1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "memory/allocation.inline.hpp"
26 #include "opto/addnode.hpp"
27 #include "opto/connode.hpp"
28 #include "opto/convertnode.hpp"
29 #include "opto/memnode.hpp"
30 #include "opto/mulnode.hpp"
31 #include "opto/phaseX.hpp"
32 #include "opto/rangeinference.hpp"
33 #include "opto/subnode.hpp"
34 #include "utilities/powerOfTwo.hpp"
35
36 // Portions of code courtesy of Clifford Click
37
38
39 //=============================================================================
40 //------------------------------hash-------------------------------------------
41 // Hash function over MulNodes. Needs to be commutative; i.e., I swap
42 // (commute) inputs to MulNodes willy-nilly so the hash function must return
43 // the same value in the presence of edge swapping.
44 uint MulNode::hash() const {
45 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode();
46 }
47
48 //------------------------------Identity---------------------------------------
49 // Multiplying a one preserves the other argument
50 Node* MulNode::Identity(PhaseGVN* phase) {
51 const Type *one = mul_id(); // The multiplicative identity
52 if( phase->type( in(1) )->higher_equal( one ) ) return in(2);
53 if( phase->type( in(2) )->higher_equal( one ) ) return in(1);
54
55 return this;
56 }
57
58 //------------------------------Ideal------------------------------------------
59 // We also canonicalize the Node, moving constants to the right input,
60 // and flatten expressions (so that 1+x+2 becomes x+3).
61 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) {
62 Node* in1 = in(1);
63 Node* in2 = in(2);
64 Node* progress = nullptr; // Progress flag
65
66 // This code is used by And nodes too, but some conversions are
67 // only valid for the actual Mul nodes.
68 uint op = Opcode();
69 bool real_mul = (op == Op_MulI) || (op == Op_MulL) ||
70 (op == Op_MulF) || (op == Op_MulD) ||
71 (op == Op_MulHF);
72
73 // Convert "(-a)*(-b)" into "a*b".
74 if (real_mul && in1->is_Sub() && in2->is_Sub()) {
75 if (phase->type(in1->in(1))->is_zero_type() &&
76 phase->type(in2->in(1))->is_zero_type()) {
77 set_req_X(1, in1->in(2), phase);
78 set_req_X(2, in2->in(2), phase);
79 in1 = in(1);
80 in2 = in(2);
81 progress = this;
82 }
83 }
84
85 // convert "max(a,b) * min(a,b)" into "a*b".
86 if ((in(1)->Opcode() == max_opcode() && in(2)->Opcode() == min_opcode())
87 || (in(1)->Opcode() == min_opcode() && in(2)->Opcode() == max_opcode())) {
88 Node *in11 = in(1)->in(1);
89 Node *in12 = in(1)->in(2);
90
91 Node *in21 = in(2)->in(1);
92 Node *in22 = in(2)->in(2);
93
94 if ((in11 == in21 && in12 == in22) ||
95 (in11 == in22 && in12 == in21)) {
96 set_req_X(1, in11, phase);
97 set_req_X(2, in12, phase);
98 in1 = in(1);
99 in2 = in(2);
100 progress = this;
101 }
102 }
103
104 const Type* t1 = phase->type(in1);
105 const Type* t2 = phase->type(in2);
106
107 // We are OK if right is a constant, or right is a load and
108 // left is a non-constant.
109 if( !(t2->singleton() ||
110 (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
111 if( t1->singleton() || // Left input is a constant?
112 // Otherwise, sort inputs (commutativity) to help value numbering.
113 (in(1)->_idx > in(2)->_idx) ) {
114 swap_edges(1, 2);
115 const Type *t = t1;
116 t1 = t2;
117 t2 = t;
118 progress = this; // Made progress
119 }
120 }
121
122 // If the right input is a constant, and the left input is a product of a
123 // constant, flatten the expression tree.
124 if( t2->singleton() && // Right input is a constant?
125 op != Op_MulF && // Float & double cannot reassociate
126 op != Op_MulD &&
127 op != Op_MulHF) {
128 if( t2 == Type::TOP ) return nullptr;
129 Node *mul1 = in(1);
130 #ifdef ASSERT
131 // Check for dead loop
132 int op1 = mul1->Opcode();
133 if ((mul1 == this) || (in(2) == this) ||
134 ((op1 == mul_opcode() || op1 == add_opcode()) &&
135 ((mul1->in(1) == this) || (mul1->in(2) == this) ||
136 (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) {
137 assert(false, "dead loop in MulNode::Ideal");
138 }
139 #endif
140
141 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply?
142 // Mul of a constant?
143 const Type *t12 = phase->type( mul1->in(2) );
144 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
145 // Compute new constant; check for overflow
146 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
147 if( tcon01->singleton() ) {
148 // The Mul of the flattened expression
149 set_req_X(1, mul1->in(1), phase);
150 set_req_X(2, phase->makecon(tcon01), phase);
151 t2 = tcon01;
152 progress = this; // Made progress
153 }
154 }
155 }
156 // If the right input is a constant, and the left input is an add of a
157 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
158 const Node *add1 = in(1);
159 if( add1->Opcode() == add_opcode() ) { // Left input is an add?
160 // Add of a constant?
161 const Type *t12 = phase->type( add1->in(2) );
162 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
163 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
164 // Compute new constant; check for overflow
165 const Type *tcon01 = mul_ring(t2,t12);
166 if( tcon01->singleton() ) {
167
168 // Convert (X+con1)*con0 into X*con0
169 Node *mul = clone(); // mul = ()*con0
170 mul->set_req(1,add1->in(1)); // mul = X*con0
171 mul = phase->transform(mul);
172
173 Node *add2 = add1->clone();
174 add2->set_req(1, mul); // X*con0 + con0*con1
175 add2->set_req(2, phase->makecon(tcon01) );
176 progress = add2;
177 }
178 }
179 } // End of is left input an add
180 } // End of is right input a Mul
181
182 return progress;
183 }
184
185 //------------------------------Value-----------------------------------------
186 const Type* MulNode::Value(PhaseGVN* phase) const {
187 const Type *t1 = phase->type( in(1) );
188 const Type *t2 = phase->type( in(2) );
189 // Either input is TOP ==> the result is TOP
190 if( t1 == Type::TOP ) return Type::TOP;
191 if( t2 == Type::TOP ) return Type::TOP;
192
193 // Either input is ZERO ==> the result is ZERO.
194 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
195 int op = Opcode();
196 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
197 const Type *zero = add_id(); // The multiplicative zero
198 if( t1->higher_equal( zero ) ) return zero;
199 if( t2->higher_equal( zero ) ) return zero;
200 }
201
202 // TODO 8350865 Still needed? Yes, I think this is from PhaseMacroExpand::expand_mh_intrinsic_return
203 // Code pattern on return from a call that returns an __Value. Can
204 // be optimized away if the return value turns out to be an oop.
205 if (op == Op_AndX &&
206 in(1) != nullptr &&
207 in(1)->Opcode() == Op_CastP2X &&
208 in(1)->in(1) != nullptr &&
209 phase->type(in(1)->in(1))->isa_oopptr() &&
210 t2->isa_intptr_t()->_lo >= 0 &&
211 t2->isa_intptr_t()->_hi <= MinObjAlignmentInBytesMask) {
212 return add_id();
213 }
214
215 // Either input is BOTTOM ==> the result is the local BOTTOM
216 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
217 return bottom_type();
218
219 return mul_ring(t1,t2); // Local flavor of type multiplication
220 }
221
222 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
223 switch (bt) {
224 case T_INT:
225 return new MulINode(in1, in2);
226 case T_LONG:
227 return new MulLNode(in1, in2);
228 default:
229 fatal("Not implemented for %s", type2name(bt));
230 }
231 return nullptr;
232 }
233
234 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
235 switch (bt) {
236 case T_INT:
237 return new AndINode(in1, in2);
238 case T_LONG:
239 return new AndLNode(in1, in2);
240 default:
241 fatal("Not implemented for %s", type2name(bt));
242 }
243 return nullptr;
244 }
245
246
247 //=============================================================================
248 //------------------------------Ideal------------------------------------------
249 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
250 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
251 const jint con = in(2)->find_int_con(0);
252 if (con == 0) {
253 // If in(2) is not a constant, call Ideal() of the parent class to
254 // try to move constant to the right side.
255 return MulNode::Ideal(phase, can_reshape);
256 }
257
258 // Now we have a constant Node on the right and the constant in con.
259 if (con == 1) {
260 // By one is handled by Identity call
261 return nullptr;
262 }
263
264 // Check for negative constant; if so negate the final result
265 bool sign_flip = false;
266
267 unsigned int abs_con = g_uabs(con);
268 if (abs_con != (unsigned int)con) {
269 sign_flip = true;
270 }
271
272 // Get low bit; check for being the only bit
273 Node *res = nullptr;
274 unsigned int bit1 = submultiple_power_of_2(abs_con);
275 if (bit1 == abs_con) { // Found a power of 2?
276 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
277 } else {
278 // Check for constant with 2 bits set
279 unsigned int bit2 = abs_con - bit1;
280 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
281 if (bit2 + bit1 == abs_con) { // Found all bits in con?
282 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
283 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
284 res = new AddINode(n2, n1);
285 } else if (is_power_of_2(abs_con + 1)) {
286 // Sleezy: power-of-2 - 1. Next time be generic.
287 unsigned int temp = abs_con + 1;
288 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
289 res = new SubINode(n1, in(1));
290 } else {
291 return MulNode::Ideal(phase, can_reshape);
292 }
293 }
294
295 if (sign_flip) { // Need to negate result?
296 res = phase->transform(res);// Transform, before making the zero con
297 res = new SubINode(phase->intcon(0),res);
298 }
299
300 return res; // Return final result
301 }
302
303 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
304 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
305 //
306 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
307 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
308 // - no overflow (i.e. product fits into x bits)
309 // - overflow (i.e. product does not fit into x bits)
310 //
311 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
312 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
313 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
314 // The cross product consists of the following four multiplications with 2x-bit results:
315 // (1) left_lo * right_lo
316 // (2) left_lo * right_hi
317 // (3) left_hi * right_lo
318 // (4) left_hi * right_hi
319 //
320 // Let's define the following two functions:
321 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
322 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
323 //
324 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
325 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
326 // Ux(p) = 0
327 // Lx(p) = p
328 //
329 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
330 // with x bits:
331 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
332 //
333 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
334 // (i.e. all values in the x-bit space could be possible):
335 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
336 //
337 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
338 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
339 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
340 // the lower x bits the same number of times:
341 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
342 //
343 // If all upper x bits are equal, we can conclude:
344 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
345 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
346 //
347 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
348 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
349 //
350 //
351 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
352 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
353 // upper x bits of the 2x-bit result are all set to ones which is minus one in two's complement. If there is an underflow,
354 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
355 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
356 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
357 // that there is no underflow.
358 //
359 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
360 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
361 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
362 // can compare the upper x bits in a unified way with N() being the normalization function:
363 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
364 template<typename NativeType>
365 class IntegerTypeMultiplication {
366
367 NativeType _lo_left;
368 NativeType _lo_right;
369 NativeType _hi_left;
370 NativeType _hi_right;
371 short _widen_left;
372 short _widen_right;
373
374 static const Type* overflow_type();
375 static NativeType multiply_high(NativeType x, NativeType y);
376 const Type* create_type(NativeType lo, NativeType hi) const;
377
378 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
379 return normalize_overflow_value(x, y, multiply_high(x, y));
380 }
381
382 bool cross_product_not_same_overflow_value() const {
383 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
384 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
385 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
386 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
387 return lo_lo_high_product != lo_hi_high_product ||
388 lo_hi_high_product != hi_lo_high_product ||
389 hi_lo_high_product != hi_hi_high_product;
390 }
391
392 bool does_product_overflow(NativeType x, NativeType y) const {
393 return multiply_high_signed_overflow_value(x, y) != 0;
394 }
395
396 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
397 return java_multiply(x, y) < 0 ? result + 1 : result;
398 }
399
400 public:
401 template<class IntegerType>
402 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
403 : _lo_left(left->_lo), _lo_right(right->_lo),
404 _hi_left(left->_hi), _hi_right(right->_hi),
405 _widen_left(left->_widen), _widen_right(right->_widen) {}
406
407 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
408 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
409 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
410 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
411 const Type* compute() const {
412 if (cross_product_not_same_overflow_value()) {
413 return overflow_type();
414 }
415
416 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
417 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
418 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
419 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
420 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
421 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
422 return create_type(min, max);
423 }
424
425 bool does_overflow() const {
426 return does_product_overflow(_lo_left, _lo_right) ||
427 does_product_overflow(_lo_left, _hi_right) ||
428 does_product_overflow(_hi_left, _lo_right) ||
429 does_product_overflow(_hi_left, _hi_right);
430 }
431 };
432
433 template <>
434 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
435 return TypeInt::INT;
436 }
437
438 template <>
439 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
440 const jlong x_64 = x;
441 const jlong y_64 = y;
442 const jlong product = x_64 * y_64;
443 return (jint)((uint64_t)product >> 32u);
444 }
445
446 template <>
447 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
448 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
449 }
450
451 template <>
452 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
453 return TypeLong::LONG;
454 }
455
456 template <>
457 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
458 return multiply_high_signed(x, y);
459 }
460
461 template <>
462 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
463 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
464 }
465
466 // Compute the product type of two integer ranges into this node.
467 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
468 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
469 return integer_multiplication.compute();
470 }
471
472 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
473 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
474 return integer_multiplication.does_overflow();
475 }
476
477 // Compute the product type of two long ranges into this node.
478 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
479 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
480 return integer_multiplication.compute();
481 }
482
483 //=============================================================================
484 //------------------------------Ideal------------------------------------------
485 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
486 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
487 const jlong con = in(2)->find_long_con(0);
488 if (con == 0) {
489 // If in(2) is not a constant, call Ideal() of the parent class to
490 // try to move constant to the right side.
491 return MulNode::Ideal(phase, can_reshape);
492 }
493
494 // Now we have a constant Node on the right and the constant in con.
495 if (con == 1) {
496 // By one is handled by Identity call
497 return nullptr;
498 }
499
500 // Check for negative constant; if so negate the final result
501 bool sign_flip = false;
502 julong abs_con = g_uabs(con);
503 if (abs_con != (julong)con) {
504 sign_flip = true;
505 }
506
507 // Get low bit; check for being the only bit
508 Node *res = nullptr;
509 julong bit1 = submultiple_power_of_2(abs_con);
510 if (bit1 == abs_con) { // Found a power of 2?
511 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
512 } else {
513
514 // Check for constant with 2 bits set
515 julong bit2 = abs_con-bit1;
516 bit2 = bit2 & (0-bit2); // Extract 2nd bit
517 if (bit2 + bit1 == abs_con) { // Found all bits in con?
518 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
519 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
520 res = new AddLNode(n2, n1);
521
522 } else if (is_power_of_2(abs_con+1)) {
523 // Sleezy: power-of-2 -1. Next time be generic.
524 julong temp = abs_con + 1;
525 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
526 res = new SubLNode(n1, in(1));
527 } else {
528 return MulNode::Ideal(phase, can_reshape);
529 }
530 }
531
532 if (sign_flip) { // Need to negate result?
533 res = phase->transform(res);// Transform, before making the zero con
534 res = new SubLNode(phase->longcon(0),res);
535 }
536
537 return res; // Return final result
538 }
539
540 //=============================================================================
541 //------------------------------mul_ring---------------------------------------
542 // Compute the product type of two double ranges into this node.
543 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
544 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
545 return TypeF::make( t0->getf() * t1->getf() );
546 }
547
548 //------------------------------Ideal---------------------------------------
549 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
550 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
551 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
552
553 // x * 2 -> x + x
554 if (t2 != nullptr && t2->getf() == 2) {
555 Node* base = in(1);
556 return new AddFNode(base, base);
557 }
558 return MulNode::Ideal(phase, can_reshape);
559 }
560
561 //=============================================================================
562 //------------------------------Ideal------------------------------------------
563 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
564 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
565 const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
566
567 // x * 2 -> x + x
568 if (t2 != nullptr && t2->getf() == 2) {
569 Node* base = in(1);
570 return new AddHFNode(base, base);
571 }
572 return MulNode::Ideal(phase, can_reshape);
573 }
574
575 // Compute the product type of two half float ranges into this node.
576 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
577 if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
578 return Type::HALF_FLOAT;
579 }
580 return TypeH::make(t0->getf() * t1->getf());
581 }
582
583 //=============================================================================
584 //------------------------------mul_ring---------------------------------------
585 // Compute the product type of two double ranges into this node.
586 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
587 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
588 // We must be multiplying 2 double constants.
589 return TypeD::make( t0->getd() * t1->getd() );
590 }
591
592 //------------------------------Ideal---------------------------------------
593 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
594 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
595 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
596
597 // x * 2 -> x + x
598 if (t2 != nullptr && t2->getd() == 2) {
599 Node* base = in(1);
600 return new AddDNode(base, base);
601 }
602
603 return MulNode::Ideal(phase, can_reshape);
604 }
605
606 //=============================================================================
607 //------------------------------Value------------------------------------------
608 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
609 const Type *t1 = phase->type( in(1) );
610 const Type *t2 = phase->type( in(2) );
611 const Type *bot = bottom_type();
612 return MulHiValue(t1, t2, bot);
613 }
614
615 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
616 const Type *t1 = phase->type( in(1) );
617 const Type *t2 = phase->type( in(2) );
618 const Type *bot = bottom_type();
619 return MulHiValue(t1, t2, bot);
620 }
621
622 // A common routine used by UMulHiLNode and MulHiLNode
623 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
624 // Either input is TOP ==> the result is TOP
625 if( t1 == Type::TOP ) return Type::TOP;
626 if( t2 == Type::TOP ) return Type::TOP;
627
628 // Either input is BOTTOM ==> the result is the local BOTTOM
629 if( (t1 == bot) || (t2 == bot) ||
630 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
631 return bot;
632
633 // It is not worth trying to constant fold this stuff!
634 return TypeLong::LONG;
635 }
636
637 //=============================================================================
638 //------------------------------mul_ring---------------------------------------
639 // Supplied function returns the product of the inputs IN THE CURRENT RING.
640 // For the logical operations the ring's MUL is really a logical AND function.
641 // This also type-checks the inputs for sanity. Guaranteed never to
642 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
643 const Type* AndINode::mul_ring(const Type* t1, const Type* t2) const {
644 return RangeInference::infer_and(t1->is_int(), t2->is_int());
645 }
646
647 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
648
649 const Type* AndINode::Value(PhaseGVN* phase) const {
650 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
651 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
652 return TypeInt::ZERO;
653 }
654
655 return MulNode::Value(phase);
656 }
657
658 //------------------------------Identity---------------------------------------
659 // Masking off the high bits of an unsigned load is not required
660 Node* AndINode::Identity(PhaseGVN* phase) {
661
662 // x & x => x
663 if (in(1) == in(2)) {
664 return in(1);
665 }
666
667 Node* in1 = in(1);
668 uint op = in1->Opcode();
669 const TypeInt* t2 = phase->type(in(2))->isa_int();
670 if (t2 && t2->is_con()) {
671 int con = t2->get_con();
672 // Masking off high bits which are always zero is useless.
673 const TypeInt* t1 = phase->type(in(1))->isa_int();
674 if (t1 != nullptr && t1->_lo >= 0) {
675 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
676 if ((t1_support & con) == t1_support)
677 return in1;
678 }
679 // Masking off the high bits of a unsigned-shift-right is not
680 // needed either.
681 if (op == Op_URShiftI) {
682 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
683 if (t12 && t12->is_con()) { // Shift is by a constant
684 int shift = t12->get_con();
685 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
686 int mask = max_juint >> shift;
687 if ((mask & con) == mask) // If AND is useless, skip it
688 return in1;
689 }
690 }
691 }
692 return MulNode::Identity(phase);
693 }
694
695 //------------------------------Ideal------------------------------------------
696 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
697 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
698 Node* progress = AndIL_sum_and_mask(phase, T_INT);
699 if (progress != nullptr) {
700 return progress;
701 }
702
703 // Convert "(~a) & (~b)" into "~(a | b)"
704 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
705 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
706 Node* tn = phase->transform(or_a_b);
707 return AddNode::make_not(phase, tn, T_INT);
708 }
709
710 // Special case constant AND mask
711 const TypeInt *t2 = phase->type( in(2) )->isa_int();
712 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
713 const int mask = t2->get_con();
714 Node *load = in(1);
715 uint lop = load->Opcode();
716
717 // Masking bits off of a Character? Hi bits are already zero.
718 if( lop == Op_LoadUS &&
719 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
720 return new AndINode(load,phase->intcon(mask&0xFFFF));
721
722 // Masking bits off of a Short? Loading a Character does some masking
723 if (can_reshape &&
724 load->outcnt() == 1 && load->unique_out() == this) {
725 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
726 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
727 ldus = phase->transform(ldus);
728 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
729 }
730
731 // Masking sign bits off of a Byte? Do an unsigned byte load plus
732 // an and.
733 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
734 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
735 ldub = phase->transform(ldub);
736 return new AndINode(ldub, phase->intcon(mask));
737 }
738 }
739
740 // Masking off sign bits? Dont make them!
741 if( lop == Op_RShiftI ) {
742 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
743 if( t12 && t12->is_con() ) { // Shift is by a constant
744 int shift = t12->get_con();
745 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
746 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
747 // If the AND'ing of the 2 masks has no bits, then only original shifted
748 // bits survive. NO sign-extension bits survive the maskings.
749 if( (sign_bits_mask & mask) == 0 ) {
750 // Use zero-fill shift instead
751 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
752 return new AndINode( zshift, in(2) );
753 }
754 }
755 }
756
757 // Check for 'negate/and-1', a pattern emitted when someone asks for
758 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
759 // plus 1) and the mask is of the low order bit. Skip the negate.
760 if( lop == Op_SubI && mask == 1 && load->in(1) &&
761 phase->type(load->in(1)) == TypeInt::ZERO )
762 return new AndINode( load->in(2), in(2) );
763
764 return MulNode::Ideal(phase, can_reshape);
765 }
766
767 //=============================================================================
768 //------------------------------mul_ring---------------------------------------
769 // Supplied function returns the product of the inputs IN THE CURRENT RING.
770 // For the logical operations the ring's MUL is really a logical AND function.
771 // This also type-checks the inputs for sanity. Guaranteed never to
772 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
773 const Type* AndLNode::mul_ring(const Type* t1, const Type* t2) const {
774 return RangeInference::infer_and(t1->is_long(), t2->is_long());
775 }
776
777 const Type* AndLNode::Value(PhaseGVN* phase) const {
778 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
779 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
780 return TypeLong::ZERO;
781 }
782
783 return MulNode::Value(phase);
784 }
785
786 //------------------------------Identity---------------------------------------
787 // Masking off the high bits of an unsigned load is not required
788 Node* AndLNode::Identity(PhaseGVN* phase) {
789
790 // x & x => x
791 if (in(1) == in(2)) {
792 return in(1);
793 }
794
795 Node *usr = in(1);
796 const TypeLong *t2 = phase->type( in(2) )->isa_long();
797 if( t2 && t2->is_con() ) {
798 jlong con = t2->get_con();
799 // Masking off high bits which are always zero is useless.
800 const TypeLong* t1 = phase->type( in(1) )->isa_long();
801 if (t1 != nullptr && t1->_lo >= 0) {
802 int bit_count = log2i_graceful(t1->_hi) + 1;
803 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
804 if ((t1_support & con) == t1_support)
805 return usr;
806 }
807 uint lop = usr->Opcode();
808 // Masking off the high bits of a unsigned-shift-right is not
809 // needed either.
810 if( lop == Op_URShiftL ) {
811 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
812 if( t12 && t12->is_con() ) { // Shift is by a constant
813 int shift = t12->get_con();
814 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
815 jlong mask = max_julong >> shift;
816 if( (mask&con) == mask ) // If AND is useless, skip it
817 return usr;
818 }
819 }
820 }
821 return MulNode::Identity(phase);
822 }
823
824 //------------------------------Ideal------------------------------------------
825 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
826 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
827 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
828 if (progress != nullptr) {
829 return progress;
830 }
831
832 // Convert "(~a) & (~b)" into "~(a | b)"
833 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
834 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
835 Node* tn = phase->transform(or_a_b);
836 return AddNode::make_not(phase, tn, T_LONG);
837 }
838
839 // Special case constant AND mask
840 const TypeLong *t2 = phase->type( in(2) )->isa_long();
841 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
842 const jlong mask = t2->get_con();
843
844 Node* in1 = in(1);
845 int op = in1->Opcode();
846
847 // Are we masking a long that was converted from an int with a mask
848 // that fits in 32-bits? Commute them and use an AndINode. Don't
849 // convert masks which would cause a sign extension of the integer
850 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
851 // would be optimized away later in Identity.
852 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
853 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
854 andi = phase->transform(andi);
855 return new ConvI2LNode(andi);
856 }
857
858 // Masking off sign bits? Dont make them!
859 if (op == Op_RShiftL) {
860 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
861 if( t12 && t12->is_con() ) { // Shift is by a constant
862 int shift = t12->get_con();
863 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
864 if (shift != 0) {
865 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
866 // If the AND'ing of the 2 masks has no bits, then only original shifted
867 // bits survive. NO sign-extension bits survive the maskings.
868 if( (sign_bits_mask & mask) == 0 ) {
869 // Use zero-fill shift instead
870 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
871 return new AndLNode(zshift, in(2));
872 }
873 }
874 }
875 }
876
877 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
878 Node* load1 = in(1);
879 Node* load2 = nullptr;
880 if (load1->is_Phi() && phase->type(load1)->isa_long()) {
881 load1 = in(1)->in(1);
882 load2 = in(1)->in(2);
883 }
884 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
885 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
886 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
887 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
888 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
889 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
890 if (mask == markWord::inline_type_pattern) {
891 if (adr_t1->is_inlinetypeptr()) {
892 set_req_X(1, in(2), phase);
893 return this;
894 } else if (!adr_t1->can_be_inline_type()) {
895 set_req_X(1, phase->longcon(0), phase);
896 return this;
897 }
898 } else if (mask == markWord::null_free_array_bit_in_place) {
899 if (adr_t1->is_null_free()) {
900 set_req_X(1, in(2), phase);
901 return this;
902 } else if (adr_t1->is_not_null_free()) {
903 set_req_X(1, phase->longcon(0), phase);
904 return this;
905 }
906 } else if (mask == markWord::flat_array_bit_in_place) {
907 if (adr_t1->is_flat()) {
908 set_req_X(1, in(2), phase);
909 return this;
910 } else if (adr_t1->is_not_flat()) {
911 set_req_X(1, phase->longcon(0), phase);
912 return this;
913 }
914 }
915 }
916 }
917
918 return MulNode::Ideal(phase, can_reshape);
919 }
920
921 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
922 switch (bt) {
923 case T_INT:
924 return new LShiftINode(in1, in2);
925 case T_LONG:
926 return new LShiftLNode(in1, in2);
927 default:
928 fatal("Not implemented for %s", type2name(bt));
929 }
930 return nullptr;
931 }
932
933 // Returns whether the shift amount is constant. If so, sets count.
934 static bool const_shift_count(PhaseGVN* phase, const Node* shift_node, int* count) {
935 const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
936 if (tcount != nullptr && tcount->is_con()) {
937 *count = tcount->get_con();
938 return true;
939 }
940 return false;
941 }
942
943 // Returns whether the shift amount is constant. If so, sets real_shift and masked_shift.
944 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, int& real_shift, uint& masked_shift) {
945 if (const_shift_count(phase, shift_node, &real_shift)) {
946 masked_shift = real_shift & (nBits - 1);
947 return true;
948 }
949 return false;
950 }
951
952 // Convenience for when we don't care about the real amount
953 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, uint& masked_shift) {
954 int real_shift;
955 return mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift);
956 }
957
958 // Use this in ::Ideal only with shiftNode == this!
959 // Sets masked_shift to the masked shift amount if constant or 0 if not constant.
960 // Returns shift_node if the shift amount input node was modified, nullptr otherwise.
961 static Node* mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits, uint& masked_shift) {
962 int real_shift;
963 if (mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift)) {
964 if (masked_shift == 0) {
965 // Let Identity() handle 0 shift count.
966 return nullptr;
967 }
968
969 if (real_shift != (int)masked_shift) {
970 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
971
972 // We need to notify the caller that the graph was reshaped, as Ideal needs
973 // to return the root of the reshaped graph if any change was made.
974 return shift_node;
975 }
976 } else {
977 // Not a shift by a constant.
978 masked_shift = 0;
979 }
980 return nullptr;
981 }
982
983 // Called with
984 // outer_shift = (_ << rhs_outer)
985 // We are looking for the pattern:
986 // outer_shift = ((X << rhs_inner) << rhs_outer)
987 // where rhs_outer and rhs_inner are constant
988 // we denote inner_shift the nested expression (X << rhs_inner)
989 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
990 // where nbits is the number of bits of the shifts
991 //
992 // There are 2 cases:
993 // if con_outer + con_inner >= nbits => 0
994 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
995 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
996 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
997 const Node* inner_shift = outer_shift->in(1);
998 if (inner_shift->Opcode() != Op_LShift(bt)) {
999 return nullptr;
1000 }
1001
1002 uint nbits = bits_per_java_integer(bt);
1003 uint con_inner;
1004 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
1005 return nullptr;
1006 }
1007
1008 if (con_inner == 0) {
1009 // We let the Identity() of the inner shift do its job.
1010 return nullptr;
1011 }
1012
1013 if (con_outer + con_inner >= nbits) {
1014 // While it might be tempting to use
1015 // phase->zerocon(bt);
1016 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
1017 // to return a new node, this or nullptr, but not an old (cached) node.
1018 return ConNode::make(TypeInteger::zero(bt));
1019 }
1020
1021 // con0 + con1 < nbits ==> actual shift happens now
1022 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
1023 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
1024 }
1025
1026 //------------------------------Identity---------------------------------------
1027 Node* LShiftINode::Identity(PhaseGVN* phase) {
1028 return IdentityIL(phase, T_INT);
1029 }
1030
1031 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1032 uint con;
1033 Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), con);
1034 if (con == 0) {
1035 return nullptr;
1036 }
1037
1038 // If the right input is a constant, and the left input is an add of a
1039 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1040 Node* add1 = in(1);
1041 int add1_op = add1->Opcode();
1042 if (add1_op == Op_Add(bt)) { // Left input is an add?
1043 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
1044
1045 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1046 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1047 if (bt != T_INT || con < 16) {
1048 // Left input is an add of the same number?
1049 if (con != (bits_per_java_integer(bt) - 1) && add1->in(1) == add1->in(2)) {
1050 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1051 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
1052 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1053 // or c0 != 63 (for LShiftL) because:
1054 // (x + x) << 63 = 2x << 63, while
1055 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1056 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1057 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1058 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1059 }
1060
1061 // Left input is an add of a constant?
1062 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1063 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1064 // Compute X << con0
1065 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1066 // Compute X<<con0 + (con1<<con0)
1067 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1068 }
1069 }
1070 }
1071 // Check for "(con0 - X) << con1"
1072 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1073 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1074 if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) { // Left input is a sub?
1075 // Left input is a sub from a constant?
1076 const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
1077 if (t11 != nullptr && t11->is_con()) {
1078 // Compute X << con0
1079 Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
1080 // Compute (con1<<con0) - (X<<con0)
1081 return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
1082 }
1083 }
1084
1085 // Check for "(x >> C1) << C2"
1086 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1087 int add1Con = 0;
1088 const_shift_count(phase, add1, &add1Con);
1089
1090 // Special case C1 == C2, which just masks off low bits
1091 if (add1Con > 0 && con == (uint)add1Con) {
1092 // Convert to "(x & -(1 << C2))"
1093 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1094 } else {
1095 // Wait until the right shift has been sharpened to the correct count
1096 if (add1Con > 0 && (uint)add1Con < bits_per_java_integer(bt)) {
1097 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1098 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1099 if (phase->is_IterGVN()) {
1100 if (con > (uint)add1Con) {
1101 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1102 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1103 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1104 } else {
1105 assert(con < (uint)add1Con, "must be (%d < %d)", con, add1Con);
1106 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1107
1108 // Handle logical and arithmetic shifts
1109 Node* rshift;
1110 if (add1_op == Op_RShift(bt)) {
1111 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1112 } else {
1113 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1114 }
1115
1116 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1117 }
1118 } else {
1119 phase->record_for_igvn(this);
1120 }
1121 }
1122 }
1123 }
1124
1125 // Check for "((x >> C1) & Y) << C2"
1126 if (add1_op == Op_And(bt)) {
1127 Node* add2 = add1->in(1);
1128 int add2_op = add2->Opcode();
1129 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1130 // Special case C1 == C2, which just masks off low bits
1131 if (add2->in(2) == in(2)) {
1132 // Convert to "(x & (Y << C2))"
1133 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1134 return MulNode::make_and(add2->in(1), y_sh, bt);
1135 }
1136
1137 int add2Con = 0;
1138 const_shift_count(phase, add2, &add2Con);
1139 if (add2Con > 0 && (uint)add2Con < bits_per_java_integer(bt)) {
1140 if (phase->is_IterGVN()) {
1141 // Convert to "((x >> C1) << C2) & (Y << C2)"
1142
1143 // Make "(x >> C1) << C2", which will get folded away by the rule above
1144 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1145 // Make "Y << C2", which will simplify when Y is a constant
1146 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1147
1148 return MulNode::make_and(x_sh, y_sh, bt);
1149 } else {
1150 phase->record_for_igvn(this);
1151 }
1152 }
1153 }
1154 }
1155
1156 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1157 // before shifting them away.
1158 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1159 assert(bt != T_INT || bits_mask == right_n_bits(bits_per_java_integer(bt)-con), "inconsistent");
1160 if (add1_op == Op_And(bt) &&
1161 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1162 return LShiftNode::make(add1->in(1), in(2), bt);
1163 }
1164
1165 // Collapse nested left-shifts with constant rhs:
1166 // (X << con1) << con2 ==> X << (con1 + con2)
1167 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1168 if (doubleShift != nullptr) {
1169 return doubleShift;
1170 }
1171
1172 return progress;
1173 }
1174
1175 //------------------------------Ideal------------------------------------------
1176 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1177 return IdealIL(phase, can_reshape, T_INT);
1178 }
1179
1180 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1181 const Type* t1 = phase->type(in(1));
1182 const Type* t2 = phase->type(in(2));
1183 // Either input is TOP ==> the result is TOP
1184 if (t1 == Type::TOP) {
1185 return Type::TOP;
1186 }
1187 if (t2 == Type::TOP) {
1188 return Type::TOP;
1189 }
1190
1191 // Left input is ZERO ==> the result is ZERO.
1192 if (t1 == TypeInteger::zero(bt)) {
1193 return TypeInteger::zero(bt);
1194 }
1195 // Shift by zero does nothing
1196 if (t2 == TypeInt::ZERO) {
1197 return t1;
1198 }
1199
1200 // Either input is BOTTOM ==> the result is BOTTOM
1201 if ((t1 == TypeInteger::bottom(bt)) || (t2 == TypeInt::INT) ||
1202 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM)) {
1203 return TypeInteger::bottom(bt);
1204 }
1205
1206 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1207 const TypeInt* r2 = t2->is_int(); // Handy access
1208
1209 if (!r2->is_con()) {
1210 return TypeInteger::bottom(bt);
1211 }
1212
1213 uint shift = r2->get_con();
1214 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1215 // Shift by a multiple of 32/64 does nothing:
1216 if (shift == 0) {
1217 return t1;
1218 }
1219
1220 // If the shift is a constant, shift the bounds of the type,
1221 // unless this could lead to an overflow.
1222 if (!r1->is_con()) {
1223 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1224 #ifdef ASSERT
1225 if (bt == T_INT) {
1226 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1227 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1228 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1229 }
1230 #endif
1231 if (java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo &&
1232 java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) {
1233 // No overflow. The range shifts up cleanly.
1234 return TypeInteger::make(java_shift_left(lo, shift, bt),
1235 java_shift_left(hi, shift, bt),
1236 MAX2(r1->_widen, r2->_widen), bt);
1237 }
1238 return TypeInteger::bottom(bt);
1239 }
1240
1241 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1242 }
1243
1244 //------------------------------Value------------------------------------------
1245 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1246 return ValueIL(phase, T_INT);
1247 }
1248
1249 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1250 int count = 0;
1251 if (const_shift_count(phase, this, &count) && (count & (bits_per_java_integer(bt) - 1)) == 0) {
1252 // Shift by a multiple of 32/64 does nothing
1253 return in(1);
1254 }
1255 return this;
1256 }
1257
1258 //=============================================================================
1259 //------------------------------Identity---------------------------------------
1260 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1261 return IdentityIL(phase, T_LONG);
1262 }
1263
1264 //------------------------------Ideal------------------------------------------
1265 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1266 return IdealIL(phase, can_reshape, T_LONG);
1267 }
1268
1269 //------------------------------Value------------------------------------------
1270 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1271 return ValueIL(phase, T_LONG);
1272 }
1273
1274 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1275 switch (bt) {
1276 case T_INT:
1277 return new RShiftINode(in1, in2);
1278 case T_LONG:
1279 return new RShiftLNode(in1, in2);
1280 default:
1281 fatal("Not implemented for %s", type2name(bt));
1282 }
1283 return nullptr;
1284 }
1285
1286
1287 //=============================================================================
1288 //------------------------------Identity---------------------------------------
1289 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1290 int count = 0;
1291 if (const_shift_count(phase, this, &count)) {
1292 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1293 // Shift by a multiple of 32/64 does nothing
1294 return in(1);
1295 }
1296 // Check for useless sign-masking
1297 int lshift_count = 0;
1298 if (in(1)->Opcode() == Op_LShift(bt) &&
1299 in(1)->req() == 3 &&
1300 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1301 // negative constant (e.g. -1 vs 31)
1302 const_shift_count(phase, in(1), &lshift_count)) {
1303 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1304 lshift_count &= bits_per_java_integer(bt) - 1;
1305 if (count == lshift_count) {
1306 // Compute masks for which this shifting doesn't change
1307 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1308 jlong hi = ~lo; // 00007FFF
1309 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1310 if (t11 == nullptr) {
1311 return this;
1312 }
1313 // Does actual value fit inside of mask?
1314 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1315 return in(1)->in(1); // Then shifting is a nop
1316 }
1317 }
1318 }
1319 }
1320 return this;
1321 }
1322
1323 Node* RShiftINode::Identity(PhaseGVN* phase) {
1324 return IdentityIL(phase, T_INT);
1325 }
1326
1327 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1328 // Inputs may be TOP if they are dead.
1329 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1330 if (t1 == nullptr) {
1331 return NodeSentinel; // Left input is an integer
1332 }
1333
1334 uint shift;
1335 Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), shift);
1336 if (shift == 0) {
1337 return NodeSentinel;
1338 }
1339
1340 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1341 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1342 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1343 const Node* and_node = in(1);
1344 if (and_node->Opcode() != Op_And(bt)) {
1345 return progress;
1346 }
1347 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1348 if (mask_t != nullptr && mask_t->is_con()) {
1349 jlong maskbits = mask_t->get_con_as_long(bt);
1350 // Convert to "(x >> shift) & (mask >> shift)"
1351 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1352 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1353 }
1354
1355 return progress;
1356 }
1357
1358 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1359 Node* progress = IdealIL(phase, can_reshape, T_INT);
1360 if (progress == NodeSentinel) {
1361 return nullptr;
1362 }
1363 if (progress != nullptr) {
1364 return progress;
1365 }
1366 uint shift;
1367 progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, shift);
1368 assert(shift != 0, "handled by IdealIL");
1369
1370 // Check for "(short[i] <<16)>>16" which simply sign-extends
1371 const Node *shl = in(1);
1372 if (shl->Opcode() != Op_LShiftI) {
1373 return progress;
1374 }
1375
1376 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1377 if (left_shift_t == nullptr) {
1378 return progress;
1379 }
1380 if (shift == 16 && left_shift_t->is_con(16)) {
1381 Node *ld = shl->in(1);
1382 if (ld->Opcode() == Op_LoadS) {
1383 // Sign extension is just useless here. Return a RShiftI of zero instead
1384 // returning 'ld' directly. We cannot return an old Node directly as
1385 // that is the job of 'Identity' calls and Identity calls only work on
1386 // direct inputs ('ld' is an extra Node removed from 'this'). The
1387 // combined optimization requires Identity only return direct inputs.
1388 set_req_X(1, ld, phase);
1389 set_req_X(2, phase->intcon(0), phase);
1390 return this;
1391 }
1392 else if (can_reshape &&
1393 ld->Opcode() == Op_LoadUS &&
1394 ld->outcnt() == 1 && ld->unique_out() == shl)
1395 // Replace zero-extension-load with sign-extension-load
1396 return ld->as_Load()->convert_to_signed_load(*phase);
1397 }
1398
1399 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1400 if (shift == 24 && left_shift_t->is_con(24)) {
1401 Node *ld = shl->in(1);
1402 if (ld->Opcode() == Op_LoadB) {
1403 // Sign extension is just useless here
1404 set_req_X(1, ld, phase);
1405 set_req_X(2, phase->intcon(0), phase);
1406 return this;
1407 }
1408 }
1409
1410 return progress;
1411 }
1412
1413 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1414 const Type* t1 = phase->type(in(1));
1415 const Type* t2 = phase->type(in(2));
1416 // Either input is TOP ==> the result is TOP
1417 if (t1 == Type::TOP) {
1418 return Type::TOP;
1419 }
1420 if (t2 == Type::TOP) {
1421 return Type::TOP;
1422 }
1423
1424 // Left input is ZERO ==> the result is ZERO.
1425 if (t1 == TypeInteger::zero(bt)) {
1426 return TypeInteger::zero(bt);
1427 }
1428 // Shift by zero does nothing
1429 if (t2 == TypeInt::ZERO) {
1430 return t1;
1431 }
1432
1433 // Either input is BOTTOM ==> the result is BOTTOM
1434 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1435 return TypeInteger::bottom(bt);
1436 }
1437
1438 const TypeInteger* r1 = t1->isa_integer(bt);
1439 const TypeInt* r2 = t2->isa_int();
1440
1441 // If the shift is a constant, just shift the bounds of the type.
1442 // For example, if the shift is 31/63, we just propagate sign bits.
1443 if (!r1->is_con() && r2->is_con()) {
1444 uint shift = r2->get_con();
1445 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1446 // Shift by a multiple of 32/64 does nothing:
1447 if (shift == 0) {
1448 return t1;
1449 }
1450 // Calculate reasonably aggressive bounds for the result.
1451 // This is necessary if we are to correctly type things
1452 // like (x<<24>>24) == ((byte)x).
1453 jlong lo = r1->lo_as_long() >> (jint)shift;
1454 jlong hi = r1->hi_as_long() >> (jint)shift;
1455 assert(lo <= hi, "must have valid bounds");
1456 #ifdef ASSERT
1457 if (bt == T_INT) {
1458 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1459 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1460 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1461 }
1462 #endif
1463 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1464 #ifdef ASSERT
1465 // Make sure we get the sign-capture idiom correct.
1466 if (shift == bits_per_java_integer(bt) - 1) {
1467 if (r1->lo_as_long() >= 0) {
1468 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1469 }
1470 if (r1->hi_as_long() < 0) {
1471 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1472 }
1473 }
1474 #endif
1475 return ti;
1476 }
1477
1478 if (!r1->is_con() || !r2->is_con()) {
1479 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1480 if (r1->lo_as_long() >= 0) {
1481 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1482 }
1483
1484 // Conversely, if the left input is negative then the result must be negative.
1485 if (r1->hi_as_long() <= -1) {
1486 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1487 }
1488
1489 return TypeInteger::bottom(bt);
1490 }
1491
1492 // Signed shift right
1493 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1494 }
1495
1496 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1497 return ValueIL(phase, T_INT);
1498 }
1499
1500 //=============================================================================
1501 //------------------------------Identity---------------------------------------
1502 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1503 return IdentityIL(phase, T_LONG);
1504 }
1505
1506 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1507 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1508 if (progress == NodeSentinel) {
1509 return nullptr;
1510 }
1511 return progress;
1512 }
1513
1514 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1515 return ValueIL(phase, T_LONG);
1516 }
1517
1518 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1519 switch (bt) {
1520 case T_INT:
1521 return new URShiftINode(in1, in2);
1522 case T_LONG:
1523 return new URShiftLNode(in1, in2);
1524 default:
1525 fatal("Not implemented for %s", type2name(bt));
1526 }
1527 return nullptr;
1528 }
1529
1530 //=============================================================================
1531 //------------------------------Identity---------------------------------------
1532 Node* URShiftINode::Identity(PhaseGVN* phase) {
1533 int count = 0;
1534 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1535 // Shift by a multiple of 32 does nothing
1536 return in(1);
1537 }
1538
1539 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1540 // Happens during new-array length computation.
1541 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1542 Node *add = in(1);
1543 if (add->Opcode() == Op_AddI) {
1544 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1545 if (t2 && t2->is_con(wordSize - 1) &&
1546 add->in(1)->Opcode() == Op_LShiftI) {
1547 // Check that shift_counts are LogBytesPerWord.
1548 Node *lshift_count = add->in(1)->in(2);
1549 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1550 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1551 t_lshift_count == phase->type(in(2))) {
1552 Node *x = add->in(1)->in(1);
1553 const TypeInt *t_x = phase->type(x)->isa_int();
1554 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1555 return x;
1556 }
1557 }
1558 }
1559 }
1560
1561 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1562 }
1563
1564 //------------------------------Ideal------------------------------------------
1565 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1566 uint con;
1567 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, con);
1568 if (con == 0) {
1569 return nullptr;
1570 }
1571
1572 // We'll be wanting the right-shift amount as a mask of that many bits
1573 const int mask = right_n_bits(BitsPerJavaInteger - con);
1574
1575 int in1_op = in(1)->Opcode();
1576
1577 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1578 if( in1_op == Op_URShiftI ) {
1579 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1580 if( t12 && t12->is_con() ) { // Right input is a constant
1581 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1582 const int con2 = t12->get_con() & 31; // Shift count is always masked
1583 const int con3 = con+con2;
1584 if( con3 < 32 ) // Only merge shifts if total is < 32
1585 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1586 }
1587 }
1588
1589 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1590 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1591 // If Q is "X << z" the rounding is useless. Look for patterns like
1592 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1593 Node *add = in(1);
1594 if (in1_op == Op_AddI) {
1595 Node *lshl = add->in(1);
1596 Node *y = add->in(2);
1597 if (lshl->Opcode() != Op_LShiftI) {
1598 lshl = add->in(2);
1599 y = add->in(1);
1600 }
1601 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1602 // negative constant (e.g. -1 vs 31)
1603 int lshl_con = 0;
1604 if (lshl->Opcode() == Op_LShiftI &&
1605 const_shift_count(phase, lshl, &lshl_con) &&
1606 (lshl_con & (BitsPerJavaInteger - 1)) == con) {
1607 Node *y_z = phase->transform(new URShiftINode(y, in(2)));
1608 Node *sum = phase->transform(new AddINode(lshl->in(1), y_z));
1609 return new AndINode(sum, phase->intcon(mask));
1610 }
1611 }
1612
1613 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1614 // This shortens the mask. Also, if we are extracting a high byte and
1615 // storing it to a buffer, the mask will be removed completely.
1616 Node *andi = in(1);
1617 if( in1_op == Op_AndI ) {
1618 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1619 if( t3 && t3->is_con() ) { // Right input is a constant
1620 jint mask2 = t3->get_con();
1621 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1622 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1623 return new AndINode(newshr, phase->intcon(mask2));
1624 // The negative values are easier to materialize than positive ones.
1625 // A typical case from address arithmetic is ((x & ~15) >> 4).
1626 // It's better to change that to ((x >> 4) & ~0) versus
1627 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1628 }
1629 }
1630
1631 // Check for "(X << z ) >>> z" which simply zero-extends
1632 Node *shl = in(1);
1633 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1634 // negative constant (e.g. -1 vs 31)
1635 int shl_con = 0;
1636 if (in1_op == Op_LShiftI &&
1637 const_shift_count(phase, shl, &shl_con) &&
1638 (shl_con & (BitsPerJavaInteger - 1)) == con)
1639 return new AndINode(shl->in(1), phase->intcon(mask));
1640
1641 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1642 const TypeInt* t2 = phase->type(in(2))->isa_int();
1643 Node *shr = in(1);
1644 if ( in1_op == Op_RShiftI ) {
1645 Node *in11 = shr->in(1);
1646 Node *in12 = shr->in(2);
1647 const TypeInt *t11 = phase->type(in11)->isa_int();
1648 const TypeInt *t12 = phase->type(in12)->isa_int();
1649 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1650 return new URShiftINode(in11, phase->intcon(31));
1651 }
1652 }
1653
1654 return progress;
1655 }
1656
1657 //------------------------------Value------------------------------------------
1658 // A URShiftINode shifts its input2 right by input1 amount.
1659 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1660 // (This is a near clone of RShiftINode::Value.)
1661 const Type *t1 = phase->type( in(1) );
1662 const Type *t2 = phase->type( in(2) );
1663 // Either input is TOP ==> the result is TOP
1664 if( t1 == Type::TOP ) return Type::TOP;
1665 if( t2 == Type::TOP ) return Type::TOP;
1666
1667 // Left input is ZERO ==> the result is ZERO.
1668 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1669 // Shift by zero does nothing
1670 if( t2 == TypeInt::ZERO ) return t1;
1671
1672 // Either input is BOTTOM ==> the result is BOTTOM
1673 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1674 return TypeInt::INT;
1675
1676 if (t2 == TypeInt::INT)
1677 return TypeInt::INT;
1678
1679 const TypeInt *r1 = t1->is_int(); // Handy access
1680 const TypeInt *r2 = t2->is_int(); // Handy access
1681
1682 if (r2->is_con()) {
1683 uint shift = r2->get_con();
1684 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1685 // Shift by a multiple of 32 does nothing:
1686 if (shift == 0) return t1;
1687 // Calculate reasonably aggressive bounds for the result.
1688 jint lo = (juint)r1->_lo >> (juint)shift;
1689 jint hi = (juint)r1->_hi >> (juint)shift;
1690 if (r1->_hi >= 0 && r1->_lo < 0) {
1691 // If the type has both negative and positive values,
1692 // there are two separate sub-domains to worry about:
1693 // The positive half and the negative half.
1694 jint neg_lo = lo;
1695 jint neg_hi = (juint)-1 >> (juint)shift;
1696 jint pos_lo = (juint) 0 >> (juint)shift;
1697 jint pos_hi = hi;
1698 lo = MIN2(neg_lo, pos_lo); // == 0
1699 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1700 }
1701 assert(lo <= hi, "must have valid bounds");
1702 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1703 #ifdef ASSERT
1704 // Make sure we get the sign-capture idiom correct.
1705 if (shift == BitsPerJavaInteger-1) {
1706 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1707 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1708 }
1709 #endif
1710 return ti;
1711 }
1712
1713 //
1714 // Do not support shifted oops in info for GC
1715 //
1716 // else if( t1->base() == Type::InstPtr ) {
1717 //
1718 // const TypeInstPtr *o = t1->is_instptr();
1719 // if( t1->singleton() )
1720 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1721 // }
1722 // else if( t1->base() == Type::KlassPtr ) {
1723 // const TypeKlassPtr *o = t1->is_klassptr();
1724 // if( t1->singleton() )
1725 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1726 // }
1727
1728 return TypeInt::INT;
1729 }
1730
1731 //=============================================================================
1732 //------------------------------Identity---------------------------------------
1733 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1734 int count = 0;
1735 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1736 // Shift by a multiple of 64 does nothing
1737 return in(1);
1738 }
1739 return this;
1740 }
1741
1742 //------------------------------Ideal------------------------------------------
1743 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1744 uint con;
1745 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong, con);
1746 if (con == 0) {
1747 return nullptr;
1748 }
1749
1750 // We'll be wanting the right-shift amount as a mask of that many bits
1751 const jlong mask = jlong(max_julong >> con);
1752
1753 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1754 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1755 // If Q is "X << z" the rounding is useless. Look for patterns like
1756 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1757 Node *add = in(1);
1758 const TypeInt *t2 = phase->type(in(2))->isa_int();
1759 if (add->Opcode() == Op_AddL) {
1760 Node *lshl = add->in(1);
1761 Node *y = add->in(2);
1762 if (lshl->Opcode() != Op_LShiftL) {
1763 lshl = add->in(2);
1764 y = add->in(1);
1765 }
1766 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1767 // negative constant (e.g. -1 vs 63)
1768 int lshl_con = 0;
1769 if (lshl->Opcode() == Op_LShiftL &&
1770 const_shift_count(phase, lshl, &lshl_con) &&
1771 (lshl_con & (BitsPerJavaLong - 1)) == con) {
1772 Node* y_z = phase->transform(new URShiftLNode(y, in(2)));
1773 Node* sum = phase->transform(new AddLNode(lshl->in(1), y_z));
1774 return new AndLNode(sum, phase->longcon(mask));
1775 }
1776 }
1777
1778 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1779 // This shortens the mask. Also, if we are extracting a high byte and
1780 // storing it to a buffer, the mask will be removed completely.
1781 Node *andi = in(1);
1782 if( andi->Opcode() == Op_AndL ) {
1783 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1784 if( t3 && t3->is_con() ) { // Right input is a constant
1785 jlong mask2 = t3->get_con();
1786 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1787 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1788 return new AndLNode(newshr, phase->longcon(mask2));
1789 }
1790 }
1791
1792 // Check for "(X << z ) >>> z" which simply zero-extends
1793 Node *shl = in(1);
1794 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1795 // negative constant (e.g. -1 vs 63)
1796 int shl_con = 0;
1797 if (shl->Opcode() == Op_LShiftL &&
1798 const_shift_count(phase, shl, &shl_con) &&
1799 (shl_con & (BitsPerJavaLong - 1)) == con) {
1800 return new AndLNode(shl->in(1), phase->longcon(mask));
1801 }
1802
1803 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1804 Node *shr = in(1);
1805 if ( shr->Opcode() == Op_RShiftL ) {
1806 Node *in11 = shr->in(1);
1807 Node *in12 = shr->in(2);
1808 const TypeLong *t11 = phase->type(in11)->isa_long();
1809 const TypeInt *t12 = phase->type(in12)->isa_int();
1810 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1811 return new URShiftLNode(in11, phase->intcon(63));
1812 }
1813 }
1814
1815 return progress;
1816 }
1817
1818 //------------------------------Value------------------------------------------
1819 // A URShiftINode shifts its input2 right by input1 amount.
1820 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1821 // (This is a near clone of RShiftLNode::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 == TypeLong::ZERO ) return TypeLong::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 TypeLong::LONG;
1836
1837 if (t2 == TypeInt::INT)
1838 return TypeLong::LONG;
1839
1840 const TypeLong *r1 = t1->is_long(); // 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 &= BitsPerJavaLong - 1; // semantics of Java shifts
1846 // Shift by a multiple of 64 does nothing:
1847 if (shift == 0) return t1;
1848 // Calculate reasonably aggressive bounds for the result.
1849 jlong lo = (julong)r1->_lo >> (juint)shift;
1850 jlong hi = (julong)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 jlong neg_lo = lo;
1856 jlong neg_hi = (julong)-1 >> (juint)shift;
1857 jlong pos_lo = (julong) 0 >> (juint)shift;
1858 jlong pos_hi = hi;
1859 //lo = MIN2(neg_lo, pos_lo); // == 0
1860 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1861 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1862 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1863 }
1864 assert(lo <= hi, "must have valid bounds");
1865 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1866 #ifdef ASSERT
1867 // Make sure we get the sign-capture idiom correct.
1868 if (shift == BitsPerJavaLong - 1) {
1869 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1870 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1871 }
1872 #endif
1873 return tl;
1874 }
1875
1876 return TypeLong::LONG; // Give up
1877 }
1878
1879 //=============================================================================
1880 //------------------------------Ideal------------------------------------------
1881 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1882 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1883 // This reduces the number of rules in the matcher, as we only need to check
1884 // for negations on the second argument, and not the symmetric case where
1885 // the first argument is negated.
1886 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1887 swap_edges(1, 2);
1888 return this;
1889 }
1890 return nullptr;
1891 }
1892
1893 //=============================================================================
1894 //------------------------------Value------------------------------------------
1895 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1896 const Type *t1 = phase->type(in(1));
1897 if (t1 == Type::TOP) return Type::TOP;
1898 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1899 const Type *t2 = phase->type(in(2));
1900 if (t2 == Type::TOP) return Type::TOP;
1901 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1902 const Type *t3 = phase->type(in(3));
1903 if (t3 == Type::TOP) return Type::TOP;
1904 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1905 #ifndef __STDC_IEC_559__
1906 return Type::DOUBLE;
1907 #else
1908 double d1 = t1->getd();
1909 double d2 = t2->getd();
1910 double d3 = t3->getd();
1911 return TypeD::make(fma(d1, d2, d3));
1912 #endif
1913 }
1914
1915 //=============================================================================
1916 //------------------------------Value------------------------------------------
1917 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1918 const Type *t1 = phase->type(in(1));
1919 if (t1 == Type::TOP) return Type::TOP;
1920 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1921 const Type *t2 = phase->type(in(2));
1922 if (t2 == Type::TOP) return Type::TOP;
1923 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1924 const Type *t3 = phase->type(in(3));
1925 if (t3 == Type::TOP) return Type::TOP;
1926 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1927 #ifndef __STDC_IEC_559__
1928 return Type::FLOAT;
1929 #else
1930 float f1 = t1->getf();
1931 float f2 = t2->getf();
1932 float f3 = t3->getf();
1933 return TypeF::make(fma(f1, f2, f3));
1934 #endif
1935 }
1936
1937 //=============================================================================
1938 //------------------------------Value------------------------------------------
1939 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
1940 const Type* t1 = phase->type(in(1));
1941 if (t1 == Type::TOP) { return Type::TOP; }
1942 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1943 const Type* t2 = phase->type(in(2));
1944 if (t2 == Type::TOP) { return Type::TOP; }
1945 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1946 const Type* t3 = phase->type(in(3));
1947 if (t3 == Type::TOP) { return Type::TOP; }
1948 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1949 #ifndef __STDC_IEC_559__
1950 return Type::HALF_FLOAT;
1951 #else
1952 float f1 = t1->getf();
1953 float f2 = t2->getf();
1954 float f3 = t3->getf();
1955 return TypeH::make(fma(f1, f2, f3));
1956 #endif
1957 }
1958
1959 //=============================================================================
1960 //------------------------------hash-------------------------------------------
1961 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1962 // The hash function must return the same value when edge swapping is performed.
1963 uint MulAddS2INode::hash() const {
1964 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1965 }
1966
1967 //------------------------------Rotate Operations ------------------------------
1968
1969 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1970 const Type* t1 = phase->type(in(1));
1971 if (t1 == Type::TOP) {
1972 return this;
1973 }
1974 int count = 0;
1975 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1976 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1977 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1978 // Rotate by a multiple of 32/64 does nothing
1979 return in(1);
1980 }
1981 return this;
1982 }
1983
1984 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1985 const Type* t1 = phase->type(in(1));
1986 const Type* t2 = phase->type(in(2));
1987 // Either input is TOP ==> the result is TOP
1988 if (t1 == Type::TOP || t2 == Type::TOP) {
1989 return Type::TOP;
1990 }
1991
1992 if (t1->isa_int()) {
1993 const TypeInt* r1 = t1->is_int();
1994 const TypeInt* r2 = t2->is_int();
1995
1996 // Left input is ZERO ==> the result is ZERO.
1997 if (r1 == TypeInt::ZERO) {
1998 return TypeInt::ZERO;
1999 }
2000 // Rotate by zero does nothing
2001 if (r2 == TypeInt::ZERO) {
2002 return r1;
2003 }
2004 if (r1->is_con() && r2->is_con()) {
2005 juint r1_con = (juint)r1->get_con();
2006 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2007 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
2008 }
2009 return TypeInt::INT;
2010 } else {
2011 assert(t1->isa_long(), "Type must be a long");
2012 const TypeLong* r1 = t1->is_long();
2013 const TypeInt* r2 = t2->is_int();
2014
2015 // Left input is ZERO ==> the result is ZERO.
2016 if (r1 == TypeLong::ZERO) {
2017 return TypeLong::ZERO;
2018 }
2019 // Rotate by zero does nothing
2020 if (r2 == TypeInt::ZERO) {
2021 return r1;
2022 }
2023 if (r1->is_con() && r2->is_con()) {
2024 julong r1_con = (julong)r1->get_con();
2025 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2026 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
2027 }
2028 return TypeLong::LONG;
2029 }
2030 }
2031
2032 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2033 const Type* t1 = phase->type(in(1));
2034 const Type* t2 = phase->type(in(2));
2035 if (t2->isa_int() && t2->is_int()->is_con()) {
2036 if (t1->isa_int()) {
2037 int lshift = t2->is_int()->get_con() & 31;
2038 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2039 } else if (t1 != Type::TOP) {
2040 assert(t1->isa_long(), "Type must be a long");
2041 int lshift = t2->is_int()->get_con() & 63;
2042 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2043 }
2044 }
2045 return nullptr;
2046 }
2047
2048 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2049 const Type* t1 = phase->type(in(1));
2050 if (t1 == Type::TOP) {
2051 return this;
2052 }
2053 int count = 0;
2054 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2055 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2056 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2057 // Rotate by a multiple of 32/64 does nothing
2058 return in(1);
2059 }
2060 return this;
2061 }
2062
2063 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2064 const Type* t1 = phase->type(in(1));
2065 const Type* t2 = phase->type(in(2));
2066 // Either input is TOP ==> the result is TOP
2067 if (t1 == Type::TOP || t2 == Type::TOP) {
2068 return Type::TOP;
2069 }
2070
2071 if (t1->isa_int()) {
2072 const TypeInt* r1 = t1->is_int();
2073 const TypeInt* r2 = t2->is_int();
2074
2075 // Left input is ZERO ==> the result is ZERO.
2076 if (r1 == TypeInt::ZERO) {
2077 return TypeInt::ZERO;
2078 }
2079 // Rotate by zero does nothing
2080 if (r2 == TypeInt::ZERO) {
2081 return r1;
2082 }
2083 if (r1->is_con() && r2->is_con()) {
2084 juint r1_con = (juint)r1->get_con();
2085 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2086 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2087 }
2088 return TypeInt::INT;
2089 } else {
2090 assert(t1->isa_long(), "Type must be a long");
2091 const TypeLong* r1 = t1->is_long();
2092 const TypeInt* r2 = t2->is_int();
2093 // Left input is ZERO ==> the result is ZERO.
2094 if (r1 == TypeLong::ZERO) {
2095 return TypeLong::ZERO;
2096 }
2097 // Rotate by zero does nothing
2098 if (r2 == TypeInt::ZERO) {
2099 return r1;
2100 }
2101 if (r1->is_con() && r2->is_con()) {
2102 julong r1_con = (julong)r1->get_con();
2103 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2104 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2105 }
2106 return TypeLong::LONG;
2107 }
2108 }
2109
2110 //------------------------------ Sum & Mask ------------------------------
2111
2112 // Returns a lower bound on the number of trailing zeros in expr.
2113 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2114 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2115 if (type == nullptr) {
2116 return 0;
2117 }
2118
2119 expr = expr->uncast();
2120 type = phase->type(expr)->isa_integer(bt);
2121 if (type == nullptr) {
2122 return 0;
2123 }
2124
2125 if (type->is_con()) {
2126 jlong con = type->get_con_as_long(bt);
2127 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2128 }
2129
2130 if (expr->Opcode() == Op_ConvI2L) {
2131 expr = expr->in(1)->uncast();
2132 bt = T_INT;
2133 type = phase->type(expr)->isa_int();
2134 }
2135
2136 // Pattern: expr = (x << shift)
2137 if (expr->Opcode() == Op_LShift(bt)) {
2138 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2139 if (shift_t == nullptr || !shift_t->is_con()) {
2140 return 0;
2141 }
2142 // We need to truncate the shift, as it may not have been canonicalized yet.
2143 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2144 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2145 // (JLS: "Shift Operators")
2146 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2147 return shift_t->get_con() & shift_mask;
2148 }
2149
2150 return 0;
2151 }
2152
2153 // Checks whether expr is neutral additive element (zero) under mask,
2154 // i.e. whether an expression of the form:
2155 // (AndX (AddX (expr addend) mask)
2156 // (expr + addend) & mask
2157 // is equivalent to
2158 // (AndX addend mask)
2159 // addend & mask
2160 // for any addend.
2161 // (The X in AndX must be I or L, depending on bt).
2162 //
2163 // We check for the sufficient condition when the lowest set bit in expr is higher than
2164 // the highest set bit in mask, i.e.:
2165 // expr: eeeeee0000000000000
2166 // mask: 000000mmmmmmmmmmmmm
2167 // <--w bits--->
2168 // We do not test for other cases.
2169 //
2170 // Correctness:
2171 // Given "expr" with at least "w" trailing zeros,
2172 // let "mod = 2^w", "suffix_mask = mod - 1"
2173 //
2174 // Since "mask" only has bits set where "suffix_mask" does, we have:
2175 // mask = suffix_mask & mask (SUFFIX_MASK)
2176 //
2177 // And since expr only has bits set above w, and suffix_mask only below:
2178 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2179 //
2180 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2181 // a power of 2, and we are computing in a ring of powers of 2, we know that
2182 // (x + y) % mod = (x % mod + y) % mod
2183 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2184 //
2185 // We can now prove the equality:
2186 // (expr + addend) & mask
2187 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2188 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2189 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2190 // = addend & mask (SUFFIX_MASK)
2191 //
2192 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2193 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2194 // When the mask is negative, it has the most significant bit set.
2195 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2196 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2197 return false;
2198 }
2199
2200 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2201 if (mask_t->hi_as_long() == 0) {
2202 assert(mask_t->lo_as_long() == 0, "checked earlier");
2203 return false;
2204 }
2205
2206 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2207 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2208 return expr_trailing_zeros >= mask_bit_width;
2209 }
2210
2211 // Reduces the pattern:
2212 // (AndX (AddX add1 add2) mask)
2213 // to
2214 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2215 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2216 Node* add = in(1);
2217 Node* mask = in(2);
2218 int addidx = 0;
2219 if (add->Opcode() == Op_Add(bt)) {
2220 addidx = 1;
2221 } else if (mask->Opcode() == Op_Add(bt)) {
2222 mask = add;
2223 addidx = 2;
2224 add = in(addidx);
2225 }
2226 if (addidx > 0) {
2227 Node* add1 = add->in(1);
2228 Node* add2 = add->in(2);
2229 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2230 set_req_X(addidx, add2, phase);
2231 return this;
2232 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2233 set_req_X(addidx, add1, phase);
2234 return this;
2235 }
2236 }
2237 return nullptr;
2238 }