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 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
898
899 // We need to notify the caller that the graph was reshaped, as Ideal needs
900 // to return the root of the reshaped graph if any change was made.
901 return shift_node;
902 }
903 }
904
905 return nullptr;
906 }
907
908 // Called with
909 // outer_shift = (_ << rhs_outer)
910 // We are looking for the pattern:
911 // outer_shift = ((X << rhs_inner) << rhs_outer)
912 // where rhs_outer and rhs_inner are constant
913 // we denote inner_shift the nested expression (X << rhs_inner)
914 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
915 // where nbits is the number of bits of the shifts
916 //
917 // There are 2 cases:
918 // if con_outer + con_inner >= nbits => 0
919 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
920 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
921 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
922 const Node* inner_shift = outer_shift->in(1);
923 if (inner_shift->Opcode() != Op_LShift(bt)) {
924 return nullptr;
925 }
926
927 uint nbits = bits_per_java_integer(bt);
928 uint con_inner;
929 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
930 return nullptr;
931 }
932
933 if (con_inner == 0) {
934 // We let the Identity() of the inner shift do its job.
935 return nullptr;
936 }
937
938 if (con_outer + con_inner >= nbits) {
939 // While it might be tempting to use
940 // phase->zerocon(bt);
941 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
942 // to return a new node, this or nullptr, but not an old (cached) node.
943 return ConNode::make(TypeInteger::zero(bt));
944 }
945
946 // con0 + con1 < nbits ==> actual shift happens now
947 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
948 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
949 }
950
951 //------------------------------Identity---------------------------------------
952 Node* LShiftINode::Identity(PhaseGVN* phase) {
953 return IdentityIL(phase, T_INT);
954 }
955
956 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
957 uint con;
958 uint num_bits = bits_per_java_integer(bt);
959 Node* progress = mask_and_replace_shift_amount(phase, this, num_bits, con);
960 if (con == 0) {
961 return nullptr;
962 }
963
964 // If the right input is a constant, and the left input is an add of a
965 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
966 Node* add1 = in(1);
967 int add1_op = add1->Opcode();
968 if (add1_op == Op_Add(bt)) { // Left input is an add?
969 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
970
971 // Transform is legal, but check for profit. Avoid breaking 'i2s'
972 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
973 if (bt != T_INT || con < 16) {
974 // Left input is an add of the same number?
975 if (con != (num_bits - 1) && add1->in(1) == add1->in(2)) {
976 // Convert "(x + x) << c0" into "x << (c0 + 1)"
977 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
978 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
979 // or c0 != 63 (for LShiftL) because:
980 // (x + x) << 63 = 2x << 63, while
981 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
982 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
983 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
984 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
985 }
986
987 // Left input is an add of a constant?
988 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
989 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
990 // Compute X << con0
991 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
992 // Compute X<<con0 + (con1<<con0)
993 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
994 }
995 }
996 }
997 // Check for "(con0 - X) << con1"
998 // Transform is legal, but check for profit. Avoid breaking 'i2s'
999 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1000 if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) { // Left input is a sub?
1001 // Left input is a sub from a constant?
1002 const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
1003 if (t11 != nullptr && t11->is_con()) {
1004 // Compute X << con0
1005 Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
1006 // Compute (con1<<con0) - (X<<con0)
1007 return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
1008 }
1009 }
1010
1011 // Check for "(x >> C1) << C2"
1012 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1013 uint add1Con;
1014 mask_shift_amount(phase, add1, num_bits, add1Con);
1015
1016 // Special case C1 == C2, which just masks off low bits
1017 if (add1Con > 0 && con == add1Con) {
1018 // Convert to "(x & -(1 << C2))"
1019 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1020 } else {
1021 // Wait until the right shift has been sharpened to the correct count
1022 if (add1Con > 0) {
1023 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1024 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1025 if (phase->is_IterGVN()) {
1026 if (con > add1Con) {
1027 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1028 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1029 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1030 } else {
1031 assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1032 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1033
1034 // Handle logical and arithmetic shifts
1035 Node* rshift;
1036 if (add1_op == Op_RShift(bt)) {
1037 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1038 } else {
1039 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1040 }
1041
1042 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1043 }
1044 } else {
1045 phase->record_for_igvn(this);
1046 }
1047 }
1048 }
1049 }
1050
1051 // Check for "((x >> C1) & Y) << C2"
1052 if (add1_op == Op_And(bt)) {
1053 Node* add2 = add1->in(1);
1054 int add2_op = add2->Opcode();
1055 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1056 // Special case C1 == C2, which just masks off low bits
1057 if (add2->in(2) == in(2)) {
1058 // Convert to "(x & (Y << C2))"
1059 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1060 return MulNode::make_and(add2->in(1), y_sh, bt);
1061 }
1062
1063 uint add2Con;
1064 if (mask_shift_amount(phase, add2, num_bits, add2Con) && add2Con > 0) {
1065 if (phase->is_IterGVN()) {
1066 // Convert to "((x >> C1) << C2) & (Y << C2)"
1067
1068 // Make "(x >> C1) << C2", which will get folded away by the rule above
1069 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1070 // Make "Y << C2", which will simplify when Y is a constant
1071 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1072
1073 return MulNode::make_and(x_sh, y_sh, bt);
1074 } else {
1075 phase->record_for_igvn(this);
1076 }
1077 }
1078 }
1079 }
1080
1081 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1082 // before shifting them away.
1083 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1084 assert(bt != T_INT || bits_mask == right_n_bits(num_bits - con), "inconsistent");
1085 if (add1_op == Op_And(bt) &&
1086 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1087 return LShiftNode::make(add1->in(1), in(2), bt);
1088 }
1089
1090 // Collapse nested left-shifts with constant rhs:
1091 // (X << con1) << con2 ==> X << (con1 + con2)
1092 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1093 if (doubleShift != nullptr) {
1094 return doubleShift;
1095 }
1096
1097 return progress;
1098 }
1099
1100 //------------------------------Ideal------------------------------------------
1101 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1102 return IdealIL(phase, can_reshape, T_INT);
1103 }
1104
1105 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1106 const Type* t1 = phase->type(in(1));
1107 const Type* t2 = phase->type(in(2));
1108 // Either input is TOP ==> the result is TOP
1109 if (t1 == Type::TOP) {
1110 return Type::TOP;
1111 }
1112 if (t2 == Type::TOP) {
1113 return Type::TOP;
1114 }
1115
1116 // Left input is ZERO ==> the result is ZERO.
1117 if (t1 == TypeInteger::zero(bt)) {
1118 return TypeInteger::zero(bt);
1119 }
1120 // Shift by zero does nothing
1121 if (t2 == TypeInt::ZERO) {
1122 return t1;
1123 }
1124
1125 // If nothing is known about the shift amount then the result is BOTTOM
1126 if (t2 == TypeInt::INT) {
1127 return TypeInteger::bottom(bt);
1128 }
1129
1130 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1131 // Since the shift semantics in Java take into account only the bottom five
1132 // bits for ints and the bottom six bits for longs, we can further constrain
1133 // the range of values of the shift amount by ANDing with the right mask based
1134 // on whether the type is int or long.
1135 const TypeInt* mask = TypeInt::make(bits_per_java_integer(bt) - 1);
1136 const TypeInt* r2 = RangeInference::infer_and(t2->is_int(), mask);
1137
1138 if (!r2->is_con()) {
1139 return TypeInteger::bottom(bt);
1140 }
1141
1142 uint shift = r2->get_con();
1143 // Shift by a multiple of 32/64 does nothing:
1144 if (shift == 0) {
1145 return t1;
1146 }
1147
1148 // If the shift is a constant, shift the bounds of the type,
1149 // unless this could lead to an overflow.
1150 if (!r1->is_con()) {
1151 #ifdef ASSERT
1152 if (bt == T_INT) {
1153 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1154 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1155 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1156 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1157 }
1158 #endif
1159
1160 if (bt == T_INT) {
1161 return RangeInference::infer_lshift(r1->is_int(), shift);
1162 }
1163
1164 return RangeInference::infer_lshift(r1->is_long(), shift);
1165 }
1166
1167 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1168 }
1169
1170 //------------------------------Value------------------------------------------
1171 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1172 return ValueIL(phase, T_INT);
1173 }
1174
1175 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1176 uint count;
1177 if (mask_shift_amount(phase, this, bits_per_java_integer(bt), count) && count == 0) {
1178 // Shift by a multiple of 32/64 does nothing
1179 return in(1);
1180 }
1181 return this;
1182 }
1183
1184 //=============================================================================
1185 //------------------------------Identity---------------------------------------
1186 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1187 return IdentityIL(phase, T_LONG);
1188 }
1189
1190 //------------------------------Ideal------------------------------------------
1191 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1192 return IdealIL(phase, can_reshape, T_LONG);
1193 }
1194
1195 //------------------------------Value------------------------------------------
1196 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1197 return ValueIL(phase, T_LONG);
1198 }
1199
1200 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1201 switch (bt) {
1202 case T_INT:
1203 return new RShiftINode(in1, in2);
1204 case T_LONG:
1205 return new RShiftLNode(in1, in2);
1206 default:
1207 fatal("Not implemented for %s", type2name(bt));
1208 }
1209 return nullptr;
1210 }
1211
1212
1213 //=============================================================================
1214 //------------------------------Identity---------------------------------------
1215 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1216 uint count;
1217 uint num_bits = bits_per_java_integer(bt);
1218 if (mask_shift_amount(phase, this, num_bits, count)) {
1219 if (count == 0) {
1220 // Shift by a multiple of 32/64 does nothing
1221 return in(1);
1222 }
1223 // Check for useless sign-masking
1224 uint lshift_count;
1225 if (in(1)->Opcode() == Op_LShift(bt) &&
1226 in(1)->req() == 3 &&
1227 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1228 // negative constant (e.g. -1 vs 31)
1229 mask_shift_amount(phase, in(1), num_bits, lshift_count)) {
1230 if (count == lshift_count) {
1231 // Compute masks for which this shifting doesn't change
1232 jlong lo = (CONST64(-1) << (num_bits - count - 1)); // FFFF8000
1233 jlong hi = ~lo; // 00007FFF
1234 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1235 if (t11 == nullptr) {
1236 return this;
1237 }
1238 // Does actual value fit inside of mask?
1239 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1240 return in(1)->in(1); // Then shifting is a nop
1241 }
1242 }
1243 }
1244 }
1245 return this;
1246 }
1247
1248 Node* RShiftINode::Identity(PhaseGVN* phase) {
1249 return IdentityIL(phase, T_INT);
1250 }
1251
1252 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1253 // Inputs may be TOP if they are dead.
1254 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1255 if (t1 == nullptr) {
1256 return NodeSentinel; // Left input is an integer
1257 }
1258
1259 uint shift;
1260 Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), shift);
1261 if (shift == 0) {
1262 return NodeSentinel;
1263 }
1264
1265 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1266 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1267 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1268 const Node* and_node = in(1);
1269 if (and_node->Opcode() != Op_And(bt)) {
1270 return progress;
1271 }
1272 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1273 if (mask_t != nullptr && mask_t->is_con()) {
1274 jlong maskbits = mask_t->get_con_as_long(bt);
1275 // Convert to "(x >> shift) & (mask >> shift)"
1276 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1277 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1278 }
1279
1280 return progress;
1281 }
1282
1283 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1284 Node* progress = IdealIL(phase, can_reshape, T_INT);
1285 if (progress == NodeSentinel) {
1286 return nullptr;
1287 }
1288 if (progress != nullptr) {
1289 return progress;
1290 }
1291 uint shift;
1292 progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, shift);
1293 assert(shift != 0, "handled by IdealIL");
1294
1295 // Check for "(short[i] <<16)>>16" which simply sign-extends
1296 const Node *shl = in(1);
1297 if (shl->Opcode() != Op_LShiftI) {
1298 return progress;
1299 }
1300
1301 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1302 if (left_shift_t == nullptr) {
1303 return progress;
1304 }
1305 if (shift == 16 && left_shift_t->is_con(16)) {
1306 Node *ld = shl->in(1);
1307 if (ld->Opcode() == Op_LoadS) {
1308 // Sign extension is just useless here. Return a RShiftI of zero instead
1309 // returning 'ld' directly. We cannot return an old Node directly as
1310 // that is the job of 'Identity' calls and Identity calls only work on
1311 // direct inputs ('ld' is an extra Node removed from 'this'). The
1312 // combined optimization requires Identity only return direct inputs.
1313 set_req_X(1, ld, phase);
1314 set_req_X(2, phase->intcon(0), phase);
1315 return this;
1316 }
1317 else if (can_reshape &&
1318 ld->Opcode() == Op_LoadUS &&
1319 ld->outcnt() == 1 && ld->unique_out() == shl)
1320 // Replace zero-extension-load with sign-extension-load
1321 return ld->as_Load()->convert_to_signed_load(*phase);
1322 }
1323
1324 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1325 if (shift == 24 && left_shift_t->is_con(24)) {
1326 Node *ld = shl->in(1);
1327 if (ld->Opcode() == Op_LoadB) {
1328 // Sign extension is just useless here
1329 set_req_X(1, ld, phase);
1330 set_req_X(2, phase->intcon(0), phase);
1331 return this;
1332 }
1333 }
1334
1335 return progress;
1336 }
1337
1338 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1339 const Type* t1 = phase->type(in(1));
1340 const Type* t2 = phase->type(in(2));
1341 // Either input is TOP ==> the result is TOP
1342 if (t1 == Type::TOP) {
1343 return Type::TOP;
1344 }
1345 if (t2 == Type::TOP) {
1346 return Type::TOP;
1347 }
1348
1349 // Left input is ZERO ==> the result is ZERO.
1350 if (t1 == TypeInteger::zero(bt)) {
1351 return TypeInteger::zero(bt);
1352 }
1353 // Shift by zero does nothing
1354 if (t2 == TypeInt::ZERO) {
1355 return t1;
1356 }
1357
1358 // Either input is BOTTOM ==> the result is BOTTOM
1359 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1360 return TypeInteger::bottom(bt);
1361 }
1362
1363 const TypeInteger* r1 = t1->isa_integer(bt);
1364 const TypeInt* r2 = t2->isa_int();
1365
1366 // If the shift is a constant, just shift the bounds of the type.
1367 // For example, if the shift is 31/63, we just propagate sign bits.
1368 if (!r1->is_con() && r2->is_con()) {
1369 uint shift = r2->get_con();
1370 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1371 // Shift by a multiple of 32/64 does nothing:
1372 if (shift == 0) {
1373 return t1;
1374 }
1375 // Calculate reasonably aggressive bounds for the result.
1376 // This is necessary if we are to correctly type things
1377 // like (x<<24>>24) == ((byte)x).
1378 jlong lo = r1->lo_as_long() >> (jint)shift;
1379 jlong hi = r1->hi_as_long() >> (jint)shift;
1380 assert(lo <= hi, "must have valid bounds");
1381 #ifdef ASSERT
1382 if (bt == T_INT) {
1383 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1384 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1385 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1386 }
1387 #endif
1388 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1389 #ifdef ASSERT
1390 // Make sure we get the sign-capture idiom correct.
1391 if (shift == bits_per_java_integer(bt) - 1) {
1392 if (r1->lo_as_long() >= 0) {
1393 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1394 }
1395 if (r1->hi_as_long() < 0) {
1396 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1397 }
1398 }
1399 #endif
1400 return ti;
1401 }
1402
1403 if (!r1->is_con() || !r2->is_con()) {
1404 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1405 if (r1->lo_as_long() >= 0) {
1406 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1407 }
1408
1409 // Conversely, if the left input is negative then the result must be negative.
1410 if (r1->hi_as_long() <= -1) {
1411 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1412 }
1413
1414 return TypeInteger::bottom(bt);
1415 }
1416
1417 // Signed shift right
1418 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1419 }
1420
1421 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1422 return ValueIL(phase, T_INT);
1423 }
1424
1425 //=============================================================================
1426 //------------------------------Identity---------------------------------------
1427 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1428 return IdentityIL(phase, T_LONG);
1429 }
1430
1431 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1432 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1433 if (progress == NodeSentinel) {
1434 return nullptr;
1435 }
1436 return progress;
1437 }
1438
1439 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1440 return ValueIL(phase, T_LONG);
1441 }
1442
1443 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1444 switch (bt) {
1445 case T_INT:
1446 return new URShiftINode(in1, in2);
1447 case T_LONG:
1448 return new URShiftLNode(in1, in2);
1449 default:
1450 fatal("Not implemented for %s", type2name(bt));
1451 }
1452 return nullptr;
1453 }
1454
1455 //=============================================================================
1456 //------------------------------Identity---------------------------------------
1457 Node* URShiftINode::Identity(PhaseGVN* phase) {
1458 uint count;
1459 if (mask_shift_amount(phase, this, BitsPerJavaInteger, count) && count == 0) {
1460 // Shift by a multiple of 32 does nothing
1461 return in(1);
1462 }
1463
1464 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1465 // Happens during new-array length computation.
1466 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1467 Node *add = in(1);
1468 if (add->Opcode() == Op_AddI) {
1469 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1470 if (t2 && t2->is_con(wordSize - 1) &&
1471 add->in(1)->Opcode() == Op_LShiftI) {
1472 // Check that shift_counts are LogBytesPerWord.
1473 Node *lshift_count = add->in(1)->in(2);
1474 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1475 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1476 t_lshift_count == phase->type(in(2))) {
1477 Node *x = add->in(1)->in(1);
1478 const TypeInt *t_x = phase->type(x)->isa_int();
1479 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1480 return x;
1481 }
1482 }
1483 }
1484 }
1485
1486 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1487 }
1488
1489 //------------------------------Ideal------------------------------------------
1490 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1491 uint con;
1492 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, con);
1493 if (con == 0) {
1494 return nullptr;
1495 }
1496
1497 // We'll be wanting the right-shift amount as a mask of that many bits
1498 const int mask = right_n_bits(BitsPerJavaInteger - con);
1499
1500 int in1_op = in(1)->Opcode();
1501
1502 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1503 if( in1_op == Op_URShiftI ) {
1504 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1505 if( t12 && t12->is_con() ) { // Right input is a constant
1506 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1507 const int con2 = t12->get_con() & 31; // Shift count is always masked
1508 const int con3 = con+con2;
1509 if( con3 < 32 ) // Only merge shifts if total is < 32
1510 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1511 }
1512 }
1513
1514 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1515 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1516 // If Q is "X << z" the rounding is useless. Look for patterns like
1517 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1518 Node *add = in(1);
1519 if (in1_op == Op_AddI) {
1520 Node *lshl = add->in(1);
1521 Node *y = add->in(2);
1522 if (lshl->Opcode() != Op_LShiftI) {
1523 lshl = add->in(2);
1524 y = add->in(1);
1525 }
1526 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1527 // negative constant (e.g. -1 vs 31)
1528 uint lshl_con;
1529 if (lshl->Opcode() == Op_LShiftI &&
1530 mask_shift_amount(phase, lshl, BitsPerJavaInteger, lshl_con) &&
1531 lshl_con == con) {
1532 Node *y_z = phase->transform(new URShiftINode(y, in(2)));
1533 Node *sum = phase->transform(new AddINode(lshl->in(1), y_z));
1534 return new AndINode(sum, phase->intcon(mask));
1535 }
1536 }
1537
1538 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1539 // This shortens the mask. Also, if we are extracting a high byte and
1540 // storing it to a buffer, the mask will be removed completely.
1541 Node *andi = in(1);
1542 if( in1_op == Op_AndI ) {
1543 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1544 if( t3 && t3->is_con() ) { // Right input is a constant
1545 jint mask2 = t3->get_con();
1546 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1547 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1548 return new AndINode(newshr, phase->intcon(mask2));
1549 // The negative values are easier to materialize than positive ones.
1550 // A typical case from address arithmetic is ((x & ~15) >> 4).
1551 // It's better to change that to ((x >> 4) & ~0) versus
1552 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
1553 }
1554 }
1555
1556 // Check for "(X << z ) >>> z" which simply zero-extends
1557 Node *shl = in(1);
1558 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1559 // negative constant (e.g. -1 vs 31)
1560 uint shl_con;
1561 if (in1_op == Op_LShiftI &&
1562 mask_shift_amount(phase, shl, BitsPerJavaInteger, shl_con) &&
1563 shl_con == con)
1564 return new AndINode(shl->in(1), phase->intcon(mask));
1565
1566 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1567 const TypeInt* t2 = phase->type(in(2))->isa_int();
1568 Node *shr = in(1);
1569 if ( in1_op == Op_RShiftI ) {
1570 Node *in11 = shr->in(1);
1571 Node *in12 = shr->in(2);
1572 const TypeInt *t11 = phase->type(in11)->isa_int();
1573 const TypeInt *t12 = phase->type(in12)->isa_int();
1574 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1575 return new URShiftINode(in11, phase->intcon(31));
1576 }
1577 }
1578
1579 return progress;
1580 }
1581
1582 //------------------------------Value------------------------------------------
1583 // A URShiftINode shifts its input2 right by input1 amount.
1584 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1585 // (This is a near clone of RShiftINode::Value.)
1586 const Type *t1 = phase->type( in(1) );
1587 const Type *t2 = phase->type( in(2) );
1588 // Either input is TOP ==> the result is TOP
1589 if( t1 == Type::TOP ) return Type::TOP;
1590 if( t2 == Type::TOP ) return Type::TOP;
1591
1592 // Left input is ZERO ==> the result is ZERO.
1593 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1594 // Shift by zero does nothing
1595 if( t2 == TypeInt::ZERO ) return t1;
1596
1597 // Either input is BOTTOM ==> the result is BOTTOM
1598 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1599 return TypeInt::INT;
1600
1601 if (t2 == TypeInt::INT)
1602 return TypeInt::INT;
1603
1604 const TypeInt *r1 = t1->is_int(); // Handy access
1605 const TypeInt *r2 = t2->is_int(); // Handy access
1606
1607 if (r2->is_con()) {
1608 uint shift = r2->get_con();
1609 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1610 // Shift by a multiple of 32 does nothing:
1611 if (shift == 0) return t1;
1612 // Calculate reasonably aggressive bounds for the result.
1613 jint lo = (juint)r1->_lo >> (juint)shift;
1614 jint hi = (juint)r1->_hi >> (juint)shift;
1615 if (r1->_hi >= 0 && r1->_lo < 0) {
1616 // If the type has both negative and positive values,
1617 // there are two separate sub-domains to worry about:
1618 // The positive half and the negative half.
1619 jint neg_lo = lo;
1620 jint neg_hi = (juint)-1 >> (juint)shift;
1621 jint pos_lo = (juint) 0 >> (juint)shift;
1622 jint pos_hi = hi;
1623 lo = MIN2(neg_lo, pos_lo); // == 0
1624 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1625 }
1626 assert(lo <= hi, "must have valid bounds");
1627 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1628 #ifdef ASSERT
1629 // Make sure we get the sign-capture idiom correct.
1630 if (shift == BitsPerJavaInteger-1) {
1631 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1632 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1633 }
1634 #endif
1635 return ti;
1636 }
1637
1638 //
1639 // Do not support shifted oops in info for GC
1640 //
1641 // else if( t1->base() == Type::InstPtr ) {
1642 //
1643 // const TypeInstPtr *o = t1->is_instptr();
1644 // if( t1->singleton() )
1645 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1646 // }
1647 // else if( t1->base() == Type::KlassPtr ) {
1648 // const TypeKlassPtr *o = t1->is_klassptr();
1649 // if( t1->singleton() )
1650 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1651 // }
1652
1653 return TypeInt::INT;
1654 }
1655
1656 //=============================================================================
1657 //------------------------------Identity---------------------------------------
1658 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1659 uint count;
1660 if (mask_shift_amount(phase, this, BitsPerJavaLong, count) && count == 0) {
1661 // Shift by a multiple of 64 does nothing
1662 return in(1);
1663 }
1664 return this;
1665 }
1666
1667 //------------------------------Ideal------------------------------------------
1668 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1669 uint con;
1670 Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong, con);
1671 if (con == 0) {
1672 return nullptr;
1673 }
1674
1675 // We'll be wanting the right-shift amount as a mask of that many bits
1676 const jlong mask = jlong(max_julong >> con);
1677
1678 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1679 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1680 // If Q is "X << z" the rounding is useless. Look for patterns like
1681 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1682 Node *add = in(1);
1683 const TypeInt *t2 = phase->type(in(2))->isa_int();
1684 if (add->Opcode() == Op_AddL) {
1685 Node *lshl = add->in(1);
1686 Node *y = add->in(2);
1687 if (lshl->Opcode() != Op_LShiftL) {
1688 lshl = add->in(2);
1689 y = add->in(1);
1690 }
1691 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1692 // negative constant (e.g. -1 vs 63)
1693 uint lshl_con;
1694 if (lshl->Opcode() == Op_LShiftL &&
1695 mask_shift_amount(phase, lshl, BitsPerJavaLong, lshl_con) &&
1696 lshl_con == con) {
1697 Node* y_z = phase->transform(new URShiftLNode(y, in(2)));
1698 Node* sum = phase->transform(new AddLNode(lshl->in(1), y_z));
1699 return new AndLNode(sum, phase->longcon(mask));
1700 }
1701 }
1702
1703 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1704 // This shortens the mask. Also, if we are extracting a high byte and
1705 // storing it to a buffer, the mask will be removed completely.
1706 Node *andi = in(1);
1707 if( andi->Opcode() == Op_AndL ) {
1708 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1709 if( t3 && t3->is_con() ) { // Right input is a constant
1710 jlong mask2 = t3->get_con();
1711 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1712 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1713 return new AndLNode(newshr, phase->longcon(mask2));
1714 }
1715 }
1716
1717 // Check for "(X << z ) >>> z" which simply zero-extends
1718 Node *shl = in(1);
1719 // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1720 // negative constant (e.g. -1 vs 63)
1721 uint shl_con;
1722 if (shl->Opcode() == Op_LShiftL &&
1723 mask_shift_amount(phase, shl, BitsPerJavaLong, shl_con) &&
1724 shl_con == con) {
1725 return new AndLNode(shl->in(1), phase->longcon(mask));
1726 }
1727
1728 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1729 Node *shr = in(1);
1730 if ( shr->Opcode() == Op_RShiftL ) {
1731 Node *in11 = shr->in(1);
1732 Node *in12 = shr->in(2);
1733 const TypeLong *t11 = phase->type(in11)->isa_long();
1734 const TypeInt *t12 = phase->type(in12)->isa_int();
1735 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1736 return new URShiftLNode(in11, phase->intcon(63));
1737 }
1738 }
1739
1740 return progress;
1741 }
1742
1743 //------------------------------Value------------------------------------------
1744 // A URShiftINode shifts its input2 right by input1 amount.
1745 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1746 // (This is a near clone of RShiftLNode::Value.)
1747 const Type *t1 = phase->type( in(1) );
1748 const Type *t2 = phase->type( in(2) );
1749 // Either input is TOP ==> the result is TOP
1750 if( t1 == Type::TOP ) return Type::TOP;
1751 if( t2 == Type::TOP ) return Type::TOP;
1752
1753 // Left input is ZERO ==> the result is ZERO.
1754 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1755 // Shift by zero does nothing
1756 if( t2 == TypeInt::ZERO ) return t1;
1757
1758 // Either input is BOTTOM ==> the result is BOTTOM
1759 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1760 return TypeLong::LONG;
1761
1762 if (t2 == TypeInt::INT)
1763 return TypeLong::LONG;
1764
1765 const TypeLong *r1 = t1->is_long(); // Handy access
1766 const TypeInt *r2 = t2->is_int (); // Handy access
1767
1768 if (r2->is_con()) {
1769 uint shift = r2->get_con();
1770 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1771 // Shift by a multiple of 64 does nothing:
1772 if (shift == 0) return t1;
1773 // Calculate reasonably aggressive bounds for the result.
1774 jlong lo = (julong)r1->_lo >> (juint)shift;
1775 jlong hi = (julong)r1->_hi >> (juint)shift;
1776 if (r1->_hi >= 0 && r1->_lo < 0) {
1777 // If the type has both negative and positive values,
1778 // there are two separate sub-domains to worry about:
1779 // The positive half and the negative half.
1780 jlong neg_lo = lo;
1781 jlong neg_hi = (julong)-1 >> (juint)shift;
1782 jlong pos_lo = (julong) 0 >> (juint)shift;
1783 jlong pos_hi = hi;
1784 //lo = MIN2(neg_lo, pos_lo); // == 0
1785 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1786 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1787 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1788 }
1789 assert(lo <= hi, "must have valid bounds");
1790 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1791 #ifdef ASSERT
1792 // Make sure we get the sign-capture idiom correct.
1793 if (shift == BitsPerJavaLong - 1) {
1794 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1795 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1796 }
1797 #endif
1798 return tl;
1799 }
1800
1801 return TypeLong::LONG; // Give up
1802 }
1803
1804 //=============================================================================
1805 //------------------------------Ideal------------------------------------------
1806 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1807 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1808 // This reduces the number of rules in the matcher, as we only need to check
1809 // for negations on the second argument, and not the symmetric case where
1810 // the first argument is negated.
1811 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1812 swap_edges(1, 2);
1813 return this;
1814 }
1815 return nullptr;
1816 }
1817
1818 //=============================================================================
1819 //------------------------------Value------------------------------------------
1820 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1821 const Type *t1 = phase->type(in(1));
1822 if (t1 == Type::TOP) return Type::TOP;
1823 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1824 const Type *t2 = phase->type(in(2));
1825 if (t2 == Type::TOP) return Type::TOP;
1826 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1827 const Type *t3 = phase->type(in(3));
1828 if (t3 == Type::TOP) return Type::TOP;
1829 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1830 #ifndef __STDC_IEC_559__
1831 return Type::DOUBLE;
1832 #else
1833 double d1 = t1->getd();
1834 double d2 = t2->getd();
1835 double d3 = t3->getd();
1836 return TypeD::make(fma(d1, d2, d3));
1837 #endif
1838 }
1839
1840 //=============================================================================
1841 //------------------------------Value------------------------------------------
1842 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1843 const Type *t1 = phase->type(in(1));
1844 if (t1 == Type::TOP) return Type::TOP;
1845 if (t1->base() != Type::FloatCon) return Type::FLOAT;
1846 const Type *t2 = phase->type(in(2));
1847 if (t2 == Type::TOP) return Type::TOP;
1848 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1849 const Type *t3 = phase->type(in(3));
1850 if (t3 == Type::TOP) return Type::TOP;
1851 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1852 #ifndef __STDC_IEC_559__
1853 return Type::FLOAT;
1854 #else
1855 float f1 = t1->getf();
1856 float f2 = t2->getf();
1857 float f3 = t3->getf();
1858 return TypeF::make(fma(f1, f2, f3));
1859 #endif
1860 }
1861
1862 //=============================================================================
1863 //------------------------------Value------------------------------------------
1864 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
1865 const Type* t1 = phase->type(in(1));
1866 if (t1 == Type::TOP) { return Type::TOP; }
1867 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1868 const Type* t2 = phase->type(in(2));
1869 if (t2 == Type::TOP) { return Type::TOP; }
1870 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1871 const Type* t3 = phase->type(in(3));
1872 if (t3 == Type::TOP) { return Type::TOP; }
1873 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1874 #ifndef __STDC_IEC_559__
1875 return Type::HALF_FLOAT;
1876 #else
1877 float f1 = t1->getf();
1878 float f2 = t2->getf();
1879 float f3 = t3->getf();
1880 return TypeH::make(fma(f1, f2, f3));
1881 #endif
1882 }
1883
1884 //=============================================================================
1885 //------------------------------hash-------------------------------------------
1886 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1887 // The hash function must return the same value when edge swapping is performed.
1888 uint MulAddS2INode::hash() const {
1889 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1890 }
1891
1892 //------------------------------Rotate Operations ------------------------------
1893
1894 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1895 const Type* t1 = phase->type(in(1));
1896 if (t1 == Type::TOP) {
1897 return this;
1898 }
1899 uint count;
1900 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1901 uint num_bits = t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong;
1902 if (mask_shift_amount(phase, this, num_bits, count) && count == 0) {
1903 // Rotate by a multiple of 32/64 does nothing
1904 return in(1);
1905 }
1906 return this;
1907 }
1908
1909 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1910 const Type* t1 = phase->type(in(1));
1911 const Type* t2 = phase->type(in(2));
1912 // Either input is TOP ==> the result is TOP
1913 if (t1 == Type::TOP || t2 == Type::TOP) {
1914 return Type::TOP;
1915 }
1916
1917 if (t1->isa_int()) {
1918 const TypeInt* r1 = t1->is_int();
1919 const TypeInt* r2 = t2->is_int();
1920
1921 // Left input is ZERO ==> the result is ZERO.
1922 if (r1 == TypeInt::ZERO) {
1923 return TypeInt::ZERO;
1924 }
1925 // Rotate by zero does nothing
1926 if (r2 == TypeInt::ZERO) {
1927 return r1;
1928 }
1929 if (r1->is_con() && r2->is_con()) {
1930 juint r1_con = (juint)r1->get_con();
1931 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1932 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
1933 }
1934 return TypeInt::INT;
1935 } else {
1936 assert(t1->isa_long(), "Type must be a long");
1937 const TypeLong* r1 = t1->is_long();
1938 const TypeInt* r2 = t2->is_int();
1939
1940 // Left input is ZERO ==> the result is ZERO.
1941 if (r1 == TypeLong::ZERO) {
1942 return TypeLong::ZERO;
1943 }
1944 // Rotate by zero does nothing
1945 if (r2 == TypeInt::ZERO) {
1946 return r1;
1947 }
1948 if (r1->is_con() && r2->is_con()) {
1949 julong r1_con = (julong)r1->get_con();
1950 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
1951 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
1952 }
1953 return TypeLong::LONG;
1954 }
1955 }
1956
1957 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1958 const Type* t1 = phase->type(in(1));
1959 const Type* t2 = phase->type(in(2));
1960 if (t2->isa_int() && t2->is_int()->is_con()) {
1961 if (t1->isa_int()) {
1962 int lshift = t2->is_int()->get_con() & 31;
1963 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
1964 } else if (t1 != Type::TOP) {
1965 assert(t1->isa_long(), "Type must be a long");
1966 int lshift = t2->is_int()->get_con() & 63;
1967 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
1968 }
1969 }
1970 return nullptr;
1971 }
1972
1973 Node* RotateRightNode::Identity(PhaseGVN* phase) {
1974 const Type* t1 = phase->type(in(1));
1975 if (t1 == Type::TOP) {
1976 return this;
1977 }
1978 uint count;
1979 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1980 uint num_bits = t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong;
1981 if (mask_shift_amount(phase, this, num_bits, count) && count == 0) {
1982 // Rotate by a multiple of 32/64 does nothing
1983 return in(1);
1984 }
1985 return this;
1986 }
1987
1988 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
1989 const Type* t1 = phase->type(in(1));
1990 const Type* t2 = phase->type(in(2));
1991 // Either input is TOP ==> the result is TOP
1992 if (t1 == Type::TOP || t2 == Type::TOP) {
1993 return Type::TOP;
1994 }
1995
1996 if (t1->isa_int()) {
1997 const TypeInt* r1 = t1->is_int();
1998 const TypeInt* r2 = t2->is_int();
1999
2000 // Left input is ZERO ==> the result is ZERO.
2001 if (r1 == TypeInt::ZERO) {
2002 return TypeInt::ZERO;
2003 }
2004 // Rotate by zero does nothing
2005 if (r2 == TypeInt::ZERO) {
2006 return r1;
2007 }
2008 if (r1->is_con() && r2->is_con()) {
2009 juint r1_con = (juint)r1->get_con();
2010 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2011 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2012 }
2013 return TypeInt::INT;
2014 } else {
2015 assert(t1->isa_long(), "Type must be a long");
2016 const TypeLong* r1 = t1->is_long();
2017 const TypeInt* r2 = t2->is_int();
2018 // Left input is ZERO ==> the result is ZERO.
2019 if (r1 == TypeLong::ZERO) {
2020 return TypeLong::ZERO;
2021 }
2022 // Rotate by zero does nothing
2023 if (r2 == TypeInt::ZERO) {
2024 return r1;
2025 }
2026 if (r1->is_con() && r2->is_con()) {
2027 julong r1_con = (julong)r1->get_con();
2028 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2029 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2030 }
2031 return TypeLong::LONG;
2032 }
2033 }
2034
2035 //------------------------------ Sum & Mask ------------------------------
2036
2037 // Returns a lower bound on the number of trailing zeros in expr.
2038 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2039 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2040 if (type == nullptr) {
2041 return 0;
2042 }
2043
2044 expr = expr->uncast();
2045 type = phase->type(expr)->isa_integer(bt);
2046 if (type == nullptr) {
2047 return 0;
2048 }
2049
2050 if (type->is_con()) {
2051 jlong con = type->get_con_as_long(bt);
2052 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2053 }
2054
2055 if (expr->Opcode() == Op_ConvI2L) {
2056 expr = expr->in(1)->uncast();
2057 bt = T_INT;
2058 type = phase->type(expr)->isa_int();
2059 }
2060
2061 // Pattern: expr = (x << shift)
2062 if (expr->Opcode() == Op_LShift(bt)) {
2063 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2064 if (shift_t == nullptr || !shift_t->is_con()) {
2065 return 0;
2066 }
2067 // We need to truncate the shift, as it may not have been canonicalized yet.
2068 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2069 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2070 // (JLS: "Shift Operators")
2071 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2072 return shift_t->get_con() & shift_mask;
2073 }
2074
2075 return 0;
2076 }
2077
2078 // Checks whether expr is neutral additive element (zero) under mask,
2079 // i.e. whether an expression of the form:
2080 // (AndX (AddX (expr addend) mask)
2081 // (expr + addend) & mask
2082 // is equivalent to
2083 // (AndX addend mask)
2084 // addend & mask
2085 // for any addend.
2086 // (The X in AndX must be I or L, depending on bt).
2087 //
2088 // We check for the sufficient condition when the lowest set bit in expr is higher than
2089 // the highest set bit in mask, i.e.:
2090 // expr: eeeeee0000000000000
2091 // mask: 000000mmmmmmmmmmmmm
2092 // <--w bits--->
2093 // We do not test for other cases.
2094 //
2095 // Correctness:
2096 // Given "expr" with at least "w" trailing zeros,
2097 // let "mod = 2^w", "suffix_mask = mod - 1"
2098 //
2099 // Since "mask" only has bits set where "suffix_mask" does, we have:
2100 // mask = suffix_mask & mask (SUFFIX_MASK)
2101 //
2102 // And since expr only has bits set above w, and suffix_mask only below:
2103 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2104 //
2105 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2106 // a power of 2, and we are computing in a ring of powers of 2, we know that
2107 // (x + y) % mod = (x % mod + y) % mod
2108 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2109 //
2110 // We can now prove the equality:
2111 // (expr + addend) & mask
2112 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2113 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2114 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2115 // = addend & mask (SUFFIX_MASK)
2116 //
2117 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2118 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2119 // When the mask is negative, it has the most significant bit set.
2120 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2121 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2122 return false;
2123 }
2124
2125 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2126 if (mask_t->hi_as_long() == 0) {
2127 assert(mask_t->lo_as_long() == 0, "checked earlier");
2128 return false;
2129 }
2130
2131 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2132 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2133 return expr_trailing_zeros >= mask_bit_width;
2134 }
2135
2136 // Reduces the pattern:
2137 // (AndX (AddX add1 add2) mask)
2138 // to
2139 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2140 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2141 Node* add = in(1);
2142 Node* mask = in(2);
2143 int addidx = 0;
2144 if (add->Opcode() == Op_Add(bt)) {
2145 addidx = 1;
2146 } else if (mask->Opcode() == Op_Add(bt)) {
2147 mask = add;
2148 addidx = 2;
2149 add = in(addidx);
2150 }
2151 if (addidx > 0) {
2152 Node* add1 = add->in(1);
2153 Node* add2 = add->in(2);
2154 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2155 set_req_X(addidx, add2, phase);
2156 return this;
2157 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2158 set_req_X(addidx, add1, phase);
2159 return this;
2160 }
2161 }
2162 return nullptr;
2163 }