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 #if defined(IA32)
218 // Can't trust native compilers to properly fold strict double
219 // multiplication with round-to-zero on this platform.
220 if (op == Op_MulD) {
221 return TypeD::DOUBLE;
222 }
223 #endif
224
225 return mul_ring(t1,t2); // Local flavor of type multiplication
226 }
227
228 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
229 switch (bt) {
230 case T_INT:
231 return new MulINode(in1, in2);
232 case T_LONG:
233 return new MulLNode(in1, in2);
234 default:
235 fatal("Not implemented for %s", type2name(bt));
236 }
237 return nullptr;
238 }
239
240 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
241 switch (bt) {
242 case T_INT:
243 return new AndINode(in1, in2);
244 case T_LONG:
245 return new AndLNode(in1, in2);
246 default:
247 fatal("Not implemented for %s", type2name(bt));
248 }
249 return nullptr;
250 }
251
252
253 //=============================================================================
254 //------------------------------Ideal------------------------------------------
255 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
256 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
257 const jint con = in(2)->find_int_con(0);
258 if (con == 0) {
259 // If in(2) is not a constant, call Ideal() of the parent class to
260 // try to move constant to the right side.
261 return MulNode::Ideal(phase, can_reshape);
262 }
263
264 // Now we have a constant Node on the right and the constant in con.
265 if (con == 1) {
266 // By one is handled by Identity call
267 return nullptr;
268 }
269
270 // Check for negative constant; if so negate the final result
271 bool sign_flip = false;
272
273 unsigned int abs_con = g_uabs(con);
274 if (abs_con != (unsigned int)con) {
275 sign_flip = true;
276 }
277
278 // Get low bit; check for being the only bit
279 Node *res = nullptr;
280 unsigned int bit1 = submultiple_power_of_2(abs_con);
281 if (bit1 == abs_con) { // Found a power of 2?
282 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
283 } else {
284 // Check for constant with 2 bits set
285 unsigned int bit2 = abs_con - bit1;
286 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
287 if (bit2 + bit1 == abs_con) { // Found all bits in con?
288 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
289 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
290 res = new AddINode(n2, n1);
291 } else if (is_power_of_2(abs_con + 1)) {
292 // Sleezy: power-of-2 - 1. Next time be generic.
293 unsigned int temp = abs_con + 1;
294 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
295 res = new SubINode(n1, in(1));
296 } else {
297 return MulNode::Ideal(phase, can_reshape);
298 }
299 }
300
301 if (sign_flip) { // Need to negate result?
302 res = phase->transform(res);// Transform, before making the zero con
303 res = new SubINode(phase->intcon(0),res);
304 }
305
306 return res; // Return final result
307 }
308
309 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
310 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
311 //
312 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
313 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
314 // - no overflow (i.e. product fits into x bits)
315 // - overflow (i.e. product does not fit into x bits)
316 //
317 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
318 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
319 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
320 // The cross product consists of the following four multiplications with 2x-bit results:
321 // (1) left_lo * right_lo
322 // (2) left_lo * right_hi
323 // (3) left_hi * right_lo
324 // (4) left_hi * right_hi
325 //
326 // Let's define the following two functions:
327 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
328 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
329 //
330 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
331 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
332 // Ux(p) = 0
333 // Lx(p) = p
334 //
335 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
336 // with x bits:
337 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
338 //
339 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
340 // (i.e. all values in the x-bit space could be possible):
341 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
342 //
343 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
344 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
345 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
346 // the lower x bits the same number of times:
347 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
348 //
349 // If all upper x bits are equal, we can conclude:
350 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
351 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
352 //
353 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
354 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
355 //
356 //
357 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
358 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
359 // 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,
360 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
361 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
362 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
363 // that there is no underflow.
364 //
365 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
366 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
367 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
368 // can compare the upper x bits in a unified way with N() being the normalization function:
369 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
370 template<typename NativeType>
371 class IntegerTypeMultiplication {
372
373 NativeType _lo_left;
374 NativeType _lo_right;
375 NativeType _hi_left;
376 NativeType _hi_right;
377 short _widen_left;
378 short _widen_right;
379
380 static const Type* overflow_type();
381 static NativeType multiply_high(NativeType x, NativeType y);
382 const Type* create_type(NativeType lo, NativeType hi) const;
383
384 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
385 return normalize_overflow_value(x, y, multiply_high(x, y));
386 }
387
388 bool cross_product_not_same_overflow_value() const {
389 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
390 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
391 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
392 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
393 return lo_lo_high_product != lo_hi_high_product ||
394 lo_hi_high_product != hi_lo_high_product ||
395 hi_lo_high_product != hi_hi_high_product;
396 }
397
398 bool does_product_overflow(NativeType x, NativeType y) const {
399 return multiply_high_signed_overflow_value(x, y) != 0;
400 }
401
402 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
403 return java_multiply(x, y) < 0 ? result + 1 : result;
404 }
405
406 public:
407 template<class IntegerType>
408 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
409 : _lo_left(left->_lo), _lo_right(right->_lo),
410 _hi_left(left->_hi), _hi_right(right->_hi),
411 _widen_left(left->_widen), _widen_right(right->_widen) {}
412
413 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
414 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
415 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
416 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
417 const Type* compute() const {
418 if (cross_product_not_same_overflow_value()) {
419 return overflow_type();
420 }
421
422 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
423 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
424 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
425 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
426 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
427 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
428 return create_type(min, max);
429 }
430
431 bool does_overflow() const {
432 return does_product_overflow(_lo_left, _lo_right) ||
433 does_product_overflow(_lo_left, _hi_right) ||
434 does_product_overflow(_hi_left, _lo_right) ||
435 does_product_overflow(_hi_left, _hi_right);
436 }
437 };
438
439 template <>
440 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
441 return TypeInt::INT;
442 }
443
444 template <>
445 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
446 const jlong x_64 = x;
447 const jlong y_64 = y;
448 const jlong product = x_64 * y_64;
449 return (jint)((uint64_t)product >> 32u);
450 }
451
452 template <>
453 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
454 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
455 }
456
457 template <>
458 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
459 return TypeLong::LONG;
460 }
461
462 template <>
463 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
464 return multiply_high_signed(x, y);
465 }
466
467 template <>
468 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
469 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
470 }
471
472 // Compute the product type of two integer ranges into this node.
473 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
474 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
475 return integer_multiplication.compute();
476 }
477
478 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
479 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
480 return integer_multiplication.does_overflow();
481 }
482
483 // Compute the product type of two long ranges into this node.
484 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
485 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
486 return integer_multiplication.compute();
487 }
488
489 //=============================================================================
490 //------------------------------Ideal------------------------------------------
491 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
492 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
493 const jlong con = in(2)->find_long_con(0);
494 if (con == 0) {
495 // If in(2) is not a constant, call Ideal() of the parent class to
496 // try to move constant to the right side.
497 return MulNode::Ideal(phase, can_reshape);
498 }
499
500 // Now we have a constant Node on the right and the constant in con.
501 if (con == 1) {
502 // By one is handled by Identity call
503 return nullptr;
504 }
505
506 // Check for negative constant; if so negate the final result
507 bool sign_flip = false;
508 julong abs_con = g_uabs(con);
509 if (abs_con != (julong)con) {
510 sign_flip = true;
511 }
512
513 // Get low bit; check for being the only bit
514 Node *res = nullptr;
515 julong bit1 = submultiple_power_of_2(abs_con);
516 if (bit1 == abs_con) { // Found a power of 2?
517 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
518 } else {
519
520 // Check for constant with 2 bits set
521 julong bit2 = abs_con-bit1;
522 bit2 = bit2 & (0-bit2); // Extract 2nd bit
523 if (bit2 + bit1 == abs_con) { // Found all bits in con?
524 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
525 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
526 res = new AddLNode(n2, n1);
527
528 } else if (is_power_of_2(abs_con+1)) {
529 // Sleezy: power-of-2 -1. Next time be generic.
530 julong temp = abs_con + 1;
531 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
532 res = new SubLNode(n1, in(1));
533 } else {
534 return MulNode::Ideal(phase, can_reshape);
535 }
536 }
537
538 if (sign_flip) { // Need to negate result?
539 res = phase->transform(res);// Transform, before making the zero con
540 res = new SubLNode(phase->longcon(0),res);
541 }
542
543 return res; // Return final result
544 }
545
546 //=============================================================================
547 //------------------------------mul_ring---------------------------------------
548 // Compute the product type of two double ranges into this node.
549 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
550 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
551 return TypeF::make( t0->getf() * t1->getf() );
552 }
553
554 //------------------------------Ideal---------------------------------------
555 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
556 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
557 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
558
559 // x * 2 -> x + x
560 if (t2 != nullptr && t2->getf() == 2) {
561 Node* base = in(1);
562 return new AddFNode(base, base);
563 }
564 return MulNode::Ideal(phase, can_reshape);
565 }
566
567 //=============================================================================
568 //------------------------------Ideal------------------------------------------
569 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
570 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
571 const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
572
573 // x * 2 -> x + x
574 if (t2 != nullptr && t2->getf() == 2) {
575 Node* base = in(1);
576 return new AddHFNode(base, base);
577 }
578 return MulNode::Ideal(phase, can_reshape);
579 }
580
581 // Compute the product type of two half float ranges into this node.
582 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
583 if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
584 return Type::HALF_FLOAT;
585 }
586 return TypeH::make(t0->getf() * t1->getf());
587 }
588
589 //=============================================================================
590 //------------------------------mul_ring---------------------------------------
591 // Compute the product type of two double ranges into this node.
592 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
593 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
594 // We must be multiplying 2 double constants.
595 return TypeD::make( t0->getd() * t1->getd() );
596 }
597
598 //------------------------------Ideal---------------------------------------
599 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
600 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
601 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
602
603 // x * 2 -> x + x
604 if (t2 != nullptr && t2->getd() == 2) {
605 Node* base = in(1);
606 return new AddDNode(base, base);
607 }
608
609 return MulNode::Ideal(phase, can_reshape);
610 }
611
612 //=============================================================================
613 //------------------------------Value------------------------------------------
614 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
615 const Type *t1 = phase->type( in(1) );
616 const Type *t2 = phase->type( in(2) );
617 const Type *bot = bottom_type();
618 return MulHiValue(t1, t2, bot);
619 }
620
621 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
622 const Type *t1 = phase->type( in(1) );
623 const Type *t2 = phase->type( in(2) );
624 const Type *bot = bottom_type();
625 return MulHiValue(t1, t2, bot);
626 }
627
628 // A common routine used by UMulHiLNode and MulHiLNode
629 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
630 // Either input is TOP ==> the result is TOP
631 if( t1 == Type::TOP ) return Type::TOP;
632 if( t2 == Type::TOP ) return Type::TOP;
633
634 // Either input is BOTTOM ==> the result is the local BOTTOM
635 if( (t1 == bot) || (t2 == bot) ||
636 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
637 return bot;
638
639 // It is not worth trying to constant fold this stuff!
640 return TypeLong::LONG;
641 }
642
643 template<typename IntegerType>
644 static const IntegerType* and_value(const IntegerType* r0, const IntegerType* r1) {
645 typedef typename IntegerType::NativeType NativeType;
646 static_assert(std::is_signed<NativeType>::value, "Native type of IntegerType must be signed!");
647
648 int widen = MAX2(r0->_widen, r1->_widen);
649
650 // If both types are constants, we can calculate a constant result.
651 if (r0->is_con() && r1->is_con()) {
652 return IntegerType::make(r0->get_con() & r1->get_con());
653 }
654
655 // If both ranges are positive, the result will range from 0 up to the hi value of the smaller range. The minimum
656 // 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.
657 if (r0->_lo >= 0 && r1->_lo >= 0) {
658 return IntegerType::make(0, MIN2(r0->_hi, r1->_hi), widen);
659 }
660
661 // If only one range is positive, the result will range from 0 up to that range's maximum value.
662 // For the operation 'x & C' where C is a positive constant, the result will be in the range [0..C]. With that observation,
663 // 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]'
664 // where c >= 0 will be in the range [0..C].
665 if (r0->_lo >= 0) {
666 return IntegerType::make(0, r0->_hi, widen);
667 }
668
669 if (r1->_lo >= 0) {
670 return IntegerType::make(0, r1->_hi, widen);
671 }
672
673 // At this point, all positive ranges will have already been handled, so the only remaining cases will be negative ranges
674 // and constants.
675
676 assert(r0->_lo < 0 && r1->_lo < 0, "positive ranges should already be handled!");
677
678 // As two's complement means that both numbers will start with leading 1s, the lower bound of both ranges will contain
679 // the common leading 1s of both minimum values. In order to count them with count_leading_zeros, the bits are inverted.
680 NativeType sel_val = ~MIN2(r0->_lo, r1->_lo);
681
682 NativeType min;
683 if (sel_val == 0) {
684 // Since count_leading_zeros is undefined at 0, we short-circuit the condition where both ranges have a minimum of -1.
685 min = -1;
686 } else {
687 // 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.
688 int shift_bits = count_leading_zeros(sel_val) - 1;
689 min = std::numeric_limits<NativeType>::min() >> shift_bits;
690 }
691
692 NativeType max;
693 if (r0->_hi < 0 && r1->_hi < 0) {
694 // If both ranges are negative, then the same optimization as both positive ranges will apply, and the smaller hi
695 // value will mask off any bits set by higher values.
696 max = MIN2(r0->_hi, r1->_hi);
697 } else {
698 // In the case of ranges that cross zero, negative values can cause the higher order bits to be set, so the maximum
699 // positive value can be as high as the larger hi value.
700 max = MAX2(r0->_hi, r1->_hi);
701 }
702
703 return IntegerType::make(min, max, widen);
704 }
705
706 //=============================================================================
707 //------------------------------mul_ring---------------------------------------
708 // Supplied function returns the product of the inputs IN THE CURRENT RING.
709 // For the logical operations the ring's MUL is really a logical AND function.
710 // This also type-checks the inputs for sanity. Guaranteed never to
711 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
712 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
713 const TypeInt* r0 = t0->is_int();
714 const TypeInt* r1 = t1->is_int();
715
716 return and_value<TypeInt>(r0, r1);
717 }
718
719 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
720
721 const Type* AndINode::Value(PhaseGVN* phase) const {
722 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
723 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
724 return TypeInt::ZERO;
725 }
726
727 return MulNode::Value(phase);
728 }
729
730 //------------------------------Identity---------------------------------------
731 // Masking off the high bits of an unsigned load is not required
732 Node* AndINode::Identity(PhaseGVN* phase) {
733
734 // x & x => x
735 if (in(1) == in(2)) {
736 return in(1);
737 }
738
739 Node* in1 = in(1);
740 uint op = in1->Opcode();
741 const TypeInt* t2 = phase->type(in(2))->isa_int();
742 if (t2 && t2->is_con()) {
743 int con = t2->get_con();
744 // Masking off high bits which are always zero is useless.
745 const TypeInt* t1 = phase->type(in(1))->isa_int();
746 if (t1 != nullptr && t1->_lo >= 0) {
747 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
748 if ((t1_support & con) == t1_support)
749 return in1;
750 }
751 // Masking off the high bits of a unsigned-shift-right is not
752 // needed either.
753 if (op == Op_URShiftI) {
754 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
755 if (t12 && t12->is_con()) { // Shift is by a constant
756 int shift = t12->get_con();
757 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
758 int mask = max_juint >> shift;
759 if ((mask & con) == mask) // If AND is useless, skip it
760 return in1;
761 }
762 }
763 }
764 return MulNode::Identity(phase);
765 }
766
767 //------------------------------Ideal------------------------------------------
768 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
769 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
770 Node* progress = AndIL_sum_and_mask(phase, T_INT);
771 if (progress != nullptr) {
772 return progress;
773 }
774
775 // Convert "(~a) & (~b)" into "~(a | b)"
776 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
777 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
778 Node* tn = phase->transform(or_a_b);
779 return AddNode::make_not(phase, tn, T_INT);
780 }
781
782 // Special case constant AND mask
783 const TypeInt *t2 = phase->type( in(2) )->isa_int();
784 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
785 const int mask = t2->get_con();
786 Node *load = in(1);
787 uint lop = load->Opcode();
788
789 // Masking bits off of a Character? Hi bits are already zero.
790 if( lop == Op_LoadUS &&
791 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
792 return new AndINode(load,phase->intcon(mask&0xFFFF));
793
794 // Masking bits off of a Short? Loading a Character does some masking
795 if (can_reshape &&
796 load->outcnt() == 1 && load->unique_out() == this) {
797 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
798 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
799 ldus = phase->transform(ldus);
800 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
801 }
802
803 // Masking sign bits off of a Byte? Do an unsigned byte load plus
804 // an and.
805 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
806 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
807 ldub = phase->transform(ldub);
808 return new AndINode(ldub, phase->intcon(mask));
809 }
810 }
811
812 // Masking off sign bits? Dont make them!
813 if( lop == Op_RShiftI ) {
814 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
815 if( t12 && t12->is_con() ) { // Shift is by a constant
816 int shift = t12->get_con();
817 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
818 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
819 // If the AND'ing of the 2 masks has no bits, then only original shifted
820 // bits survive. NO sign-extension bits survive the maskings.
821 if( (sign_bits_mask & mask) == 0 ) {
822 // Use zero-fill shift instead
823 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
824 return new AndINode( zshift, in(2) );
825 }
826 }
827 }
828
829 // Check for 'negate/and-1', a pattern emitted when someone asks for
830 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
831 // plus 1) and the mask is of the low order bit. Skip the negate.
832 if( lop == Op_SubI && mask == 1 && load->in(1) &&
833 phase->type(load->in(1)) == TypeInt::ZERO )
834 return new AndINode( load->in(2), in(2) );
835
836 return MulNode::Ideal(phase, can_reshape);
837 }
838
839 //=============================================================================
840 //------------------------------mul_ring---------------------------------------
841 // Supplied function returns the product of the inputs IN THE CURRENT RING.
842 // For the logical operations the ring's MUL is really a logical AND function.
843 // This also type-checks the inputs for sanity. Guaranteed never to
844 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
845 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
846 const TypeLong* r0 = t0->is_long();
847 const TypeLong* r1 = t1->is_long();
848
849 return and_value<TypeLong>(r0, r1);
850 }
851
852 const Type* AndLNode::Value(PhaseGVN* phase) const {
853 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
854 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
855 return TypeLong::ZERO;
856 }
857
858 return MulNode::Value(phase);
859 }
860
861 //------------------------------Identity---------------------------------------
862 // Masking off the high bits of an unsigned load is not required
863 Node* AndLNode::Identity(PhaseGVN* phase) {
864
865 // x & x => x
866 if (in(1) == in(2)) {
867 return in(1);
868 }
869
870 Node *usr = in(1);
871 const TypeLong *t2 = phase->type( in(2) )->isa_long();
872 if( t2 && t2->is_con() ) {
873 jlong con = t2->get_con();
874 // Masking off high bits which are always zero is useless.
875 const TypeLong* t1 = phase->type( in(1) )->isa_long();
876 if (t1 != nullptr && t1->_lo >= 0) {
877 int bit_count = log2i_graceful(t1->_hi) + 1;
878 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
879 if ((t1_support & con) == t1_support)
880 return usr;
881 }
882 uint lop = usr->Opcode();
883 // Masking off the high bits of a unsigned-shift-right is not
884 // needed either.
885 if( lop == Op_URShiftL ) {
886 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
887 if( t12 && t12->is_con() ) { // Shift is by a constant
888 int shift = t12->get_con();
889 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
890 jlong mask = max_julong >> shift;
891 if( (mask&con) == mask ) // If AND is useless, skip it
892 return usr;
893 }
894 }
895 }
896 return MulNode::Identity(phase);
897 }
898
899 //------------------------------Ideal------------------------------------------
900 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
901 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
902 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
903 if (progress != nullptr) {
904 return progress;
905 }
906
907 // Convert "(~a) & (~b)" into "~(a | b)"
908 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
909 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
910 Node* tn = phase->transform(or_a_b);
911 return AddNode::make_not(phase, tn, T_LONG);
912 }
913
914 // Special case constant AND mask
915 const TypeLong *t2 = phase->type( in(2) )->isa_long();
916 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
917 const jlong mask = t2->get_con();
918
919 Node* in1 = in(1);
920 int op = in1->Opcode();
921
922 // Are we masking a long that was converted from an int with a mask
923 // that fits in 32-bits? Commute them and use an AndINode. Don't
924 // convert masks which would cause a sign extension of the integer
925 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
926 // would be optimized away later in Identity.
927 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
928 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
929 andi = phase->transform(andi);
930 return new ConvI2LNode(andi);
931 }
932
933 // Masking off sign bits? Dont make them!
934 if (op == Op_RShiftL) {
935 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
936 if( t12 && t12->is_con() ) { // Shift is by a constant
937 int shift = t12->get_con();
938 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
939 if (shift != 0) {
940 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
941 // If the AND'ing of the 2 masks has no bits, then only original shifted
942 // bits survive. NO sign-extension bits survive the maskings.
943 if( (sign_bits_mask & mask) == 0 ) {
944 // Use zero-fill shift instead
945 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
946 return new AndLNode(zshift, in(2));
947 }
948 }
949 }
950 }
951
952 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
953 Node* load1 = in(1);
954 Node* load2 = nullptr;
955 if (load1->is_Phi() && phase->type(load1)->isa_long()) {
956 load1 = in(1)->in(1);
957 load2 = in(1)->in(2);
958 }
959 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
960 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
961 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
962 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
963 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
964 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
965 if (mask == markWord::inline_type_pattern) {
966 if (adr_t1->is_inlinetypeptr()) {
967 set_req_X(1, in(2), phase);
968 return this;
969 } else if (!adr_t1->can_be_inline_type()) {
970 set_req_X(1, phase->longcon(0), phase);
971 return this;
972 }
973 } else if (mask == markWord::null_free_array_bit_in_place) {
974 if (adr_t1->is_null_free()) {
975 set_req_X(1, in(2), phase);
976 return this;
977 } else if (adr_t1->is_not_null_free()) {
978 set_req_X(1, phase->longcon(0), phase);
979 return this;
980 }
981 } else if (mask == markWord::flat_array_bit_in_place) {
982 if (adr_t1->is_flat()) {
983 set_req_X(1, in(2), phase);
984 return this;
985 } else if (adr_t1->is_not_flat()) {
986 set_req_X(1, phase->longcon(0), phase);
987 return this;
988 }
989 }
990 }
991 }
992
993 return MulNode::Ideal(phase, can_reshape);
994 }
995
996 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
997 switch (bt) {
998 case T_INT:
999 return new LShiftINode(in1, in2);
1000 case T_LONG:
1001 return new LShiftLNode(in1, in2);
1002 default:
1003 fatal("Not implemented for %s", type2name(bt));
1004 }
1005 return nullptr;
1006 }
1007
1008 // Returns whether the shift amount is constant. If so, sets count.
1009 static bool const_shift_count(PhaseGVN* phase, const Node* shift_node, int* count) {
1010 const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
1011 if (tcount != nullptr && tcount->is_con()) {
1012 *count = tcount->get_con();
1013 return true;
1014 }
1015 return false;
1016 }
1017
1018 // Returns whether the shift amount is constant. If so, sets real_shift and masked_shift.
1019 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, int& real_shift, uint& masked_shift) {
1020 if (const_shift_count(phase, shift_node, &real_shift)) {
1021 masked_shift = real_shift & (nBits - 1);
1022 return true;
1023 }
1024 return false;
1025 }
1026
1027 // Convenience for when we don't care about the real amount
1028 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, uint& masked_shift) {
1029 int real_shift;
1030 return mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift);
1031 }
1032
1033 // Use this in ::Ideal only with shiftNode == this!
1034 // Returns the masked shift amount if constant or 0 if not constant.
1035 static uint mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits) {
1036 int real_shift;
1037 uint masked_shift;
1038 if (mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift)) {
1039 if (masked_shift == 0) {
1040 // Let Identity() handle 0 shift count.
1041 return 0;
1042 }
1043
1044 if (real_shift != (int)masked_shift) {
1045 PhaseIterGVN* igvn = phase->is_IterGVN();
1046 if (igvn != nullptr) {
1047 igvn->_worklist.push(shift_node);
1048 }
1049 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
1050 }
1051 return masked_shift;
1052 }
1053 // Not a shift by a constant.
1054 return 0;
1055 }
1056
1057 // Called with
1058 // outer_shift = (_ << rhs_outer)
1059 // We are looking for the pattern:
1060 // outer_shift = ((X << rhs_inner) << rhs_outer)
1061 // where rhs_outer and rhs_inner are constant
1062 // we denote inner_shift the nested expression (X << rhs_inner)
1063 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
1064 // where nbits is the number of bits of the shifts
1065 //
1066 // There are 2 cases:
1067 // if con_outer + con_inner >= nbits => 0
1068 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
1069 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
1070 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
1071 const Node* inner_shift = outer_shift->in(1);
1072 if (inner_shift->Opcode() != Op_LShift(bt)) {
1073 return nullptr;
1074 }
1075
1076 uint nbits = bits_per_java_integer(bt);
1077 uint con_inner;
1078 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
1079 return nullptr;
1080 }
1081
1082 if (con_inner == 0) {
1083 // We let the Identity() of the inner shift do its job.
1084 return nullptr;
1085 }
1086
1087 if (con_outer + con_inner >= nbits) {
1088 // While it might be tempting to use
1089 // phase->zerocon(bt);
1090 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
1091 // to return a new node, this or nullptr, but not an old (cached) node.
1092 return ConNode::make(TypeInteger::zero(bt));
1093 }
1094
1095 // con0 + con1 < nbits ==> actual shift happens now
1096 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
1097 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
1098 }
1099
1100 //------------------------------Identity---------------------------------------
1101 Node* LShiftINode::Identity(PhaseGVN* phase) {
1102 return IdentityIL(phase, T_INT);
1103 }
1104
1105 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1106 uint con = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1107 if (con == 0) {
1108 return nullptr;
1109 }
1110
1111 // If the right input is a constant, and the left input is an add of a
1112 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1113 Node* add1 = in(1);
1114 int add1_op = add1->Opcode();
1115 if (add1_op == Op_Add(bt)) { // Left input is an add?
1116 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
1117
1118 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1119 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1120 if (bt != T_INT || con < 16) {
1121 // Left input is an add of the same number?
1122 if (con != (bits_per_java_integer(bt) - 1) && add1->in(1) == add1->in(2)) {
1123 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1124 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
1125 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1126 // or c0 != 63 (for LShiftL) because:
1127 // (x + x) << 63 = 2x << 63, while
1128 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1129 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1130 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1131 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1132 }
1133
1134 // Left input is an add of a constant?
1135 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1136 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1137 // Compute X << con0
1138 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1139 // Compute X<<con0 + (con1<<con0)
1140 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1141 }
1142 }
1143 }
1144
1145 // Check for "(x >> C1) << C2"
1146 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1147 int add1Con = 0;
1148 const_shift_count(phase, add1, &add1Con);
1149
1150 // Special case C1 == C2, which just masks off low bits
1151 if (add1Con > 0 && con == (uint)add1Con) {
1152 // Convert to "(x & -(1 << C2))"
1153 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1154 } else {
1155 // Wait until the right shift has been sharpened to the correct count
1156 if (add1Con > 0 && (uint)add1Con < bits_per_java_integer(bt)) {
1157 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1158 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1159 if (phase->is_IterGVN()) {
1160 if (con > (uint)add1Con) {
1161 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1162 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1163 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1164 } else {
1165 assert(con < (uint)add1Con, "must be (%d < %d)", con, add1Con);
1166 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1167
1168 // Handle logical and arithmetic shifts
1169 Node* rshift;
1170 if (add1_op == Op_RShift(bt)) {
1171 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1172 } else {
1173 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1174 }
1175
1176 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1177 }
1178 } else {
1179 phase->record_for_igvn(this);
1180 }
1181 }
1182 }
1183 }
1184
1185 // Check for "((x >> C1) & Y) << C2"
1186 if (add1_op == Op_And(bt)) {
1187 Node* add2 = add1->in(1);
1188 int add2_op = add2->Opcode();
1189 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1190 // Special case C1 == C2, which just masks off low bits
1191 if (add2->in(2) == in(2)) {
1192 // Convert to "(x & (Y << C2))"
1193 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1194 return MulNode::make_and(add2->in(1), y_sh, bt);
1195 }
1196
1197 int add2Con = 0;
1198 const_shift_count(phase, add2, &add2Con);
1199 if (add2Con > 0 && (uint)add2Con < bits_per_java_integer(bt)) {
1200 if (phase->is_IterGVN()) {
1201 // Convert to "((x >> C1) << C2) & (Y << C2)"
1202
1203 // Make "(x >> C1) << C2", which will get folded away by the rule above
1204 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1205 // Make "Y << C2", which will simplify when Y is a constant
1206 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1207
1208 return MulNode::make_and(x_sh, y_sh, bt);
1209 } else {
1210 phase->record_for_igvn(this);
1211 }
1212 }
1213 }
1214 }
1215
1216 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1217 // before shifting them away.
1218 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1219 assert(bt != T_INT || bits_mask == right_n_bits(bits_per_java_integer(bt)-con), "inconsistent");
1220 if (add1_op == Op_And(bt) &&
1221 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1222 return LShiftNode::make(add1->in(1), in(2), bt);
1223 }
1224
1225 // Collapse nested left-shifts with constant rhs:
1226 // (X << con1) << con2 ==> X << (con1 + con2)
1227 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1228 if (doubleShift != nullptr) {
1229 return doubleShift;
1230 }
1231
1232 return nullptr;
1233 }
1234
1235 //------------------------------Ideal------------------------------------------
1236 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1237 return IdealIL(phase, can_reshape, T_INT);
1238 }
1239
1240 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1241 const Type* t1 = phase->type(in(1));
1242 const Type* t2 = phase->type(in(2));
1243 // Either input is TOP ==> the result is TOP
1244 if (t1 == Type::TOP) {
1245 return Type::TOP;
1246 }
1247 if (t2 == Type::TOP) {
1248 return Type::TOP;
1249 }
1250
1251 // Left input is ZERO ==> the result is ZERO.
1252 if (t1 == TypeInteger::zero(bt)) {
1253 return TypeInteger::zero(bt);
1254 }
1255 // Shift by zero does nothing
1256 if (t2 == TypeInt::ZERO) {
1257 return t1;
1258 }
1259
1260 // Either input is BOTTOM ==> the result is BOTTOM
1261 if ((t1 == TypeInteger::bottom(bt)) || (t2 == TypeInt::INT) ||
1262 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM)) {
1263 return TypeInteger::bottom(bt);
1264 }
1265
1266 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1267 const TypeInt* r2 = t2->is_int(); // Handy access
1268
1269 if (!r2->is_con()) {
1270 return TypeInteger::bottom(bt);
1271 }
1272
1273 uint shift = r2->get_con();
1274 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1275 // Shift by a multiple of 32/64 does nothing:
1276 if (shift == 0) {
1277 return t1;
1278 }
1279
1280 // If the shift is a constant, shift the bounds of the type,
1281 // unless this could lead to an overflow.
1282 if (!r1->is_con()) {
1283 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1284 #ifdef ASSERT
1285 if (bt == T_INT) {
1286 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1287 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1288 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1289 }
1290 #endif
1291 if (java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo &&
1292 java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) {
1293 // No overflow. The range shifts up cleanly.
1294 return TypeInteger::make(java_shift_left(lo, shift, bt),
1295 java_shift_left(hi, shift, bt),
1296 MAX2(r1->_widen, r2->_widen), bt);
1297 }
1298 return TypeInteger::bottom(bt);
1299 }
1300
1301 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1302 }
1303
1304 //------------------------------Value------------------------------------------
1305 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1306 return ValueIL(phase, T_INT);
1307 }
1308
1309 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1310 int count = 0;
1311 if (const_shift_count(phase, this, &count) && (count & (bits_per_java_integer(bt) - 1)) == 0) {
1312 // Shift by a multiple of 32/64 does nothing
1313 return in(1);
1314 }
1315 return this;
1316 }
1317
1318 //=============================================================================
1319 //------------------------------Identity---------------------------------------
1320 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1321 return IdentityIL(phase, T_LONG);
1322 }
1323
1324 //------------------------------Ideal------------------------------------------
1325 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1326 return IdealIL(phase, can_reshape, T_LONG);
1327 }
1328
1329 //------------------------------Value------------------------------------------
1330 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1331 return ValueIL(phase, T_LONG);
1332 }
1333
1334 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1335 switch (bt) {
1336 case T_INT:
1337 return new RShiftINode(in1, in2);
1338 case T_LONG:
1339 return new RShiftLNode(in1, in2);
1340 default:
1341 fatal("Not implemented for %s", type2name(bt));
1342 }
1343 return nullptr;
1344 }
1345
1346
1347 //=============================================================================
1348 //------------------------------Identity---------------------------------------
1349 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1350 int count = 0;
1351 if (const_shift_count(phase, this, &count)) {
1352 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1353 // Shift by a multiple of 32/64 does nothing
1354 return in(1);
1355 }
1356 // Check for useless sign-masking
1357 if (in(1)->Opcode() == Op_LShift(bt) &&
1358 in(1)->req() == 3 &&
1359 in(1)->in(2) == in(2)) {
1360 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1361 // Compute masks for which this shifting doesn't change
1362 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1363 jlong hi = ~lo; // 00007FFF
1364 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1365 if (t11 == nullptr) {
1366 return this;
1367 }
1368 // Does actual value fit inside of mask?
1369 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1370 return in(1)->in(1); // Then shifting is a nop
1371 }
1372 }
1373 }
1374 return this;
1375 }
1376
1377 Node* RShiftINode::Identity(PhaseGVN* phase) {
1378 return IdentityIL(phase, T_INT);
1379 }
1380
1381 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1382 // Inputs may be TOP if they are dead.
1383 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1384 if (t1 == nullptr) {
1385 return NodeSentinel; // Left input is an integer
1386 }
1387 int shift = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1388 if (shift == 0) {
1389 return NodeSentinel;
1390 }
1391
1392 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1393 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1394 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1395 const Node* and_node = in(1);
1396 if (and_node->Opcode() != Op_And(bt)) {
1397 return nullptr;
1398 }
1399 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1400 if (mask_t != nullptr && mask_t->is_con()) {
1401 jlong maskbits = mask_t->get_con_as_long(bt);
1402 // Convert to "(x >> shift) & (mask >> shift)"
1403 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1404 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1405 }
1406 return nullptr;
1407 }
1408
1409 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1410 Node* progress = IdealIL(phase, can_reshape, T_INT);
1411 if (progress == NodeSentinel) {
1412 return nullptr;
1413 }
1414 if (progress != nullptr) {
1415 return progress;
1416 }
1417 int shift = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1418 assert(shift != 0, "handled by IdealIL");
1419
1420 // Check for "(short[i] <<16)>>16" which simply sign-extends
1421 const Node *shl = in(1);
1422 if (shl->Opcode() != Op_LShiftI) {
1423 return nullptr;
1424 }
1425
1426 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1427 if (left_shift_t == nullptr) {
1428 return nullptr;
1429 }
1430 if (shift == 16 && left_shift_t->is_con(16)) {
1431 Node *ld = shl->in(1);
1432 if (ld->Opcode() == Op_LoadS) {
1433 // Sign extension is just useless here. Return a RShiftI of zero instead
1434 // returning 'ld' directly. We cannot return an old Node directly as
1435 // that is the job of 'Identity' calls and Identity calls only work on
1436 // direct inputs ('ld' is an extra Node removed from 'this'). The
1437 // combined optimization requires Identity only return direct inputs.
1438 set_req_X(1, ld, phase);
1439 set_req_X(2, phase->intcon(0), phase);
1440 return this;
1441 }
1442 else if (can_reshape &&
1443 ld->Opcode() == Op_LoadUS &&
1444 ld->outcnt() == 1 && ld->unique_out() == shl)
1445 // Replace zero-extension-load with sign-extension-load
1446 return ld->as_Load()->convert_to_signed_load(*phase);
1447 }
1448
1449 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1450 if (shift == 24 && left_shift_t->is_con(24)) {
1451 Node *ld = shl->in(1);
1452 if (ld->Opcode() == Op_LoadB) {
1453 // Sign extension is just useless here
1454 set_req_X(1, ld, phase);
1455 set_req_X(2, phase->intcon(0), phase);
1456 return this;
1457 }
1458 }
1459
1460 return nullptr;
1461 }
1462
1463 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1464 const Type* t1 = phase->type(in(1));
1465 const Type* t2 = phase->type(in(2));
1466 // Either input is TOP ==> the result is TOP
1467 if (t1 == Type::TOP) {
1468 return Type::TOP;
1469 }
1470 if (t2 == Type::TOP) {
1471 return Type::TOP;
1472 }
1473
1474 // Left input is ZERO ==> the result is ZERO.
1475 if (t1 == TypeInteger::zero(bt)) {
1476 return TypeInteger::zero(bt);
1477 }
1478 // Shift by zero does nothing
1479 if (t2 == TypeInt::ZERO) {
1480 return t1;
1481 }
1482
1483 // Either input is BOTTOM ==> the result is BOTTOM
1484 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1485 return TypeInteger::bottom(bt);
1486 }
1487
1488 const TypeInteger* r1 = t1->isa_integer(bt);
1489 const TypeInt* r2 = t2->isa_int();
1490
1491 // If the shift is a constant, just shift the bounds of the type.
1492 // For example, if the shift is 31/63, we just propagate sign bits.
1493 if (!r1->is_con() && r2->is_con()) {
1494 uint shift = r2->get_con();
1495 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1496 // Shift by a multiple of 32/64 does nothing:
1497 if (shift == 0) {
1498 return t1;
1499 }
1500 // Calculate reasonably aggressive bounds for the result.
1501 // This is necessary if we are to correctly type things
1502 // like (x<<24>>24) == ((byte)x).
1503 jlong lo = r1->lo_as_long() >> (jint)shift;
1504 jlong hi = r1->hi_as_long() >> (jint)shift;
1505 assert(lo <= hi, "must have valid bounds");
1506 #ifdef ASSERT
1507 if (bt == T_INT) {
1508 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1509 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1510 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1511 }
1512 #endif
1513 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1514 #ifdef ASSERT
1515 // Make sure we get the sign-capture idiom correct.
1516 if (shift == bits_per_java_integer(bt) - 1) {
1517 if (r1->lo_as_long() >= 0) {
1518 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1519 }
1520 if (r1->hi_as_long() < 0) {
1521 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1522 }
1523 }
1524 #endif
1525 return ti;
1526 }
1527
1528 if (!r1->is_con() || !r2->is_con()) {
1529 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1530 if (r1->lo_as_long() >= 0) {
1531 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1532 }
1533
1534 // Conversely, if the left input is negative then the result must be negative.
1535 if (r1->hi_as_long() <= -1) {
1536 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1537 }
1538
1539 return TypeInteger::bottom(bt);
1540 }
1541
1542 // Signed shift right
1543 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1544 }
1545
1546 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1547 return ValueIL(phase, T_INT);
1548 }
1549
1550 //=============================================================================
1551 //------------------------------Identity---------------------------------------
1552 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1553 return IdentityIL(phase, T_LONG);
1554 }
1555
1556 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1557 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1558 if (progress == NodeSentinel) {
1559 return nullptr;
1560 }
1561 return progress;
1562 }
1563
1564 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1565 return ValueIL(phase, T_LONG);
1566 }
1567
1568 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1569 switch (bt) {
1570 case T_INT:
1571 return new URShiftINode(in1, in2);
1572 case T_LONG:
1573 return new URShiftLNode(in1, in2);
1574 default:
1575 fatal("Not implemented for %s", type2name(bt));
1576 }
1577 return nullptr;
1578 }
1579
1580 //=============================================================================
1581 //------------------------------Identity---------------------------------------
1582 Node* URShiftINode::Identity(PhaseGVN* phase) {
1583 int count = 0;
1584 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1585 // Shift by a multiple of 32 does nothing
1586 return in(1);
1587 }
1588
1589 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1590 // Happens during new-array length computation.
1591 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1592 Node *add = in(1);
1593 if (add->Opcode() == Op_AddI) {
1594 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1595 if (t2 && t2->is_con(wordSize - 1) &&
1596 add->in(1)->Opcode() == Op_LShiftI) {
1597 // Check that shift_counts are LogBytesPerWord.
1598 Node *lshift_count = add->in(1)->in(2);
1599 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1600 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1601 t_lshift_count == phase->type(in(2))) {
1602 Node *x = add->in(1)->in(1);
1603 const TypeInt *t_x = phase->type(x)->isa_int();
1604 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1605 return x;
1606 }
1607 }
1608 }
1609 }
1610
1611 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1612 }
1613
1614 //------------------------------Ideal------------------------------------------
1615 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1616 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1617 if (con == 0) {
1618 return nullptr;
1619 }
1620
1621 // We'll be wanting the right-shift amount as a mask of that many bits
1622 const int mask = right_n_bits(BitsPerJavaInteger - con);
1623
1624 int in1_op = in(1)->Opcode();
1625
1626 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1627 if( in1_op == Op_URShiftI ) {
1628 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1629 if( t12 && t12->is_con() ) { // Right input is a constant
1630 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1631 const int con2 = t12->get_con() & 31; // Shift count is always masked
1632 const int con3 = con+con2;
1633 if( con3 < 32 ) // Only merge shifts if total is < 32
1634 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1635 }
1636 }
1637
1638 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1639 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1640 // If Q is "X << z" the rounding is useless. Look for patterns like
1641 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1642 Node *add = in(1);
1643 const TypeInt *t2 = phase->type(in(2))->isa_int();
1644 if (in1_op == Op_AddI) {
1645 Node *lshl = add->in(1);
1646 if( lshl->Opcode() == Op_LShiftI &&
1647 phase->type(lshl->in(2)) == t2 ) {
1648 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1649 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1650 return new AndINode( sum, phase->intcon(mask) );
1651 }
1652 }
1653
1654 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1655 // This shortens the mask. Also, if we are extracting a high byte and
1656 // storing it to a buffer, the mask will be removed completely.
1657 Node *andi = in(1);
1658 if( in1_op == Op_AndI ) {
1659 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1660 if( t3 && t3->is_con() ) { // Right input is a constant
1661 jint mask2 = t3->get_con();
1662 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1663 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1664 return new AndINode(newshr, phase->intcon(mask2));
1665 // The negative values are easier to materialize than positive ones.
1666 // A typical case from address arithmetic is ((x & ~15) >> 4).
1667 // It's better to change that to ((x >> 4) & ~0) versus
1668 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1669 }
1670 }
1671
1672 // Check for "(X << z ) >>> z" which simply zero-extends
1673 Node *shl = in(1);
1674 if( in1_op == Op_LShiftI &&
1675 phase->type(shl->in(2)) == t2 )
1676 return new AndINode( shl->in(1), phase->intcon(mask) );
1677
1678 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1679 Node *shr = in(1);
1680 if ( in1_op == Op_RShiftI ) {
1681 Node *in11 = shr->in(1);
1682 Node *in12 = shr->in(2);
1683 const TypeInt *t11 = phase->type(in11)->isa_int();
1684 const TypeInt *t12 = phase->type(in12)->isa_int();
1685 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1686 return new URShiftINode(in11, phase->intcon(31));
1687 }
1688 }
1689
1690 return nullptr;
1691 }
1692
1693 //------------------------------Value------------------------------------------
1694 // A URShiftINode shifts its input2 right by input1 amount.
1695 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1696 // (This is a near clone of RShiftINode::Value.)
1697 const Type *t1 = phase->type( in(1) );
1698 const Type *t2 = phase->type( in(2) );
1699 // Either input is TOP ==> the result is TOP
1700 if( t1 == Type::TOP ) return Type::TOP;
1701 if( t2 == Type::TOP ) return Type::TOP;
1702
1703 // Left input is ZERO ==> the result is ZERO.
1704 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1705 // Shift by zero does nothing
1706 if( t2 == TypeInt::ZERO ) return t1;
1707
1708 // Either input is BOTTOM ==> the result is BOTTOM
1709 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1710 return TypeInt::INT;
1711
1712 if (t2 == TypeInt::INT)
1713 return TypeInt::INT;
1714
1715 const TypeInt *r1 = t1->is_int(); // Handy access
1716 const TypeInt *r2 = t2->is_int(); // Handy access
1717
1718 if (r2->is_con()) {
1719 uint shift = r2->get_con();
1720 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1721 // Shift by a multiple of 32 does nothing:
1722 if (shift == 0) return t1;
1723 // Calculate reasonably aggressive bounds for the result.
1724 jint lo = (juint)r1->_lo >> (juint)shift;
1725 jint hi = (juint)r1->_hi >> (juint)shift;
1726 if (r1->_hi >= 0 && r1->_lo < 0) {
1727 // If the type has both negative and positive values,
1728 // there are two separate sub-domains to worry about:
1729 // The positive half and the negative half.
1730 jint neg_lo = lo;
1731 jint neg_hi = (juint)-1 >> (juint)shift;
1732 jint pos_lo = (juint) 0 >> (juint)shift;
1733 jint pos_hi = hi;
1734 lo = MIN2(neg_lo, pos_lo); // == 0
1735 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1736 }
1737 assert(lo <= hi, "must have valid bounds");
1738 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1739 #ifdef ASSERT
1740 // Make sure we get the sign-capture idiom correct.
1741 if (shift == BitsPerJavaInteger-1) {
1742 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1743 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1744 }
1745 #endif
1746 return ti;
1747 }
1748
1749 //
1750 // Do not support shifted oops in info for GC
1751 //
1752 // else if( t1->base() == Type::InstPtr ) {
1753 //
1754 // const TypeInstPtr *o = t1->is_instptr();
1755 // if( t1->singleton() )
1756 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1757 // }
1758 // else if( t1->base() == Type::KlassPtr ) {
1759 // const TypeKlassPtr *o = t1->is_klassptr();
1760 // if( t1->singleton() )
1761 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1762 // }
1763
1764 return TypeInt::INT;
1765 }
1766
1767 //=============================================================================
1768 //------------------------------Identity---------------------------------------
1769 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1770 int count = 0;
1771 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1772 // Shift by a multiple of 64 does nothing
1773 return in(1);
1774 }
1775 return this;
1776 }
1777
1778 //------------------------------Ideal------------------------------------------
1779 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1780 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong);
1781 if (con == 0) {
1782 return nullptr;
1783 }
1784
1785 // We'll be wanting the right-shift amount as a mask of that many bits
1786 const jlong mask = jlong(max_julong >> con);
1787
1788 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1789 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1790 // If Q is "X << z" the rounding is useless. Look for patterns like
1791 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1792 Node *add = in(1);
1793 const TypeInt *t2 = phase->type(in(2))->isa_int();
1794 if (add->Opcode() == Op_AddL) {
1795 Node *lshl = add->in(1);
1796 if( lshl->Opcode() == Op_LShiftL &&
1797 phase->type(lshl->in(2)) == t2 ) {
1798 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1799 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1800 return new AndLNode( sum, phase->longcon(mask) );
1801 }
1802 }
1803
1804 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1805 // This shortens the mask. Also, if we are extracting a high byte and
1806 // storing it to a buffer, the mask will be removed completely.
1807 Node *andi = in(1);
1808 if( andi->Opcode() == Op_AndL ) {
1809 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1810 if( t3 && t3->is_con() ) { // Right input is a constant
1811 jlong mask2 = t3->get_con();
1812 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1813 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1814 return new AndLNode(newshr, phase->longcon(mask2));
1815 }
1816 }
1817
1818 // Check for "(X << z ) >>> z" which simply zero-extends
1819 Node *shl = in(1);
1820 if( shl->Opcode() == Op_LShiftL &&
1821 phase->type(shl->in(2)) == t2 )
1822 return new AndLNode( shl->in(1), phase->longcon(mask) );
1823
1824 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1825 Node *shr = in(1);
1826 if ( shr->Opcode() == Op_RShiftL ) {
1827 Node *in11 = shr->in(1);
1828 Node *in12 = shr->in(2);
1829 const TypeLong *t11 = phase->type(in11)->isa_long();
1830 const TypeInt *t12 = phase->type(in12)->isa_int();
1831 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1832 return new URShiftLNode(in11, phase->intcon(63));
1833 }
1834 }
1835 return nullptr;
1836 }
1837
1838 //------------------------------Value------------------------------------------
1839 // A URShiftINode shifts its input2 right by input1 amount.
1840 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1841 // (This is a near clone of RShiftLNode::Value.)
1842 const Type *t1 = phase->type( in(1) );
1843 const Type *t2 = phase->type( in(2) );
1844 // Either input is TOP ==> the result is TOP
1845 if( t1 == Type::TOP ) return Type::TOP;
1846 if( t2 == Type::TOP ) return Type::TOP;
1847
1848 // Left input is ZERO ==> the result is ZERO.
1849 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1850 // Shift by zero does nothing
1851 if( t2 == TypeInt::ZERO ) return t1;
1852
1853 // Either input is BOTTOM ==> the result is BOTTOM
1854 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1855 return TypeLong::LONG;
1856
1857 if (t2 == TypeInt::INT)
1858 return TypeLong::LONG;
1859
1860 const TypeLong *r1 = t1->is_long(); // Handy access
1861 const TypeInt *r2 = t2->is_int (); // Handy access
1862
1863 if (r2->is_con()) {
1864 uint shift = r2->get_con();
1865 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1866 // Shift by a multiple of 64 does nothing:
1867 if (shift == 0) return t1;
1868 // Calculate reasonably aggressive bounds for the result.
1869 jlong lo = (julong)r1->_lo >> (juint)shift;
1870 jlong hi = (julong)r1->_hi >> (juint)shift;
1871 if (r1->_hi >= 0 && r1->_lo < 0) {
1872 // If the type has both negative and positive values,
1873 // there are two separate sub-domains to worry about:
1874 // The positive half and the negative half.
1875 jlong neg_lo = lo;
1876 jlong neg_hi = (julong)-1 >> (juint)shift;
1877 jlong pos_lo = (julong) 0 >> (juint)shift;
1878 jlong pos_hi = hi;
1879 //lo = MIN2(neg_lo, pos_lo); // == 0
1880 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1881 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1882 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1883 }
1884 assert(lo <= hi, "must have valid bounds");
1885 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1886 #ifdef ASSERT
1887 // Make sure we get the sign-capture idiom correct.
1888 if (shift == BitsPerJavaLong - 1) {
1889 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1890 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1891 }
1892 #endif
1893 return tl;
1894 }
1895
1896 return TypeLong::LONG; // Give up
1897 }
1898
1899 //=============================================================================
1900 //------------------------------Ideal------------------------------------------
1901 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1902 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1903 // This reduces the number of rules in the matcher, as we only need to check
1904 // for negations on the second argument, and not the symmetric case where
1905 // the first argument is negated.
1906 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1907 swap_edges(1, 2);
1908 return this;
1909 }
1910 return nullptr;
1911 }
1912
1913 //=============================================================================
1914 //------------------------------Value------------------------------------------
1915 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1916 const Type *t1 = phase->type(in(1));
1917 if (t1 == Type::TOP) return Type::TOP;
1918 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1919 const Type *t2 = phase->type(in(2));
1920 if (t2 == Type::TOP) return Type::TOP;
1921 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1922 const Type *t3 = phase->type(in(3));
1923 if (t3 == Type::TOP) return Type::TOP;
1924 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1925 #ifndef __STDC_IEC_559__
1926 return Type::DOUBLE;
1927 #else
1928 double d1 = t1->getd();
1929 double d2 = t2->getd();
1930 double d3 = t3->getd();
1931 return TypeD::make(fma(d1, d2, d3));
1932 #endif
1933 }
1934
1935 //=============================================================================
1936 //------------------------------Value------------------------------------------
1937 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1938 const Type *t1 = phase->type(in(1));
1939 if (t1 == Type::TOP) return Type::TOP;
1940 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1941 const Type *t2 = phase->type(in(2));
1942 if (t2 == Type::TOP) return Type::TOP;
1943 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1944 const Type *t3 = phase->type(in(3));
1945 if (t3 == Type::TOP) return Type::TOP;
1946 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1947 #ifndef __STDC_IEC_559__
1948 return Type::FLOAT;
1949 #else
1950 float f1 = t1->getf();
1951 float f2 = t2->getf();
1952 float f3 = t3->getf();
1953 return TypeF::make(fma(f1, f2, f3));
1954 #endif
1955 }
1956
1957 //=============================================================================
1958 //------------------------------Value------------------------------------------
1959 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
1960 const Type* t1 = phase->type(in(1));
1961 if (t1 == Type::TOP) { return Type::TOP; }
1962 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1963 const Type* t2 = phase->type(in(2));
1964 if (t2 == Type::TOP) { return Type::TOP; }
1965 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1966 const Type* t3 = phase->type(in(3));
1967 if (t3 == Type::TOP) { return Type::TOP; }
1968 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1969 #ifndef __STDC_IEC_559__
1970 return Type::HALF_FLOAT;
1971 #else
1972 float f1 = t1->getf();
1973 float f2 = t2->getf();
1974 float f3 = t3->getf();
1975 return TypeH::make(fma(f1, f2, f3));
1976 #endif
1977 }
1978
1979 //=============================================================================
1980 //------------------------------hash-------------------------------------------
1981 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1982 // The hash function must return the same value when edge swapping is performed.
1983 uint MulAddS2INode::hash() const {
1984 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1985 }
1986
1987 //------------------------------Rotate Operations ------------------------------
1988
1989 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1990 const Type* t1 = phase->type(in(1));
1991 if (t1 == Type::TOP) {
1992 return this;
1993 }
1994 int count = 0;
1995 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1996 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1997 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1998 // Rotate by a multiple of 32/64 does nothing
1999 return in(1);
2000 }
2001 return this;
2002 }
2003
2004 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
2005 const Type* t1 = phase->type(in(1));
2006 const Type* t2 = phase->type(in(2));
2007 // Either input is TOP ==> the result is TOP
2008 if (t1 == Type::TOP || t2 == Type::TOP) {
2009 return Type::TOP;
2010 }
2011
2012 if (t1->isa_int()) {
2013 const TypeInt* r1 = t1->is_int();
2014 const TypeInt* r2 = t2->is_int();
2015
2016 // Left input is ZERO ==> the result is ZERO.
2017 if (r1 == TypeInt::ZERO) {
2018 return TypeInt::ZERO;
2019 }
2020 // Rotate by zero does nothing
2021 if (r2 == TypeInt::ZERO) {
2022 return r1;
2023 }
2024 if (r1->is_con() && r2->is_con()) {
2025 juint r1_con = (juint)r1->get_con();
2026 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2027 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2028 }
2029 return TypeInt::INT;
2030 } else {
2031 assert(t1->isa_long(), "Type must be a long");
2032 const TypeLong* r1 = t1->is_long();
2033 const TypeInt* r2 = t2->is_int();
2034
2035 // Left input is ZERO ==> the result is ZERO.
2036 if (r1 == TypeLong::ZERO) {
2037 return TypeLong::ZERO;
2038 }
2039 // Rotate by zero does nothing
2040 if (r2 == TypeInt::ZERO) {
2041 return r1;
2042 }
2043 if (r1->is_con() && r2->is_con()) {
2044 julong r1_con = (julong)r1->get_con();
2045 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2046 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2047 }
2048 return TypeLong::LONG;
2049 }
2050 }
2051
2052 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2053 const Type* t1 = phase->type(in(1));
2054 const Type* t2 = phase->type(in(2));
2055 if (t2->isa_int() && t2->is_int()->is_con()) {
2056 if (t1->isa_int()) {
2057 int lshift = t2->is_int()->get_con() & 31;
2058 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2059 } else if (t1 != Type::TOP) {
2060 assert(t1->isa_long(), "Type must be a long");
2061 int lshift = t2->is_int()->get_con() & 63;
2062 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2063 }
2064 }
2065 return nullptr;
2066 }
2067
2068 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2069 const Type* t1 = phase->type(in(1));
2070 if (t1 == Type::TOP) {
2071 return this;
2072 }
2073 int count = 0;
2074 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2075 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2076 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2077 // Rotate by a multiple of 32/64 does nothing
2078 return in(1);
2079 }
2080 return this;
2081 }
2082
2083 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2084 const Type* t1 = phase->type(in(1));
2085 const Type* t2 = phase->type(in(2));
2086 // Either input is TOP ==> the result is TOP
2087 if (t1 == Type::TOP || t2 == Type::TOP) {
2088 return Type::TOP;
2089 }
2090
2091 if (t1->isa_int()) {
2092 const TypeInt* r1 = t1->is_int();
2093 const TypeInt* r2 = t2->is_int();
2094
2095 // Left input is ZERO ==> the result is ZERO.
2096 if (r1 == TypeInt::ZERO) {
2097 return TypeInt::ZERO;
2098 }
2099 // Rotate by zero does nothing
2100 if (r2 == TypeInt::ZERO) {
2101 return r1;
2102 }
2103 if (r1->is_con() && r2->is_con()) {
2104 juint r1_con = (juint)r1->get_con();
2105 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2106 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2107 }
2108 return TypeInt::INT;
2109 } else {
2110 assert(t1->isa_long(), "Type must be a long");
2111 const TypeLong* r1 = t1->is_long();
2112 const TypeInt* r2 = t2->is_int();
2113 // Left input is ZERO ==> the result is ZERO.
2114 if (r1 == TypeLong::ZERO) {
2115 return TypeLong::ZERO;
2116 }
2117 // Rotate by zero does nothing
2118 if (r2 == TypeInt::ZERO) {
2119 return r1;
2120 }
2121 if (r1->is_con() && r2->is_con()) {
2122 julong r1_con = (julong)r1->get_con();
2123 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2124 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2125 }
2126 return TypeLong::LONG;
2127 }
2128 }
2129
2130 //------------------------------ Sum & Mask ------------------------------
2131
2132 // Returns a lower bound on the number of trailing zeros in expr.
2133 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2134 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2135 if (type == nullptr) {
2136 return 0;
2137 }
2138
2139 expr = expr->uncast();
2140 type = phase->type(expr)->isa_integer(bt);
2141 if (type == nullptr) {
2142 return 0;
2143 }
2144
2145 if (type->is_con()) {
2146 jlong con = type->get_con_as_long(bt);
2147 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2148 }
2149
2150 if (expr->Opcode() == Op_ConvI2L) {
2151 expr = expr->in(1)->uncast();
2152 bt = T_INT;
2153 type = phase->type(expr)->isa_int();
2154 }
2155
2156 // Pattern: expr = (x << shift)
2157 if (expr->Opcode() == Op_LShift(bt)) {
2158 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2159 if (shift_t == nullptr || !shift_t->is_con()) {
2160 return 0;
2161 }
2162 // We need to truncate the shift, as it may not have been canonicalized yet.
2163 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2164 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2165 // (JLS: "Shift Operators")
2166 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2167 return shift_t->get_con() & shift_mask;
2168 }
2169
2170 return 0;
2171 }
2172
2173 // Checks whether expr is neutral additive element (zero) under mask,
2174 // i.e. whether an expression of the form:
2175 // (AndX (AddX (expr addend) mask)
2176 // (expr + addend) & mask
2177 // is equivalent to
2178 // (AndX addend mask)
2179 // addend & mask
2180 // for any addend.
2181 // (The X in AndX must be I or L, depending on bt).
2182 //
2183 // We check for the sufficient condition when the lowest set bit in expr is higher than
2184 // the highest set bit in mask, i.e.:
2185 // expr: eeeeee0000000000000
2186 // mask: 000000mmmmmmmmmmmmm
2187 // <--w bits--->
2188 // We do not test for other cases.
2189 //
2190 // Correctness:
2191 // Given "expr" with at least "w" trailing zeros,
2192 // let "mod = 2^w", "suffix_mask = mod - 1"
2193 //
2194 // Since "mask" only has bits set where "suffix_mask" does, we have:
2195 // mask = suffix_mask & mask (SUFFIX_MASK)
2196 //
2197 // And since expr only has bits set above w, and suffix_mask only below:
2198 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2199 //
2200 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2201 // a power of 2, and we are computing in a ring of powers of 2, we know that
2202 // (x + y) % mod = (x % mod + y) % mod
2203 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2204 //
2205 // We can now prove the equality:
2206 // (expr + addend) & mask
2207 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2208 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2209 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2210 // = addend & mask (SUFFIX_MASK)
2211 //
2212 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2213 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2214 // When the mask is negative, it has the most significant bit set.
2215 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2216 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2217 return false;
2218 }
2219
2220 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2221 if (mask_t->hi_as_long() == 0) {
2222 assert(mask_t->lo_as_long() == 0, "checked earlier");
2223 return false;
2224 }
2225
2226 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2227 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2228 return expr_trailing_zeros >= mask_bit_width;
2229 }
2230
2231 // Reduces the pattern:
2232 // (AndX (AddX add1 add2) mask)
2233 // to
2234 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2235 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2236 Node* add = in(1);
2237 Node* mask = in(2);
2238 int addidx = 0;
2239 if (add->Opcode() == Op_Add(bt)) {
2240 addidx = 1;
2241 } else if (mask->Opcode() == Op_Add(bt)) {
2242 mask = add;
2243 addidx = 2;
2244 add = in(addidx);
2245 }
2246 if (addidx > 0) {
2247 Node* add1 = add->in(1);
2248 Node* add2 = add->in(2);
2249 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2250 set_req_X(addidx, add2, phase);
2251 return this;
2252 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2253 set_req_X(addidx, add1, phase);
2254 return this;
2255 }
2256 }
2257 return nullptr;
2258 }