1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "memory/allocation.inline.hpp"
26 #include "opto/addnode.hpp"
27 #include "opto/connode.hpp"
28 #include "opto/convertnode.hpp"
29 #include "opto/memnode.hpp"
30 #include "opto/mulnode.hpp"
31 #include "opto/phaseX.hpp"
32 #include "opto/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 Node* in1 = in(1);
655 uint op = in1->Opcode();
656 const TypeInt* t2 = phase->type(in(2))->isa_int();
657 if (t2 && t2->is_con()) {
658 int con = t2->get_con();
659 // Masking off high bits which are always zero is useless.
660 const TypeInt* t1 = phase->type(in(1))->isa_int();
661 if (t1 != nullptr && t1->_lo >= 0) {
662 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi));
663 if ((t1_support & con) == t1_support)
664 return in1;
665 }
666 // Masking off the high bits of a unsigned-shift-right is not
667 // needed either.
668 if (op == Op_URShiftI) {
669 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
670 if (t12 && t12->is_con()) { // Shift is by a constant
671 int shift = t12->get_con();
672 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
673 int mask = max_juint >> shift;
674 if ((mask & con) == mask) // If AND is useless, skip it
675 return in1;
676 }
677 }
678 }
679 return MulNode::Identity(phase);
680 }
681
682 //------------------------------Ideal------------------------------------------
683 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
684 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
685 Node* progress = AndIL_sum_and_mask(phase, T_INT);
686 if (progress != nullptr) {
687 return progress;
688 }
689
690 // Convert "(~a) & (~b)" into "~(a | b)"
691 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
692 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
693 Node* tn = phase->transform(or_a_b);
694 return AddNode::make_not(phase, tn, T_INT);
695 }
696
697 // Special case constant AND mask
698 const TypeInt *t2 = phase->type( in(2) )->isa_int();
699 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
700 const int mask = t2->get_con();
701 Node *load = in(1);
702 uint lop = load->Opcode();
703
704 // Masking bits off of a Character? Hi bits are already zero.
705 if( lop == Op_LoadUS &&
706 (mask & 0xFFFF0000) ) // Can we make a smaller mask?
707 return new AndINode(load,phase->intcon(mask&0xFFFF));
708
709 // Masking bits off of a Short? Loading a Character does some masking
710 if (can_reshape &&
711 load->outcnt() == 1 && load->unique_out() == this) {
712 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
713 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
714 ldus = phase->transform(ldus);
715 return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
716 }
717
718 // Masking sign bits off of a Byte? Do an unsigned byte load plus
719 // an and.
720 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
721 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
722 ldub = phase->transform(ldub);
723 return new AndINode(ldub, phase->intcon(mask));
724 }
725 }
726
727 // Masking off sign bits? Dont make them!
728 if( lop == Op_RShiftI ) {
729 const TypeInt *t12 = phase->type(load->in(2))->isa_int();
730 if( t12 && t12->is_con() ) { // Shift is by a constant
731 int shift = t12->get_con();
732 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
733 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
734 // If the AND'ing of the 2 masks has no bits, then only original shifted
735 // bits survive. NO sign-extension bits survive the maskings.
736 if( (sign_bits_mask & mask) == 0 ) {
737 // Use zero-fill shift instead
738 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
739 return new AndINode( zshift, in(2) );
740 }
741 }
742 }
743
744 // Check for 'negate/and-1', a pattern emitted when someone asks for
745 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement
746 // plus 1) and the mask is of the low order bit. Skip the negate.
747 if( lop == Op_SubI && mask == 1 && load->in(1) &&
748 phase->type(load->in(1)) == TypeInt::ZERO )
749 return new AndINode( load->in(2), in(2) );
750
751 return MulNode::Ideal(phase, can_reshape);
752 }
753
754 //=============================================================================
755 //------------------------------mul_ring---------------------------------------
756 // Supplied function returns the product of the inputs IN THE CURRENT RING.
757 // For the logical operations the ring's MUL is really a logical AND function.
758 // This also type-checks the inputs for sanity. Guaranteed never to
759 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
760 const Type* AndLNode::mul_ring(const Type* t1, const Type* t2) const {
761 return RangeInference::infer_and(t1->is_long(), t2->is_long());
762 }
763
764 const Type* AndLNode::Value(PhaseGVN* phase) const {
765 if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
766 AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
767 return TypeLong::ZERO;
768 }
769
770 return MulNode::Value(phase);
771 }
772
773 //------------------------------Identity---------------------------------------
774 // Masking off the high bits of an unsigned load is not required
775 Node* AndLNode::Identity(PhaseGVN* phase) {
776
777 // x & x => x
778 if (in(1) == in(2)) {
779 return in(1);
780 }
781
782 Node *usr = in(1);
783 const TypeLong *t2 = phase->type( in(2) )->isa_long();
784 if( t2 && t2->is_con() ) {
785 jlong con = t2->get_con();
786 // Masking off high bits which are always zero is useless.
787 const TypeLong* t1 = phase->type( in(1) )->isa_long();
788 if (t1 != nullptr && t1->_lo >= 0) {
789 int bit_count = log2i_graceful(t1->_hi) + 1;
790 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
791 if ((t1_support & con) == t1_support)
792 return usr;
793 }
794 uint lop = usr->Opcode();
795 // Masking off the high bits of a unsigned-shift-right is not
796 // needed either.
797 if( lop == Op_URShiftL ) {
798 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
799 if( t12 && t12->is_con() ) { // Shift is by a constant
800 int shift = t12->get_con();
801 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
802 jlong mask = max_julong >> shift;
803 if( (mask&con) == mask ) // If AND is useless, skip it
804 return usr;
805 }
806 }
807 }
808 return MulNode::Identity(phase);
809 }
810
811 //------------------------------Ideal------------------------------------------
812 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
813 // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
814 Node* progress = AndIL_sum_and_mask(phase, T_LONG);
815 if (progress != nullptr) {
816 return progress;
817 }
818
819 // Convert "(~a) & (~b)" into "~(a | b)"
820 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
821 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
822 Node* tn = phase->transform(or_a_b);
823 return AddNode::make_not(phase, tn, T_LONG);
824 }
825
826 // Special case constant AND mask
827 const TypeLong *t2 = phase->type( in(2) )->isa_long();
828 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
829 const jlong mask = t2->get_con();
830
831 Node* in1 = in(1);
832 int op = in1->Opcode();
833
834 // Are we masking a long that was converted from an int with a mask
835 // that fits in 32-bits? Commute them and use an AndINode. Don't
836 // convert masks which would cause a sign extension of the integer
837 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which
838 // would be optimized away later in Identity.
839 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
840 Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
841 andi = phase->transform(andi);
842 return new ConvI2LNode(andi);
843 }
844
845 // Masking off sign bits? Dont make them!
846 if (op == Op_RShiftL) {
847 const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
848 if( t12 && t12->is_con() ) { // Shift is by a constant
849 int shift = t12->get_con();
850 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
851 if (shift != 0) {
852 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
853 // If the AND'ing of the 2 masks has no bits, then only original shifted
854 // bits survive. NO sign-extension bits survive the maskings.
855 if( (sign_bits_mask & mask) == 0 ) {
856 // Use zero-fill shift instead
857 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
858 return new AndLNode(zshift, in(2));
859 }
860 }
861 }
862 }
863
864 return MulNode::Ideal(phase, can_reshape);
865 }
866
867 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
868 switch (bt) {
869 case T_INT:
870 return new LShiftINode(in1, in2);
871 case T_LONG:
872 return new LShiftLNode(in1, in2);
873 default:
874 fatal("Not implemented for %s", type2name(bt));
875 }
876 return nullptr;
877 }
878
879 // Returns whether the shift amount is constant. If so, sets count.
880 static bool const_shift_count(PhaseGVN* phase, const Node* shift_node, int* count) {
881 const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
882 if (tcount != nullptr && tcount->is_con()) {
883 *count = tcount->get_con();
884 return true;
885 }
886 return false;
887 }
888
889 // Returns whether the shift amount is constant. If so, sets real_shift and masked_shift.
890 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, int& real_shift, uint& masked_shift) {
891 if (const_shift_count(phase, shift_node, &real_shift)) {
892 masked_shift = real_shift & (nBits - 1);
893 return true;
894 }
895 return false;
896 }
897
898 // Convenience for when we don't care about the real amount
899 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint nBits, uint& masked_shift) {
900 int real_shift;
901 return mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift);
902 }
903
904 // Use this in ::Ideal only with shiftNode == this!
905 // Returns the masked shift amount if constant or 0 if not constant.
906 static uint mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint nBits) {
907 int real_shift;
908 uint masked_shift;
909 if (mask_shift_amount(phase, shift_node, nBits, real_shift, masked_shift)) {
910 if (masked_shift == 0) {
911 // Let Identity() handle 0 shift count.
912 return 0;
913 }
914
915 if (real_shift != (int)masked_shift) {
916 PhaseIterGVN* igvn = phase->is_IterGVN();
917 if (igvn != nullptr) {
918 igvn->_worklist.push(shift_node);
919 }
920 shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
921 }
922 return masked_shift;
923 }
924 // Not a shift by a constant.
925 return 0;
926 }
927
928 // Called with
929 // outer_shift = (_ << rhs_outer)
930 // We are looking for the pattern:
931 // outer_shift = ((X << rhs_inner) << rhs_outer)
932 // where rhs_outer and rhs_inner are constant
933 // we denote inner_shift the nested expression (X << rhs_inner)
934 // con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
935 // where nbits is the number of bits of the shifts
936 //
937 // There are 2 cases:
938 // if con_outer + con_inner >= nbits => 0
939 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
940 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
941 assert(bt == T_LONG || bt == T_INT, "Unexpected type");
942 const Node* inner_shift = outer_shift->in(1);
943 if (inner_shift->Opcode() != Op_LShift(bt)) {
944 return nullptr;
945 }
946
947 uint nbits = bits_per_java_integer(bt);
948 uint con_inner;
949 if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
950 return nullptr;
951 }
952
953 if (con_inner == 0) {
954 // We let the Identity() of the inner shift do its job.
955 return nullptr;
956 }
957
958 if (con_outer + con_inner >= nbits) {
959 // While it might be tempting to use
960 // phase->zerocon(bt);
961 // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
962 // to return a new node, this or nullptr, but not an old (cached) node.
963 return ConNode::make(TypeInteger::zero(bt));
964 }
965
966 // con0 + con1 < nbits ==> actual shift happens now
967 Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
968 return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
969 }
970
971 //------------------------------Identity---------------------------------------
972 Node* LShiftINode::Identity(PhaseGVN* phase) {
973 return IdentityIL(phase, T_INT);
974 }
975
976 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
977 uint con = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
978 if (con == 0) {
979 return nullptr;
980 }
981
982 // If the right input is a constant, and the left input is an add of a
983 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
984 Node* add1 = in(1);
985 int add1_op = add1->Opcode();
986 if (add1_op == Op_Add(bt)) { // Left input is an add?
987 assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
988
989 // Transform is legal, but check for profit. Avoid breaking 'i2s'
990 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
991 if (bt != T_INT || con < 16) {
992 // Left input is an add of the same number?
993 if (con != (bits_per_java_integer(bt) - 1) && add1->in(1) == add1->in(2)) {
994 // Convert "(x + x) << c0" into "x << (c0 + 1)"
995 // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
996 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
997 // or c0 != 63 (for LShiftL) because:
998 // (x + x) << 63 = 2x << 63, while
999 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1000 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1001 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1002 return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1003 }
1004
1005 // Left input is an add of a constant?
1006 const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1007 if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1008 // Compute X << con0
1009 Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1010 // Compute X<<con0 + (con1<<con0)
1011 return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1012 }
1013 }
1014 }
1015 // Check for "(con0 - X) << con1"
1016 // Transform is legal, but check for profit. Avoid breaking 'i2s'
1017 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1018 if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) { // Left input is a sub?
1019 // Left input is a sub from a constant?
1020 const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
1021 if (t11 != nullptr && t11->is_con()) {
1022 // Compute X << con0
1023 Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
1024 // Compute (con1<<con0) - (X<<con0)
1025 return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
1026 }
1027 }
1028
1029 // Check for "(x >> C1) << C2"
1030 if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1031 int add1Con = 0;
1032 const_shift_count(phase, add1, &add1Con);
1033
1034 // Special case C1 == C2, which just masks off low bits
1035 if (add1Con > 0 && con == (uint)add1Con) {
1036 // Convert to "(x & -(1 << C2))"
1037 return MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1038 } else {
1039 // Wait until the right shift has been sharpened to the correct count
1040 if (add1Con > 0 && (uint)add1Con < bits_per_java_integer(bt)) {
1041 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1042 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1043 if (phase->is_IterGVN()) {
1044 if (con > (uint)add1Con) {
1045 // Creates "(x << (C2 - C1)) & -(1 << C2)"
1046 Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1047 return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1048 } else {
1049 assert(con < (uint)add1Con, "must be (%d < %d)", con, add1Con);
1050 // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1051
1052 // Handle logical and arithmetic shifts
1053 Node* rshift;
1054 if (add1_op == Op_RShift(bt)) {
1055 rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1056 } else {
1057 rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1058 }
1059
1060 return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1, con, bt)), bt), bt);
1061 }
1062 } else {
1063 phase->record_for_igvn(this);
1064 }
1065 }
1066 }
1067 }
1068
1069 // Check for "((x >> C1) & Y) << C2"
1070 if (add1_op == Op_And(bt)) {
1071 Node* add2 = add1->in(1);
1072 int add2_op = add2->Opcode();
1073 if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1074 // Special case C1 == C2, which just masks off low bits
1075 if (add2->in(2) == in(2)) {
1076 // Convert to "(x & (Y << C2))"
1077 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1078 return MulNode::make_and(add2->in(1), y_sh, bt);
1079 }
1080
1081 int add2Con = 0;
1082 const_shift_count(phase, add2, &add2Con);
1083 if (add2Con > 0 && (uint)add2Con < bits_per_java_integer(bt)) {
1084 if (phase->is_IterGVN()) {
1085 // Convert to "((x >> C1) << C2) & (Y << C2)"
1086
1087 // Make "(x >> C1) << C2", which will get folded away by the rule above
1088 Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1089 // Make "Y << C2", which will simplify when Y is a constant
1090 Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1091
1092 return MulNode::make_and(x_sh, y_sh, bt);
1093 } else {
1094 phase->record_for_igvn(this);
1095 }
1096 }
1097 }
1098 }
1099
1100 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1101 // before shifting them away.
1102 const jlong bits_mask = max_unsigned_integer(bt) >> con;
1103 assert(bt != T_INT || bits_mask == right_n_bits(bits_per_java_integer(bt)-con), "inconsistent");
1104 if (add1_op == Op_And(bt) &&
1105 phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1106 return LShiftNode::make(add1->in(1), in(2), bt);
1107 }
1108
1109 // Collapse nested left-shifts with constant rhs:
1110 // (X << con1) << con2 ==> X << (con1 + con2)
1111 Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1112 if (doubleShift != nullptr) {
1113 return doubleShift;
1114 }
1115
1116 return nullptr;
1117 }
1118
1119 //------------------------------Ideal------------------------------------------
1120 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1121 return IdealIL(phase, can_reshape, T_INT);
1122 }
1123
1124 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1125 const Type* t1 = phase->type(in(1));
1126 const Type* t2 = phase->type(in(2));
1127 // Either input is TOP ==> the result is TOP
1128 if (t1 == Type::TOP) {
1129 return Type::TOP;
1130 }
1131 if (t2 == Type::TOP) {
1132 return Type::TOP;
1133 }
1134
1135 // Left input is ZERO ==> the result is ZERO.
1136 if (t1 == TypeInteger::zero(bt)) {
1137 return TypeInteger::zero(bt);
1138 }
1139 // Shift by zero does nothing
1140 if (t2 == TypeInt::ZERO) {
1141 return t1;
1142 }
1143
1144 // Either input is BOTTOM ==> the result is BOTTOM
1145 if ((t1 == TypeInteger::bottom(bt)) || (t2 == TypeInt::INT) ||
1146 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM)) {
1147 return TypeInteger::bottom(bt);
1148 }
1149
1150 const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1151 const TypeInt* r2 = t2->is_int(); // Handy access
1152
1153 if (!r2->is_con()) {
1154 return TypeInteger::bottom(bt);
1155 }
1156
1157 uint shift = r2->get_con();
1158 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1159 // Shift by a multiple of 32/64 does nothing:
1160 if (shift == 0) {
1161 return t1;
1162 }
1163
1164 // If the shift is a constant, shift the bounds of the type,
1165 // unless this could lead to an overflow.
1166 if (!r1->is_con()) {
1167 jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1168 #ifdef ASSERT
1169 if (bt == T_INT) {
1170 jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1171 assert((java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1172 assert((java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1173 }
1174 #endif
1175 if (java_shift_right(java_shift_left(lo, shift, bt), shift, bt) == lo &&
1176 java_shift_right(java_shift_left(hi, shift, bt), shift, bt) == hi) {
1177 // No overflow. The range shifts up cleanly.
1178 return TypeInteger::make(java_shift_left(lo, shift, bt),
1179 java_shift_left(hi, shift, bt),
1180 MAX2(r1->_widen, r2->_widen), bt);
1181 }
1182 return TypeInteger::bottom(bt);
1183 }
1184
1185 return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1186 }
1187
1188 //------------------------------Value------------------------------------------
1189 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1190 return ValueIL(phase, T_INT);
1191 }
1192
1193 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1194 int count = 0;
1195 if (const_shift_count(phase, this, &count) && (count & (bits_per_java_integer(bt) - 1)) == 0) {
1196 // Shift by a multiple of 32/64 does nothing
1197 return in(1);
1198 }
1199 return this;
1200 }
1201
1202 //=============================================================================
1203 //------------------------------Identity---------------------------------------
1204 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1205 return IdentityIL(phase, T_LONG);
1206 }
1207
1208 //------------------------------Ideal------------------------------------------
1209 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1210 return IdealIL(phase, can_reshape, T_LONG);
1211 }
1212
1213 //------------------------------Value------------------------------------------
1214 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1215 return ValueIL(phase, T_LONG);
1216 }
1217
1218 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1219 switch (bt) {
1220 case T_INT:
1221 return new RShiftINode(in1, in2);
1222 case T_LONG:
1223 return new RShiftLNode(in1, in2);
1224 default:
1225 fatal("Not implemented for %s", type2name(bt));
1226 }
1227 return nullptr;
1228 }
1229
1230
1231 //=============================================================================
1232 //------------------------------Identity---------------------------------------
1233 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1234 int count = 0;
1235 if (const_shift_count(phase, this, &count)) {
1236 if ((count & (bits_per_java_integer(bt) - 1)) == 0) {
1237 // Shift by a multiple of 32/64 does nothing
1238 return in(1);
1239 }
1240 // Check for useless sign-masking
1241 if (in(1)->Opcode() == Op_LShift(bt) &&
1242 in(1)->req() == 3 &&
1243 in(1)->in(2) == in(2)) {
1244 count &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1245 // Compute masks for which this shifting doesn't change
1246 jlong lo = (CONST64(-1) << (bits_per_java_integer(bt) - ((uint)count)-1)); // FFFF8000
1247 jlong hi = ~lo; // 00007FFF
1248 const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1249 if (t11 == nullptr) {
1250 return this;
1251 }
1252 // Does actual value fit inside of mask?
1253 if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1254 return in(1)->in(1); // Then shifting is a nop
1255 }
1256 }
1257 }
1258 return this;
1259 }
1260
1261 Node* RShiftINode::Identity(PhaseGVN* phase) {
1262 return IdentityIL(phase, T_INT);
1263 }
1264
1265 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1266 // Inputs may be TOP if they are dead.
1267 const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1268 if (t1 == nullptr) {
1269 return NodeSentinel; // Left input is an integer
1270 }
1271 int shift = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt));
1272 if (shift == 0) {
1273 return NodeSentinel;
1274 }
1275
1276 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1277 // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1278 // Such expressions arise normally from shift chains like (byte)(x >> 24).
1279 const Node* and_node = in(1);
1280 if (and_node->Opcode() != Op_And(bt)) {
1281 return nullptr;
1282 }
1283 const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1284 if (mask_t != nullptr && mask_t->is_con()) {
1285 jlong maskbits = mask_t->get_con_as_long(bt);
1286 // Convert to "(x >> shift) & (mask >> shift)"
1287 Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1288 return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1289 }
1290 return nullptr;
1291 }
1292
1293 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1294 Node* progress = IdealIL(phase, can_reshape, T_INT);
1295 if (progress == NodeSentinel) {
1296 return nullptr;
1297 }
1298 if (progress != nullptr) {
1299 return progress;
1300 }
1301 int shift = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1302 assert(shift != 0, "handled by IdealIL");
1303
1304 // Check for "(short[i] <<16)>>16" which simply sign-extends
1305 const Node *shl = in(1);
1306 if (shl->Opcode() != Op_LShiftI) {
1307 return nullptr;
1308 }
1309
1310 const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1311 if (left_shift_t == nullptr) {
1312 return nullptr;
1313 }
1314 if (shift == 16 && left_shift_t->is_con(16)) {
1315 Node *ld = shl->in(1);
1316 if (ld->Opcode() == Op_LoadS) {
1317 // Sign extension is just useless here. Return a RShiftI of zero instead
1318 // returning 'ld' directly. We cannot return an old Node directly as
1319 // that is the job of 'Identity' calls and Identity calls only work on
1320 // direct inputs ('ld' is an extra Node removed from 'this'). The
1321 // combined optimization requires Identity only return direct inputs.
1322 set_req_X(1, ld, phase);
1323 set_req_X(2, phase->intcon(0), phase);
1324 return this;
1325 }
1326 else if (can_reshape &&
1327 ld->Opcode() == Op_LoadUS &&
1328 ld->outcnt() == 1 && ld->unique_out() == shl)
1329 // Replace zero-extension-load with sign-extension-load
1330 return ld->as_Load()->convert_to_signed_load(*phase);
1331 }
1332
1333 // Check for "(byte[i] <<24)>>24" which simply sign-extends
1334 if (shift == 24 && left_shift_t->is_con(24)) {
1335 Node *ld = shl->in(1);
1336 if (ld->Opcode() == Op_LoadB) {
1337 // Sign extension is just useless here
1338 set_req_X(1, ld, phase);
1339 set_req_X(2, phase->intcon(0), phase);
1340 return this;
1341 }
1342 }
1343
1344 return nullptr;
1345 }
1346
1347 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1348 const Type* t1 = phase->type(in(1));
1349 const Type* t2 = phase->type(in(2));
1350 // Either input is TOP ==> the result is TOP
1351 if (t1 == Type::TOP) {
1352 return Type::TOP;
1353 }
1354 if (t2 == Type::TOP) {
1355 return Type::TOP;
1356 }
1357
1358 // Left input is ZERO ==> the result is ZERO.
1359 if (t1 == TypeInteger::zero(bt)) {
1360 return TypeInteger::zero(bt);
1361 }
1362 // Shift by zero does nothing
1363 if (t2 == TypeInt::ZERO) {
1364 return t1;
1365 }
1366
1367 // Either input is BOTTOM ==> the result is BOTTOM
1368 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1369 return TypeInteger::bottom(bt);
1370 }
1371
1372 const TypeInteger* r1 = t1->isa_integer(bt);
1373 const TypeInt* r2 = t2->isa_int();
1374
1375 // If the shift is a constant, just shift the bounds of the type.
1376 // For example, if the shift is 31/63, we just propagate sign bits.
1377 if (!r1->is_con() && r2->is_con()) {
1378 uint shift = r2->get_con();
1379 shift &= bits_per_java_integer(bt) - 1; // semantics of Java shifts
1380 // Shift by a multiple of 32/64 does nothing:
1381 if (shift == 0) {
1382 return t1;
1383 }
1384 // Calculate reasonably aggressive bounds for the result.
1385 // This is necessary if we are to correctly type things
1386 // like (x<<24>>24) == ((byte)x).
1387 jlong lo = r1->lo_as_long() >> (jint)shift;
1388 jlong hi = r1->hi_as_long() >> (jint)shift;
1389 assert(lo <= hi, "must have valid bounds");
1390 #ifdef ASSERT
1391 if (bt == T_INT) {
1392 jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1393 jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1394 assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1395 }
1396 #endif
1397 const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1398 #ifdef ASSERT
1399 // Make sure we get the sign-capture idiom correct.
1400 if (shift == bits_per_java_integer(bt) - 1) {
1401 if (r1->lo_as_long() >= 0) {
1402 assert(ti == TypeInteger::zero(bt), ">>31/63 of + is 0");
1403 }
1404 if (r1->hi_as_long() < 0) {
1405 assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1406 }
1407 }
1408 #endif
1409 return ti;
1410 }
1411
1412 if (!r1->is_con() || !r2->is_con()) {
1413 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1414 if (r1->lo_as_long() >= 0) {
1415 return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1416 }
1417
1418 // Conversely, if the left input is negative then the result must be negative.
1419 if (r1->hi_as_long() <= -1) {
1420 return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1421 }
1422
1423 return TypeInteger::bottom(bt);
1424 }
1425
1426 // Signed shift right
1427 return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1428 }
1429
1430 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1431 return ValueIL(phase, T_INT);
1432 }
1433
1434 //=============================================================================
1435 //------------------------------Identity---------------------------------------
1436 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1437 return IdentityIL(phase, T_LONG);
1438 }
1439
1440 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1441 Node* progress = IdealIL(phase, can_reshape, T_LONG);
1442 if (progress == NodeSentinel) {
1443 return nullptr;
1444 }
1445 return progress;
1446 }
1447
1448 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1449 return ValueIL(phase, T_LONG);
1450 }
1451
1452 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1453 switch (bt) {
1454 case T_INT:
1455 return new URShiftINode(in1, in2);
1456 case T_LONG:
1457 return new URShiftLNode(in1, in2);
1458 default:
1459 fatal("Not implemented for %s", type2name(bt));
1460 }
1461 return nullptr;
1462 }
1463
1464 //=============================================================================
1465 //------------------------------Identity---------------------------------------
1466 Node* URShiftINode::Identity(PhaseGVN* phase) {
1467 int count = 0;
1468 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) {
1469 // Shift by a multiple of 32 does nothing
1470 return in(1);
1471 }
1472
1473 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1474 // Happens during new-array length computation.
1475 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1476 Node *add = in(1);
1477 if (add->Opcode() == Op_AddI) {
1478 const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1479 if (t2 && t2->is_con(wordSize - 1) &&
1480 add->in(1)->Opcode() == Op_LShiftI) {
1481 // Check that shift_counts are LogBytesPerWord.
1482 Node *lshift_count = add->in(1)->in(2);
1483 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1484 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1485 t_lshift_count == phase->type(in(2))) {
1486 Node *x = add->in(1)->in(1);
1487 const TypeInt *t_x = phase->type(x)->isa_int();
1488 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1489 return x;
1490 }
1491 }
1492 }
1493 }
1494
1495 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1496 }
1497
1498 //------------------------------Ideal------------------------------------------
1499 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1500 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger);
1501 if (con == 0) {
1502 return nullptr;
1503 }
1504
1505 // We'll be wanting the right-shift amount as a mask of that many bits
1506 const int mask = right_n_bits(BitsPerJavaInteger - con);
1507
1508 int in1_op = in(1)->Opcode();
1509
1510 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1511 if( in1_op == Op_URShiftI ) {
1512 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1513 if( t12 && t12->is_con() ) { // Right input is a constant
1514 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1515 const int con2 = t12->get_con() & 31; // Shift count is always masked
1516 const int con3 = con+con2;
1517 if( con3 < 32 ) // Only merge shifts if total is < 32
1518 return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1519 }
1520 }
1521
1522 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1523 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1524 // If Q is "X << z" the rounding is useless. Look for patterns like
1525 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1526 Node *add = in(1);
1527 const TypeInt *t2 = phase->type(in(2))->isa_int();
1528 if (in1_op == Op_AddI) {
1529 Node *lshl = add->in(1);
1530 if( lshl->Opcode() == Op_LShiftI &&
1531 phase->type(lshl->in(2)) == t2 ) {
1532 Node *y_z = phase->transform( new URShiftINode(add->in(2),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 if( in1_op == Op_LShiftI &&
1559 phase->type(shl->in(2)) == t2 )
1560 return new AndINode( shl->in(1), phase->intcon(mask) );
1561
1562 // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1563 Node *shr = in(1);
1564 if ( in1_op == Op_RShiftI ) {
1565 Node *in11 = shr->in(1);
1566 Node *in12 = shr->in(2);
1567 const TypeInt *t11 = phase->type(in11)->isa_int();
1568 const TypeInt *t12 = phase->type(in12)->isa_int();
1569 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1570 return new URShiftINode(in11, phase->intcon(31));
1571 }
1572 }
1573
1574 return nullptr;
1575 }
1576
1577 //------------------------------Value------------------------------------------
1578 // A URShiftINode shifts its input2 right by input1 amount.
1579 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1580 // (This is a near clone of RShiftINode::Value.)
1581 const Type *t1 = phase->type( in(1) );
1582 const Type *t2 = phase->type( in(2) );
1583 // Either input is TOP ==> the result is TOP
1584 if( t1 == Type::TOP ) return Type::TOP;
1585 if( t2 == Type::TOP ) return Type::TOP;
1586
1587 // Left input is ZERO ==> the result is ZERO.
1588 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1589 // Shift by zero does nothing
1590 if( t2 == TypeInt::ZERO ) return t1;
1591
1592 // Either input is BOTTOM ==> the result is BOTTOM
1593 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1594 return TypeInt::INT;
1595
1596 if (t2 == TypeInt::INT)
1597 return TypeInt::INT;
1598
1599 const TypeInt *r1 = t1->is_int(); // Handy access
1600 const TypeInt *r2 = t2->is_int(); // Handy access
1601
1602 if (r2->is_con()) {
1603 uint shift = r2->get_con();
1604 shift &= BitsPerJavaInteger-1; // semantics of Java shifts
1605 // Shift by a multiple of 32 does nothing:
1606 if (shift == 0) return t1;
1607 // Calculate reasonably aggressive bounds for the result.
1608 jint lo = (juint)r1->_lo >> (juint)shift;
1609 jint hi = (juint)r1->_hi >> (juint)shift;
1610 if (r1->_hi >= 0 && r1->_lo < 0) {
1611 // If the type has both negative and positive values,
1612 // there are two separate sub-domains to worry about:
1613 // The positive half and the negative half.
1614 jint neg_lo = lo;
1615 jint neg_hi = (juint)-1 >> (juint)shift;
1616 jint pos_lo = (juint) 0 >> (juint)shift;
1617 jint pos_hi = hi;
1618 lo = MIN2(neg_lo, pos_lo); // == 0
1619 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1620 }
1621 assert(lo <= hi, "must have valid bounds");
1622 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1623 #ifdef ASSERT
1624 // Make sure we get the sign-capture idiom correct.
1625 if (shift == BitsPerJavaInteger-1) {
1626 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1627 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
1628 }
1629 #endif
1630 return ti;
1631 }
1632
1633 //
1634 // Do not support shifted oops in info for GC
1635 //
1636 // else if( t1->base() == Type::InstPtr ) {
1637 //
1638 // const TypeInstPtr *o = t1->is_instptr();
1639 // if( t1->singleton() )
1640 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1641 // }
1642 // else if( t1->base() == Type::KlassPtr ) {
1643 // const TypeKlassPtr *o = t1->is_klassptr();
1644 // if( t1->singleton() )
1645 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1646 // }
1647
1648 return TypeInt::INT;
1649 }
1650
1651 //=============================================================================
1652 //------------------------------Identity---------------------------------------
1653 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1654 int count = 0;
1655 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) {
1656 // Shift by a multiple of 64 does nothing
1657 return in(1);
1658 }
1659 return this;
1660 }
1661
1662 //------------------------------Ideal------------------------------------------
1663 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1664 int con = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong);
1665 if (con == 0) {
1666 return nullptr;
1667 }
1668
1669 // We'll be wanting the right-shift amount as a mask of that many bits
1670 const jlong mask = jlong(max_julong >> con);
1671
1672 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
1673 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1674 // If Q is "X << z" the rounding is useless. Look for patterns like
1675 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
1676 Node *add = in(1);
1677 const TypeInt *t2 = phase->type(in(2))->isa_int();
1678 if (add->Opcode() == Op_AddL) {
1679 Node *lshl = add->in(1);
1680 if( lshl->Opcode() == Op_LShiftL &&
1681 phase->type(lshl->in(2)) == t2 ) {
1682 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) );
1683 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) );
1684 return new AndLNode( sum, phase->longcon(mask) );
1685 }
1686 }
1687
1688 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
1689 // This shortens the mask. Also, if we are extracting a high byte and
1690 // storing it to a buffer, the mask will be removed completely.
1691 Node *andi = in(1);
1692 if( andi->Opcode() == Op_AndL ) {
1693 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1694 if( t3 && t3->is_con() ) { // Right input is a constant
1695 jlong mask2 = t3->get_con();
1696 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
1697 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1698 return new AndLNode(newshr, phase->longcon(mask2));
1699 }
1700 }
1701
1702 // Check for "(X << z ) >>> z" which simply zero-extends
1703 Node *shl = in(1);
1704 if( shl->Opcode() == Op_LShiftL &&
1705 phase->type(shl->in(2)) == t2 )
1706 return new AndLNode( shl->in(1), phase->longcon(mask) );
1707
1708 // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1709 Node *shr = in(1);
1710 if ( shr->Opcode() == Op_RShiftL ) {
1711 Node *in11 = shr->in(1);
1712 Node *in12 = shr->in(2);
1713 const TypeLong *t11 = phase->type(in11)->isa_long();
1714 const TypeInt *t12 = phase->type(in12)->isa_int();
1715 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1716 return new URShiftLNode(in11, phase->intcon(63));
1717 }
1718 }
1719 return nullptr;
1720 }
1721
1722 //------------------------------Value------------------------------------------
1723 // A URShiftINode shifts its input2 right by input1 amount.
1724 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1725 // (This is a near clone of RShiftLNode::Value.)
1726 const Type *t1 = phase->type( in(1) );
1727 const Type *t2 = phase->type( in(2) );
1728 // Either input is TOP ==> the result is TOP
1729 if( t1 == Type::TOP ) return Type::TOP;
1730 if( t2 == Type::TOP ) return Type::TOP;
1731
1732 // Left input is ZERO ==> the result is ZERO.
1733 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1734 // Shift by zero does nothing
1735 if( t2 == TypeInt::ZERO ) return t1;
1736
1737 // Either input is BOTTOM ==> the result is BOTTOM
1738 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1739 return TypeLong::LONG;
1740
1741 if (t2 == TypeInt::INT)
1742 return TypeLong::LONG;
1743
1744 const TypeLong *r1 = t1->is_long(); // Handy access
1745 const TypeInt *r2 = t2->is_int (); // Handy access
1746
1747 if (r2->is_con()) {
1748 uint shift = r2->get_con();
1749 shift &= BitsPerJavaLong - 1; // semantics of Java shifts
1750 // Shift by a multiple of 64 does nothing:
1751 if (shift == 0) return t1;
1752 // Calculate reasonably aggressive bounds for the result.
1753 jlong lo = (julong)r1->_lo >> (juint)shift;
1754 jlong hi = (julong)r1->_hi >> (juint)shift;
1755 if (r1->_hi >= 0 && r1->_lo < 0) {
1756 // If the type has both negative and positive values,
1757 // there are two separate sub-domains to worry about:
1758 // The positive half and the negative half.
1759 jlong neg_lo = lo;
1760 jlong neg_hi = (julong)-1 >> (juint)shift;
1761 jlong pos_lo = (julong) 0 >> (juint)shift;
1762 jlong pos_hi = hi;
1763 //lo = MIN2(neg_lo, pos_lo); // == 0
1764 lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1765 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
1766 hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1767 }
1768 assert(lo <= hi, "must have valid bounds");
1769 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1770 #ifdef ASSERT
1771 // Make sure we get the sign-capture idiom correct.
1772 if (shift == BitsPerJavaLong - 1) {
1773 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1774 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
1775 }
1776 #endif
1777 return tl;
1778 }
1779
1780 return TypeLong::LONG; // Give up
1781 }
1782
1783 //=============================================================================
1784 //------------------------------Ideal------------------------------------------
1785 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1786 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1787 // This reduces the number of rules in the matcher, as we only need to check
1788 // for negations on the second argument, and not the symmetric case where
1789 // the first argument is negated.
1790 if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1791 swap_edges(1, 2);
1792 return this;
1793 }
1794 return nullptr;
1795 }
1796
1797 //=============================================================================
1798 //------------------------------Value------------------------------------------
1799 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1800 const Type *t1 = phase->type(in(1));
1801 if (t1 == Type::TOP) return Type::TOP;
1802 if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1803 const Type *t2 = phase->type(in(2));
1804 if (t2 == Type::TOP) return Type::TOP;
1805 if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1806 const Type *t3 = phase->type(in(3));
1807 if (t3 == Type::TOP) return Type::TOP;
1808 if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1809 #ifndef __STDC_IEC_559__
1810 return Type::DOUBLE;
1811 #else
1812 double d1 = t1->getd();
1813 double d2 = t2->getd();
1814 double d3 = t3->getd();
1815 return TypeD::make(fma(d1, d2, d3));
1816 #endif
1817 }
1818
1819 //=============================================================================
1820 //------------------------------Value------------------------------------------
1821 const Type* FmaFNode::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::FloatCon) return Type::FLOAT;
1825 const Type *t2 = phase->type(in(2));
1826 if (t2 == Type::TOP) return Type::TOP;
1827 if (t2->base() != Type::FloatCon) return Type::FLOAT;
1828 const Type *t3 = phase->type(in(3));
1829 if (t3 == Type::TOP) return Type::TOP;
1830 if (t3->base() != Type::FloatCon) return Type::FLOAT;
1831 #ifndef __STDC_IEC_559__
1832 return Type::FLOAT;
1833 #else
1834 float f1 = t1->getf();
1835 float f2 = t2->getf();
1836 float f3 = t3->getf();
1837 return TypeF::make(fma(f1, f2, f3));
1838 #endif
1839 }
1840
1841 //=============================================================================
1842 //------------------------------Value------------------------------------------
1843 const Type* FmaHFNode::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::HalfFloatCon) { return Type::HALF_FLOAT; }
1847 const Type* t2 = phase->type(in(2));
1848 if (t2 == Type::TOP) { return Type::TOP; }
1849 if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1850 const Type* t3 = phase->type(in(3));
1851 if (t3 == Type::TOP) { return Type::TOP; }
1852 if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1853 #ifndef __STDC_IEC_559__
1854 return Type::HALF_FLOAT;
1855 #else
1856 float f1 = t1->getf();
1857 float f2 = t2->getf();
1858 float f3 = t3->getf();
1859 return TypeH::make(fma(f1, f2, f3));
1860 #endif
1861 }
1862
1863 //=============================================================================
1864 //------------------------------hash-------------------------------------------
1865 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs.
1866 // The hash function must return the same value when edge swapping is performed.
1867 uint MulAddS2INode::hash() const {
1868 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1869 }
1870
1871 //------------------------------Rotate Operations ------------------------------
1872
1873 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1874 const Type* t1 = phase->type(in(1));
1875 if (t1 == Type::TOP) {
1876 return this;
1877 }
1878 int count = 0;
1879 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1880 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1881 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1882 // Rotate by a multiple of 32/64 does nothing
1883 return in(1);
1884 }
1885 return this;
1886 }
1887
1888 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1889 const Type* t1 = phase->type(in(1));
1890 const Type* t2 = phase->type(in(2));
1891 // Either input is TOP ==> the result is TOP
1892 if (t1 == Type::TOP || t2 == Type::TOP) {
1893 return Type::TOP;
1894 }
1895
1896 if (t1->isa_int()) {
1897 const TypeInt* r1 = t1->is_int();
1898 const TypeInt* r2 = t2->is_int();
1899
1900 // Left input is ZERO ==> the result is ZERO.
1901 if (r1 == TypeInt::ZERO) {
1902 return TypeInt::ZERO;
1903 }
1904 // Rotate by zero does nothing
1905 if (r2 == TypeInt::ZERO) {
1906 return r1;
1907 }
1908 if (r1->is_con() && r2->is_con()) {
1909 juint r1_con = (juint)r1->get_con();
1910 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1911 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
1912 }
1913 return TypeInt::INT;
1914 } else {
1915 assert(t1->isa_long(), "Type must be a long");
1916 const TypeLong* r1 = t1->is_long();
1917 const TypeInt* r2 = t2->is_int();
1918
1919 // Left input is ZERO ==> the result is ZERO.
1920 if (r1 == TypeLong::ZERO) {
1921 return TypeLong::ZERO;
1922 }
1923 // Rotate by zero does nothing
1924 if (r2 == TypeInt::ZERO) {
1925 return r1;
1926 }
1927 if (r1->is_con() && r2->is_con()) {
1928 julong r1_con = (julong)r1->get_con();
1929 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
1930 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
1931 }
1932 return TypeLong::LONG;
1933 }
1934 }
1935
1936 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1937 const Type* t1 = phase->type(in(1));
1938 const Type* t2 = phase->type(in(2));
1939 if (t2->isa_int() && t2->is_int()->is_con()) {
1940 if (t1->isa_int()) {
1941 int lshift = t2->is_int()->get_con() & 31;
1942 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
1943 } else if (t1 != Type::TOP) {
1944 assert(t1->isa_long(), "Type must be a long");
1945 int lshift = t2->is_int()->get_con() & 63;
1946 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
1947 }
1948 }
1949 return nullptr;
1950 }
1951
1952 Node* RotateRightNode::Identity(PhaseGVN* phase) {
1953 const Type* t1 = phase->type(in(1));
1954 if (t1 == Type::TOP) {
1955 return this;
1956 }
1957 int count = 0;
1958 assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1959 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1;
1960 if (const_shift_count(phase, this, &count) && (count & mask) == 0) {
1961 // Rotate by a multiple of 32/64 does nothing
1962 return in(1);
1963 }
1964 return this;
1965 }
1966
1967 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
1968 const Type* t1 = phase->type(in(1));
1969 const Type* t2 = phase->type(in(2));
1970 // Either input is TOP ==> the result is TOP
1971 if (t1 == Type::TOP || t2 == Type::TOP) {
1972 return Type::TOP;
1973 }
1974
1975 if (t1->isa_int()) {
1976 const TypeInt* r1 = t1->is_int();
1977 const TypeInt* r2 = t2->is_int();
1978
1979 // Left input is ZERO ==> the result is ZERO.
1980 if (r1 == TypeInt::ZERO) {
1981 return TypeInt::ZERO;
1982 }
1983 // Rotate by zero does nothing
1984 if (r2 == TypeInt::ZERO) {
1985 return r1;
1986 }
1987 if (r1->is_con() && r2->is_con()) {
1988 juint r1_con = (juint)r1->get_con();
1989 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1990 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
1991 }
1992 return TypeInt::INT;
1993 } else {
1994 assert(t1->isa_long(), "Type must be a long");
1995 const TypeLong* r1 = t1->is_long();
1996 const TypeInt* r2 = t2->is_int();
1997 // Left input is ZERO ==> the result is ZERO.
1998 if (r1 == TypeLong::ZERO) {
1999 return TypeLong::ZERO;
2000 }
2001 // Rotate by zero does nothing
2002 if (r2 == TypeInt::ZERO) {
2003 return r1;
2004 }
2005 if (r1->is_con() && r2->is_con()) {
2006 julong r1_con = (julong)r1->get_con();
2007 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2008 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2009 }
2010 return TypeLong::LONG;
2011 }
2012 }
2013
2014 //------------------------------ Sum & Mask ------------------------------
2015
2016 // Returns a lower bound on the number of trailing zeros in expr.
2017 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2018 const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2019 if (type == nullptr) {
2020 return 0;
2021 }
2022
2023 expr = expr->uncast();
2024 type = phase->type(expr)->isa_integer(bt);
2025 if (type == nullptr) {
2026 return 0;
2027 }
2028
2029 if (type->is_con()) {
2030 jlong con = type->get_con_as_long(bt);
2031 return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2032 }
2033
2034 if (expr->Opcode() == Op_ConvI2L) {
2035 expr = expr->in(1)->uncast();
2036 bt = T_INT;
2037 type = phase->type(expr)->isa_int();
2038 }
2039
2040 // Pattern: expr = (x << shift)
2041 if (expr->Opcode() == Op_LShift(bt)) {
2042 const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2043 if (shift_t == nullptr || !shift_t->is_con()) {
2044 return 0;
2045 }
2046 // We need to truncate the shift, as it may not have been canonicalized yet.
2047 // T_INT: 0..31 -> shift_mask = 4 * 8 - 1 = 31
2048 // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2049 // (JLS: "Shift Operators")
2050 jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2051 return shift_t->get_con() & shift_mask;
2052 }
2053
2054 return 0;
2055 }
2056
2057 // Checks whether expr is neutral additive element (zero) under mask,
2058 // i.e. whether an expression of the form:
2059 // (AndX (AddX (expr addend) mask)
2060 // (expr + addend) & mask
2061 // is equivalent to
2062 // (AndX addend mask)
2063 // addend & mask
2064 // for any addend.
2065 // (The X in AndX must be I or L, depending on bt).
2066 //
2067 // We check for the sufficient condition when the lowest set bit in expr is higher than
2068 // the highest set bit in mask, i.e.:
2069 // expr: eeeeee0000000000000
2070 // mask: 000000mmmmmmmmmmmmm
2071 // <--w bits--->
2072 // We do not test for other cases.
2073 //
2074 // Correctness:
2075 // Given "expr" with at least "w" trailing zeros,
2076 // let "mod = 2^w", "suffix_mask = mod - 1"
2077 //
2078 // Since "mask" only has bits set where "suffix_mask" does, we have:
2079 // mask = suffix_mask & mask (SUFFIX_MASK)
2080 //
2081 // And since expr only has bits set above w, and suffix_mask only below:
2082 // expr & suffix_mask == 0 (NO_BIT_OVERLAP)
2083 //
2084 // From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2085 // a power of 2, and we are computing in a ring of powers of 2, we know that
2086 // (x + y) % mod = (x % mod + y) % mod
2087 // (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask (MOD_ARITH)
2088 //
2089 // We can now prove the equality:
2090 // (expr + addend) & mask
2091 // = (expr + addend) & suffix_mask & mask (SUFFIX_MASK)
2092 // = (expr & suffix_mask + addend) & suffix_mask & mask (MOD_ARITH)
2093 // = (0 + addend) & suffix_mask & mask (NO_BIT_OVERLAP)
2094 // = addend & mask (SUFFIX_MASK)
2095 //
2096 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2097 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2098 // When the mask is negative, it has the most significant bit set.
2099 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2100 if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2101 return false;
2102 }
2103
2104 // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2105 if (mask_t->hi_as_long() == 0) {
2106 assert(mask_t->lo_as_long() == 0, "checked earlier");
2107 return false;
2108 }
2109
2110 jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2111 jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2112 return expr_trailing_zeros >= mask_bit_width;
2113 }
2114
2115 // Reduces the pattern:
2116 // (AndX (AddX add1 add2) mask)
2117 // to
2118 // (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2119 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2120 Node* add = in(1);
2121 Node* mask = in(2);
2122 int addidx = 0;
2123 if (add->Opcode() == Op_Add(bt)) {
2124 addidx = 1;
2125 } else if (mask->Opcode() == Op_Add(bt)) {
2126 mask = add;
2127 addidx = 2;
2128 add = in(addidx);
2129 }
2130 if (addidx > 0) {
2131 Node* add1 = add->in(1);
2132 Node* add2 = add->in(2);
2133 if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2134 set_req_X(addidx, add2, phase);
2135 return this;
2136 } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2137 set_req_X(addidx, add1, phase);
2138 return this;
2139 }
2140 }
2141 return nullptr;
2142 }