ZERO;
2145 }
2146 // Rotate by zero does nothing
2147 if (r2 == TypeInt::ZERO) {
2148 return r1;
2149 }
2150 if (r1->is_con() && r2->is_con()) {
2151 juint r1_con = (juint)r1->get_con();
2152 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2153 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2154 }
2155 return TypeInt::INT;
2156 } else {
2157 assert(t1->isa_long(), "Type must be a long");
2158 const TypeLong* r1 = t1->is_long();
2159 const TypeInt* r2 = t2->is_int();
2160 // Left input is ZERO ==> the result is ZERO.
2161 if (r1 == TypeLong::ZERO) {
2162 return TypeLong::ZERO;
2163 }
2164 // Rotate by zero does nothing
2165 if (r2 == TypeInt::ZERO) {
2166 return r1;
2167 }
2168 if (r1->is_con() && r2->is_con()) {
2169 julong r1_con = (julong)r1->get_con();
2170 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2171 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2172 }
2173 return TypeLong::LONG;
2174 }
2175 }
2176
2177 //------------------------------ Sum & Mask ------------------------------
2178
2179 // Returns a lower bound on the number of trailing zeros in expr.
2180 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2181 expr = expr->uncast();
2182 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2183 if (type == nullptr) {
2184 return 0;
2185 }
2186
2187 if (type->is_con()) {
2188 jlong con = type->get_con_as_long(bt);
2189 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2190 }
2191
2192 if (expr->Opcode() == Op_ConvI2L) {
2193 expr = expr->in(1)->uncast();
2194 bt = T_INT;
2195 type = phase->type(expr)->isa_int();
2196 }
2197
2198 // Pattern: expr = (x << shift)
2199 if (expr->Opcode() == Op_LShift(bt)) {
2200 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2201 if (shift_t == nullptr || !shift_t->is_con()) {
2202 return 0;
2203 }
2204 // We need to truncate the shift, as it may not have been canonicalized yet.
2205 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2206 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2207 // (JLS: "Shift Operators")
2208 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2209 return shift_t->get_con() & shift_mask;
2210 }
2211
2212 return 0;
2213 }
2214
2215 // Checks whether expr is neutral additive element (zero) under mask,
2216 // i.e. whether an expression of the form:
2217 // (AndX (AddX (expr addend) mask)
2218 // (expr + addend) & mask
2219 // is equivalent to
2220 // (AndX addend mask)
2221 // addend & mask
2222 // for any addend.
2223 // (The X in AndX must be I or L, depending on bt).
2224 //
2225 // We check for the sufficient condition when the lowest set bit in expr is higher than
2226 // the highest set bit in mask, i.e.:
2227 // expr: eeeeee0000000000000
2228 // mask: 000000mmmmmmmmmmmmm
2229 // <--w bits--->
2230 // We do not test for other cases.
2231 //
2232 // Correctness:
2233 // Given "expr" with at least "w" trailing zeros,
2234 // let "mod = 2^w", "suffix_mask = mod - 1"
2235 //
2236 // Since "mask" only has bits set where "suffix_mask" does, we have:
2237 // mask = suffix_mask & mask (SUFFIX_MASK)
2238 //
2239 // And since expr only has bits set above w, and suffix_mask only below:
2240 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2241 //
2242 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2243 // a power of 2, and we are computing in a ring of powers of 2, we know that
2244 // (x + y) % mod = (x % mod + y) % mod
2245 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2246 //
2247 // We can now prove the equality:
2248 // (expr + addend) & mask
2249 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2250 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2251 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2252 // = addend & mask (SUFFIX_MASK)
2253 //
2254 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2255 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2256 // When the mask is negative, it has the most significant bit set.
2257 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2258 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2259 return false;
2260 }
2261
2262 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2263 if (mask_t->hi_as_long() == 0) {
2264 assert(mask_t->lo_as_long() == 0, "checked earlier");
2265 return false;
2266 }
2267
2268 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2269 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2270 return expr_trailing_zeros >= mask_bit_width;
2271 }
2272
2273 // Reduces the pattern:
2274 // (AndX (AddX add1 add2) mask)
2275 // to
2276 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2277 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2278 Node* add = in(1);
2279 Node* mask = in(2);
2280 int addidx = 0;
2281 if (add->Opcode() == Op_Add(bt)) {
2282 addidx = 1;
2283 } else if (mask->Opcode() == Op_Add(bt)) {
2284 mask = add;
2285 addidx = 2;
2286 add = in(addidx);
2287 }
2288 if (addidx > 0) {
2289 Node* add1 = add->in(1);
2290 Node* add2 = add->in(2);
2291 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2292 set_req_X(addidx, add2, phase);
2293 return this;
2294 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2295 set_req_X(addidx, add1, phase);
2296 return this;
2297 }
2298 }
2299 return nullptr;
2300 }