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 = 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 = 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, int& 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, int& 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 int mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits) {
1036 int real_shift;
1037 int 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 != 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, int 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 int nbits = static_cast<int>(bits_per_java_integer(bt));
1077 int 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 int count = 0;
1103 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1104 // Shift by a multiple of 32 does nothing
1105 return in(1);
1106 }
1107 return this;
1108 }
1109
1110 //------------------------------Ideal------------------------------------------
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 //
1114 // Also collapse nested left-shifts with constant rhs:
1115 // (X << con1) << con2 ==> X << (con1 + con2)
1116 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1117 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1118 if (con == 0) {
1119 return nullptr;
1120 }
1121
1122 // Left input is an add?
1123 Node *add1 = in(1);
1124 int add1_op = add1->Opcode();
1125 if( add1_op == Op_AddI ) { // Left input is an add?
1126 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" );
1127
1128 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1129 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1130 if( con < 16 ) {
1131 // Left input is an add of the same number?
1132 if (add1->in(1) == add1->in(2)) {
1133 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1134 // In general, this optimization cannot be applied for c0 == 31 since
1135 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1136 return new LShiftINode(add1->in(1), phase->intcon(con + 1));
1137 }
1138
1139 // Left input is an add of a constant?
1140 const TypeInt *t12 = phase->type(add1->in(2))->isa_int();
1141 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1142 // Compute X << con0
1143 Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) );
1144 // Compute X<<con0 + (con1<<con0)
1145 return new AddINode( lsh, phase->intcon(t12->get_con() << con));
1146 }
1147 }
1148 }
1149
1150 // Check for "(x >> C1) << C2"
1151 if (add1_op == Op_RShiftI || add1_op == Op_URShiftI) {
1152 int add1Con = 0;
1153 const_shift_count(phase, add1, &add1Con);
1154
1155 // Special case C1 == C2, which just masks off low bits
1156 if (add1Con > 0 && con == add1Con) {
1157 // Convert to "(x & -(1 << C2))"
1158 return new AndINode(add1->in(1), phase->intcon(java_negate(jint(1 << con))));
1159 } else {
1160 // Wait until the right shift has been sharpened to the correct count
1161 if (add1Con > 0 && add1Con < BitsPerJavaInteger) {
1162 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1163 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1164 if (phase->is_IterGVN()) {
1165 if (con > add1Con) {
1166 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1167 Node* lshift = phase->transform(new LShiftINode(add1->in(1), phase->intcon(con - add1Con)));
1168 return new AndINode(lshift, phase->intcon(java_negate(jint(1 << con))));
1169 } else {
1170 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1171 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1172
1173 // Handle logical and arithmetic shifts
1174 Node* rshift;
1175 if (add1_op == Op_RShiftI) {
1176 rshift = phase->transform(new RShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1177 } else {
1178 rshift = phase->transform(new URShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1179 }
1180
1181 return new AndINode(rshift, phase->intcon(java_negate(jint(1 << con))));
1182 }
1183 } else {
1184 phase->record_for_igvn(this);
1185 }
1186 }
1187 }
1188 }
1189
1190 // Check for "((x >> C1) & Y) << C2"
1191 if (add1_op == Op_AndI) {
1192 Node *add2 = add1->in(1);
1193 int add2_op = add2->Opcode();
1194 if (add2_op == Op_RShiftI || add2_op == Op_URShiftI) {
1195 // Special case C1 == C2, which just masks off low bits
1196 if (add2->in(2) == in(2)) {
1197 // Convert to "(x & (Y << C2))"
1198 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1199 return new AndINode(add2->in(1), y_sh);
1200 }
1201
1202 int add2Con = 0;
1203 const_shift_count(phase, add2, &add2Con);
1204 if (add2Con > 0 && add2Con < BitsPerJavaInteger) {
1205 if (phase->is_IterGVN()) {
1206 // Convert to "((x >> C1) << C2) & (Y << C2)"
1207
1208 // Make "(x >> C1) << C2", which will get folded away by the rule above
1209 Node* x_sh = phase->transform(new LShiftINode(add2, phase->intcon(con)));
1210 // Make "Y << C2", which will simplify when Y is a constant
1211 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1212
1213 return new AndINode(x_sh, y_sh);
1214 } else {
1215 phase->record_for_igvn(this);
1216 }
1217 }
1218 }
1219 }
1220
1221 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1222 // before shifting them away.
1223 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con);
1224 if( add1_op == Op_AndI &&
1225 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) )
1226 return new LShiftINode( add1->in(1), in(2) );
1227
1228 // Performs:
1229 // (X << con1) << con2 ==> X << (con1 + con2)
1230 Node* doubleShift = collapse_nested_shift_left(phase, this, con, T_INT);
1231 if (doubleShift != nullptr) {
1232 return doubleShift;
1233 }
1234
1235 return nullptr;
1236 }
1237
1238 //------------------------------Value------------------------------------------
1239 // A LShiftINode shifts its input2 left by input1 amount.
1240 const Type* LShiftINode::Value(PhaseGVN* phase) 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 ) return Type::TOP;
1245 if( t2 == Type::TOP ) return Type::TOP;
1246
1247 // Left input is ZERO ==> the result is ZERO.
1248 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1249 // Shift by zero does nothing
1250 if( t2 == TypeInt::ZERO ) return t1;
1251
1252 // Either input is BOTTOM ==> the result is BOTTOM
1253 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) ||
1254 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1255 return TypeInt::INT;
1256
1257 const TypeInt *r1 = t1->is_int(); // Handy access
1258 const TypeInt *r2 = t2->is_int(); // Handy access
1259
1260 if (!r2->is_con())
1261 return TypeInt::INT;
1262
1263 uint shift = r2->get_con();
1264 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1265 // Shift by a multiple of 32 does nothing:
1266 if (shift == 0) return t1;
1267
1268 // If the shift is a constant, shift the bounds of the type,
1269 // unless this could lead to an overflow.
1270 if (!r1->is_con()) {
1271 jint lo = r1->_lo, hi = r1->_hi;
1272 if (((lo << shift) >> shift) == lo &&
1273 ((hi << shift) >> shift) == hi) {
1274 // No overflow. The range shifts up cleanly.
1275 return TypeInt::make((jint)lo << (jint)shift,
1276 (jint)hi << (jint)shift,
1277 MAX2(r1->_widen,r2->_widen));
1278 }
1279 return TypeInt::INT;
1280 }
1281
1282 return TypeInt::make( (jint)r1->get_con() << (jint)shift );
1283 }
1284
1285 //=============================================================================
1286 //------------------------------Identity---------------------------------------
1287 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1288 int count = 0;
1289 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1290 // Shift by a multiple of 64 does nothing
1291 return in(1);
1292 }
1293 return this;
1294 }
1295
1296 //------------------------------Ideal------------------------------------------
1297 // If the right input is a constant, and the left input is an add of a
1298 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1299 //
1300 // Also collapse nested left-shifts with constant rhs:
1301 // (X << con1) << con2 ==> X << (con1 + con2)
1302 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1303 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong);
1304 if (con == 0) {
1305 return nullptr;
1306 }
1307
1308 // Left input is an add?
1309 Node *add1 = in(1);
1310 int add1_op = add1->Opcode();
1311 if( add1_op == Op_AddL ) { // Left input is an add?
1312 // Avoid dead data cycles from dead loops
1313 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" );
1314
1315 // Left input is an add of the same number?
1316 if (con != (BitsPerJavaLong - 1) && add1->in(1) == add1->in(2)) {
1317 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1318 // Can only be applied if c0 != 63 because:
1319 // (x + x) << 63 = 2x << 63, while
1320 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1321 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1322 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1323 return new LShiftLNode(add1->in(1), phase->intcon(con + 1));
1324 }
1325
1326 // Left input is an add of a constant?
1327 const TypeLong *t12 = phase->type(add1->in(2))->isa_long();
1328 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1329 // Compute X << con0
1330 Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) );
1331 // Compute X<<con0 + (con1<<con0)
1332 return new AddLNode( lsh, phase->longcon(t12->get_con() << con));
1333 }
1334 }
1335
1336 // Check for "(x >> C1) << C2"
1337 if (add1_op == Op_RShiftL || add1_op == Op_URShiftL) {
1338 int add1Con = 0;
1339 const_shift_count(phase, add1, &add1Con);
1340
1341 // Special case C1 == C2, which just masks off low bits
1342 if (add1Con > 0 && con == add1Con) {
1343 // Convert to "(x & -(1 << C2))"
1344 return new AndLNode(add1->in(1), phase->longcon(java_negate(jlong(CONST64(1) << con))));
1345 } else {
1346 // Wait until the right shift has been sharpened to the correct count
1347 if (add1Con > 0 && add1Con < BitsPerJavaLong) {
1348 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1349 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1350 if (phase->is_IterGVN()) {
1351 if (con > add1Con) {
1352 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1353 Node* lshift = phase->transform(new LShiftLNode(add1->in(1), phase->intcon(con - add1Con)));
1354 return new AndLNode(lshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1355 } else {
1356 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1357 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1358
1359 // Handle logical and arithmetic shifts
1360 Node* rshift;
1361 if (add1_op == Op_RShiftL) {
1362 rshift = phase->transform(new RShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1363 } else {
1364 rshift = phase->transform(new URShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1365 }
1366
1367 return new AndLNode(rshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1368 }
1369 } else {
1370 phase->record_for_igvn(this);
1371 }
1372 }
1373 }
1374 }
1375
1376 // Check for "((x >> C1) & Y) << C2"
1377 if (add1_op == Op_AndL) {
1378 Node* add2 = add1->in(1);
1379 int add2_op = add2->Opcode();
1380 if (add2_op == Op_RShiftL || add2_op == Op_URShiftL) {
1381 // Special case C1 == C2, which just masks off low bits
1382 if (add2->in(2) == in(2)) {
1383 // Convert to "(x & (Y << C2))"
1384 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1385 return new AndLNode(add2->in(1), y_sh);
1386 }
1387
1388 int add2Con = 0;
1389 const_shift_count(phase, add2, &add2Con);
1390 if (add2Con > 0 && add2Con < BitsPerJavaLong) {
1391 if (phase->is_IterGVN()) {
1392 // Convert to "((x >> C1) << C2) & (Y << C2)"
1393
1394 // Make "(x >> C1) << C2", which will get folded away by the rule above
1395 Node* x_sh = phase->transform(new LShiftLNode(add2, phase->intcon(con)));
1396 // Make "Y << C2", which will simplify when Y is a constant
1397 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1398
1399 return new AndLNode(x_sh, y_sh);
1400 } else {
1401 phase->record_for_igvn(this);
1402 }
1403 }
1404 }
1405 }
1406
1407 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits
1408 // before shifting them away.
1409 const jlong bits_mask = jlong(max_julong >> con);
1410 if( add1_op == Op_AndL &&
1411 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) )
1412 return new LShiftLNode( add1->in(1), in(2) );
1413
1414 // Performs:
1415 // (X << con1) << con2 ==> X << (con1 + con2)
1416 Node* doubleShift = collapse_nested_shift_left(phase, this, con, T_LONG);
1417 if (doubleShift != nullptr) {
1418 return doubleShift;
1419 }
1420
1421 return nullptr;
1422 }
1423
1424 //------------------------------Value------------------------------------------
1425 // A LShiftLNode shifts its input2 left by input1 amount.
1426 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1427 const Type *t1 = phase->type( in(1) );
1428 const Type *t2 = phase->type( in(2) );
1429 // Either input is TOP ==> the result is TOP
1430 if( t1 == Type::TOP ) return Type::TOP;
1431 if( t2 == Type::TOP ) return Type::TOP;
1432
1433 // Left input is ZERO ==> the result is ZERO.
1434 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1435 // Shift by zero does nothing
1436 if( t2 == TypeInt::ZERO ) return t1;
1437
1438 // Either input is BOTTOM ==> the result is BOTTOM
1439 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) ||
1440 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1441 return TypeLong::LONG;
1442
1443 const TypeLong *r1 = t1->is_long(); // Handy access
1444 const TypeInt *r2 = t2->is_int(); // Handy access
1445
1446 if (!r2->is_con())
1447 return TypeLong::LONG;
1448
1449 uint shift = r2->get_con();
1450 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1451 // Shift by a multiple of 64 does nothing:
1452 if (shift == 0) return t1;
1453
1454 // If the shift is a constant, shift the bounds of the type,
1455 // unless this could lead to an overflow.
1456 if (!r1->is_con()) {
1457 jlong lo = r1->_lo, hi = r1->_hi;
1458 if (((lo << shift) >> shift) == lo &&
1459 ((hi << shift) >> shift) == hi) {
1460 // No overflow. The range shifts up cleanly.
1461 return TypeLong::make((jlong)lo << (jint)shift,
1462 (jlong)hi << (jint)shift,
1463 MAX2(r1->_widen,r2->_widen));
1464 }
1465 return TypeLong::LONG;
1466 }
1467
1468 return TypeLong::make( (jlong)r1->get_con() << (jint)shift );
1469 }
1470
1471 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1472 switch (bt) {
1473 case T_INT:
1474 return new RShiftINode(in1, in2);
1475 case T_LONG:
1476 return new RShiftLNode(in1, in2);
1477 default:
1478 fatal("Not implemented for %s", type2name(bt));
1479 }
1480 return nullptr;
1481 }
1482
1483
1484 //=============================================================================
1485 //------------------------------Identity---------------------------------------
1486 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1487 int count = 0;
1488 if (const_shift_count(phase, this, &count)) {
1489 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1490 // Shift by a multiple of 32/64 does nothing
1491 return in(1);
1492 }
1493 // Check for useless sign-masking
1494 if (in(1)->Opcode() == Op_LShift(bt) &&
1495 in(1)->req() == 3 &&
1496 in(1)->in(2) == in(2)) {
1497 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1498 // Compute masks for which this shifting doesn't change
1499 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1500 jlong hi = ~lo; // 00007FFF
1501 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1502 if (t11 == nullptr) {
1503 return this;
1504 }
1505 // Does actual value fit inside of mask?
1506 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1507 return in(1)->in(1); // Then shifting is a nop
1508 }
1509 }
1510 }
1511 return this;
1512 }
1513
1514 Node* RShiftINode::Identity(PhaseGVN* phase) {
1515 return IdentityIL(phase, T_INT);
1516 }
1517
1518 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1519 // Inputs may be TOP if they are dead.
1520 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1521 if (t1 == nullptr) {
1522 return NodeSentinel; // Left input is an integer
1523 }
1524 int shift = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1525 if (shift == 0) {
1526 return NodeSentinel;
1527 }
1528
1529 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1530 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1531 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1532 const Node* and_node = in(1);
1533 if (and_node->Opcode() != Op_And(bt)) {
1534 return nullptr;
1535 }
1536 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1537 if (mask_t != nullptr && mask_t->is_con()) {
1538 jlong maskbits = mask_t->get_con_as_long(bt);
1539 // Convert to "(x >> shift) & (mask >> shift)"
1540 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1541 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1542 }
1543 return nullptr;
1544 }
1545
1546 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1547 Node* progress = IdealIL(phase, can_reshape, T_INT);
1548 if (progress == NodeSentinel) {
1549 return nullptr;
1550 }
1551 if (progress != nullptr) {
1552 return progress;
1553 }
1554 int shift = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1555 assert(shift != 0, "handled by IdealIL");
1556
1557 // Check for "(short[i] <<16)>>16" which simply sign-extends
1558 const Node *shl = in(1);
1559 if (shl->Opcode() != Op_LShiftI) {
1560 return nullptr;
1561 }
1562
1563 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1564 if (left_shift_t == nullptr) {
1565 return nullptr;
1566 }
1567 if (shift == 16 && left_shift_t->is_con(16)) {
1568 Node *ld = shl->in(1);
1569 if (ld->Opcode() == Op_LoadS) {
1570 // Sign extension is just useless here. Return a RShiftI of zero instead
1571 // returning 'ld' directly. We cannot return an old Node directly as
1572 // that is the job of 'Identity' calls and Identity calls only work on
1573 // direct inputs ('ld' is an extra Node removed from 'this'). The
1574 // combined optimization requires Identity only return direct inputs.
1575 set_req_X(1, ld, phase);
1576 set_req_X(2, phase->intcon(0), phase);
1577 return this;
1578 }
1579 else if (can_reshape &&
1580 ld->Opcode() == Op_LoadUS &&
1581 ld->outcnt() == 1 && ld->unique_out() == shl)
1582 // Replace zero-extension-load with sign-extension-load
1583 return ld->as_Load()->convert_to_signed_load(*phase);
1584 }
1585
1586 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1587 if (shift == 24 && left_shift_t->is_con(24)) {
1588 Node *ld = shl->in(1);
1589 if (ld->Opcode() == Op_LoadB) {
1590 // Sign extension is just useless here
1591 set_req_X(1, ld, phase);
1592 set_req_X(2, phase->intcon(0), phase);
1593 return this;
1594 }
1595 }
1596
1597 return nullptr;
1598 }
1599
1600 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1601 const Type* t1 = phase->type(in(1));
1602 const Type* t2 = phase->type(in(2));
1603 // Either input is TOP ==> the result is TOP
1604 if (t1 == Type::TOP) {
1605 return Type::TOP;
1606 }
1607 if (t2 == Type::TOP) {
1608 return Type::TOP;
1609 }
1610
1611 // Left input is ZERO ==> the result is ZERO.
1612 if (t1 == TypeInteger::zero(bt)) {
1613 return TypeInteger::zero(bt);
1614 }
1615 // Shift by zero does nothing
1616 if (t2 == TypeInt::ZERO) {
1617 return t1;
1618 }
1619
1620 // Either input is BOTTOM ==> the result is BOTTOM
1621 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1622 return TypeInteger::bottom(bt);
1623 }
1624
1625 const TypeInteger* r1 = t1->isa_integer(bt);
1626 const TypeInt* r2 = t2->isa_int();
1627
1628 // If the shift is a constant, just shift the bounds of the type.
1629 // For example, if the shift is 31/63, we just propagate sign bits.
1630 if (!r1->is_con() && r2->is_con()) {
1631 uint shift = r2->get_con();
1632 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1633 // Shift by a multiple of 32/64 does nothing:
1634 if (shift == 0) {
1635 return t1;
1636 }
1637 // Calculate reasonably aggressive bounds for the result.
1638 // This is necessary if we are to correctly type things
1639 // like (x<<24>>24) == ((byte)x).
1640 jlong lo = r1->lo_as_long() >> (jint)shift;
1641 jlong hi = r1->hi_as_long() >> (jint)shift;
1642 assert(lo <= hi, "must have valid bounds");
1643 #ifdef ASSERT
1644 if (bt == T_INT) {
1645 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1646 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1647 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1648 }
1649 #endif
1650 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1651 #ifdef ASSERT
1652 // Make sure we get the sign-capture idiom correct.
1653 if (shift == bits_per_java_integer(bt) - 1) {
1654 if (r1->lo_as_long() >= 0) {
1655 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1656 }
1657 if (r1->hi_as_long() < 0) {
1658 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1659 }
1660 }
1661 #endif
1662 return ti;
1663 }
1664
1665 if (!r1->is_con() || !r2->is_con()) {
1666 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1667 if (r1->lo_as_long() >= 0) {
1668 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1669 }
1670
1671 // Conversely, if the left input is negative then the result must be negative.
1672 if (r1->hi_as_long() <= -1) {
1673 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1674 }
1675
1676 return TypeInteger::bottom(bt);
1677 }
1678
1679 // Signed shift right
1680 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1681 }
1682
1683 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1684 return ValueIL(phase, T_INT);
1685 }
1686
1687 //=============================================================================
1688 //------------------------------Identity---------------------------------------
1689 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1690 return IdentityIL(phase, T_LONG);
1691 }
1692
1693 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1694 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1695 if (progress == NodeSentinel) {
1696 return nullptr;
1697 }
1698 return progress;
1699 }
1700
1701 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1702 return ValueIL(phase, T_LONG);
1703 }
1704
1705 //=============================================================================
1706 //------------------------------Identity---------------------------------------
1707 Node* URShiftINode::Identity(PhaseGVN* phase) {
1708 int count = 0;
1709 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1710 // Shift by a multiple of 32 does nothing
1711 return in(1);
1712 }
1713
1714 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1715 // Happens during new-array length computation.
1716 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1717 Node *add = in(1);
1718 if (add->Opcode() == Op_AddI) {
1719 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1720 if (t2 && t2->is_con(wordSize - 1) &&
1721 add->in(1)->Opcode() == Op_LShiftI) {
1722 // Check that shift_counts are LogBytesPerWord.
1723 Node *lshift_count = add->in(1)->in(2);
1724 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1725 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1726 t_lshift_count == phase->type(in(2))) {
1727 Node *x = add->in(1)->in(1);
1728 const TypeInt *t_x = phase->type(x)->isa_int();
1729 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1730 return x;
1731 }
1732 }
1733 }
1734 }
1735
1736 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1737 }
1738
1739 //------------------------------Ideal------------------------------------------
1740 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1741 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1742 if (con == 0) {
1743 return nullptr;
1744 }
1745
1746 // We'll be wanting the right-shift amount as a mask of that many bits
1747 const int mask = right_n_bits(BitsPerJavaInteger - con);
1748
1749 int in1_op = in(1)->Opcode();
1750
1751 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1752 if( in1_op == Op_URShiftI ) {
1753 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1754 if( t12 && t12->is_con() ) { // Right input is a constant
1755 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1756 const int con2 = t12->get_con() & 31; // Shift count is always masked
1757 const int con3 = con+con2;
1758 if( con3 < 32 ) // Only merge shifts if total is < 32
1759 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1760 }
1761 }
1762
1763 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1764 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1765 // If Q is "X << z" the rounding is useless. Look for patterns like
1766 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1767 Node *add = in(1);
1768 const TypeInt *t2 = phase->type(in(2))->isa_int();
1769 if (in1_op == Op_AddI) {
1770 Node *lshl = add->in(1);
1771 if( lshl->Opcode() == Op_LShiftI &&
1772 phase->type(lshl->in(2)) == t2 ) {
1773 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1774 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1775 return new AndINode( sum, phase->intcon(mask) );
1776 }
1777 }
1778
1779 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1780 // This shortens the mask. Also, if we are extracting a high byte and
1781 // storing it to a buffer, the mask will be removed completely.
1782 Node *andi = in(1);
1783 if( in1_op == Op_AndI ) {
1784 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1785 if( t3 && t3->is_con() ) { // Right input is a constant
1786 jint mask2 = t3->get_con();
1787 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1788 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1789 return new AndINode(newshr, phase->intcon(mask2));
1790 // The negative values are easier to materialize than positive ones.
1791 // A typical case from address arithmetic is ((x & ~15) >> 4).
1792 // It's better to change that to ((x >> 4) & ~0) versus
1793 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1794 }
1795 }
1796
1797 // Check for "(X << z ) >>> z" which simply zero-extends
1798 Node *shl = in(1);
1799 if( in1_op == Op_LShiftI &&
1800 phase->type(shl->in(2)) == t2 )
1801 return new AndINode( shl->in(1), phase->intcon(mask) );
1802
1803 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1804 Node *shr = in(1);
1805 if ( in1_op == Op_RShiftI ) {
1806 Node *in11 = shr->in(1);
1807 Node *in12 = shr->in(2);
1808 const TypeInt *t11 = phase->type(in11)->isa_int();
1809 const TypeInt *t12 = phase->type(in12)->isa_int();
1810 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1811 return new URShiftINode(in11, phase->intcon(31));
1812 }
1813 }
1814
1815 return nullptr;
1816 }
1817
1818 //------------------------------Value------------------------------------------
1819 // A URShiftINode shifts its input2 right by input1 amount.
1820 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1821 // (This is a near clone of RShiftINode::Value.)
1822 const Type *t1 = phase->type( in(1) );
1823 const Type *t2 = phase->type( in(2) );
1824 // Either input is TOP ==> the result is TOP
1825 if( t1 == Type::TOP ) return Type::TOP;
1826 if( t2 == Type::TOP ) return Type::TOP;
1827
1828 // Left input is ZERO ==> the result is ZERO.
1829 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1830 // Shift by zero does nothing
1831 if( t2 == TypeInt::ZERO ) return t1;
1832
1833 // Either input is BOTTOM ==> the result is BOTTOM
1834 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1835 return TypeInt::INT;
1836
1837 if (t2 == TypeInt::INT)
1838 return TypeInt::INT;
1839
1840 const TypeInt *r1 = t1->is_int(); // Handy access
1841 const TypeInt *r2 = t2->is_int(); // Handy access
1842
1843 if (r2->is_con()) {
1844 uint shift = r2->get_con();
1845 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1846 // Shift by a multiple of 32 does nothing:
1847 if (shift == 0) return t1;
1848 // Calculate reasonably aggressive bounds for the result.
1849 jint lo = (juint)r1->_lo >> (juint)shift;
1850 jint hi = (juint)r1->_hi >> (juint)shift;
1851 if (r1->_hi >= 0 && r1->_lo < 0) {
1852 // If the type has both negative and positive values,
1853 // there are two separate sub-domains to worry about:
1854 // The positive half and the negative half.
1855 jint neg_lo = lo;
1856 jint neg_hi = (juint)-1 >> (juint)shift;
1857 jint pos_lo = (juint) 0 >> (juint)shift;
1858 jint pos_hi = hi;
1859 lo = MIN2(neg_lo, pos_lo); // == 0
1860 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1861 }
1862 assert(lo <= hi, "must have valid bounds");
1863 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1864 #ifdef ASSERT
1865 // Make sure we get the sign-capture idiom correct.
1866 if (shift == BitsPerJavaInteger-1) {
1867 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1868 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1869 }
1870 #endif
1871 return ti;
1872 }
1873
1874 //
1875 // Do not support shifted oops in info for GC
1876 //
1877 // else if( t1->base() == Type::InstPtr ) {
1878 //
1879 // const TypeInstPtr *o = t1->is_instptr();
1880 // if( t1->singleton() )
1881 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1882 // }
1883 // else if( t1->base() == Type::KlassPtr ) {
1884 // const TypeKlassPtr *o = t1->is_klassptr();
1885 // if( t1->singleton() )
1886 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1887 // }
1888
1889 return TypeInt::INT;
1890 }
1891
1892 //=============================================================================
1893 //------------------------------Identity---------------------------------------
1894 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1895 int count = 0;
1896 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1897 // Shift by a multiple of 64 does nothing
1898 return in(1);
1899 }
1900 return this;
1901 }
1902
1903 //------------------------------Ideal------------------------------------------
1904 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1905 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong);
1906 if (con == 0) {
1907 return nullptr;
1908 }
1909
1910 // We'll be wanting the right-shift amount as a mask of that many bits
1911 const jlong mask = jlong(max_julong >> con);
1912
1913 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1914 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1915 // If Q is "X << z" the rounding is useless. Look for patterns like
1916 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1917 Node *add = in(1);
1918 const TypeInt *t2 = phase->type(in(2))->isa_int();
1919 if (add->Opcode() == Op_AddL) {
1920 Node *lshl = add->in(1);
1921 if( lshl->Opcode() == Op_LShiftL &&
1922 phase->type(lshl->in(2)) == t2 ) {
1923 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1924 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1925 return new AndLNode( sum, phase->longcon(mask) );
1926 }
1927 }
1928
1929 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1930 // This shortens the mask. Also, if we are extracting a high byte and
1931 // storing it to a buffer, the mask will be removed completely.
1932 Node *andi = in(1);
1933 if( andi->Opcode() == Op_AndL ) {
1934 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1935 if( t3 && t3->is_con() ) { // Right input is a constant
1936 jlong mask2 = t3->get_con();
1937 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1938 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1939 return new AndLNode(newshr, phase->longcon(mask2));
1940 }
1941 }
1942
1943 // Check for "(X << z ) >>> z" which simply zero-extends
1944 Node *shl = in(1);
1945 if( shl->Opcode() == Op_LShiftL &&
1946 phase->type(shl->in(2)) == t2 )
1947 return new AndLNode( shl->in(1), phase->longcon(mask) );
1948
1949 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1950 Node *shr = in(1);
1951 if ( shr->Opcode() == Op_RShiftL ) {
1952 Node *in11 = shr->in(1);
1953 Node *in12 = shr->in(2);
1954 const TypeLong *t11 = phase->type(in11)->isa_long();
1955 const TypeInt *t12 = phase->type(in12)->isa_int();
1956 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1957 return new URShiftLNode(in11, phase->intcon(63));
1958 }
1959 }
1960 return nullptr;
1961 }
1962
1963 //------------------------------Value------------------------------------------
1964 // A URShiftINode shifts its input2 right by input1 amount.
1965 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1966 // (This is a near clone of RShiftLNode::Value.)
1967 const Type *t1 = phase->type( in(1) );
1968 const Type *t2 = phase->type( in(2) );
1969 // Either input is TOP ==> the result is TOP
1970 if( t1 == Type::TOP ) return Type::TOP;
1971 if( t2 == Type::TOP ) return Type::TOP;
1972
1973 // Left input is ZERO ==> the result is ZERO.
1974 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1975 // Shift by zero does nothing
1976 if( t2 == TypeInt::ZERO ) return t1;
1977
1978 // Either input is BOTTOM ==> the result is BOTTOM
1979 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1980 return TypeLong::LONG;
1981
1982 if (t2 == TypeInt::INT)
1983 return TypeLong::LONG;
1984
1985 const TypeLong *r1 = t1->is_long(); // Handy access
1986 const TypeInt *r2 = t2->is_int (); // Handy access
1987
1988 if (r2->is_con()) {
1989 uint shift = r2->get_con();
1990 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1991 // Shift by a multiple of 64 does nothing:
1992 if (shift == 0) return t1;
1993 // Calculate reasonably aggressive bounds for the result.
1994 jlong lo = (julong)r1->_lo >> (juint)shift;
1995 jlong hi = (julong)r1->_hi >> (juint)shift;
1996 if (r1->_hi >= 0 && r1->_lo < 0) {
1997 // If the type has both negative and positive values,
1998 // there are two separate sub-domains to worry about:
1999 // The positive half and the negative half.
2000 jlong neg_lo = lo;
2001 jlong neg_hi = (julong)-1 >> (juint)shift;
2002 jlong pos_lo = (julong) 0 >> (juint)shift;
2003 jlong pos_hi = hi;
2004 //lo = MIN2(neg_lo, pos_lo); // == 0
2005 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
2006 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
2007 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
2008 }
2009 assert(lo <= hi, "must have valid bounds");
2010 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
2011 #ifdef ASSERT
2012 // Make sure we get the sign-capture idiom correct.
2013 if (shift == BitsPerJavaLong - 1) {
2014 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
2015 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
2016 }
2017 #endif
2018 return tl;
2019 }
2020
2021 return TypeLong::LONG; // Give up
2022 }
2023
2024 //=============================================================================
2025 //------------------------------Ideal------------------------------------------
2026 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2027 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
2028 // This reduces the number of rules in the matcher, as we only need to check
2029 // for negations on the second argument, and not the symmetric case where
2030 // the first argument is negated.
2031 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
2032 swap_edges(1, 2);
2033 return this;
2034 }
2035 return nullptr;
2036 }
2037
2038 //=============================================================================
2039 //------------------------------Value------------------------------------------
2040 const Type* FmaDNode::Value(PhaseGVN* phase) const {
2041 const Type *t1 = phase->type(in(1));
2042 if (t1 == Type::TOP) return Type::TOP;
2043 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
2044 const Type *t2 = phase->type(in(2));
2045 if (t2 == Type::TOP) return Type::TOP;
2046 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
2047 const Type *t3 = phase->type(in(3));
2048 if (t3 == Type::TOP) return Type::TOP;
2049 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
2050 #ifndef __STDC_IEC_559__
2051 return Type::DOUBLE;
2052 #else
2053 double d1 = t1->getd();
2054 double d2 = t2->getd();
2055 double d3 = t3->getd();
2056 return TypeD::make(fma(d1, d2, d3));
2057 #endif
2058 }
2059
2060 //=============================================================================
2061 //------------------------------Value------------------------------------------
2062 const Type* FmaFNode::Value(PhaseGVN* phase) const {
2063 const Type *t1 = phase->type(in(1));
2064 if (t1 == Type::TOP) return Type::TOP;
2065 if (t1->base() != Type::FloatCon) return Type::FLOAT;
2066 const Type *t2 = phase->type(in(2));
2067 if (t2 == Type::TOP) return Type::TOP;
2068 if (t2->base() != Type::FloatCon) return Type::FLOAT;
2069 const Type *t3 = phase->type(in(3));
2070 if (t3 == Type::TOP) return Type::TOP;
2071 if (t3->base() != Type::FloatCon) return Type::FLOAT;
2072 #ifndef __STDC_IEC_559__
2073 return Type::FLOAT;
2074 #else
2075 float f1 = t1->getf();
2076 float f2 = t2->getf();
2077 float f3 = t3->getf();
2078 return TypeF::make(fma(f1, f2, f3));
2079 #endif
2080 }
2081
2082 //=============================================================================
2083 //------------------------------Value------------------------------------------
2084 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
2085 const Type* t1 = phase->type(in(1));
2086 if (t1 == Type::TOP) { return Type::TOP; }
2087 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2088 const Type* t2 = phase->type(in(2));
2089 if (t2 == Type::TOP) { return Type::TOP; }
2090 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2091 const Type* t3 = phase->type(in(3));
2092 if (t3 == Type::TOP) { return Type::TOP; }
2093 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
2094 #ifndef __STDC_IEC_559__
2095 return Type::HALF_FLOAT;
2096 #else
2097 float f1 = t1->getf();
2098 float f2 = t2->getf();
2099 float f3 = t3->getf();
2100 return TypeH::make(fma(f1, f2, f3));
2101 #endif
2102 }
2103
2104 //=============================================================================
2105 //------------------------------hash-------------------------------------------
2106 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
2107 // The hash function must return the same value when edge swapping is performed.
2108 uint MulAddS2INode::hash() const {
2109 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
2110 }
2111
2112 //------------------------------Rotate Operations ------------------------------
2113
2114 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
2115 const Type* t1 = phase->type(in(1));
2116 if (t1 == Type::TOP) {
2117 return this;
2118 }
2119 int count = 0;
2120 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2121 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2122 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2123 // Rotate by a multiple of 32/64 does nothing
2124 return in(1);
2125 }
2126 return this;
2127 }
2128
2129 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
2130 const Type* t1 = phase->type(in(1));
2131 const Type* t2 = phase->type(in(2));
2132 // Either input is TOP ==> the result is TOP
2133 if (t1 == Type::TOP || t2 == Type::TOP) {
2134 return Type::TOP;
2135 }
2136
2137 if (t1->isa_int()) {
2138 const TypeInt* r1 = t1->is_int();
2139 const TypeInt* r2 = t2->is_int();
2140
2141 // Left input is ZERO ==> the result is ZERO.
2142 if (r1 == TypeInt::ZERO) {
2143 return TypeInt::ZERO;
2144 }
2145 // Rotate by zero does nothing
2146 if (r2 == TypeInt::ZERO) {
2147 return r1;
2148 }
2149 if (r1->is_con() && r2->is_con()) {
2150 juint r1_con = (juint)r1->get_con();
2151 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2152 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2153 }
2154 return TypeInt::INT;
2155 } else {
2156 assert(t1->isa_long(), "Type must be a long");
2157 const TypeLong* r1 = t1->is_long();
2158 const TypeInt* r2 = t2->is_int();
2159
2160 // Left input is ZERO ==> the result is ZERO.
2161 if (r1 == TypeLong::ZERO) {
2162 return TypeLong::ZERO;
2163 }
2164 // Rotate by zero does nothing
2165 if (r2 == TypeInt::ZERO) {
2166 return r1;
2167 }
2168 if (r1->is_con() && r2->is_con()) {
2169 julong r1_con = (julong)r1->get_con();
2170 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2171 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2172 }
2173 return TypeLong::LONG;
2174 }
2175 }
2176
2177 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2178 const Type* t1 = phase->type(in(1));
2179 const Type* t2 = phase->type(in(2));
2180 if (t2->isa_int() && t2->is_int()->is_con()) {
2181 if (t1->isa_int()) {
2182 int lshift = t2->is_int()->get_con() & 31;
2183 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2184 } else if (t1 != Type::TOP) {
2185 assert(t1->isa_long(), "Type must be a long");
2186 int lshift = t2->is_int()->get_con() & 63;
2187 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2188 }
2189 }
2190 return nullptr;
2191 }
2192
2193 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2194 const Type* t1 = phase->type(in(1));
2195 if (t1 == Type::TOP) {
2196 return this;
2197 }
2198 int count = 0;
2199 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2200 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2201 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2202 // Rotate by a multiple of 32/64 does nothing
2203 return in(1);
2204 }
2205 return this;
2206 }
2207
2208 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2209 const Type* t1 = phase->type(in(1));
2210 const Type* t2 = phase->type(in(2));
2211 // Either input is TOP ==> the result is TOP
2212 if (t1 == Type::TOP || t2 == Type::TOP) {
2213 return Type::TOP;
2214 }
2215
2216 if (t1->isa_int()) {
2217 const TypeInt* r1 = t1->is_int();
2218 const TypeInt* r2 = t2->is_int();
2219
2220 // Left input is ZERO ==> the result is ZERO.
2221 if (r1 == TypeInt::ZERO) {
2222 return TypeInt::ZERO;
2223 }
2224 // Rotate by zero does nothing
2225 if (r2 == TypeInt::ZERO) {
2226 return r1;
2227 }
2228 if (r1->is_con() && r2->is_con()) {
2229 juint r1_con = (juint)r1->get_con();
2230 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2231 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2232 }
2233 return TypeInt::INT;
2234 } else {
2235 assert(t1->isa_long(), "Type must be a long");
2236 const TypeLong* r1 = t1->is_long();
2237 const TypeInt* r2 = t2->is_int();
2238 // Left input is ZERO ==> the result is ZERO.
2239 if (r1 == TypeLong::ZERO) {
2240 return TypeLong::ZERO;
2241 }
2242 // Rotate by zero does nothing
2243 if (r2 == TypeInt::ZERO) {
2244 return r1;
2245 }
2246 if (r1->is_con() && r2->is_con()) {
2247 julong r1_con = (julong)r1->get_con();
2248 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2249 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2250 }
2251 return TypeLong::LONG;
2252 }
2253 }
2254
2255 //------------------------------ Sum & Mask ------------------------------
2256
2257 // Returns a lower bound on the number of trailing zeros in expr.
2258 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2259 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2260 if (type == nullptr) {
2261 return 0;
2262 }
2263
2264 expr = expr->uncast();
2265 type = phase->type(expr)->isa_integer(bt);
2266 if (type == nullptr) {
2267 return 0;
2268 }
2269
2270 if (type->is_con()) {
2271 jlong con = type->get_con_as_long(bt);
2272 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2273 }
2274
2275 if (expr->Opcode() == Op_ConvI2L) {
2276 expr = expr->in(1)->uncast();
2277 bt = T_INT;
2278 type = phase->type(expr)->isa_int();
2279 }
2280
2281 // Pattern: expr = (x << shift)
2282 if (expr->Opcode() == Op_LShift(bt)) {
2283 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2284 if (shift_t == nullptr || !shift_t->is_con()) {
2285 return 0;
2286 }
2287 // We need to truncate the shift, as it may not have been canonicalized yet.
2288 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2289 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2290 // (JLS: "Shift Operators")
2291 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2292 return shift_t->get_con() & shift_mask;
2293 }
2294
2295 return 0;
2296 }
2297
2298 // Checks whether expr is neutral additive element (zero) under mask,
2299 // i.e. whether an expression of the form:
2300 // (AndX (AddX (expr addend) mask)
2301 // (expr + addend) & mask
2302 // is equivalent to
2303 // (AndX addend mask)
2304 // addend & mask
2305 // for any addend.
2306 // (The X in AndX must be I or L, depending on bt).
2307 //
2308 // We check for the sufficient condition when the lowest set bit in expr is higher than
2309 // the highest set bit in mask, i.e.:
2310 // expr: eeeeee0000000000000
2311 // mask: 000000mmmmmmmmmmmmm
2312 // <--w bits--->
2313 // We do not test for other cases.
2314 //
2315 // Correctness:
2316 // Given "expr" with at least "w" trailing zeros,
2317 // let "mod = 2^w", "suffix_mask = mod - 1"
2318 //
2319 // Since "mask" only has bits set where "suffix_mask" does, we have:
2320 // mask = suffix_mask & mask (SUFFIX_MASK)
2321 //
2322 // And since expr only has bits set above w, and suffix_mask only below:
2323 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2324 //
2325 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2326 // a power of 2, and we are computing in a ring of powers of 2, we know that
2327 // (x + y) % mod = (x % mod + y) % mod
2328 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2329 //
2330 // We can now prove the equality:
2331 // (expr + addend) & mask
2332 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2333 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2334 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2335 // = addend & mask (SUFFIX_MASK)
2336 //
2337 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2338 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2339 // When the mask is negative, it has the most significant bit set.
2340 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2341 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2342 return false;
2343 }
2344
2345 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2346 if (mask_t->hi_as_long() == 0) {
2347 assert(mask_t->lo_as_long() == 0, "checked earlier");
2348 return false;
2349 }
2350
2351 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2352 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2353 return expr_trailing_zeros >= mask_bit_width;
2354 }
2355
2356 // Reduces the pattern:
2357 // (AndX (AddX add1 add2) mask)
2358 // to
2359 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2360 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2361 Node* add = in(1);
2362 Node* mask = in(2);
2363 int addidx = 0;
2364 if (add->Opcode() == Op_Add(bt)) {
2365 addidx = 1;
2366 } else if (mask->Opcode() == Op_Add(bt)) {
2367 mask = add;
2368 addidx = 2;
2369 add = in(addidx);
2370 }
2371 if (addidx > 0) {
2372 Node* add1 = add->in(1);
2373 Node* add2 = add->in(2);
2374 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2375 set_req_X(addidx, add2, phase);
2376 return this;
2377 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2378 set_req_X(addidx, add1, phase);
2379 return this;
2380 }
2381 }
2382 return nullptr;
2383 }