1 /*
2 * Copyright (c) 1997, 2024, 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 "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/connode.hpp"
29 #include "opto/convertnode.hpp"
30 #include "opto/memnode.hpp"
31 #include "opto/mulnode.hpp"
32 #include "opto/phaseX.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
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 if( t2 == Type::TOP ) return nullptr;
127 Node *mul1 = in(1);
128 #ifdef ASSERT
129 // Check for dead loop
130 int op1 = mul1->Opcode();
131 if ((mul1 == this) || (in(2) == this) ||
132 ((op1 == mul_opcode() || op1 == add_opcode()) &&
133 ((mul1->in(1) == this) || (mul1->in(2) == this) ||
134 (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) {
135 assert(false, "dead loop in MulNode::Ideal");
136 }
137 #endif
138
139 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply?
140 // Mul of a constant?
141 const Type *t12 = phase->type( mul1->in(2) );
142 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
143 // Compute new constant; check for overflow
144 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
145 if( tcon01->singleton() ) {
146 // The Mul of the flattened expression
147 set_req_X(1, mul1->in(1), phase);
148 set_req_X(2, phase->makecon(tcon01), phase);
149 t2 = tcon01;
150 progress = this; // Made progress
151 }
152 }
153 }
154 // If the right input is a constant, and the left input is an add of a
155 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
156 const Node *add1 = in(1);
157 if( add1->Opcode() == add_opcode() ) { // Left input is an add?
158 // Add of a constant?
159 const Type *t12 = phase->type( add1->in(2) );
160 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
161 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
162 // Compute new constant; check for overflow
163 const Type *tcon01 = mul_ring(t2,t12);
164 if( tcon01->singleton() ) {
165
166 // Convert (X+con1)*con0 into X*con0
167 Node *mul = clone(); // mul = ()*con0
168 mul->set_req(1,add1->in(1)); // mul = X*con0
169 mul = phase->transform(mul);
170
171 Node *add2 = add1->clone();
172 add2->set_req(1, mul); // X*con0 + con0*con1
173 add2->set_req(2, phase->makecon(tcon01) );
174 progress = add2;
175 }
176 }
177 } // End of is left input an add
178 } // End of is right input a Mul
179
180 return progress;
181 }
182
183 //------------------------------Value-----------------------------------------
184 const Type* MulNode::Value(PhaseGVN* phase) const {
185 const Type *t1 = phase->type( in(1) );
186 const Type *t2 = phase->type( in(2) );
187 // Either input is TOP ==> the result is TOP
188 if( t1 == Type::TOP ) return Type::TOP;
189 if( t2 == Type::TOP ) return Type::TOP;
190
191 // Either input is ZERO ==> the result is ZERO.
192 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
193 int op = Opcode();
194 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
195 const Type *zero = add_id(); // The multiplicative zero
196 if( t1->higher_equal( zero ) ) return zero;
197 if( t2->higher_equal( zero ) ) return zero;
198 }
199
200 // Either input is BOTTOM ==> the result is the local BOTTOM
201 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
202 return bottom_type();
203
204 #if defined(IA32)
205 // Can't trust native compilers to properly fold strict double
206 // multiplication with round-to-zero on this platform.
207 if (op == Op_MulD) {
208 return TypeD::DOUBLE;
209 }
210 #endif
211
212 return mul_ring(t1,t2); // Local flavor of type multiplication
213 }
214
215 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
216 switch (bt) {
217 case T_INT:
218 return new MulINode(in1, in2);
219 case T_LONG:
220 return new MulLNode(in1, in2);
221 default:
222 fatal("Not implemented for %s", type2name(bt));
223 }
224 return nullptr;
225 }
226
227
228 //=============================================================================
229 //------------------------------Ideal------------------------------------------
230 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
231 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
232 const jint con = in(2)->find_int_con(0);
233 if (con == 0) {
234 // If in(2) is not a constant, call Ideal() of the parent class to
235 // try to move constant to the right side.
236 return MulNode::Ideal(phase, can_reshape);
237 }
238
239 // Now we have a constant Node on the right and the constant in con.
240 if (con == 1) {
241 // By one is handled by Identity call
242 return nullptr;
243 }
244
245 // Check for negative constant; if so negate the final result
246 bool sign_flip = false;
247
248 unsigned int abs_con = uabs(con);
249 if (abs_con != (unsigned int)con) {
250 sign_flip = true;
251 }
252
253 // Get low bit; check for being the only bit
254 Node *res = nullptr;
255 unsigned int bit1 = submultiple_power_of_2(abs_con);
256 if (bit1 == abs_con) { // Found a power of 2?
257 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
258 } else {
259 // Check for constant with 2 bits set
260 unsigned int bit2 = abs_con - bit1;
261 bit2 = bit2 & (0 - bit2); // Extract 2nd bit
262 if (bit2 + bit1 == abs_con) { // Found all bits in con?
263 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
264 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
265 res = new AddINode(n2, n1);
266 } else if (is_power_of_2(abs_con + 1)) {
267 // Sleezy: power-of-2 - 1. Next time be generic.
268 unsigned int temp = abs_con + 1;
269 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
270 res = new SubINode(n1, in(1));
271 } else {
272 return MulNode::Ideal(phase, can_reshape);
273 }
274 }
275
276 if (sign_flip) { // Need to negate result?
277 res = phase->transform(res);// Transform, before making the zero con
278 res = new SubINode(phase->intcon(0),res);
279 }
280
281 return res; // Return final result
282 }
283
284 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
285 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
286 //
287 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
288 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
289 // - no overflow (i.e. product fits into x bits)
290 // - overflow (i.e. product does not fit into x bits)
291 //
292 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
293 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
294 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
295 // The cross product consists of the following four multiplications with 2x-bit results:
296 // (1) left_lo * right_lo
297 // (2) left_lo * right_hi
298 // (3) left_hi * right_lo
299 // (4) left_hi * right_hi
300 //
301 // Let's define the following two functions:
302 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
303 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
304 //
305 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
306 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
307 // Ux(p) = 0
308 // Lx(p) = p
309 //
310 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
311 // with x bits:
312 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
313 //
314 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
315 // (i.e. all values in the x-bit space could be possible):
316 // [result_lo,result_hi] = [NativeType_min,NativeType_max]
317 //
318 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
319 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
320 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
321 // the lower x bits the same number of times:
322 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
323 //
324 // If all upper x bits are equal, we can conclude:
325 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
326 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
327 //
328 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
329 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
330 //
331 //
332 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
333 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
334 // 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,
335 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
336 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
337 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
338 // that there is no underflow.
339 //
340 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
341 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
342 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
343 // can compare the upper x bits in a unified way with N() being the normalization function:
344 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
345 template<typename NativeType>
346 class IntegerTypeMultiplication {
347
348 NativeType _lo_left;
349 NativeType _lo_right;
350 NativeType _hi_left;
351 NativeType _hi_right;
352 short _widen_left;
353 short _widen_right;
354
355 static const Type* overflow_type();
356 static NativeType multiply_high(NativeType x, NativeType y);
357 const Type* create_type(NativeType lo, NativeType hi) const;
358
359 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
360 return normalize_overflow_value(x, y, multiply_high(x, y));
361 }
362
363 bool cross_product_not_same_overflow_value() const {
364 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
365 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
366 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
367 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
368 return lo_lo_high_product != lo_hi_high_product ||
369 lo_hi_high_product != hi_lo_high_product ||
370 hi_lo_high_product != hi_hi_high_product;
371 }
372
373 bool does_product_overflow(NativeType x, NativeType y) const {
374 return multiply_high_signed_overflow_value(x, y) != 0;
375 }
376
377 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
378 return java_multiply(x, y) < 0 ? result + 1 : result;
379 }
380
381 public:
382 template<class IntegerType>
383 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
384 : _lo_left(left->_lo), _lo_right(right->_lo),
385 _hi_left(left->_hi), _hi_right(right->_hi),
386 _widen_left(left->_widen), _widen_right(right->_widen) {}
387
388 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
389 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
390 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
391 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
392 const Type* compute() const {
393 if (cross_product_not_same_overflow_value()) {
394 return overflow_type();
395 }
396
397 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
398 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
399 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
400 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
401 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
402 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
403 return create_type(min, max);
404 }
405
406 bool does_overflow() const {
407 return does_product_overflow(_lo_left, _lo_right) ||
408 does_product_overflow(_lo_left, _hi_right) ||
409 does_product_overflow(_hi_left, _lo_right) ||
410 does_product_overflow(_hi_left, _hi_right);
411 }
412 };
413
414 template <>
415 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
416 return TypeInt::INT;
417 }
418
419 template <>
420 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
421 const jlong x_64 = x;
422 const jlong y_64 = y;
423 const jlong product = x_64 * y_64;
424 return (jint)((uint64_t)product >> 32u);
425 }
426
427 template <>
428 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
429 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
430 }
431
432 template <>
433 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
434 return TypeLong::LONG;
435 }
436
437 template <>
438 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
439 return multiply_high_signed(x, y);
440 }
441
442 template <>
443 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
444 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
445 }
446
447 // Compute the product type of two integer ranges into this node.
448 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
449 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
450 return integer_multiplication.compute();
451 }
452
453 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
454 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
455 return integer_multiplication.does_overflow();
456 }
457
458 // Compute the product type of two long ranges into this node.
459 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
460 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
461 return integer_multiplication.compute();
462 }
463
464 //=============================================================================
465 //------------------------------Ideal------------------------------------------
466 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
467 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
468 const jlong con = in(2)->find_long_con(0);
469 if (con == 0) {
470 // If in(2) is not a constant, call Ideal() of the parent class to
471 // try to move constant to the right side.
472 return MulNode::Ideal(phase, can_reshape);
473 }
474
475 // Now we have a constant Node on the right and the constant in con.
476 if (con == 1) {
477 // By one is handled by Identity call
478 return nullptr;
479 }
480
481 // Check for negative constant; if so negate the final result
482 bool sign_flip = false;
483 julong abs_con = uabs(con);
484 if (abs_con != (julong)con) {
485 sign_flip = true;
486 }
487
488 // Get low bit; check for being the only bit
489 Node *res = nullptr;
490 julong bit1 = submultiple_power_of_2(abs_con);
491 if (bit1 == abs_con) { // Found a power of 2?
492 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
493 } else {
494
495 // Check for constant with 2 bits set
496 julong bit2 = abs_con-bit1;
497 bit2 = bit2 & (0-bit2); // Extract 2nd bit
498 if (bit2 + bit1 == abs_con) { // Found all bits in con?
499 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
500 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
501 res = new AddLNode(n2, n1);
502
503 } else if (is_power_of_2(abs_con+1)) {
504 // Sleezy: power-of-2 -1. Next time be generic.
505 julong temp = abs_con + 1;
506 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
507 res = new SubLNode(n1, in(1));
508 } else {
509 return MulNode::Ideal(phase, can_reshape);
510 }
511 }
512
513 if (sign_flip) { // Need to negate result?
514 res = phase->transform(res);// Transform, before making the zero con
515 res = new SubLNode(phase->longcon(0),res);
516 }
517
518 return res; // Return final result
519 }
520
521 //=============================================================================
522 //------------------------------mul_ring---------------------------------------
523 // Compute the product type of two double ranges into this node.
524 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
525 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
526 return TypeF::make( t0->getf() * t1->getf() );
527 }
528
529 //------------------------------Ideal---------------------------------------
530 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
531 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
532 const TypeF *t2 = phase->type(in(2))->isa_float_constant();
533
534 // x * 2 -> x + x
535 if (t2 != nullptr && t2->getf() == 2) {
536 Node* base = in(1);
537 return new AddFNode(base, base);
538 }
539
540 return MulNode::Ideal(phase, can_reshape);
541 }
542
543 //=============================================================================
544 //------------------------------mul_ring---------------------------------------
545 // Compute the product type of two double ranges into this node.
546 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
547 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
548 // We must be multiplying 2 double constants.
549 return TypeD::make( t0->getd() * t1->getd() );
550 }
551
552 //------------------------------Ideal---------------------------------------
553 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
554 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
555 const TypeD *t2 = phase->type(in(2))->isa_double_constant();
556
557 // x * 2 -> x + x
558 if (t2 != nullptr && t2->getd() == 2) {
559 Node* base = in(1);
560 return new AddDNode(base, base);
561 }
562
563 return MulNode::Ideal(phase, can_reshape);
564 }
565
566 //=============================================================================
567 //------------------------------Value------------------------------------------
568 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
569 const Type *t1 = phase->type( in(1) );
570 const Type *t2 = phase->type( in(2) );
571 const Type *bot = bottom_type();
572 return MulHiValue(t1, t2, bot);
573 }
574
575 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
576 const Type *t1 = phase->type( in(1) );
577 const Type *t2 = phase->type( in(2) );
578 const Type *bot = bottom_type();
579 return MulHiValue(t1, t2, bot);
580 }
581
582 // A common routine used by UMulHiLNode and MulHiLNode
583 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
584 // Either input is TOP ==> the result is TOP
585 if( t1 == Type::TOP ) return Type::TOP;
586 if( t2 == Type::TOP ) return Type::TOP;
587
588 // Either input is BOTTOM ==> the result is the local BOTTOM
589 if( (t1 == bot) || (t2 == bot) ||
590 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
591 return bot;
592
593 // It is not worth trying to constant fold this stuff!
594 return TypeLong::LONG;
595 }
596
597 template<typename IntegerType>
598 static const IntegerType* and_value(const IntegerType* r0, const IntegerType* r1) {
599 typedef typename IntegerType::NativeType NativeType;
600 static_assert(std::is_signed<NativeType>::value, "Native type of IntegerType must be signed!");
601
602 int widen = MAX2(r0->_widen, r1->_widen);
603
604 // If both types are constants, we can calculate a constant result.
605 if (r0->is_con() && r1->is_con()) {
606 return IntegerType::make(r0->get_con() & r1->get_con());
607 }
608
609 // If both ranges are positive, the result will range from 0 up to the hi value of the smaller range. The minimum
610 // 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.
611 if (r0->_lo >= 0 && r1->_lo >= 0) {
612 return IntegerType::make(0, MIN2(r0->_hi, r1->_hi), widen);
613 }
614
615 // If only one range is positive, the result will range from 0 up to that range's maximum value.
616 // For the operation 'x & C' where C is a positive constant, the result will be in the range [0..C]. With that observation,
617 // 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]'
618 // where c >= 0 will be in the range [0..C].
619 if (r0->_lo >= 0) {
620 return IntegerType::make(0, r0->_hi, widen);
621 }
622
623 if (r1->_lo >= 0) {
624 return IntegerType::make(0, r1->_hi, widen);
625 }
626
627 // At this point, all positive ranges will have already been handled, so the only remaining cases will be negative ranges
628 // and constants.
629
630 assert(r0->_lo < 0 && r1->_lo < 0, "positive ranges should already be handled!");
631
632 // As two's complement means that both numbers will start with leading 1s, the lower bound of both ranges will contain
633 // the common leading 1s of both minimum values. In order to count them with count_leading_zeros, the bits are inverted.
634 NativeType sel_val = ~MIN2(r0->_lo, r1->_lo);
635
636 NativeType min;
637 if (sel_val == 0) {
638 // Since count_leading_zeros is undefined at 0, we short-circuit the condition where both ranges have a minimum of -1.
639 min = -1;
640 } else {
641 // 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.
642 int shift_bits = count_leading_zeros(sel_val) - 1;
643 min = std::numeric_limits<NativeType>::min() >> shift_bits;
644 }
645
646 NativeType max;
647 if (r0->_hi < 0 && r1->_hi < 0) {
648 // If both ranges are negative, then the same optimization as both positive ranges will apply, and the smaller hi
649 // value will mask off any bits set by higher values.
650 max = MIN2(r0->_hi, r1->_hi);
651 } else {
652 // In the case of ranges that cross zero, negative values can cause the higher order bits to be set, so the maximum
653 // positive value can be as high as the larger hi value.
654 max = MAX2(r0->_hi, r1->_hi);
655 }
656
657 return IntegerType::make(min, max, widen);
658 }
659
660 //=============================================================================
661 //------------------------------mul_ring---------------------------------------
662 // Supplied function returns the product of the inputs IN THE CURRENT RING.
663 // For the logical operations the ring's MUL is really a logical AND function.
664 // This also type-checks the inputs for sanity. Guaranteed never to
665 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
666 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
667 const TypeInt* r0 = t0->is_int();
668 const TypeInt* r1 = t1->is_int();
669
670 return and_value<TypeInt>(r0, r1);
671 }
672
673 const Type* AndINode::Value(PhaseGVN* phase) const {
674 // patterns similar to (v << 2) & 3
675 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_INT, true)) {
676 return TypeInt::ZERO;
677 }
678
679 return MulNode::Value(phase);
680 }
681
682 //------------------------------Identity---------------------------------------
683 // Masking off the high bits of an unsigned load is not required
684 Node* AndINode::Identity(PhaseGVN* phase) {
685
686 // x & x => x
687 if (in(1) == in(2)) {
688 return in(1);
689 }
690
691 Node* in1 = in(1);
692 uint op = in1->Opcode();
693 const TypeInt* t2 = phase->type(in(2))->isa_int();
694 if (t2 && t2->is_con()) {
695 int con = t2->get_con();
696 // Masking off high bits which are always zero is useless.
697 const TypeInt* t1 = phase->type(in(1))->isa_int();
698 if (t1 != nullptr && t1->_lo >= 0) {
699 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
700 if ((t1_support & con) == t1_support)
701 return in1;
702 }
703 // Masking off the high bits of a unsigned-shift-right is not
704 // needed either.
705 if (op == Op_URShiftI) {
706 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
707 if (t12 && t12->is_con()) { // Shift is by a constant
708 int shift = t12->get_con();
709 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
710 int mask = max_juint >> shift;
711 if ((mask & con) == mask) // If AND is useless, skip it
712 return in1;
713 }
714 }
715 }
716 return MulNode::Identity(phase);
717 }
718
719 //------------------------------Ideal------------------------------------------
720 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
721 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3
722 Node* progress = AndIL_add_shift_and_mask(phase, T_INT);
723 if (progress != nullptr) {
724 return progress;
725 }
726
727 // Convert "(~a) & (~b)" into "~(a | b)"
728 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
729 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
730 Node* tn = phase->transform(or_a_b);
731 return AddNode::make_not(phase, tn, T_INT);
732 }
733
734 // Special case constant AND mask
735 const TypeInt *t2 = phase->type( in(2) )->isa_int();
736 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
737 const int mask = t2->get_con();
738 Node *load = in(1);
739 uint lop = load->Opcode();
740
741 // Masking bits off of a Character? Hi bits are already zero.
742 if( lop == Op_LoadUS &&
743 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
744 return new AndINode(load,phase->intcon(mask&0xFFFF));
745
746 // Masking bits off of a Short? Loading a Character does some masking
747 if (can_reshape &&
748 load->outcnt() == 1 && load->unique_out() == this) {
749 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
750 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
751 ldus = phase->transform(ldus);
752 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
753 }
754
755 // Masking sign bits off of a Byte? Do an unsigned byte load plus
756 // an and.
757 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
758 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
759 ldub = phase->transform(ldub);
760 return new AndINode(ldub, phase->intcon(mask));
761 }
762 }
763
764 // Masking off sign bits? Dont make them!
765 if( lop == Op_RShiftI ) {
766 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
767 if( t12 && t12->is_con() ) { // Shift is by a constant
768 int shift = t12->get_con();
769 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
770 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
771 // If the AND'ing of the 2 masks has no bits, then only original shifted
772 // bits survive. NO sign-extension bits survive the maskings.
773 if( (sign_bits_mask & mask) == 0 ) {
774 // Use zero-fill shift instead
775 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
776 return new AndINode( zshift, in(2) );
777 }
778 }
779 }
780
781 // Check for 'negate/and-1', a pattern emitted when someone asks for
782 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
783 // plus 1) and the mask is of the low order bit. Skip the negate.
784 if( lop == Op_SubI && mask == 1 && load->in(1) &&
785 phase->type(load->in(1)) == TypeInt::ZERO )
786 return new AndINode( load->in(2), in(2) );
787
788 return MulNode::Ideal(phase, can_reshape);
789 }
790
791 //=============================================================================
792 //------------------------------mul_ring---------------------------------------
793 // Supplied function returns the product of the inputs IN THE CURRENT RING.
794 // For the logical operations the ring's MUL is really a logical AND function.
795 // This also type-checks the inputs for sanity. Guaranteed never to
796 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
797 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
798 const TypeLong* r0 = t0->is_long();
799 const TypeLong* r1 = t1->is_long();
800
801 return and_value<TypeLong>(r0, r1);
802 }
803
804 const Type* AndLNode::Value(PhaseGVN* phase) const {
805 // patterns similar to (v << 2) & 3
806 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_LONG, true)) {
807 return TypeLong::ZERO;
808 }
809
810 return MulNode::Value(phase);
811 }
812
813 //------------------------------Identity---------------------------------------
814 // Masking off the high bits of an unsigned load is not required
815 Node* AndLNode::Identity(PhaseGVN* phase) {
816
817 // x & x => x
818 if (in(1) == in(2)) {
819 return in(1);
820 }
821
822 Node *usr = in(1);
823 const TypeLong *t2 = phase->type( in(2) )->isa_long();
824 if( t2 && t2->is_con() ) {
825 jlong con = t2->get_con();
826 // Masking off high bits which are always zero is useless.
827 const TypeLong* t1 = phase->type( in(1) )->isa_long();
828 if (t1 != nullptr && t1->_lo >= 0) {
829 int bit_count = log2i_graceful(t1->_hi) + 1;
830 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
831 if ((t1_support & con) == t1_support)
832 return usr;
833 }
834 uint lop = usr->Opcode();
835 // Masking off the high bits of a unsigned-shift-right is not
836 // needed either.
837 if( lop == Op_URShiftL ) {
838 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
839 if( t12 && t12->is_con() ) { // Shift is by a constant
840 int shift = t12->get_con();
841 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
842 jlong mask = max_julong >> shift;
843 if( (mask&con) == mask ) // If AND is useless, skip it
844 return usr;
845 }
846 }
847 }
848 return MulNode::Identity(phase);
849 }
850
851 //------------------------------Ideal------------------------------------------
852 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
853 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3
854 Node* progress = AndIL_add_shift_and_mask(phase, T_LONG);
855 if (progress != nullptr) {
856 return progress;
857 }
858
859 // Convert "(~a) & (~b)" into "~(a | b)"
860 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
861 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
862 Node* tn = phase->transform(or_a_b);
863 return AddNode::make_not(phase, tn, T_LONG);
864 }
865
866 // Special case constant AND mask
867 const TypeLong *t2 = phase->type( in(2) )->isa_long();
868 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
869 const jlong mask = t2->get_con();
870
871 Node* in1 = in(1);
872 int op = in1->Opcode();
873
874 // Are we masking a long that was converted from an int with a mask
875 // that fits in 32-bits? Commute them and use an AndINode. Don't
876 // convert masks which would cause a sign extension of the integer
877 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
878 // would be optimized away later in Identity.
879 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
880 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
881 andi = phase->transform(andi);
882 return new ConvI2LNode(andi);
883 }
884
885 // Masking off sign bits? Dont make them!
886 if (op == Op_RShiftL) {
887 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
888 if( t12 && t12->is_con() ) { // Shift is by a constant
889 int shift = t12->get_con();
890 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
891 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
892 // If the AND'ing of the 2 masks has no bits, then only original shifted
893 // bits survive. NO sign-extension bits survive the maskings.
894 if( (sign_bits_mask & mask) == 0 ) {
895 // Use zero-fill shift instead
896 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
897 return new AndLNode(zshift, in(2));
898 }
899 }
900 }
901
902 return MulNode::Ideal(phase, can_reshape);
903 }
904
905 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
906 switch (bt) {
907 case T_INT:
908 return new LShiftINode(in1, in2);
909 case T_LONG:
910 return new LShiftLNode(in1, in2);
911 default:
912 fatal("Not implemented for %s", type2name(bt));
913 }
914 return nullptr;
915 }
916
917 //=============================================================================
918
919 static bool const_shift_count(PhaseGVN* phase, Node* shiftNode, int* count) {
920 const TypeInt* tcount = phase->type(shiftNode->in(2))->isa_int();
921 if (tcount != nullptr && tcount->is_con()) {
922 *count = tcount->get_con();
923 return true;
924 }
925 return false;
926 }
927
928 static int maskShiftAmount(PhaseGVN* phase, Node* shiftNode, int nBits) {
929 int count = 0;
930 if (const_shift_count(phase, shiftNode, &count)) {
931 int maskedShift = count & (nBits - 1);
932 if (maskedShift == 0) {
933 // Let Identity() handle 0 shift count.
934 return 0;
935 }
936
937 if (count != maskedShift) {
938 shiftNode->set_req(2, phase->intcon(maskedShift)); // Replace shift count with masked value.
939 PhaseIterGVN* igvn = phase->is_IterGVN();
940 if (igvn) {
941 igvn->rehash_node_delayed(shiftNode);
942 }
943 }
944 return maskedShift;
945 }
946 return 0;
947 }
948
949 //------------------------------Identity---------------------------------------
950 Node* LShiftINode::Identity(PhaseGVN* phase) {
951 int count = 0;
952 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
953 // Shift by a multiple of 32 does nothing
954 return in(1);
955 }
956 return this;
957 }
958
959 //------------------------------Ideal------------------------------------------
960 // If the right input is a constant, and the left input is an add of a
961 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
962 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
963 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
964 if (con == 0) {
965 return nullptr;
966 }
967
968 // Left input is an add?
969 Node *add1 = in(1);
970 int add1_op = add1->Opcode();
971 if( add1_op == Op_AddI ) { // Left input is an add?
972 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" );
973
974 // Transform is legal, but check for profit. Avoid breaking 'i2s'
975 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
976 if( con < 16 ) {
977 // Left input is an add of the same number?
978 if (add1->in(1) == add1->in(2)) {
979 // Convert "(x + x) << c0" into "x << (c0 + 1)"
980 // In general, this optimization cannot be applied for c0 == 31 since
981 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
982 return new LShiftINode(add1->in(1), phase->intcon(con + 1));
983 }
984
985 // Left input is an add of a constant?
986 const TypeInt *t12 = phase->type(add1->in(2))->isa_int();
987 if( t12 && t12->is_con() ){ // Left input is an add of a con?
988 // Compute X << con0
989 Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) );
990 // Compute X<<con0 + (con1<<con0)
991 return new AddINode( lsh, phase->intcon(t12->get_con() << con));
992 }
993 }
994 }
995
996 // Check for "(x >> C1) << C2"
997 if (add1_op == Op_RShiftI || add1_op == Op_URShiftI) {
998 int add1Con = 0;
999 const_shift_count(phase, add1, &add1Con);
1000
1001 // Special case C1 == C2, which just masks off low bits
1002 if (add1Con > 0 && con == add1Con) {
1003 // Convert to "(x & -(1 << C2))"
1004 return new AndINode(add1->in(1), phase->intcon(java_negate(jint(1 << con))));
1005 } else {
1006 // Wait until the right shift has been sharpened to the correct count
1007 if (add1Con > 0 && add1Con < BitsPerJavaInteger) {
1008 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1009 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1010 if (phase->is_IterGVN()) {
1011 if (con > add1Con) {
1012 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1013 Node* lshift = phase->transform(new LShiftINode(add1->in(1), phase->intcon(con - add1Con)));
1014 return new AndINode(lshift, phase->intcon(java_negate(jint(1 << con))));
1015 } else {
1016 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1017 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1018
1019 // Handle logical and arithmetic shifts
1020 Node* rshift;
1021 if (add1_op == Op_RShiftI) {
1022 rshift = phase->transform(new RShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1023 } else {
1024 rshift = phase->transform(new URShiftINode(add1->in(1), phase->intcon(add1Con - con)));
1025 }
1026
1027 return new AndINode(rshift, phase->intcon(java_negate(jint(1 << con))));
1028 }
1029 } else {
1030 phase->record_for_igvn(this);
1031 }
1032 }
1033 }
1034 }
1035
1036 // Check for "((x >> C1) & Y) << C2"
1037 if (add1_op == Op_AndI) {
1038 Node *add2 = add1->in(1);
1039 int add2_op = add2->Opcode();
1040 if (add2_op == Op_RShiftI || add2_op == Op_URShiftI) {
1041 // Special case C1 == C2, which just masks off low bits
1042 if (add2->in(2) == in(2)) {
1043 // Convert to "(x & (Y << C2))"
1044 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1045 return new AndINode(add2->in(1), y_sh);
1046 }
1047
1048 int add2Con = 0;
1049 const_shift_count(phase, add2, &add2Con);
1050 if (add2Con > 0 && add2Con < BitsPerJavaInteger) {
1051 if (phase->is_IterGVN()) {
1052 // Convert to "((x >> C1) << C2) & (Y << C2)"
1053
1054 // Make "(x >> C1) << C2", which will get folded away by the rule above
1055 Node* x_sh = phase->transform(new LShiftINode(add2, phase->intcon(con)));
1056 // Make "Y << C2", which will simplify when Y is a constant
1057 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con)));
1058
1059 return new AndINode(x_sh, y_sh);
1060 } else {
1061 phase->record_for_igvn(this);
1062 }
1063 }
1064 }
1065 }
1066
1067 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1068 // before shifting them away.
1069 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con);
1070 if( add1_op == Op_AndI &&
1071 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) )
1072 return new LShiftINode( add1->in(1), in(2) );
1073
1074 return nullptr;
1075 }
1076
1077 //------------------------------Value------------------------------------------
1078 // A LShiftINode shifts its input2 left by input1 amount.
1079 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1080 const Type *t1 = phase->type( in(1) );
1081 const Type *t2 = phase->type( in(2) );
1082 // Either input is TOP ==> the result is TOP
1083 if( t1 == Type::TOP ) return Type::TOP;
1084 if( t2 == Type::TOP ) return Type::TOP;
1085
1086 // Left input is ZERO ==> the result is ZERO.
1087 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1088 // Shift by zero does nothing
1089 if( t2 == TypeInt::ZERO ) return t1;
1090
1091 // Either input is BOTTOM ==> the result is BOTTOM
1092 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) ||
1093 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1094 return TypeInt::INT;
1095
1096 const TypeInt *r1 = t1->is_int(); // Handy access
1097 const TypeInt *r2 = t2->is_int(); // Handy access
1098
1099 if (!r2->is_con())
1100 return TypeInt::INT;
1101
1102 uint shift = r2->get_con();
1103 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1104 // Shift by a multiple of 32 does nothing:
1105 if (shift == 0) return t1;
1106
1107 // If the shift is a constant, shift the bounds of the type,
1108 // unless this could lead to an overflow.
1109 if (!r1->is_con()) {
1110 jint lo = r1->_lo, hi = r1->_hi;
1111 if (((lo << shift) >> shift) == lo &&
1112 ((hi << shift) >> shift) == hi) {
1113 // No overflow. The range shifts up cleanly.
1114 return TypeInt::make((jint)lo << (jint)shift,
1115 (jint)hi << (jint)shift,
1116 MAX2(r1->_widen,r2->_widen));
1117 }
1118 return TypeInt::INT;
1119 }
1120
1121 return TypeInt::make( (jint)r1->get_con() << (jint)shift );
1122 }
1123
1124 //=============================================================================
1125 //------------------------------Identity---------------------------------------
1126 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1127 int count = 0;
1128 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1129 // Shift by a multiple of 64 does nothing
1130 return in(1);
1131 }
1132 return this;
1133 }
1134
1135 //------------------------------Ideal------------------------------------------
1136 // If the right input is a constant, and the left input is an add of a
1137 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1138 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1139 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1140 if (con == 0) {
1141 return nullptr;
1142 }
1143
1144 // Left input is an add?
1145 Node *add1 = in(1);
1146 int add1_op = add1->Opcode();
1147 if( add1_op == Op_AddL ) { // Left input is an add?
1148 // Avoid dead data cycles from dead loops
1149 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" );
1150
1151 // Left input is an add of the same number?
1152 if (con != (BitsPerJavaLong - 1) && add1->in(1) == add1->in(2)) {
1153 // Convert "(x + x) << c0" into "x << (c0 + 1)"
1154 // Can only be applied if c0 != 63 because:
1155 // (x + x) << 63 = 2x << 63, while
1156 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1157 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1158 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1159 return new LShiftLNode(add1->in(1), phase->intcon(con + 1));
1160 }
1161
1162 // Left input is an add of a constant?
1163 const TypeLong *t12 = phase->type(add1->in(2))->isa_long();
1164 if( t12 && t12->is_con() ){ // Left input is an add of a con?
1165 // Compute X << con0
1166 Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) );
1167 // Compute X<<con0 + (con1<<con0)
1168 return new AddLNode( lsh, phase->longcon(t12->get_con() << con));
1169 }
1170 }
1171
1172 // Check for "(x >> C1) << C2"
1173 if (add1_op == Op_RShiftL || add1_op == Op_URShiftL) {
1174 int add1Con = 0;
1175 const_shift_count(phase, add1, &add1Con);
1176
1177 // Special case C1 == C2, which just masks off low bits
1178 if (add1Con > 0 && con == add1Con) {
1179 // Convert to "(x & -(1 << C2))"
1180 return new AndLNode(add1->in(1), phase->longcon(java_negate(jlong(CONST64(1) << con))));
1181 } else {
1182 // Wait until the right shift has been sharpened to the correct count
1183 if (add1Con > 0 && add1Con < BitsPerJavaLong) {
1184 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1185 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1186 if (phase->is_IterGVN()) {
1187 if (con > add1Con) {
1188 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1189 Node* lshift = phase->transform(new LShiftLNode(add1->in(1), phase->intcon(con - add1Con)));
1190 return new AndLNode(lshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1191 } else {
1192 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1193 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1194
1195 // Handle logical and arithmetic shifts
1196 Node* rshift;
1197 if (add1_op == Op_RShiftL) {
1198 rshift = phase->transform(new RShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1199 } else {
1200 rshift = phase->transform(new URShiftLNode(add1->in(1), phase->intcon(add1Con - con)));
1201 }
1202
1203 return new AndLNode(rshift, phase->longcon(java_negate(jlong(CONST64(1) << con))));
1204 }
1205 } else {
1206 phase->record_for_igvn(this);
1207 }
1208 }
1209 }
1210 }
1211
1212 // Check for "((x >> C1) & Y) << C2"
1213 if (add1_op == Op_AndL) {
1214 Node* add2 = add1->in(1);
1215 int add2_op = add2->Opcode();
1216 if (add2_op == Op_RShiftL || add2_op == Op_URShiftL) {
1217 // Special case C1 == C2, which just masks off low bits
1218 if (add2->in(2) == in(2)) {
1219 // Convert to "(x & (Y << C2))"
1220 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1221 return new AndLNode(add2->in(1), y_sh);
1222 }
1223
1224 int add2Con = 0;
1225 const_shift_count(phase, add2, &add2Con);
1226 if (add2Con > 0 && add2Con < BitsPerJavaLong) {
1227 if (phase->is_IterGVN()) {
1228 // Convert to "((x >> C1) << C2) & (Y << C2)"
1229
1230 // Make "(x >> C1) << C2", which will get folded away by the rule above
1231 Node* x_sh = phase->transform(new LShiftLNode(add2, phase->intcon(con)));
1232 // Make "Y << C2", which will simplify when Y is a constant
1233 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con)));
1234
1235 return new AndLNode(x_sh, y_sh);
1236 } else {
1237 phase->record_for_igvn(this);
1238 }
1239 }
1240 }
1241 }
1242
1243 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits
1244 // before shifting them away.
1245 const jlong bits_mask = jlong(max_julong >> con);
1246 if( add1_op == Op_AndL &&
1247 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) )
1248 return new LShiftLNode( add1->in(1), in(2) );
1249
1250 return nullptr;
1251 }
1252
1253 //------------------------------Value------------------------------------------
1254 // A LShiftLNode shifts its input2 left by input1 amount.
1255 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1256 const Type *t1 = phase->type( in(1) );
1257 const Type *t2 = phase->type( in(2) );
1258 // Either input is TOP ==> the result is TOP
1259 if( t1 == Type::TOP ) return Type::TOP;
1260 if( t2 == Type::TOP ) return Type::TOP;
1261
1262 // Left input is ZERO ==> the result is ZERO.
1263 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1264 // Shift by zero does nothing
1265 if( t2 == TypeInt::ZERO ) return t1;
1266
1267 // Either input is BOTTOM ==> the result is BOTTOM
1268 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) ||
1269 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
1270 return TypeLong::LONG;
1271
1272 const TypeLong *r1 = t1->is_long(); // Handy access
1273 const TypeInt *r2 = t2->is_int(); // Handy access
1274
1275 if (!r2->is_con())
1276 return TypeLong::LONG;
1277
1278 uint shift = r2->get_con();
1279 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1280 // Shift by a multiple of 64 does nothing:
1281 if (shift == 0) return t1;
1282
1283 // If the shift is a constant, shift the bounds of the type,
1284 // unless this could lead to an overflow.
1285 if (!r1->is_con()) {
1286 jlong lo = r1->_lo, hi = r1->_hi;
1287 if (((lo << shift) >> shift) == lo &&
1288 ((hi << shift) >> shift) == hi) {
1289 // No overflow. The range shifts up cleanly.
1290 return TypeLong::make((jlong)lo << (jint)shift,
1291 (jlong)hi << (jint)shift,
1292 MAX2(r1->_widen,r2->_widen));
1293 }
1294 return TypeLong::LONG;
1295 }
1296
1297 return TypeLong::make( (jlong)r1->get_con() << (jint)shift );
1298 }
1299
1300 //=============================================================================
1301 //------------------------------Identity---------------------------------------
1302 Node* RShiftINode::Identity(PhaseGVN* phase) {
1303 int count = 0;
1304 if (const_shift_count(phase, this, &count)) {
1305 if ((count & (BitsPerJavaInteger - 1)) == 0) {
1306 // Shift by a multiple of 32 does nothing
1307 return in(1);
1308 }
1309 // Check for useless sign-masking
1310 if (in(1)->Opcode() == Op_LShiftI &&
1311 in(1)->req() == 3 &&
1312 in(1)->in(2) == in(2)) {
1313 count &= BitsPerJavaInteger-1; // semantics of Java shifts
1314 // Compute masks for which this shifting doesn't change
1315 int lo = (-1 << (BitsPerJavaInteger - ((uint)count)-1)); // FFFF8000
1316 int hi = ~lo; // 00007FFF
1317 const TypeInt* t11 = phase->type(in(1)->in(1))->isa_int();
1318 if (t11 == nullptr) {
1319 return this;
1320 }
1321 // Does actual value fit inside of mask?
1322 if (lo <= t11->_lo && t11->_hi <= hi) {
1323 return in(1)->in(1); // Then shifting is a nop
1324 }
1325 }
1326 }
1327 return this;
1328 }
1329
1330 //------------------------------Ideal------------------------------------------
1331 Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1332 // Inputs may be TOP if they are dead.
1333 const TypeInt *t1 = phase->type(in(1))->isa_int();
1334 if (!t1) return nullptr; // Left input is an integer
1335 const TypeInt *t3; // type of in(1).in(2)
1336 int shift = maskShiftAmount(phase, this, BitsPerJavaInteger);
1337 if (shift == 0) {
1338 return nullptr;
1339 }
1340
1341 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1342 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1343 const Node *mask = in(1);
1344 if( mask->Opcode() == Op_AndI &&
1345 (t3 = phase->type(mask->in(2))->isa_int()) &&
1346 t3->is_con() ) {
1347 Node *x = mask->in(1);
1348 jint maskbits = t3->get_con();
1349 // Convert to "(x >> shift) & (mask >> shift)"
1350 Node *shr_nomask = phase->transform( new RShiftINode(mask->in(1), in(2)) );
1351 return new AndINode(shr_nomask, phase->intcon( maskbits >> shift));
1352 }
1353
1354 // Check for "(short[i] <<16)>>16" which simply sign-extends
1355 const Node *shl = in(1);
1356 if( shl->Opcode() != Op_LShiftI ) return nullptr;
1357
1358 if( shift == 16 &&
1359 (t3 = phase->type(shl->in(2))->isa_int()) &&
1360 t3->is_con(16) ) {
1361 Node *ld = shl->in(1);
1362 if( ld->Opcode() == Op_LoadS ) {
1363 // Sign extension is just useless here. Return a RShiftI of zero instead
1364 // returning 'ld' directly. We cannot return an old Node directly as
1365 // that is the job of 'Identity' calls and Identity calls only work on
1366 // direct inputs ('ld' is an extra Node removed from 'this'). The
1367 // combined optimization requires Identity only return direct inputs.
1368 set_req_X(1, ld, phase);
1369 set_req_X(2, phase->intcon(0), phase);
1370 return this;
1371 }
1372 else if (can_reshape &&
1373 ld->Opcode() == Op_LoadUS &&
1374 ld->outcnt() == 1 && ld->unique_out() == shl)
1375 // Replace zero-extension-load with sign-extension-load
1376 return ld->as_Load()->convert_to_signed_load(*phase);
1377 }
1378
1379 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1380 if( shift == 24 &&
1381 (t3 = phase->type(shl->in(2))->isa_int()) &&
1382 t3->is_con(24) ) {
1383 Node *ld = shl->in(1);
1384 if (ld->Opcode() == Op_LoadB) {
1385 // Sign extension is just useless here
1386 set_req_X(1, ld, phase);
1387 set_req_X(2, phase->intcon(0), phase);
1388 return this;
1389 }
1390 }
1391
1392 return nullptr;
1393 }
1394
1395 //------------------------------Value------------------------------------------
1396 // A RShiftINode shifts its input2 right by input1 amount.
1397 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1398 const Type *t1 = phase->type( in(1) );
1399 const Type *t2 = phase->type( in(2) );
1400 // Either input is TOP ==> the result is TOP
1401 if( t1 == Type::TOP ) return Type::TOP;
1402 if( t2 == Type::TOP ) return Type::TOP;
1403
1404 // Left input is ZERO ==> the result is ZERO.
1405 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1406 // Shift by zero does nothing
1407 if( t2 == TypeInt::ZERO ) return t1;
1408
1409 // Either input is BOTTOM ==> the result is BOTTOM
1410 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1411 return TypeInt::INT;
1412
1413 const TypeInt *r1 = t1->is_int(); // Handy access
1414 const TypeInt *r2 = t2->is_int(); // Handy access
1415
1416 // If the shift is a constant, just shift the bounds of the type.
1417 // For example, if the shift is 31, we just propagate sign bits.
1418 if (!r1->is_con() && r2->is_con()) {
1419 uint shift = r2->get_con();
1420 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1421 // Shift by a multiple of 32 does nothing:
1422 if (shift == 0) return t1;
1423 // Calculate reasonably aggressive bounds for the result.
1424 // This is necessary if we are to correctly type things
1425 // like (x<<24>>24) == ((byte)x).
1426 jint lo = (jint)r1->_lo >> (jint)shift;
1427 jint hi = (jint)r1->_hi >> (jint)shift;
1428 assert(lo <= hi, "must have valid bounds");
1429 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1430 #ifdef ASSERT
1431 // Make sure we get the sign-capture idiom correct.
1432 if (shift == BitsPerJavaInteger-1) {
1433 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0");
1434 if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1");
1435 }
1436 #endif
1437 return ti;
1438 }
1439
1440 if (!r1->is_con() || !r2->is_con()) {
1441 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1442 if (r1->_lo >= 0) {
1443 return TypeInt::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen));
1444 }
1445
1446 // Conversely, if the left input is negative then the result must be negative.
1447 if (r1->_hi <= -1) {
1448 return TypeInt::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen));
1449 }
1450
1451 return TypeInt::INT;
1452 }
1453
1454 // Signed shift right
1455 return TypeInt::make(r1->get_con() >> (r2->get_con() & 31));
1456 }
1457
1458 //=============================================================================
1459 //------------------------------Identity---------------------------------------
1460 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1461 const TypeInt *ti = phase->type(in(2))->isa_int(); // Shift count is an int.
1462 return (ti && ti->is_con() && (ti->get_con() & (BitsPerJavaLong - 1)) == 0) ? in(1) : this;
1463 }
1464
1465 //------------------------------Value------------------------------------------
1466 // A RShiftLNode shifts its input2 right by input1 amount.
1467 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1468 const Type *t1 = phase->type( in(1) );
1469 const Type *t2 = phase->type( in(2) );
1470 // Either input is TOP ==> the result is TOP
1471 if( t1 == Type::TOP ) return Type::TOP;
1472 if( t2 == Type::TOP ) return Type::TOP;
1473
1474 // Left input is ZERO ==> the result is ZERO.
1475 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1476 // Shift by zero does nothing
1477 if( t2 == TypeInt::ZERO ) return t1;
1478
1479 // Either input is BOTTOM ==> the result is BOTTOM
1480 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1481 return TypeLong::LONG;
1482
1483 const TypeLong *r1 = t1->is_long(); // Handy access
1484 const TypeInt *r2 = t2->is_int (); // Handy access
1485
1486 // If the shift is a constant, just shift the bounds of the type.
1487 // For example, if the shift is 63, we just propagate sign bits.
1488 if (!r1->is_con() && r2->is_con()) {
1489 uint shift = r2->get_con();
1490 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts
1491 // Shift by a multiple of 64 does nothing:
1492 if (shift == 0) return t1;
1493 // Calculate reasonably aggressive bounds for the result.
1494 // This is necessary if we are to correctly type things
1495 // like (x<<24>>24) == ((byte)x).
1496 jlong lo = (jlong)r1->_lo >> (jlong)shift;
1497 jlong hi = (jlong)r1->_hi >> (jlong)shift;
1498 assert(lo <= hi, "must have valid bounds");
1499 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1500 #ifdef ASSERT
1501 // Make sure we get the sign-capture idiom correct.
1502 if (shift == (2*BitsPerJavaInteger)-1) {
1503 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0");
1504 if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1");
1505 }
1506 #endif
1507 return tl;
1508 }
1509
1510 if (!r1->is_con() || !r2->is_con()) {
1511 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1512 if (r1->_lo >= 0) {
1513 return TypeLong::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen));
1514 }
1515
1516 // Conversely, if the left input is negative then the result must be negative.
1517 if (r1->_hi <= -1) {
1518 return TypeLong::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen));
1519 }
1520
1521 return TypeLong::LONG;
1522 }
1523
1524 return TypeLong::make(r1->get_con() >> (r2->get_con() & 63));
1525 }
1526
1527 //=============================================================================
1528 //------------------------------Identity---------------------------------------
1529 Node* URShiftINode::Identity(PhaseGVN* phase) {
1530 int count = 0;
1531 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1532 // Shift by a multiple of 32 does nothing
1533 return in(1);
1534 }
1535
1536 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1537 // Happens during new-array length computation.
1538 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1539 Node *add = in(1);
1540 if (add->Opcode() == Op_AddI) {
1541 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1542 if (t2 && t2->is_con(wordSize - 1) &&
1543 add->in(1)->Opcode() == Op_LShiftI) {
1544 // Check that shift_counts are LogBytesPerWord.
1545 Node *lshift_count = add->in(1)->in(2);
1546 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1547 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1548 t_lshift_count == phase->type(in(2))) {
1549 Node *x = add->in(1)->in(1);
1550 const TypeInt *t_x = phase->type(x)->isa_int();
1551 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1552 return x;
1553 }
1554 }
1555 }
1556 }
1557
1558 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1559 }
1560
1561 //------------------------------Ideal------------------------------------------
1562 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1563 int con = maskShiftAmount(phase, this, BitsPerJavaInteger);
1564 if (con == 0) {
1565 return nullptr;
1566 }
1567
1568 // We'll be wanting the right-shift amount as a mask of that many bits
1569 const int mask = right_n_bits(BitsPerJavaInteger - con);
1570
1571 int in1_op = in(1)->Opcode();
1572
1573 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1574 if( in1_op == Op_URShiftI ) {
1575 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1576 if( t12 && t12->is_con() ) { // Right input is a constant
1577 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1578 const int con2 = t12->get_con() & 31; // Shift count is always masked
1579 const int con3 = con+con2;
1580 if( con3 < 32 ) // Only merge shifts if total is < 32
1581 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1582 }
1583 }
1584
1585 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1586 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1587 // If Q is "X << z" the rounding is useless. Look for patterns like
1588 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1589 Node *add = in(1);
1590 const TypeInt *t2 = phase->type(in(2))->isa_int();
1591 if (in1_op == Op_AddI) {
1592 Node *lshl = add->in(1);
1593 if( lshl->Opcode() == Op_LShiftI &&
1594 phase->type(lshl->in(2)) == t2 ) {
1595 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) );
1596 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) );
1597 return new AndINode( sum, phase->intcon(mask) );
1598 }
1599 }
1600
1601 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1602 // This shortens the mask. Also, if we are extracting a high byte and
1603 // storing it to a buffer, the mask will be removed completely.
1604 Node *andi = in(1);
1605 if( in1_op == Op_AndI ) {
1606 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1607 if( t3 && t3->is_con() ) { // Right input is a constant
1608 jint mask2 = t3->get_con();
1609 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1610 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1611 return new AndINode(newshr, phase->intcon(mask2));
1612 // The negative values are easier to materialize than positive ones.
1613 // A typical case from address arithmetic is ((x & ~15) >> 4).
1614 // It's better to change that to ((x >> 4) & ~0) versus
1615 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1616 }
1617 }
1618
1619 // Check for "(X << z ) >>> z" which simply zero-extends
1620 Node *shl = in(1);
1621 if( in1_op == Op_LShiftI &&
1622 phase->type(shl->in(2)) == t2 )
1623 return new AndINode( shl->in(1), phase->intcon(mask) );
1624
1625 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1626 Node *shr = in(1);
1627 if ( in1_op == Op_RShiftI ) {
1628 Node *in11 = shr->in(1);
1629 Node *in12 = shr->in(2);
1630 const TypeInt *t11 = phase->type(in11)->isa_int();
1631 const TypeInt *t12 = phase->type(in12)->isa_int();
1632 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1633 return new URShiftINode(in11, phase->intcon(31));
1634 }
1635 }
1636
1637 return nullptr;
1638 }
1639
1640 //------------------------------Value------------------------------------------
1641 // A URShiftINode shifts its input2 right by input1 amount.
1642 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1643 // (This is a near clone of RShiftINode::Value.)
1644 const Type *t1 = phase->type( in(1) );
1645 const Type *t2 = phase->type( in(2) );
1646 // Either input is TOP ==> the result is TOP
1647 if( t1 == Type::TOP ) return Type::TOP;
1648 if( t2 == Type::TOP ) return Type::TOP;
1649
1650 // Left input is ZERO ==> the result is ZERO.
1651 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1652 // Shift by zero does nothing
1653 if( t2 == TypeInt::ZERO ) return t1;
1654
1655 // Either input is BOTTOM ==> the result is BOTTOM
1656 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1657 return TypeInt::INT;
1658
1659 if (t2 == TypeInt::INT)
1660 return TypeInt::INT;
1661
1662 const TypeInt *r1 = t1->is_int(); // Handy access
1663 const TypeInt *r2 = t2->is_int(); // Handy access
1664
1665 if (r2->is_con()) {
1666 uint shift = r2->get_con();
1667 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1668 // Shift by a multiple of 32 does nothing:
1669 if (shift == 0) return t1;
1670 // Calculate reasonably aggressive bounds for the result.
1671 jint lo = (juint)r1->_lo >> (juint)shift;
1672 jint hi = (juint)r1->_hi >> (juint)shift;
1673 if (r1->_hi >= 0 && r1->_lo < 0) {
1674 // If the type has both negative and positive values,
1675 // there are two separate sub-domains to worry about:
1676 // The positive half and the negative half.
1677 jint neg_lo = lo;
1678 jint neg_hi = (juint)-1 >> (juint)shift;
1679 jint pos_lo = (juint) 0 >> (juint)shift;
1680 jint pos_hi = hi;
1681 lo = MIN2(neg_lo, pos_lo); // == 0
1682 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1683 }
1684 assert(lo <= hi, "must have valid bounds");
1685 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1686 #ifdef ASSERT
1687 // Make sure we get the sign-capture idiom correct.
1688 if (shift == BitsPerJavaInteger-1) {
1689 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1690 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1691 }
1692 #endif
1693 return ti;
1694 }
1695
1696 //
1697 // Do not support shifted oops in info for GC
1698 //
1699 // else if( t1->base() == Type::InstPtr ) {
1700 //
1701 // const TypeInstPtr *o = t1->is_instptr();
1702 // if( t1->singleton() )
1703 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1704 // }
1705 // else if( t1->base() == Type::KlassPtr ) {
1706 // const TypeKlassPtr *o = t1->is_klassptr();
1707 // if( t1->singleton() )
1708 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1709 // }
1710
1711 return TypeInt::INT;
1712 }
1713
1714 //=============================================================================
1715 //------------------------------Identity---------------------------------------
1716 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1717 int count = 0;
1718 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1719 // Shift by a multiple of 64 does nothing
1720 return in(1);
1721 }
1722 return this;
1723 }
1724
1725 //------------------------------Ideal------------------------------------------
1726 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1727 int con = maskShiftAmount(phase, this, BitsPerJavaLong);
1728 if (con == 0) {
1729 return nullptr;
1730 }
1731
1732 // We'll be wanting the right-shift amount as a mask of that many bits
1733 const jlong mask = jlong(max_julong >> con);
1734
1735 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1736 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1737 // If Q is "X << z" the rounding is useless. Look for patterns like
1738 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1739 Node *add = in(1);
1740 const TypeInt *t2 = phase->type(in(2))->isa_int();
1741 if (add->Opcode() == Op_AddL) {
1742 Node *lshl = add->in(1);
1743 if( lshl->Opcode() == Op_LShiftL &&
1744 phase->type(lshl->in(2)) == t2 ) {
1745 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1746 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1747 return new AndLNode( sum, phase->longcon(mask) );
1748 }
1749 }
1750
1751 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1752 // This shortens the mask. Also, if we are extracting a high byte and
1753 // storing it to a buffer, the mask will be removed completely.
1754 Node *andi = in(1);
1755 if( andi->Opcode() == Op_AndL ) {
1756 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1757 if( t3 && t3->is_con() ) { // Right input is a constant
1758 jlong mask2 = t3->get_con();
1759 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1760 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1761 return new AndLNode(newshr, phase->longcon(mask2));
1762 }
1763 }
1764
1765 // Check for "(X << z ) >>> z" which simply zero-extends
1766 Node *shl = in(1);
1767 if( shl->Opcode() == Op_LShiftL &&
1768 phase->type(shl->in(2)) == t2 )
1769 return new AndLNode( shl->in(1), phase->longcon(mask) );
1770
1771 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1772 Node *shr = in(1);
1773 if ( shr->Opcode() == Op_RShiftL ) {
1774 Node *in11 = shr->in(1);
1775 Node *in12 = shr->in(2);
1776 const TypeLong *t11 = phase->type(in11)->isa_long();
1777 const TypeInt *t12 = phase->type(in12)->isa_int();
1778 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1779 return new URShiftLNode(in11, phase->intcon(63));
1780 }
1781 }
1782 return nullptr;
1783 }
1784
1785 //------------------------------Value------------------------------------------
1786 // A URShiftINode shifts its input2 right by input1 amount.
1787 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1788 // (This is a near clone of RShiftLNode::Value.)
1789 const Type *t1 = phase->type( in(1) );
1790 const Type *t2 = phase->type( in(2) );
1791 // Either input is TOP ==> the result is TOP
1792 if( t1 == Type::TOP ) return Type::TOP;
1793 if( t2 == Type::TOP ) return Type::TOP;
1794
1795 // Left input is ZERO ==> the result is ZERO.
1796 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1797 // Shift by zero does nothing
1798 if( t2 == TypeInt::ZERO ) return t1;
1799
1800 // Either input is BOTTOM ==> the result is BOTTOM
1801 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1802 return TypeLong::LONG;
1803
1804 if (t2 == TypeInt::INT)
1805 return TypeLong::LONG;
1806
1807 const TypeLong *r1 = t1->is_long(); // Handy access
1808 const TypeInt *r2 = t2->is_int (); // Handy access
1809
1810 if (r2->is_con()) {
1811 uint shift = r2->get_con();
1812 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1813 // Shift by a multiple of 64 does nothing:
1814 if (shift == 0) return t1;
1815 // Calculate reasonably aggressive bounds for the result.
1816 jlong lo = (julong)r1->_lo >> (juint)shift;
1817 jlong hi = (julong)r1->_hi >> (juint)shift;
1818 if (r1->_hi >= 0 && r1->_lo < 0) {
1819 // If the type has both negative and positive values,
1820 // there are two separate sub-domains to worry about:
1821 // The positive half and the negative half.
1822 jlong neg_lo = lo;
1823 jlong neg_hi = (julong)-1 >> (juint)shift;
1824 jlong pos_lo = (julong) 0 >> (juint)shift;
1825 jlong pos_hi = hi;
1826 //lo = MIN2(neg_lo, pos_lo); // == 0
1827 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1828 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1829 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1830 }
1831 assert(lo <= hi, "must have valid bounds");
1832 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1833 #ifdef ASSERT
1834 // Make sure we get the sign-capture idiom correct.
1835 if (shift == BitsPerJavaLong - 1) {
1836 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1837 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1838 }
1839 #endif
1840 return tl;
1841 }
1842
1843 return TypeLong::LONG; // Give up
1844 }
1845
1846 //=============================================================================
1847 //------------------------------Ideal------------------------------------------
1848 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1849 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1850 // This reduces the number of rules in the matcher, as we only need to check
1851 // for negations on the second argument, and not the symmetric case where
1852 // the first argument is negated.
1853 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1854 swap_edges(1, 2);
1855 return this;
1856 }
1857 return nullptr;
1858 }
1859
1860 //=============================================================================
1861 //------------------------------Value------------------------------------------
1862 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1863 const Type *t1 = phase->type(in(1));
1864 if (t1 == Type::TOP) return Type::TOP;
1865 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1866 const Type *t2 = phase->type(in(2));
1867 if (t2 == Type::TOP) return Type::TOP;
1868 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1869 const Type *t3 = phase->type(in(3));
1870 if (t3 == Type::TOP) return Type::TOP;
1871 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1872 #ifndef __STDC_IEC_559__
1873 return Type::DOUBLE;
1874 #else
1875 double d1 = t1->getd();
1876 double d2 = t2->getd();
1877 double d3 = t3->getd();
1878 return TypeD::make(fma(d1, d2, d3));
1879 #endif
1880 }
1881
1882 //=============================================================================
1883 //------------------------------Value------------------------------------------
1884 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1885 const Type *t1 = phase->type(in(1));
1886 if (t1 == Type::TOP) return Type::TOP;
1887 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1888 const Type *t2 = phase->type(in(2));
1889 if (t2 == Type::TOP) return Type::TOP;
1890 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1891 const Type *t3 = phase->type(in(3));
1892 if (t3 == Type::TOP) return Type::TOP;
1893 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1894 #ifndef __STDC_IEC_559__
1895 return Type::FLOAT;
1896 #else
1897 float f1 = t1->getf();
1898 float f2 = t2->getf();
1899 float f3 = t3->getf();
1900 return TypeF::make(fma(f1, f2, f3));
1901 #endif
1902 }
1903
1904 //=============================================================================
1905 //------------------------------hash-------------------------------------------
1906 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1907 // The hash function must return the same value when edge swapping is performed.
1908 uint MulAddS2INode::hash() const {
1909 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1910 }
1911
1912 //------------------------------Rotate Operations ------------------------------
1913
1914 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1915 const Type* t1 = phase->type(in(1));
1916 if (t1 == Type::TOP) {
1917 return this;
1918 }
1919 int count = 0;
1920 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1921 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1922 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1923 // Rotate by a multiple of 32/64 does nothing
1924 return in(1);
1925 }
1926 return this;
1927 }
1928
1929 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1930 const Type* t1 = phase->type(in(1));
1931 const Type* t2 = phase->type(in(2));
1932 // Either input is TOP ==> the result is TOP
1933 if (t1 == Type::TOP || t2 == Type::TOP) {
1934 return Type::TOP;
1935 }
1936
1937 if (t1->isa_int()) {
1938 const TypeInt* r1 = t1->is_int();
1939 const TypeInt* r2 = t2->is_int();
1940
1941 // Left input is ZERO ==> the result is ZERO.
1942 if (r1 == TypeInt::ZERO) {
1943 return TypeInt::ZERO;
1944 }
1945 // Rotate by zero does nothing
1946 if (r2 == TypeInt::ZERO) {
1947 return r1;
1948 }
1949 if (r1->is_con() && r2->is_con()) {
1950 juint r1_con = (juint)r1->get_con();
1951 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1952 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
1953 }
1954 return TypeInt::INT;
1955 } else {
1956 assert(t1->isa_long(), "Type must be a long");
1957 const TypeLong* r1 = t1->is_long();
1958 const TypeInt* r2 = t2->is_int();
1959
1960 // Left input is ZERO ==> the result is ZERO.
1961 if (r1 == TypeLong::ZERO) {
1962 return TypeLong::ZERO;
1963 }
1964 // Rotate by zero does nothing
1965 if (r2 == TypeInt::ZERO) {
1966 return r1;
1967 }
1968 if (r1->is_con() && r2->is_con()) {
1969 julong r1_con = (julong)r1->get_con();
1970 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
1971 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
1972 }
1973 return TypeLong::LONG;
1974 }
1975 }
1976
1977 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1978 const Type* t1 = phase->type(in(1));
1979 const Type* t2 = phase->type(in(2));
1980 if (t2->isa_int() && t2->is_int()->is_con()) {
1981 if (t1->isa_int()) {
1982 int lshift = t2->is_int()->get_con() & 31;
1983 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
1984 } else if (t1 != Type::TOP) {
1985 assert(t1->isa_long(), "Type must be a long");
1986 int lshift = t2->is_int()->get_con() & 63;
1987 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
1988 }
1989 }
1990 return nullptr;
1991 }
1992
1993 Node* RotateRightNode::Identity(PhaseGVN* phase) {
1994 const Type* t1 = phase->type(in(1));
1995 if (t1 == Type::TOP) {
1996 return this;
1997 }
1998 int count = 0;
1999 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2000 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
2001 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
2002 // Rotate by a multiple of 32/64 does nothing
2003 return in(1);
2004 }
2005 return this;
2006 }
2007
2008 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2009 const Type* t1 = phase->type(in(1));
2010 const Type* t2 = phase->type(in(2));
2011 // Either input is TOP ==> the result is TOP
2012 if (t1 == Type::TOP || t2 == Type::TOP) {
2013 return Type::TOP;
2014 }
2015
2016 if (t1->isa_int()) {
2017 const TypeInt* r1 = t1->is_int();
2018 const TypeInt* r2 = t2->is_int();
2019
2020 // Left input is ZERO ==> the result is ZERO.
2021 if (r1 == TypeInt::ZERO) {
2022 return TypeInt::ZERO;
2023 }
2024 // Rotate by zero does nothing
2025 if (r2 == TypeInt::ZERO) {
2026 return r1;
2027 }
2028 if (r1->is_con() && r2->is_con()) {
2029 juint r1_con = (juint)r1->get_con();
2030 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2031 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2032 }
2033 return TypeInt::INT;
2034 } else {
2035 assert(t1->isa_long(), "Type must be a long");
2036 const TypeLong* r1 = t1->is_long();
2037 const TypeInt* r2 = t2->is_int();
2038 // Left input is ZERO ==> the result is ZERO.
2039 if (r1 == TypeLong::ZERO) {
2040 return TypeLong::ZERO;
2041 }
2042 // Rotate by zero does nothing
2043 if (r2 == TypeInt::ZERO) {
2044 return r1;
2045 }
2046 if (r1->is_con() && r2->is_con()) {
2047 julong r1_con = (julong)r1->get_con();
2048 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2049 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2050 }
2051 return TypeLong::LONG;
2052 }
2053 }
2054
2055 // Given an expression (AndX shift mask) or (AndX mask shift),
2056 // determine if the AndX must always produce zero, because the
2057 // the shift (x<<N) is bitwise disjoint from the mask #M.
2058 // The X in AndX must be I or L, depending on bt.
2059 // Specifically, the following cases fold to zero,
2060 // when the shift value N is large enough to zero out
2061 // all the set positions of the and-mask M.
2062 // (AndI (LShiftI _ #N) #M) => #0
2063 // (AndL (LShiftL _ #N) #M) => #0
2064 // (AndL (ConvI2L (LShiftI _ #N)) #M) => #0
2065 // The M and N values must satisfy ((-1 << N) & M) == 0.
2066 // Because the optimization might work for a non-constant
2067 // mask M, we check the AndX for both operand orders.
2068 bool MulNode::AndIL_shift_and_mask_is_always_zero(PhaseGVN* phase, Node* shift, Node* mask, BasicType bt, bool check_reverse) {
2069 if (mask == nullptr || shift == nullptr) {
2070 return false;
2071 }
2072 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2073 if (mask_t == nullptr || phase->type(shift)->isa_integer(bt) == nullptr) {
2074 return false;
2075 }
2076 shift = shift->uncast();
2077 if (shift == nullptr) {
2078 return false;
2079 }
2080 if (phase->type(shift)->isa_integer(bt) == nullptr) {
2081 return false;
2082 }
2083 BasicType shift_bt = bt;
2084 if (bt == T_LONG && shift->Opcode() == Op_ConvI2L) {
2085 bt = T_INT;
2086 Node* val = shift->in(1);
2087 if (val == nullptr) {
2088 return false;
2089 }
2090 val = val->uncast();
2091 if (val == nullptr) {
2092 return false;
2093 }
2094 if (val->Opcode() == Op_LShiftI) {
2095 shift_bt = T_INT;
2096 shift = val;
2097 if (phase->type(shift)->isa_integer(bt) == nullptr) {
2098 return false;
2099 }
2100 }
2101 }
2102 if (shift->Opcode() != Op_LShift(shift_bt)) {
2103 if (check_reverse &&
2104 (mask->Opcode() == Op_LShift(bt) ||
2105 (bt == T_LONG && mask->Opcode() == Op_ConvI2L))) {
2106 // try it the other way around
2107 return AndIL_shift_and_mask_is_always_zero(phase, mask, shift, bt, false);
2108 }
2109 return false;
2110 }
2111 Node* shift2 = shift->in(2);
2112 if (shift2 == nullptr) {
2113 return false;
2114 }
2115 const Type* shift2_t = phase->type(shift2);
2116 if (!shift2_t->isa_int() || !shift2_t->is_int()->is_con()) {
2117 return false;
2118 }
2119
2120 jint shift_con = shift2_t->is_int()->get_con() & ((shift_bt == T_INT ? BitsPerJavaInteger : BitsPerJavaLong) - 1);
2121 if ((((jlong)1) << shift_con) > mask_t->hi_as_long() && mask_t->lo_as_long() >= 0) {
2122 return true;
2123 }
2124
2125 return false;
2126 }
2127
2128 // Given an expression (AndX (AddX v1 (LShiftX v2 #N)) #M)
2129 // determine if the AndX must always produce (AndX v1 #M),
2130 // because the shift (v2<<N) is bitwise disjoint from the mask #M.
2131 // The X in AndX will be I or L, depending on bt.
2132 // Specifically, the following cases fold,
2133 // when the shift value N is large enough to zero out
2134 // all the set positions of the and-mask M.
2135 // (AndI (AddI v1 (LShiftI _ #N)) #M) => (AndI v1 #M)
2136 // (AndL (AddI v1 (LShiftL _ #N)) #M) => (AndL v1 #M)
2137 // (AndL (AddL v1 (ConvI2L (LShiftI _ #N))) #M) => (AndL v1 #M)
2138 // The M and N values must satisfy ((-1 << N) & M) == 0.
2139 // Because the optimization might work for a non-constant
2140 // mask M, and because the AddX operands can come in either
2141 // order, we check for every operand order.
2142 Node* MulNode::AndIL_add_shift_and_mask(PhaseGVN* phase, BasicType bt) {
2143 Node* add = in(1);
2144 Node* mask = in(2);
2145 if (add == nullptr || mask == nullptr) {
2146 return nullptr;
2147 }
2148 int addidx = 0;
2149 if (add->Opcode() == Op_Add(bt)) {
2150 addidx = 1;
2151 } else if (mask->Opcode() == Op_Add(bt)) {
2152 mask = add;
2153 addidx = 2;
2154 add = in(addidx);
2155 }
2156 if (addidx > 0) {
2157 Node* add1 = add->in(1);
2158 Node* add2 = add->in(2);
2159 if (add1 != nullptr && add2 != nullptr) {
2160 if (AndIL_shift_and_mask_is_always_zero(phase, add1, mask, bt, false)) {
2161 set_req_X(addidx, add2, phase);
2162 return this;
2163 } else if (AndIL_shift_and_mask_is_always_zero(phase, add2, mask, bt, false)) {
2164 set_req_X(addidx, add1, phase);
2165 return this;
2166 }
2167 }
2168 }
2169 return nullptr;
2170 }