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