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 // Code pattern on return from a call that returns an __Value. Can
202 // be optimized away if the return value turns out to be an oop.
203 if (op == Op_AndX &&
204 in(1) != nullptr &&
205 in(1)->Opcode() == Op_CastP2X &&
206 in(1)->in(1) != nullptr &&
207 phase->type(in(1)->in(1))->isa_oopptr() &&
208 t2->isa_intptr_t()->_lo >= 0 &&
209 t2->isa_intptr_t()->_hi <= MinObjAlignmentInBytesMask) {
210 return add_id();
211 }
212
213 // Either input is BOTTOM ==> the result is the local BOTTOM
214 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
215 return bottom_type();
216
217 return mul_ring(t1,t2); // Local flavor of type multiplication
218 }
219
220 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
221 switch (bt) {
222 case T_INT:
223 return new MulINode(in1, in2);
224 case T_LONG:
225 return new MulLNode(in1, in2);
226 default:
227 fatal("Not implemented for %s", type2name(bt));
228 }
229 return nullptr;
230 }
231
232 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
233 switch (bt) {
234 case T_INT:
235 return new AndINode(in1, in2);
236 case T_LONG:
237 return new AndLNode(in1, in2);
238 default:
239 fatal("Not implemented for %s", type2name(bt));
240 }
241 return nullptr;
242 }
243
244
245 //=============================================================================
246 //------------------------------Ideal------------------------------------------
247 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
248 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
249 const jint con = in(2)->find_int_con(0);
250 if (con == 0) {
251 // If in(2) is not a constant, call Ideal() of the parent class to
252 // try to move constant to the right side.
253 return MulNode::Ideal(phase, can_reshape);
254 }
255
256 // Now we have a constant Node on the right and the constant in con.
257 if (con == 1) {
258 // By one is handled by Identity call
259 return nullptr;
260 }
261
262 // Check for negative constant; if so negate the final result
263 bool sign_flip = false;
264
265 unsigned int abs_con = g_uabs(con);
266 if (abs_con != (unsigned int)con) {
267 sign_flip = true;
268 }
269
270 // Get low bit; check for being the only bit
271 Node *res = nullptr;
272 unsigned int bit1 = submultiple_power_of_2(abs_con);
273 if (bit1 == abs_con) { // Found a power of 2?
274 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
275 } else {
276 // Check for constant with 2 bits set
277 unsigned int bit2 = abs_con - bit1;
278 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
279 if (bit2 + bit1 == abs_con) { // Found all bits in con?
280 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
281 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
282 res = new AddINode(n2, n1);
283 } else if (is_power_of_2(abs_con + 1)) {
284 // Sleezy: power-of-2 - 1. Next time be generic.
285 unsigned int temp = abs_con + 1;
286 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
287 res = new SubINode(n1, in(1));
288 } else {
289 return MulNode::Ideal(phase, can_reshape);
290 }
291 }
292
293 if (sign_flip) { // Need to negate result?
294 res = phase->transform(res);// Transform, before making the zero con
295 res = new SubINode(phase->intcon(0),res);
296 }
297
298 return res; // Return final result
299 }
300
301 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
302 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
303 //
304 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
305 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
306 // - no overflow (i.e. product fits into x bits)
307 // - overflow (i.e. product does not fit into x bits)
308 //
309 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
310 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
311 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
312 // The cross product consists of the following four multiplications with 2x-bit results:
313 // (1) left_lo * right_lo
314 // (2) left_lo * right_hi
315 // (3) left_hi * right_lo
316 // (4) left_hi * right_hi
317 //
318 // Let's define the following two functions:
319 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
320 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
321 //
322 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
323 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
324 // Ux(p) = 0
325 // Lx(p) = p
326 //
327 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
328 // with x bits:
329 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
330 //
331 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
332 // (i.e. all values in the x-bit space could be possible):
333 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
334 //
335 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
336 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
337 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
338 // the lower x bits the same number of times:
339 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
340 //
341 // If all upper x bits are equal, we can conclude:
342 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
343 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
344 //
345 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
346 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
347 //
348 //
349 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
350 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
351 // 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,
352 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
353 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
354 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
355 // that there is no underflow.
356 //
357 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
358 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
359 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
360 // can compare the upper x bits in a unified way with N() being the normalization function:
361 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
362 template<typename NativeType>
363 class IntegerTypeMultiplication {
364
365 NativeType _lo_left;
366 NativeType _lo_right;
367 NativeType _hi_left;
368 NativeType _hi_right;
369 short _widen_left;
370 short _widen_right;
371
372 static const Type* overflow_type();
373 static NativeType multiply_high(NativeType x, NativeType y);
374 const Type* create_type(NativeType lo, NativeType hi) const;
375
376 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
377 return normalize_overflow_value(x, y, multiply_high(x, y));
378 }
379
380 bool cross_product_not_same_overflow_value() const {
381 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
382 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
383 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
384 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
385 return lo_lo_high_product != lo_hi_high_product ||
386 lo_hi_high_product != hi_lo_high_product ||
387 hi_lo_high_product != hi_hi_high_product;
388 }
389
390 bool does_product_overflow(NativeType x, NativeType y) const {
391 return multiply_high_signed_overflow_value(x, y) != 0;
392 }
393
394 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
395 return java_multiply(x, y) < 0 ? result + 1 : result;
396 }
397
398 public:
399 template<class IntegerType>
400 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
401 : _lo_left(left->_lo), _lo_right(right->_lo),
402 _hi_left(left->_hi), _hi_right(right->_hi),
403 _widen_left(left->_widen), _widen_right(right->_widen) {}
404
405 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
406 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
407 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
408 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
409 const Type* compute() const {
410 if (cross_product_not_same_overflow_value()) {
411 return overflow_type();
412 }
413
414 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
415 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
416 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
417 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
418 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
419 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
420 return create_type(min, max);
421 }
422
423 bool does_overflow() const {
424 return does_product_overflow(_lo_left, _lo_right) ||
425 does_product_overflow(_lo_left, _hi_right) ||
426 does_product_overflow(_hi_left, _lo_right) ||
427 does_product_overflow(_hi_left, _hi_right);
428 }
429 };
430
431 template <>
432 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
433 return TypeInt::INT;
434 }
435
436 template <>
437 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
438 const jlong x_64 = x;
439 const jlong y_64 = y;
440 const jlong product = x_64 * y_64;
441 return (jint)((uint64_t)product >> 32u);
442 }
443
444 template <>
445 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
446 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
447 }
448
449 template <>
450 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
451 return TypeLong::LONG;
452 }
453
454 template <>
455 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
456 return multiply_high_signed(x, y);
457 }
458
459 template <>
460 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
461 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
462 }
463
464 // Compute the product type of two integer ranges into this node.
465 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
466 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
467 return integer_multiplication.compute();
468 }
469
470 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
471 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
472 return integer_multiplication.does_overflow();
473 }
474
475 // Compute the product type of two long ranges into this node.
476 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
477 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
478 return integer_multiplication.compute();
479 }
480
481 //=============================================================================
482 //------------------------------Ideal------------------------------------------
483 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
484 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
485 const jlong con = in(2)->find_long_con(0);
486 if (con == 0) {
487 // If in(2) is not a constant, call Ideal() of the parent class to
488 // try to move constant to the right side.
489 return MulNode::Ideal(phase, can_reshape);
490 }
491
492 // Now we have a constant Node on the right and the constant in con.
493 if (con == 1) {
494 // By one is handled by Identity call
495 return nullptr;
496 }
497
498 // Check for negative constant; if so negate the final result
499 bool sign_flip = false;
500 julong abs_con = g_uabs(con);
501 if (abs_con != (julong)con) {
502 sign_flip = true;
503 }
504
505 // Get low bit; check for being the only bit
506 Node *res = nullptr;
507 julong bit1 = submultiple_power_of_2(abs_con);
508 if (bit1 == abs_con) { // Found a power of 2?
509 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
510 } else {
511
512 // Check for constant with 2 bits set
513 julong bit2 = abs_con-bit1;
514 bit2 = bit2 & (0-bit2); // Extract 2nd bit
515 if (bit2 + bit1 == abs_con) { // Found all bits in con?
516 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
517 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
518 res = new AddLNode(n2, n1);
519
520 } else if (is_power_of_2(abs_con+1)) {
521 // Sleezy: power-of-2 -1. Next time be generic.
522 julong temp = abs_con + 1;
523 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
524 res = new SubLNode(n1, in(1));
525 } else {
526 return MulNode::Ideal(phase, can_reshape);
527 }
528 }
529
530 if (sign_flip) { // Need to negate result?
531 res = phase->transform(res);// Transform, before making the zero con
532 res = new SubLNode(phase->longcon(0),res);
533 }
534
535 return res; // Return final result
536 }
537
538 //=============================================================================
539 //------------------------------mul_ring---------------------------------------
540 // Compute the product type of two double ranges into this node.
541 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
542 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
543 return TypeF::make( t0->getf() * t1->getf() );
544 }
545
546 //------------------------------Ideal---------------------------------------
547 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
548 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
549 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
550
551 // x * 2 -> x + x
552 if (t2 != nullptr && t2->getf() == 2) {
553 Node* base = in(1);
554 return new AddFNode(base, base);
555 }
556 return MulNode::Ideal(phase, can_reshape);
557 }
558
559 //=============================================================================
560 //------------------------------Ideal------------------------------------------
561 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
562 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
563 const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
564
565 // x * 2 -> x + x
566 if (t2 != nullptr && t2->getf() == 2) {
567 Node* base = in(1);
568 return new AddHFNode(base, base);
569 }
570 return MulNode::Ideal(phase, can_reshape);
571 }
572
573 // Compute the product type of two half float ranges into this node.
574 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
575 if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
576 return Type::HALF_FLOAT;
577 }
578 return TypeH::make(t0->getf() * t1->getf());
579 }
580
581 //=============================================================================
582 //------------------------------mul_ring---------------------------------------
583 // Compute the product type of two double ranges into this node.
584 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
585 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
586 // We must be multiplying 2 double constants.
587 return TypeD::make( t0->getd() * t1->getd() );
588 }
589
590 //------------------------------Ideal---------------------------------------
591 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
592 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
593 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
594
595 // x * 2 -> x + x
596 if (t2 != nullptr && t2->getd() == 2) {
597 Node* base = in(1);
598 return new AddDNode(base, base);
599 }
600
601 return MulNode::Ideal(phase, can_reshape);
602 }
603
604 //=============================================================================
605 //------------------------------Value------------------------------------------
606 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
607 const Type *t1 = phase->type( in(1) );
608 const Type *t2 = phase->type( in(2) );
609 const Type *bot = bottom_type();
610 return MulHiValue(t1, t2, bot);
611 }
612
613 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
614 const Type *t1 = phase->type( in(1) );
615 const Type *t2 = phase->type( in(2) );
616 const Type *bot = bottom_type();
617 return MulHiValue(t1, t2, bot);
618 }
619
620 // A common routine used by UMulHiLNode and MulHiLNode
621 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
622 // Either input is TOP ==> the result is TOP
623 if( t1 == Type::TOP ) return Type::TOP;
624 if( t2 == Type::TOP ) return Type::TOP;
625
626 // Either input is BOTTOM ==> the result is the local BOTTOM
627 if( (t1 == bot) || (t2 == bot) ||
628 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
629 return bot;
630
631 // It is not worth trying to constant fold this stuff!
632 return TypeLong::LONG;
633 }
634
635 template<typename IntegerType>
636 static const IntegerType* and_value(const IntegerType* r0, const IntegerType* r1) {
637 typedef typename IntegerType::NativeType NativeType;
638 static_assert(std::is_signed<NativeType>::value, "Native type of IntegerType must be signed!");
639
640 int widen = MAX2(r0->_widen, r1->_widen);
641
642 // If both types are constants, we can calculate a constant result.
643 if (r0->is_con() && r1->is_con()) {
644 return IntegerType::make(r0->get_con() & r1->get_con());
645 }
646
647 // If both ranges are positive, the result will range from 0 up to the hi value of the smaller range. The minimum
648 // 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.
649 if (r0->_lo >= 0 && r1->_lo >= 0) {
650 return IntegerType::make(0, MIN2(r0->_hi, r1->_hi), widen);
651 }
652
653 // If only one range is positive, the result will range from 0 up to that range's maximum value.
654 // For the operation 'x & C' where C is a positive constant, the result will be in the range [0..C]. With that observation,
655 // 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]'
656 // where c >= 0 will be in the range [0..C].
657 if (r0->_lo >= 0) {
658 return IntegerType::make(0, r0->_hi, widen);
659 }
660
661 if (r1->_lo >= 0) {
662 return IntegerType::make(0, r1->_hi, widen);
663 }
664
665 // At this point, all positive ranges will have already been handled, so the only remaining cases will be negative ranges
666 // and constants.
667
668 assert(r0->_lo < 0 && r1->_lo < 0, "positive ranges should already be handled!");
669
670 // As two's complement means that both numbers will start with leading 1s, the lower bound of both ranges will contain
671 // the common leading 1s of both minimum values. In order to count them with count_leading_zeros, the bits are inverted.
672 NativeType sel_val = ~MIN2(r0->_lo, r1->_lo);
673
674 NativeType min;
675 if (sel_val == 0) {
676 // Since count_leading_zeros is undefined at 0, we short-circuit the condition where both ranges have a minimum of -1.
677 min = -1;
678 } else {
679 // 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.
680 int shift_bits = count_leading_zeros(sel_val) - 1;
681 min = std::numeric_limits<NativeType>::min() >> shift_bits;
682 }
683
684 NativeType max;
685 if (r0->_hi < 0 && r1->_hi < 0) {
686 // If both ranges are negative, then the same optimization as both positive ranges will apply, and the smaller hi
687 // value will mask off any bits set by higher values.
688 max = MIN2(r0->_hi, r1->_hi);
689 } else {
690 // In the case of ranges that cross zero, negative values can cause the higher order bits to be set, so the maximum
691 // positive value can be as high as the larger hi value.
692 max = MAX2(r0->_hi, r1->_hi);
693 }
694
695 return IntegerType::make(min, max, widen);
696 }
697
698 //=============================================================================
699 //------------------------------mul_ring---------------------------------------
700 // Supplied function returns the product of the inputs IN THE CURRENT RING.
701 // For the logical operations the ring's MUL is really a logical AND function.
702 // This also type-checks the inputs for sanity. Guaranteed never to
703 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
704 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
705 const TypeInt* r0 = t0->is_int();
706 const TypeInt* r1 = t1->is_int();
707
708 return and_value<TypeInt>(r0, r1);
709 }
710
711 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
712
713 const Type* AndINode::Value(PhaseGVN* phase) const {
714 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
715 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
716 return TypeInt::ZERO;
717 }
718
719 return MulNode::Value(phase);
720 }
721
722 //------------------------------Identity---------------------------------------
723 // Masking off the high bits of an unsigned load is not required
724 Node* AndINode::Identity(PhaseGVN* phase) {
725
726 // x & x => x
727 if (in(1) == in(2)) {
728 return in(1);
729 }
730
731 Node* in1 = in(1);
732 uint op = in1->Opcode();
733 const TypeInt* t2 = phase->type(in(2))->isa_int();
734 if (t2 && t2->is_con()) {
735 int con = t2->get_con();
736 // Masking off high bits which are always zero is useless.
737 const TypeInt* t1 = phase->type(in(1))->isa_int();
738 if (t1 != nullptr && t1->_lo >= 0) {
739 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
740 if ((t1_support & con) == t1_support)
741 return in1;
742 }
743 // Masking off the high bits of a unsigned-shift-right is not
744 // needed either.
745 if (op == Op_URShiftI) {
746 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
747 if (t12 && t12->is_con()) { // Shift is by a constant
748 int shift = t12->get_con();
749 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
750 int mask = max_juint >> shift;
751 if ((mask & con) == mask) // If AND is useless, skip it
752 return in1;
753 }
754 }
755 }
756 return MulNode::Identity(phase);
757 }
758
759 //------------------------------Ideal------------------------------------------
760 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
761 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
762 Node* progress = AndIL_sum_and_mask(phase, T_INT);
763 if (progress != nullptr) {
764 return progress;
765 }
766
767 // Convert "(~a) & (~b)" into "~(a | b)"
768 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
769 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
770 Node* tn = phase->transform(or_a_b);
771 return AddNode::make_not(phase, tn, T_INT);
772 }
773
774 // Special case constant AND mask
775 const TypeInt *t2 = phase->type( in(2) )->isa_int();
776 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
777 const int mask = t2->get_con();
778 Node *load = in(1);
779 uint lop = load->Opcode();
780
781 // Masking bits off of a Character? Hi bits are already zero.
782 if( lop == Op_LoadUS &&
783 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
784 return new AndINode(load,phase->intcon(mask&0xFFFF));
785
786 // Masking bits off of a Short? Loading a Character does some masking
787 if (can_reshape &&
788 load->outcnt() == 1 && load->unique_out() == this) {
789 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
790 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
791 ldus = phase->transform(ldus);
792 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
793 }
794
795 // Masking sign bits off of a Byte? Do an unsigned byte load plus
796 // an and.
797 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
798 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
799 ldub = phase->transform(ldub);
800 return new AndINode(ldub, phase->intcon(mask));
801 }
802 }
803
804 // Masking off sign bits? Dont make them!
805 if( lop == Op_RShiftI ) {
806 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
807 if( t12 && t12->is_con() ) { // Shift is by a constant
808 int shift = t12->get_con();
809 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
810 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
811 // If the AND'ing of the 2 masks has no bits, then only original shifted
812 // bits survive. NO sign-extension bits survive the maskings.
813 if( (sign_bits_mask & mask) == 0 ) {
814 // Use zero-fill shift instead
815 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
816 return new AndINode( zshift, in(2) );
817 }
818 }
819 }
820
821 // Check for 'negate/and-1', a pattern emitted when someone asks for
822 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
823 // plus 1) and the mask is of the low order bit. Skip the negate.
824 if( lop == Op_SubI && mask == 1 && load->in(1) &&
825 phase->type(load->in(1)) == TypeInt::ZERO )
826 return new AndINode( load->in(2), in(2) );
827
828 return MulNode::Ideal(phase, can_reshape);
829 }
830
831 //=============================================================================
832 //------------------------------mul_ring---------------------------------------
833 // Supplied function returns the product of the inputs IN THE CURRENT RING.
834 // For the logical operations the ring's MUL is really a logical AND function.
835 // This also type-checks the inputs for sanity. Guaranteed never to
836 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
837 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
838 const TypeLong* r0 = t0->is_long();
839 const TypeLong* r1 = t1->is_long();
840
841 return and_value<TypeLong>(r0, r1);
842 }
843
844 const Type* AndLNode::Value(PhaseGVN* phase) const {
845 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
846 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
847 return TypeLong::ZERO;
848 }
849
850 return MulNode::Value(phase);
851 }
852
853 //------------------------------Identity---------------------------------------
854 // Masking off the high bits of an unsigned load is not required
855 Node* AndLNode::Identity(PhaseGVN* phase) {
856
857 // x & x => x
858 if (in(1) == in(2)) {
859 return in(1);
860 }
861
862 Node *usr = in(1);
863 const TypeLong *t2 = phase->type( in(2) )->isa_long();
864 if( t2 && t2->is_con() ) {
865 jlong con = t2->get_con();
866 // Masking off high bits which are always zero is useless.
867 const TypeLong* t1 = phase->type( in(1) )->isa_long();
868 if (t1 != nullptr && t1->_lo >= 0) {
869 int bit_count = log2i_graceful(t1->_hi) + 1;
870 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
871 if ((t1_support & con) == t1_support)
872 return usr;
873 }
874 uint lop = usr->Opcode();
875 // Masking off the high bits of a unsigned-shift-right is not
876 // needed either.
877 if( lop == Op_URShiftL ) {
878 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
879 if( t12 && t12->is_con() ) { // Shift is by a constant
880 int shift = t12->get_con();
881 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
882 jlong mask = max_julong >> shift;
883 if( (mask&con) == mask ) // If AND is useless, skip it
884 return usr;
885 }
886 }
887 }
888 return MulNode::Identity(phase);
889 }
890
891 //------------------------------Ideal------------------------------------------
892 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
893 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
894 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
895 if (progress != nullptr) {
896 return progress;
897 }
898
899 // Convert "(~a) & (~b)" into "~(a | b)"
900 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
901 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
902 Node* tn = phase->transform(or_a_b);
903 return AddNode::make_not(phase, tn, T_LONG);
904 }
905
906 // Special case constant AND mask
907 const TypeLong *t2 = phase->type( in(2) )->isa_long();
908 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
909 const jlong mask = t2->get_con();
910
911 Node* in1 = in(1);
912 int op = in1->Opcode();
913
914 // Are we masking a long that was converted from an int with a mask
915 // that fits in 32-bits? Commute them and use an AndINode. Don't
916 // convert masks which would cause a sign extension of the integer
917 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
918 // would be optimized away later in Identity.
919 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
920 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
921 andi = phase->transform(andi);
922 return new ConvI2LNode(andi);
923 }
924
925 // Masking off sign bits? Dont make them!
926 if (op == Op_RShiftL) {
927 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
928 if( t12 && t12->is_con() ) { // Shift is by a constant
929 int shift = t12->get_con();
930 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
931 if (shift != 0) {
932 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
933 // If the AND'ing of the 2 masks has no bits, then only original shifted
934 // bits survive. NO sign-extension bits survive the maskings.
935 if( (sign_bits_mask & mask) == 0 ) {
936 // Use zero-fill shift instead
937 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
938 return new AndLNode(zshift, in(2));
939 }
940 }
941 }
942 }
943
944 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
945 Node* load1 = in(1);
946 Node* load2 = nullptr;
947 if (load1->is_Phi() && phase->type(load1)->isa_long()) {
948 load1 = in(1)->in(1);
949 load2 = in(1)->in(2);
950 }
951 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
952 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
953 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
954 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
955 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
956 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
957 if (mask == markWord::inline_type_pattern) {
958 if (adr_t1->is_inlinetypeptr()) {
959 set_req_X(1, in(2), phase);
960 return this;
961 } else if (!adr_t1->can_be_inline_type()) {
962 set_req_X(1, phase->longcon(0), phase);
963 return this;
964 }
965 } else if (mask == markWord::null_free_array_bit_in_place) {
966 if (adr_t1->is_null_free()) {
967 set_req_X(1, in(2), phase);
968 return this;
969 } else if (adr_t1->is_not_null_free()) {
970 set_req_X(1, phase->longcon(0), phase);
971 return this;
972 }
973 } else if (mask == markWord::flat_array_bit_in_place) {
974 if (adr_t1->is_flat()) {
975 set_req_X(1, in(2), phase);
976 return this;
977 } else if (adr_t1->is_not_flat()) {
978 set_req_X(1, phase->longcon(0), phase);
979 return this;
980 }
981 }
982 }
983 }
984
985 return MulNode::Ideal(phase, can_reshape);
986 }
987
988 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
989 switch (bt) {
990 case T_INT:
991 return new LShiftINode(in1, in2);
992 case T_LONG:
993 return new LShiftLNode(in1, in2);
994 default:
995 fatal("Not implemented for %s", type2name(bt));
996 }
997 return nullptr;
998 }
999
1000 // Returns whether the shift amount is constant. If so, sets count.
1001 static bool const_shift_count(PhaseGVN* phase, const Node* shift_node, int* count) {
1002 const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
1003 if (tcount != nullptr && tcount->is_con()) {
1004 *count = tcount->get_con();
1005 return true;
1006 }
1007 return false;
1008 }
1009
1010 // Returns whether the shift amount is constant. If so, sets real_shift and masked_shift.
1011 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, int& real_shift, uint& masked_shift) {
1012 if (const_shift_count(phase, shift_node, &real_shift)) {
1013 masked_shift = real_shift & (nBits - 1);
1014 return true;
1015 }
1016 return false;
1017 }
1018
1019 // Convenience for when we don't care about the real amount
1020 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, uint& masked_shift) {
1021 int real_shift;
1022 return mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift);
1023 }
1024
1025 // Use this in ::Ideal only with shiftNode == this!
1026 // Returns the masked shift amount if constant or 0 if not constant.
1027 static uint mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits) {
1028 int real_shift;
1029 uint masked_shift;
1030 if (mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift)) {
1031 if (masked_shift == 0) {
1032 // Let Identity() handle 0 shift count.
1033 return 0;
1034 }
1035
1036 if (real_shift != (int)masked_shift) {
1037 PhaseIterGVN* igvn = phase->is_IterGVN();
1038 if (igvn != nullptr) {
1039 igvn->_worklist.push(shift_node);
1040 }
1041 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
1042 }
1043 return masked_shift;
1044 }
1045 // Not a shift by a constant.
1046 return 0;
1047 }
1048
1049 // Called with
1050 // outer_shift = (_ << rhs_outer)
1051 // We are looking for the pattern:
1052 // outer_shift = ((X << rhs_inner) << rhs_outer)
1053 // where rhs_outer and rhs_inner are constant
1054 // we denote inner_shift the nested expression (X << rhs_inner)
1055 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
1056 // where nbits is the number of bits of the shifts
1057 //
1058 // There are 2 cases:
1059 // if con_outer + con_inner >= nbits => 0
1060 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
1061 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
1062 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
1063 const Node* inner_shift = outer_shift->in(1);
1064 if (inner_shift->Opcode() != Op_LShift(bt)) {
1065 return nullptr;
1066 }
1067
1068 uint nbits = bits_per_java_integer(bt);
1069 uint con_inner;
1070 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
1071 return nullptr;
1072 }
1073
1074 if (con_inner == 0) {
1075 // We let the Identity() of the inner shift do its job.
1076 return nullptr;
1077 }
1078
1079 if (con_outer + con_inner >= nbits) {
1080 // While it might be tempting to use
1081 // phase->zerocon(bt);
1082 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
1083 // to return a new node, this or nullptr, but not an old (cached) node.
1084 return ConNode::make(TypeInteger::zero(bt));
1085 }
1086
1087 // con0 + con1 < nbits ==> actual shift happens now
1088 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
1089 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
1090 }
1091
1092 //------------------------------Identity---------------------------------------
1093 Node* LShiftINode::Identity(PhaseGVN* phase) {
1094 return IdentityIL(phase, T_INT);
1095 }
1096
1097 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1098 uint con = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1099 if (con == 0) {
1100 return nullptr;
1101 }
1102
1103 // If the right input is a constant, and the left input is an add of a
1104 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1105 Node* add1 = in(1);
1106 int add1_op = add1->Opcode();
1107 if (add1_op == Op_Add(bt)) { // Left input is an add?
1108 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
1109
1110 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1111 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1112 if (bt != T_INT || con < 16) {
1113 // Left input is an add of the same number?
1114 if (con != (bits_per_java_integer(bt) - 1) && add1->in(1) == add1->in(2)) {
1115 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1116 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
1117 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1118 // or c0 != 63 (for LShiftL) because:
1119 // (x + x) << 63 = 2x << 63, while
1120 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1121 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1122 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1123 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1124 }
1125
1126 // Left input is an add of a constant?
1127 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1128 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1129 // Compute X << con0
1130 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1131 // Compute X<<con0 + (con1<<con0)
1132 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1133 }
1134 }
1135 }
1136
1137 // Check for "(x >> C1) << C2"
1138 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1139 int add1Con = 0;
1140 const_shift_count(phase, add1, &add1Con);
1141
1142 // Special case C1 == C2, which just masks off low bits
1143 if (add1Con > 0 && con == (uint)add1Con) {
1144 // Convert to "(x & -(1 << C2))"
1145 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1146 } else {
1147 // Wait until the right shift has been sharpened to the correct count
1148 if (add1Con > 0 && (uint)add1Con < bits_per_java_integer(bt)) {
1149 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1150 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1151 if (phase->is_IterGVN()) {
1152 if (con > (uint)add1Con) {
1153 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1154 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1155 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1156 } else {
1157 assert(con < (uint)add1Con, "must be (%d < %d)", con, add1Con);
1158 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1159
1160 // Handle logical and arithmetic shifts
1161 Node* rshift;
1162 if (add1_op == Op_RShift(bt)) {
1163 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1164 } else {
1165 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1166 }
1167
1168 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1169 }
1170 } else {
1171 phase->record_for_igvn(this);
1172 }
1173 }
1174 }
1175 }
1176
1177 // Check for "((x >> C1) & Y) << C2"
1178 if (add1_op == Op_And(bt)) {
1179 Node* add2 = add1->in(1);
1180 int add2_op = add2->Opcode();
1181 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1182 // Special case C1 == C2, which just masks off low bits
1183 if (add2->in(2) == in(2)) {
1184 // Convert to "(x & (Y << C2))"
1185 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1186 return MulNode::make_and(add2->in(1), y_sh, bt);
1187 }
1188
1189 int add2Con = 0;
1190 const_shift_count(phase, add2, &add2Con);
1191 if (add2Con > 0 && (uint)add2Con < bits_per_java_integer(bt)) {
1192 if (phase->is_IterGVN()) {
1193 // Convert to "((x >> C1) << C2) & (Y << C2)"
1194
1195 // Make "(x >> C1) << C2", which will get folded away by the rule above
1196 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1197 // Make "Y << C2", which will simplify when Y is a constant
1198 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1199
1200 return MulNode::make_and(x_sh, y_sh, bt);
1201 } else {
1202 phase->record_for_igvn(this);
1203 }
1204 }
1205 }
1206 }
1207
1208 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1209 // before shifting them away.
1210 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1211 assert(bt != T_INT || bits_mask == right_n_bits(bits_per_java_integer(bt)-con), "inconsistent");
1212 if (add1_op == Op_And(bt) &&
1213 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1214 return LShiftNode::make(add1->in(1), in(2), bt);
1215 }
1216
1217 // Collapse nested left-shifts with constant rhs:
1218 // (X << con1) << con2 ==> X << (con1 + con2)
1219 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1220 if (doubleShift != nullptr) {
1221 return doubleShift;
1222 }
1223
1224 return nullptr;
1225 }
1226
1227 //------------------------------Ideal------------------------------------------
1228 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1229 return IdealIL(phase, can_reshape, T_INT);
1230 }
1231
1232 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1233 const Type* t1 = phase->type(in(1));
1234 const Type* t2 = phase->type(in(2));
1235 // Either input is TOP ==> the result is TOP
1236 if (t1 == Type::TOP) {
1237 return Type::TOP;
1238 }
1239 if (t2 == Type::TOP) {
1240 return Type::TOP;
1241 }
1242
1243 // Left input is ZERO ==> the result is ZERO.
1244 if (t1 == TypeInteger::zero(bt)) {
1245 return TypeInteger::zero(bt);
1246 }
1247 // Shift by zero does nothing
1248 if (t2 == TypeInt::ZERO) {
1249 return t1;
1250 }
1251
1252 // Either input is BOTTOM ==> the result is BOTTOM
1253 if ((t1 == TypeInteger::bottom(bt)) || (t2 == TypeInt::INT) ||
1254 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM)) {
1255 return TypeInteger::bottom(bt);
1256 }
1257
1258 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1259 const TypeInt* r2 = t2->is_int(); // Handy access
1260
1261 if (!r2->is_con()) {
1262 return TypeInteger::bottom(bt);
1263 }
1264
1265 uint shift = r2->get_con();
1266 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1267 // Shift by a multiple of 32/64 does nothing:
1268 if (shift == 0) {
1269 return t1;
1270 }
1271
1272 // If the shift is a constant, shift the bounds of the type,
1273 // unless this could lead to an overflow.
1274 if (!r1->is_con()) {
1275 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1276 #ifdef ASSERT
1277 if (bt == T_INT) {
1278 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1279 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1280 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1281 }
1282 #endif
1283 if (java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo &&
1284 java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) {
1285 // No overflow. The range shifts up cleanly.
1286 return TypeInteger::make(java_shift_left(lo, shift, bt),
1287 java_shift_left(hi, shift, bt),
1288 MAX2(r1->_widen, r2->_widen), bt);
1289 }
1290 return TypeInteger::bottom(bt);
1291 }
1292
1293 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1294 }
1295
1296 //------------------------------Value------------------------------------------
1297 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1298 return ValueIL(phase, T_INT);
1299 }
1300
1301 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1302 int count = 0;
1303 if (const_shift_count(phase, this, &count) && (count & (bits_per_java_integer(bt) - 1)) == 0) {
1304 // Shift by a multiple of 32/64 does nothing
1305 return in(1);
1306 }
1307 return this;
1308 }
1309
1310 //=============================================================================
1311 //------------------------------Identity---------------------------------------
1312 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1313 return IdentityIL(phase, T_LONG);
1314 }
1315
1316 //------------------------------Ideal------------------------------------------
1317 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1318 return IdealIL(phase, can_reshape, T_LONG);
1319 }
1320
1321 //------------------------------Value------------------------------------------
1322 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1323 return ValueIL(phase, T_LONG);
1324 }
1325
1326 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1327 switch (bt) {
1328 case T_INT:
1329 return new RShiftINode(in1, in2);
1330 case T_LONG:
1331 return new RShiftLNode(in1, in2);
1332 default:
1333 fatal("Not implemented for %s", type2name(bt));
1334 }
1335 return nullptr;
1336 }
1337
1338
1339 //=============================================================================
1340 //------------------------------Identity---------------------------------------
1341 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1342 int count = 0;
1343 if (const_shift_count(phase, this, &count)) {
1344 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1345 // Shift by a multiple of 32/64 does nothing
1346 return in(1);
1347 }
1348 // Check for useless sign-masking
1349 if (in(1)->Opcode() == Op_LShift(bt) &&
1350 in(1)->req() == 3 &&
1351 in(1)->in(2) == in(2)) {
1352 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1353 // Compute masks for which this shifting doesn't change
1354 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1355 jlong hi = ~lo; // 00007FFF
1356 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1357 if (t11 == nullptr) {
1358 return this;
1359 }
1360 // Does actual value fit inside of mask?
1361 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1362 return in(1)->in(1); // Then shifting is a nop
1363 }
1364 }
1365 }
1366 return this;
1367 }
1368
1369 Node* RShiftINode::Identity(PhaseGVN* phase) {
1370 return IdentityIL(phase, T_INT);
1371 }
1372
1373 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1374 // Inputs may be TOP if they are dead.
1375 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1376 if (t1 == nullptr) {
1377 return NodeSentinel; // Left input is an integer
1378 }
1379 int shift = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1380 if (shift == 0) {
1381 return NodeSentinel;
1382 }
1383
1384 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1385 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1386 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1387 const Node* and_node = in(1);
1388 if (and_node->Opcode() != Op_And(bt)) {
1389 return nullptr;
1390 }
1391 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1392 if (mask_t != nullptr && mask_t->is_con()) {
1393 jlong maskbits = mask_t->get_con_as_long(bt);
1394 // Convert to "(x >> shift) & (mask >> shift)"
1395 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1396 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1397 }
1398 return nullptr;
1399 }
1400
1401 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1402 Node* progress = IdealIL(phase, can_reshape, T_INT);
1403 if (progress == NodeSentinel) {
1404 return nullptr;
1405 }
1406 if (progress != nullptr) {
1407 return progress;
1408 }
1409 int shift = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1410 assert(shift != 0, "handled by IdealIL");
1411
1412 // Check for "(short[i] <<16)>>16" which simply sign-extends
1413 const Node *shl = in(1);
1414 if (shl->Opcode() != Op_LShiftI) {
1415 return nullptr;
1416 }
1417
1418 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1419 if (left_shift_t == nullptr) {
1420 return nullptr;
1421 }
1422 if (shift == 16 && left_shift_t->is_con(16)) {
1423 Node *ld = shl->in(1);
1424 if (ld->Opcode() == Op_LoadS) {
1425 // Sign extension is just useless here. Return a RShiftI of zero instead
1426 // returning 'ld' directly. We cannot return an old Node directly as
1427 // that is the job of 'Identity' calls and Identity calls only work on
1428 // direct inputs ('ld' is an extra Node removed from 'this'). The
1429 // combined optimization requires Identity only return direct inputs.
1430 set_req_X(1, ld, phase);
1431 set_req_X(2, phase->intcon(0), phase);
1432 return this;
1433 }
1434 else if (can_reshape &&
1435 ld->Opcode() == Op_LoadUS &&
1436 ld->outcnt() == 1 && ld->unique_out() == shl)
1437 // Replace zero-extension-load with sign-extension-load
1438 return ld->as_Load()->convert_to_signed_load(*phase);
1439 }
1440
1441 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1442 if (shift == 24 && left_shift_t->is_con(24)) {
1443 Node *ld = shl->in(1);
1444 if (ld->Opcode() == Op_LoadB) {
1445 // Sign extension is just useless here
1446 set_req_X(1, ld, phase);
1447 set_req_X(2, phase->intcon(0), phase);
1448 return this;
1449 }
1450 }
1451
1452 return nullptr;
1453 }
1454
1455 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1456 const Type* t1 = phase->type(in(1));
1457 const Type* t2 = phase->type(in(2));
1458 // Either input is TOP ==> the result is TOP
1459 if (t1 == Type::TOP) {
1460 return Type::TOP;
1461 }
1462 if (t2 == Type::TOP) {
1463 return Type::TOP;
1464 }
1465
1466 // Left input is ZERO ==> the result is ZERO.
1467 if (t1 == TypeInteger::zero(bt)) {
1468 return TypeInteger::zero(bt);
1469 }
1470 // Shift by zero does nothing
1471 if (t2 == TypeInt::ZERO) {
1472 return t1;
1473 }
1474
1475 // Either input is BOTTOM ==> the result is BOTTOM
1476 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1477 return TypeInteger::bottom(bt);
1478 }
1479
1480 const TypeInteger* r1 = t1->isa_integer(bt);
1481 const TypeInt* r2 = t2->isa_int();
1482
1483 // If the shift is a constant, just shift the bounds of the type.
1484 // For example, if the shift is 31/63, we just propagate sign bits.
1485 if (!r1->is_con() && r2->is_con()) {
1486 uint shift = r2->get_con();
1487 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1488 // Shift by a multiple of 32/64 does nothing:
1489 if (shift == 0) {
1490 return t1;
1491 }
1492 // Calculate reasonably aggressive bounds for the result.
1493 // This is necessary if we are to correctly type things
1494 // like (x<<24>>24) == ((byte)x).
1495 jlong lo = r1->lo_as_long() >> (jint)shift;
1496 jlong hi = r1->hi_as_long() >> (jint)shift;
1497 assert(lo <= hi, "must have valid bounds");
1498 #ifdef ASSERT
1499 if (bt == T_INT) {
1500 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1501 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1502 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1503 }
1504 #endif
1505 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1506 #ifdef ASSERT
1507 // Make sure we get the sign-capture idiom correct.
1508 if (shift == bits_per_java_integer(bt) - 1) {
1509 if (r1->lo_as_long() >= 0) {
1510 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1511 }
1512 if (r1->hi_as_long() < 0) {
1513 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1514 }
1515 }
1516 #endif
1517 return ti;
1518 }
1519
1520 if (!r1->is_con() || !r2->is_con()) {
1521 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1522 if (r1->lo_as_long() >= 0) {
1523 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1524 }
1525
1526 // Conversely, if the left input is negative then the result must be negative.
1527 if (r1->hi_as_long() <= -1) {
1528 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1529 }
1530
1531 return TypeInteger::bottom(bt);
1532 }
1533
1534 // Signed shift right
1535 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1536 }
1537
1538 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1539 return ValueIL(phase, T_INT);
1540 }
1541
1542 //=============================================================================
1543 //------------------------------Identity---------------------------------------
1544 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1545 return IdentityIL(phase, T_LONG);
1546 }
1547
1548 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1549 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1550 if (progress == NodeSentinel) {
1551 return nullptr;
1552 }
1553 return progress;
1554 }
1555
1556 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1557 return ValueIL(phase, T_LONG);
1558 }
1559
1560 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1561 switch (bt) {
1562 case T_INT:
1563 return new URShiftINode(in1, in2);
1564 case T_LONG:
1565 return new URShiftLNode(in1, in2);
1566 default:
1567 fatal("Not implemented for %s", type2name(bt));
1568 }
1569 return nullptr;
1570 }
1571
1572 //=============================================================================
1573 //------------------------------Identity---------------------------------------
1574 Node* URShiftINode::Identity(PhaseGVN* phase) {
1575 int count = 0;
1576 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1577 // Shift by a multiple of 32 does nothing
1578 return in(1);
1579 }
1580
1581 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1582 // Happens during new-array length computation.
1583 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1584 Node *add = in(1);
1585 if (add->Opcode() == Op_AddI) {
1586 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1587 if (t2 && t2->is_con(wordSize - 1) &&
1588 add->in(1)->Opcode() == Op_LShiftI) {
1589 // Check that shift_counts are LogBytesPerWord.
1590 Node *lshift_count = add->in(1)->in(2);
1591 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1592 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1593 t_lshift_count == phase->type(in(2))) {
1594 Node *x = add->in(1)->in(1);
1595 const TypeInt *t_x = phase->type(x)->isa_int();
1596 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1597 return x;
1598 }
1599 }
1600 }
1601 }
1602
1603 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1604 }
1605
1606 //------------------------------Ideal------------------------------------------
1607 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1608 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1609 if (con == 0) {
1610 return nullptr;
1611 }
1612
1613 // We'll be wanting the right-shift amount as a mask of that many bits
1614 const int mask = right_n_bits(BitsPerJavaInteger - con);
1615
1616 int in1_op = in(1)->Opcode();
1617
1618 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1619 if( in1_op == Op_URShiftI ) {
1620 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1621 if( t12 && t12->is_con() ) { // Right input is a constant
1622 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1623 const int con2 = t12->get_con() & 31; // Shift count is always masked
1624 const int con3 = con+con2;
1625 if( con3 < 32 ) // Only merge shifts if total is < 32
1626 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1627 }
1628 }
1629
1630 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1631 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1632 // If Q is "X << z" the rounding is useless. Look for patterns like
1633 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1634 Node *add = in(1);
1635 const TypeInt *t2 = phase->type(in(2))->isa_int();
1636 if (in1_op == Op_AddI) {
1637 Node *lshl = add->in(1);
1638 if( lshl->Opcode() == Op_LShiftI &&
1639 phase->type(lshl->in(2)) == t2 ) {
1640 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1641 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1642 return new AndINode( sum, phase->intcon(mask) );
1643 }
1644 }
1645
1646 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1647 // This shortens the mask. Also, if we are extracting a high byte and
1648 // storing it to a buffer, the mask will be removed completely.
1649 Node *andi = in(1);
1650 if( in1_op == Op_AndI ) {
1651 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1652 if( t3 && t3->is_con() ) { // Right input is a constant
1653 jint mask2 = t3->get_con();
1654 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1655 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1656 return new AndINode(newshr, phase->intcon(mask2));
1657 // The negative values are easier to materialize than positive ones.
1658 // A typical case from address arithmetic is ((x & ~15) >> 4).
1659 // It's better to change that to ((x >> 4) & ~0) versus
1660 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1661 }
1662 }
1663
1664 // Check for "(X << z ) >>> z" which simply zero-extends
1665 Node *shl = in(1);
1666 if( in1_op == Op_LShiftI &&
1667 phase->type(shl->in(2)) == t2 )
1668 return new AndINode( shl->in(1), phase->intcon(mask) );
1669
1670 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1671 Node *shr = in(1);
1672 if ( in1_op == Op_RShiftI ) {
1673 Node *in11 = shr->in(1);
1674 Node *in12 = shr->in(2);
1675 const TypeInt *t11 = phase->type(in11)->isa_int();
1676 const TypeInt *t12 = phase->type(in12)->isa_int();
1677 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1678 return new URShiftINode(in11, phase->intcon(31));
1679 }
1680 }
1681
1682 return nullptr;
1683 }
1684
1685 //------------------------------Value------------------------------------------
1686 // A URShiftINode shifts its input2 right by input1 amount.
1687 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1688 // (This is a near clone of RShiftINode::Value.)
1689 const Type *t1 = phase->type( in(1) );
1690 const Type *t2 = phase->type( in(2) );
1691 // Either input is TOP ==> the result is TOP
1692 if( t1 == Type::TOP ) return Type::TOP;
1693 if( t2 == Type::TOP ) return Type::TOP;
1694
1695 // Left input is ZERO ==> the result is ZERO.
1696 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1697 // Shift by zero does nothing
1698 if( t2 == TypeInt::ZERO ) return t1;
1699
1700 // Either input is BOTTOM ==> the result is BOTTOM
1701 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1702 return TypeInt::INT;
1703
1704 if (t2 == TypeInt::INT)
1705 return TypeInt::INT;
1706
1707 const TypeInt *r1 = t1->is_int(); // Handy access
1708 const TypeInt *r2 = t2->is_int(); // Handy access
1709
1710 if (r2->is_con()) {
1711 uint shift = r2->get_con();
1712 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1713 // Shift by a multiple of 32 does nothing:
1714 if (shift == 0) return t1;
1715 // Calculate reasonably aggressive bounds for the result.
1716 jint lo = (juint)r1->_lo >> (juint)shift;
1717 jint hi = (juint)r1->_hi >> (juint)shift;
1718 if (r1->_hi >= 0 && r1->_lo < 0) {
1719 // If the type has both negative and positive values,
1720 // there are two separate sub-domains to worry about:
1721 // The positive half and the negative half.
1722 jint neg_lo = lo;
1723 jint neg_hi = (juint)-1 >> (juint)shift;
1724 jint pos_lo = (juint) 0 >> (juint)shift;
1725 jint pos_hi = hi;
1726 lo = MIN2(neg_lo, pos_lo); // == 0
1727 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1728 }
1729 assert(lo <= hi, "must have valid bounds");
1730 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1731 #ifdef ASSERT
1732 // Make sure we get the sign-capture idiom correct.
1733 if (shift == BitsPerJavaInteger-1) {
1734 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1735 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1736 }
1737 #endif
1738 return ti;
1739 }
1740
1741 //
1742 // Do not support shifted oops in info for GC
1743 //
1744 // else if( t1->base() == Type::InstPtr ) {
1745 //
1746 // const TypeInstPtr *o = t1->is_instptr();
1747 // if( t1->singleton() )
1748 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1749 // }
1750 // else if( t1->base() == Type::KlassPtr ) {
1751 // const TypeKlassPtr *o = t1->is_klassptr();
1752 // if( t1->singleton() )
1753 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1754 // }
1755
1756 return TypeInt::INT;
1757 }
1758
1759 //=============================================================================
1760 //------------------------------Identity---------------------------------------
1761 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1762 int count = 0;
1763 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1764 // Shift by a multiple of 64 does nothing
1765 return in(1);
1766 }
1767 return this;
1768 }
1769
1770 //------------------------------Ideal------------------------------------------
1771 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1772 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong);
1773 if (con == 0) {
1774 return nullptr;
1775 }
1776
1777 // We'll be wanting the right-shift amount as a mask of that many bits
1778 const jlong mask = jlong(max_julong >> con);
1779
1780 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1781 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1782 // If Q is "X << z" the rounding is useless. Look for patterns like
1783 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1784 Node *add = in(1);
1785 const TypeInt *t2 = phase->type(in(2))->isa_int();
1786 if (add->Opcode() == Op_AddL) {
1787 Node *lshl = add->in(1);
1788 if( lshl->Opcode() == Op_LShiftL &&
1789 phase->type(lshl->in(2)) == t2 ) {
1790 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1791 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1792 return new AndLNode( sum, phase->longcon(mask) );
1793 }
1794 }
1795
1796 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1797 // This shortens the mask. Also, if we are extracting a high byte and
1798 // storing it to a buffer, the mask will be removed completely.
1799 Node *andi = in(1);
1800 if( andi->Opcode() == Op_AndL ) {
1801 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1802 if( t3 && t3->is_con() ) { // Right input is a constant
1803 jlong mask2 = t3->get_con();
1804 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1805 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1806 return new AndLNode(newshr, phase->longcon(mask2));
1807 }
1808 }
1809
1810 // Check for "(X << z ) >>> z" which simply zero-extends
1811 Node *shl = in(1);
1812 if( shl->Opcode() == Op_LShiftL &&
1813 phase->type(shl->in(2)) == t2 )
1814 return new AndLNode( shl->in(1), phase->longcon(mask) );
1815
1816 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1817 Node *shr = in(1);
1818 if ( shr->Opcode() == Op_RShiftL ) {
1819 Node *in11 = shr->in(1);
1820 Node *in12 = shr->in(2);
1821 const TypeLong *t11 = phase->type(in11)->isa_long();
1822 const TypeInt *t12 = phase->type(in12)->isa_int();
1823 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1824 return new URShiftLNode(in11, phase->intcon(63));
1825 }
1826 }
1827 return nullptr;
1828 }
1829
1830 //------------------------------Value------------------------------------------
1831 // A URShiftINode shifts its input2 right by input1 amount.
1832 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1833 // (This is a near clone of RShiftLNode::Value.)
1834 const Type *t1 = phase->type( in(1) );
1835 const Type *t2 = phase->type( in(2) );
1836 // Either input is TOP ==> the result is TOP
1837 if( t1 == Type::TOP ) return Type::TOP;
1838 if( t2 == Type::TOP ) return Type::TOP;
1839
1840 // Left input is ZERO ==> the result is ZERO.
1841 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1842 // Shift by zero does nothing
1843 if( t2 == TypeInt::ZERO ) return t1;
1844
1845 // Either input is BOTTOM ==> the result is BOTTOM
1846 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1847 return TypeLong::LONG;
1848
1849 if (t2 == TypeInt::INT)
1850 return TypeLong::LONG;
1851
1852 const TypeLong *r1 = t1->is_long(); // Handy access
1853 const TypeInt *r2 = t2->is_int (); // Handy access
1854
1855 if (r2->is_con()) {
1856 uint shift = r2->get_con();
1857 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1858 // Shift by a multiple of 64 does nothing:
1859 if (shift == 0) return t1;
1860 // Calculate reasonably aggressive bounds for the result.
1861 jlong lo = (julong)r1->_lo >> (juint)shift;
1862 jlong hi = (julong)r1->_hi >> (juint)shift;
1863 if (r1->_hi >= 0 && r1->_lo < 0) {
1864 // If the type has both negative and positive values,
1865 // there are two separate sub-domains to worry about:
1866 // The positive half and the negative half.
1867 jlong neg_lo = lo;
1868 jlong neg_hi = (julong)-1 >> (juint)shift;
1869 jlong pos_lo = (julong) 0 >> (juint)shift;
1870 jlong pos_hi = hi;
1871 //lo = MIN2(neg_lo, pos_lo); // == 0
1872 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1873 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1874 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1875 }
1876 assert(lo <= hi, "must have valid bounds");
1877 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1878 #ifdef ASSERT
1879 // Make sure we get the sign-capture idiom correct.
1880 if (shift == BitsPerJavaLong - 1) {
1881 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1882 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1883 }
1884 #endif
1885 return tl;
1886 }
1887
1888 return TypeLong::LONG; // Give up
1889 }
1890
1891 //=============================================================================
1892 //------------------------------Ideal------------------------------------------
1893 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1894 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1895 // This reduces the number of rules in the matcher, as we only need to check
1896 // for negations on the second argument, and not the symmetric case where
1897 // the first argument is negated.
1898 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1899 swap_edges(1, 2);
1900 return this;
1901 }
1902 return nullptr;
1903 }
1904
1905 //=============================================================================
1906 //------------------------------Value------------------------------------------
1907 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1908 const Type *t1 = phase->type(in(1));
1909 if (t1 == Type::TOP) return Type::TOP;
1910 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1911 const Type *t2 = phase->type(in(2));
1912 if (t2 == Type::TOP) return Type::TOP;
1913 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1914 const Type *t3 = phase->type(in(3));
1915 if (t3 == Type::TOP) return Type::TOP;
1916 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1917 #ifndef __STDC_IEC_559__
1918 return Type::DOUBLE;
1919 #else
1920 double d1 = t1->getd();
1921 double d2 = t2->getd();
1922 double d3 = t3->getd();
1923 return TypeD::make(fma(d1, d2, d3));
1924 #endif
1925 }
1926
1927 //=============================================================================
1928 //------------------------------Value------------------------------------------
1929 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1930 const Type *t1 = phase->type(in(1));
1931 if (t1 == Type::TOP) return Type::TOP;
1932 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1933 const Type *t2 = phase->type(in(2));
1934 if (t2 == Type::TOP) return Type::TOP;
1935 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1936 const Type *t3 = phase->type(in(3));
1937 if (t3 == Type::TOP) return Type::TOP;
1938 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1939 #ifndef __STDC_IEC_559__
1940 return Type::FLOAT;
1941 #else
1942 float f1 = t1->getf();
1943 float f2 = t2->getf();
1944 float f3 = t3->getf();
1945 return TypeF::make(fma(f1, f2, f3));
1946 #endif
1947 }
1948
1949 //=============================================================================
1950 //------------------------------Value------------------------------------------
1951 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
1952 const Type* t1 = phase->type(in(1));
1953 if (t1 == Type::TOP) { return Type::TOP; }
1954 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1955 const Type* t2 = phase->type(in(2));
1956 if (t2 == Type::TOP) { return Type::TOP; }
1957 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1958 const Type* t3 = phase->type(in(3));
1959 if (t3 == Type::TOP) { return Type::TOP; }
1960 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1961 #ifndef __STDC_IEC_559__
1962 return Type::HALF_FLOAT;
1963 #else
1964 float f1 = t1->getf();
1965 float f2 = t2->getf();
1966 float f3 = t3->getf();
1967 return TypeH::make(fma(f1, f2, f3));
1968 #endif
1969 }
1970
1971 //=============================================================================
1972 //------------------------------hash-------------------------------------------
1973 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1974 // The hash function must return the same value when edge swapping is performed.
1975 uint MulAddS2INode::hash() const {
1976 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1977 }
1978
1979 //------------------------------Rotate Operations ------------------------------
1980
1981 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1982 const Type* t1 = phase->type(in(1));
1983 if (t1 == Type::TOP) {
1984 return this;
1985 }
1986 int count = 0;
1987 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1988 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1989 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1990 // Rotate by a multiple of 32/64 does nothing
1991 return in(1);
1992 }
1993 return this;
1994 }
1995
1996 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1997 const Type* t1 = phase->type(in(1));
1998 const Type* t2 = phase->type(in(2));
1999 // Either input is TOP ==> the result is TOP
2000 if (t1 == Type::TOP || t2 == Type::TOP) {
2001 return Type::TOP;
2002 }
2003
2004 if (t1->isa_int()) {
2005 const TypeInt* r1 = t1->is_int();
2006 const TypeInt* r2 = t2->is_int();
2007
2008 // Left input is ZERO ==> the result is ZERO.
2009 if (r1 == TypeInt::ZERO) {
2010 return TypeInt::ZERO;
2011 }
2012 // Rotate by zero does nothing
2013 if (r2 == TypeInt::ZERO) {
2014 return r1;
2015 }
2016 if (r1->is_con() && r2->is_con()) {
2017 juint r1_con = (juint)r1->get_con();
2018 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2019 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2020 }
2021 return TypeInt::INT;
2022 } else {
2023 assert(t1->isa_long(), "Type must be a long");
2024 const TypeLong* r1 = t1->is_long();
2025 const TypeInt* r2 = t2->is_int();
2026
2027 // Left input is ZERO ==> the result is ZERO.
2028 if (r1 == TypeLong::ZERO) {
2029 return TypeLong::ZERO;
2030 }
2031 // Rotate by zero does nothing
2032 if (r2 == TypeInt::ZERO) {
2033 return r1;
2034 }
2035 if (r1->is_con() && r2->is_con()) {
2036 julong r1_con = (julong)r1->get_con();
2037 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2038 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2039 }
2040 return TypeLong::LONG;
2041 }
2042 }
2043
2044 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2045 const Type* t1 = phase->type(in(1));
2046 const Type* t2 = phase->type(in(2));
2047 if (t2->isa_int() && t2->is_int()->is_con()) {
2048 if (t1->isa_int()) {
2049 int lshift = t2->is_int()->get_con() & 31;
2050 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2051 } else if (t1 != Type::TOP) {
2052 assert(t1->isa_long(), "Type must be a long");
2053 int lshift = t2->is_int()->get_con() & 63;
2054 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2055 }
2056 }
2057 return nullptr;
2058 }
2059
2060 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2061 const Type* t1 = phase->type(in(1));
2062 if (t1 == Type::TOP) {
2063 return this;
2064 }
2065 int count = 0;
2066 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2067 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2068 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2069 // Rotate by a multiple of 32/64 does nothing
2070 return in(1);
2071 }
2072 return this;
2073 }
2074
2075 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2076 const Type* t1 = phase->type(in(1));
2077 const Type* t2 = phase->type(in(2));
2078 // Either input is TOP ==> the result is TOP
2079 if (t1 == Type::TOP || t2 == Type::TOP) {
2080 return Type::TOP;
2081 }
2082
2083 if (t1->isa_int()) {
2084 const TypeInt* r1 = t1->is_int();
2085 const TypeInt* r2 = t2->is_int();
2086
2087 // Left input is ZERO ==> the result is ZERO.
2088 if (r1 == TypeInt::ZERO) {
2089 return TypeInt::ZERO;
2090 }
2091 // Rotate by zero does nothing
2092 if (r2 == TypeInt::ZERO) {
2093 return r1;
2094 }
2095 if (r1->is_con() && r2->is_con()) {
2096 juint r1_con = (juint)r1->get_con();
2097 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2098 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2099 }
2100 return TypeInt::INT;
2101 } else {
2102 assert(t1->isa_long(), "Type must be a long");
2103 const TypeLong* r1 = t1->is_long();
2104 const TypeInt* r2 = t2->is_int();
2105 // Left input is ZERO ==> the result is ZERO.
2106 if (r1 == TypeLong::ZERO) {
2107 return TypeLong::ZERO;
2108 }
2109 // Rotate by zero does nothing
2110 if (r2 == TypeInt::ZERO) {
2111 return r1;
2112 }
2113 if (r1->is_con() && r2->is_con()) {
2114 julong r1_con = (julong)r1->get_con();
2115 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2116 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2117 }
2118 return TypeLong::LONG;
2119 }
2120 }
2121
2122 //------------------------------ Sum & Mask ------------------------------
2123
2124 // Returns a lower bound on the number of trailing zeros in expr.
2125 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2126 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2127 if (type == nullptr) {
2128 return 0;
2129 }
2130
2131 expr = expr->uncast();
2132 type = phase->type(expr)->isa_integer(bt);
2133 if (type == nullptr) {
2134 return 0;
2135 }
2136
2137 if (type->is_con()) {
2138 jlong con = type->get_con_as_long(bt);
2139 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2140 }
2141
2142 if (expr->Opcode() == Op_ConvI2L) {
2143 expr = expr->in(1)->uncast();
2144 bt = T_INT;
2145 type = phase->type(expr)->isa_int();
2146 }
2147
2148 // Pattern: expr = (x << shift)
2149 if (expr->Opcode() == Op_LShift(bt)) {
2150 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2151 if (shift_t == nullptr || !shift_t->is_con()) {
2152 return 0;
2153 }
2154 // We need to truncate the shift, as it may not have been canonicalized yet.
2155 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2156 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2157 // (JLS: "Shift Operators")
2158 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2159 return shift_t->get_con() & shift_mask;
2160 }
2161
2162 return 0;
2163 }
2164
2165 // Checks whether expr is neutral additive element (zero) under mask,
2166 // i.e. whether an expression of the form:
2167 // (AndX (AddX (expr addend) mask)
2168 // (expr + addend) & mask
2169 // is equivalent to
2170 // (AndX addend mask)
2171 // addend & mask
2172 // for any addend.
2173 // (The X in AndX must be I or L, depending on bt).
2174 //
2175 // We check for the sufficient condition when the lowest set bit in expr is higher than
2176 // the highest set bit in mask, i.e.:
2177 // expr: eeeeee0000000000000
2178 // mask: 000000mmmmmmmmmmmmm
2179 // <--w bits--->
2180 // We do not test for other cases.
2181 //
2182 // Correctness:
2183 // Given "expr" with at least "w" trailing zeros,
2184 // let "mod = 2^w", "suffix_mask = mod - 1"
2185 //
2186 // Since "mask" only has bits set where "suffix_mask" does, we have:
2187 // mask = suffix_mask & mask (SUFFIX_MASK)
2188 //
2189 // And since expr only has bits set above w, and suffix_mask only below:
2190 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2191 //
2192 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2193 // a power of 2, and we are computing in a ring of powers of 2, we know that
2194 // (x + y) % mod = (x % mod + y) % mod
2195 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2196 //
2197 // We can now prove the equality:
2198 // (expr + addend) & mask
2199 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2200 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2201 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2202 // = addend & mask (SUFFIX_MASK)
2203 //
2204 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2205 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2206 // When the mask is negative, it has the most significant bit set.
2207 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2208 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2209 return false;
2210 }
2211
2212 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2213 if (mask_t->hi_as_long() == 0) {
2214 assert(mask_t->lo_as_long() == 0, "checked earlier");
2215 return false;
2216 }
2217
2218 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2219 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2220 return expr_trailing_zeros >= mask_bit_width;
2221 }
2222
2223 // Reduces the pattern:
2224 // (AndX (AddX add1 add2) mask)
2225 // to
2226 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2227 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2228 Node* add = in(1);
2229 Node* mask = in(2);
2230 int addidx = 0;
2231 if (add->Opcode() == Op_Add(bt)) {
2232 addidx = 1;
2233 } else if (mask->Opcode() == Op_Add(bt)) {
2234 mask = add;
2235 addidx = 2;
2236 add = in(addidx);
2237 }
2238 if (addidx > 0) {
2239 Node* add1 = add->in(1);
2240 Node* add2 = add->in(2);
2241 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2242 set_req_X(addidx, add2, phase);
2243 return this;
2244 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2245 set_req_X(addidx, add1, phase);
2246 return this;
2247 }
2248 }
2249 return nullptr;
2250 }