1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/loopnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/movenode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/subnode.hpp"
40 #include "runtime/sharedRuntime.hpp"
41
42 // Portions of code courtesy of Clifford Click
43
44 // Optimization - Graph Style
45
46 #include "math.h"
47
48 //=============================================================================
49 //------------------------------Identity---------------------------------------
50 // If right input is a constant 0, return the left input.
Identity(PhaseGVN * phase)51 Node* SubNode::Identity(PhaseGVN* phase) {
52 assert(in(1) != this, "Must already have called Value");
53 assert(in(2) != this, "Must already have called Value");
54
55 // Remove double negation
56 const Type *zero = add_id();
57 if( phase->type( in(1) )->higher_equal( zero ) &&
58 in(2)->Opcode() == Opcode() &&
59 phase->type( in(2)->in(1) )->higher_equal( zero ) ) {
60 return in(2)->in(2);
61 }
62
63 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
64 if( in(1)->Opcode() == Op_AddI ) {
65 if( phase->eqv(in(1)->in(2),in(2)) )
66 return in(1)->in(1);
67 if (phase->eqv(in(1)->in(1),in(2)))
68 return in(1)->in(2);
69
70 // Also catch: "(X + Opaque2(Y)) - Y". In this case, 'Y' is a loop-varying
71 // trip counter and X is likely to be loop-invariant (that's how O2 Nodes
72 // are originally used, although the optimizer sometimes jiggers things).
73 // This folding through an O2 removes a loop-exit use of a loop-varying
74 // value and generally lowers register pressure in and around the loop.
75 if( in(1)->in(2)->Opcode() == Op_Opaque2 &&
76 phase->eqv(in(1)->in(2)->in(1),in(2)) )
77 return in(1)->in(1);
78 }
79
80 return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this;
81 }
82
83 //------------------------------Value------------------------------------------
84 // A subtract node differences it's two inputs.
Value_common(PhaseTransform * phase) const85 const Type* SubNode::Value_common(PhaseTransform *phase) const {
86 const Node* in1 = in(1);
87 const Node* in2 = in(2);
88 // Either input is TOP ==> the result is TOP
89 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
90 if( t1 == Type::TOP ) return Type::TOP;
91 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
92 if( t2 == Type::TOP ) return Type::TOP;
93
94 // Not correct for SubFnode and AddFNode (must check for infinity)
95 // Equal? Subtract is zero
96 if (in1->eqv_uncast(in2)) return add_id();
97
98 // Either input is BOTTOM ==> the result is the local BOTTOM
99 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
100 return bottom_type();
101
102 return NULL;
103 }
104
Value(PhaseGVN * phase) const105 const Type* SubNode::Value(PhaseGVN* phase) const {
106 const Type* t = Value_common(phase);
107 if (t != NULL) {
108 return t;
109 }
110 const Type* t1 = phase->type(in(1));
111 const Type* t2 = phase->type(in(2));
112 return sub(t1,t2); // Local flavor of type subtraction
113
114 }
115
116 //=============================================================================
117 //------------------------------Helper function--------------------------------
118
is_cloop_increment(Node * inc)119 static bool is_cloop_increment(Node* inc) {
120 precond(inc->Opcode() == Op_AddI || inc->Opcode() == Op_AddL);
121
122 if (!inc->in(1)->is_Phi()) {
123 return false;
124 }
125 const PhiNode* phi = inc->in(1)->as_Phi();
126
127 if (phi->is_copy() || !phi->region()->is_CountedLoop()) {
128 return false;
129 }
130
131 return inc == phi->region()->as_CountedLoop()->incr();
132 }
133
134 // Given the expression '(x + C) - v', or
135 // 'v - (x + C)', we examine nodes '+' and 'v':
136 //
137 // 1. Do not convert if '+' is a counted-loop increment, because the '-' is
138 // loop invariant and converting extends the live-range of 'x' to overlap
139 // with the '+', forcing another register to be used in the loop.
140 //
141 // 2. Do not convert if 'v' is a counted-loop induction variable, because
142 // 'x' might be invariant.
143 //
ok_to_convert(Node * inc,Node * var)144 static bool ok_to_convert(Node* inc, Node* var) {
145 return !(is_cloop_increment(inc) || var->is_cloop_ind_var());
146 }
147
148 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)149 Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){
150 Node *in1 = in(1);
151 Node *in2 = in(2);
152 uint op1 = in1->Opcode();
153 uint op2 = in2->Opcode();
154
155 #ifdef ASSERT
156 // Check for dead loop
157 if( phase->eqv( in1, this ) || phase->eqv( in2, this ) ||
158 ( ( op1 == Op_AddI || op1 == Op_SubI ) &&
159 ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) ||
160 phase->eqv( in1->in(1), in1 ) || phase->eqv( in1->in(2), in1 ) ) ) )
161 assert(false, "dead loop in SubINode::Ideal");
162 #endif
163
164 const Type *t2 = phase->type( in2 );
165 if( t2 == Type::TOP ) return NULL;
166 // Convert "x-c0" into "x+ -c0".
167 if( t2->base() == Type::Int ){ // Might be bottom or top...
168 const TypeInt *i = t2->is_int();
169 if( i->is_con() )
170 return new AddINode(in1, phase->intcon(-i->get_con()));
171 }
172
173 // Convert "(x+c0) - y" into (x-y) + c0"
174 // Do not collapse (x+c0)-y if "+" is a loop increment or
175 // if "y" is a loop induction variable.
176 if( op1 == Op_AddI && ok_to_convert(in1, in2) ) {
177 const Type *tadd = phase->type( in1->in(2) );
178 if( tadd->singleton() && tadd != Type::TOP ) {
179 Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 ));
180 return new AddINode( sub2, in1->in(2) );
181 }
182 }
183
184
185 // Convert "x - (y+c0)" into "(x-y) - c0"
186 // Need the same check as in above optimization but reversed.
187 if (op2 == Op_AddI && ok_to_convert(in2, in1)) {
188 Node* in21 = in2->in(1);
189 Node* in22 = in2->in(2);
190 const TypeInt* tcon = phase->type(in22)->isa_int();
191 if (tcon != NULL && tcon->is_con()) {
192 Node* sub2 = phase->transform( new SubINode(in1, in21) );
193 Node* neg_c0 = phase->intcon(- tcon->get_con());
194 return new AddINode(sub2, neg_c0);
195 }
196 }
197
198 const Type *t1 = phase->type( in1 );
199 if( t1 == Type::TOP ) return NULL;
200
201 #ifdef ASSERT
202 // Check for dead loop
203 if( ( op2 == Op_AddI || op2 == Op_SubI ) &&
204 ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) ||
205 phase->eqv( in2->in(1), in2 ) || phase->eqv( in2->in(2), in2 ) ) )
206 assert(false, "dead loop in SubINode::Ideal");
207 #endif
208
209 // Convert "x - (x+y)" into "-y"
210 if( op2 == Op_AddI &&
211 phase->eqv( in1, in2->in(1) ) )
212 return new SubINode( phase->intcon(0),in2->in(2));
213 // Convert "(x-y) - x" into "-y"
214 if( op1 == Op_SubI &&
215 phase->eqv( in1->in(1), in2 ) )
216 return new SubINode( phase->intcon(0),in1->in(2));
217 // Convert "x - (y+x)" into "-y"
218 if( op2 == Op_AddI &&
219 phase->eqv( in1, in2->in(2) ) )
220 return new SubINode( phase->intcon(0),in2->in(1));
221
222 // Convert "0 - (x-y)" into "y-x"
223 if( t1 == TypeInt::ZERO && op2 == Op_SubI )
224 return new SubINode( in2->in(2), in2->in(1) );
225
226 // Convert "0 - (x+con)" into "-con-x"
227 jint con;
228 if( t1 == TypeInt::ZERO && op2 == Op_AddI &&
229 (con = in2->in(2)->find_int_con(0)) != 0 )
230 return new SubINode( phase->intcon(-con), in2->in(1) );
231
232 // Convert "(X+A) - (X+B)" into "A - B"
233 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) )
234 return new SubINode( in1->in(2), in2->in(2) );
235
236 // Convert "(A+X) - (B+X)" into "A - B"
237 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) )
238 return new SubINode( in1->in(1), in2->in(1) );
239
240 // Convert "(A+X) - (X+B)" into "A - B"
241 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) )
242 return new SubINode( in1->in(1), in2->in(2) );
243
244 // Convert "(X+A) - (B+X)" into "A - B"
245 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) )
246 return new SubINode( in1->in(2), in2->in(1) );
247
248 // Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally
249 // nicer to optimize than subtract.
250 if( op2 == Op_SubI && in2->outcnt() == 1) {
251 Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) );
252 return new SubINode( add1, in2->in(1) );
253 }
254
255 // Convert "0-(A>>31)" into "(A>>>31)"
256 if ( op2 == Op_RShiftI ) {
257 Node *in21 = in2->in(1);
258 Node *in22 = in2->in(2);
259 const TypeInt *zero = phase->type(in1)->isa_int();
260 const TypeInt *t21 = phase->type(in21)->isa_int();
261 const TypeInt *t22 = phase->type(in22)->isa_int();
262 if ( t21 && t22 && zero == TypeInt::ZERO && t22->is_con(31) ) {
263 return new URShiftINode(in21, in22);
264 }
265 }
266
267 return NULL;
268 }
269
270 //------------------------------sub--------------------------------------------
271 // A subtract node differences it's two inputs.
sub(const Type * t1,const Type * t2) const272 const Type *SubINode::sub( const Type *t1, const Type *t2 ) const {
273 const TypeInt *r0 = t1->is_int(); // Handy access
274 const TypeInt *r1 = t2->is_int();
275 int32_t lo = java_subtract(r0->_lo, r1->_hi);
276 int32_t hi = java_subtract(r0->_hi, r1->_lo);
277
278 // We next check for 32-bit overflow.
279 // If that happens, we just assume all integers are possible.
280 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR
281 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND
282 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR
283 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs
284 return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen));
285 else // Overflow; assume all integers
286 return TypeInt::INT;
287 }
288
289 //=============================================================================
290 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)291 Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
292 Node *in1 = in(1);
293 Node *in2 = in(2);
294 uint op1 = in1->Opcode();
295 uint op2 = in2->Opcode();
296
297 #ifdef ASSERT
298 // Check for dead loop
299 if( phase->eqv( in1, this ) || phase->eqv( in2, this ) ||
300 ( ( op1 == Op_AddL || op1 == Op_SubL ) &&
301 ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) ||
302 phase->eqv( in1->in(1), in1 ) || phase->eqv( in1->in(2), in1 ) ) ) )
303 assert(false, "dead loop in SubLNode::Ideal");
304 #endif
305
306 if( phase->type( in2 ) == Type::TOP ) return NULL;
307 const TypeLong *i = phase->type( in2 )->isa_long();
308 // Convert "x-c0" into "x+ -c0".
309 if( i && // Might be bottom or top...
310 i->is_con() )
311 return new AddLNode(in1, phase->longcon(-i->get_con()));
312
313 // Convert "(x+c0) - y" into (x-y) + c0"
314 // Do not collapse (x+c0)-y if "+" is a loop increment or
315 // if "y" is a loop induction variable.
316 if( op1 == Op_AddL && ok_to_convert(in1, in2) ) {
317 Node *in11 = in1->in(1);
318 const Type *tadd = phase->type( in1->in(2) );
319 if( tadd->singleton() && tadd != Type::TOP ) {
320 Node *sub2 = phase->transform( new SubLNode( in11, in2 ));
321 return new AddLNode( sub2, in1->in(2) );
322 }
323 }
324
325 // Convert "x - (y+c0)" into "(x-y) - c0"
326 // Need the same check as in above optimization but reversed.
327 if (op2 == Op_AddL && ok_to_convert(in2, in1)) {
328 Node* in21 = in2->in(1);
329 Node* in22 = in2->in(2);
330 const TypeLong* tcon = phase->type(in22)->isa_long();
331 if (tcon != NULL && tcon->is_con()) {
332 Node* sub2 = phase->transform( new SubLNode(in1, in21) );
333 Node* neg_c0 = phase->longcon(- tcon->get_con());
334 return new AddLNode(sub2, neg_c0);
335 }
336 }
337
338 const Type *t1 = phase->type( in1 );
339 if( t1 == Type::TOP ) return NULL;
340
341 #ifdef ASSERT
342 // Check for dead loop
343 if( ( op2 == Op_AddL || op2 == Op_SubL ) &&
344 ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) ||
345 phase->eqv( in2->in(1), in2 ) || phase->eqv( in2->in(2), in2 ) ) )
346 assert(false, "dead loop in SubLNode::Ideal");
347 #endif
348
349 // Convert "x - (x+y)" into "-y"
350 if( op2 == Op_AddL &&
351 phase->eqv( in1, in2->in(1) ) )
352 return new SubLNode( phase->makecon(TypeLong::ZERO), in2->in(2));
353 // Convert "x - (y+x)" into "-y"
354 if( op2 == Op_AddL &&
355 phase->eqv( in1, in2->in(2) ) )
356 return new SubLNode( phase->makecon(TypeLong::ZERO),in2->in(1));
357
358 // Convert "0 - (x-y)" into "y-x"
359 if( phase->type( in1 ) == TypeLong::ZERO && op2 == Op_SubL )
360 return new SubLNode( in2->in(2), in2->in(1) );
361
362 // Convert "(X+A) - (X+B)" into "A - B"
363 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) )
364 return new SubLNode( in1->in(2), in2->in(2) );
365
366 // Convert "(A+X) - (B+X)" into "A - B"
367 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) )
368 return new SubLNode( in1->in(1), in2->in(1) );
369
370 // Convert "A-(B-C)" into (A+C)-B"
371 if( op2 == Op_SubL && in2->outcnt() == 1) {
372 Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) );
373 return new SubLNode( add1, in2->in(1) );
374 }
375
376 // Convert "0L-(A>>63)" into "(A>>>63)"
377 if ( op2 == Op_RShiftL ) {
378 Node *in21 = in2->in(1);
379 Node *in22 = in2->in(2);
380 const TypeLong *zero = phase->type(in1)->isa_long();
381 const TypeLong *t21 = phase->type(in21)->isa_long();
382 const TypeInt *t22 = phase->type(in22)->isa_int();
383 if ( t21 && t22 && zero == TypeLong::ZERO && t22->is_con(63) ) {
384 return new URShiftLNode(in21, in22);
385 }
386 }
387
388 return NULL;
389 }
390
391 //------------------------------sub--------------------------------------------
392 // A subtract node differences it's two inputs.
sub(const Type * t1,const Type * t2) const393 const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const {
394 const TypeLong *r0 = t1->is_long(); // Handy access
395 const TypeLong *r1 = t2->is_long();
396 jlong lo = java_subtract(r0->_lo, r1->_hi);
397 jlong hi = java_subtract(r0->_hi, r1->_lo);
398
399 // We next check for 32-bit overflow.
400 // If that happens, we just assume all integers are possible.
401 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR
402 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND
403 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR
404 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs
405 return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen));
406 else // Overflow; assume all integers
407 return TypeLong::LONG;
408 }
409
410 //=============================================================================
411 //------------------------------Value------------------------------------------
412 // A subtract node differences its two inputs.
Value(PhaseGVN * phase) const413 const Type* SubFPNode::Value(PhaseGVN* phase) const {
414 const Node* in1 = in(1);
415 const Node* in2 = in(2);
416 // Either input is TOP ==> the result is TOP
417 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
418 if( t1 == Type::TOP ) return Type::TOP;
419 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
420 if( t2 == Type::TOP ) return Type::TOP;
421
422 // if both operands are infinity of same sign, the result is NaN; do
423 // not replace with zero
424 if( (t1->is_finite() && t2->is_finite()) ) {
425 if( phase->eqv(in1, in2) ) return add_id();
426 }
427
428 // Either input is BOTTOM ==> the result is the local BOTTOM
429 const Type *bot = bottom_type();
430 if( (t1 == bot) || (t2 == bot) ||
431 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
432 return bot;
433
434 return sub(t1,t2); // Local flavor of type subtraction
435 }
436
437
438 //=============================================================================
439 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)440 Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
441 const Type *t2 = phase->type( in(2) );
442 // Convert "x-c0" into "x+ -c0".
443 if( t2->base() == Type::FloatCon ) { // Might be bottom or top...
444 // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) );
445 }
446
447 // Not associative because of boundary conditions (infinity)
448 if( IdealizedNumerics && !phase->C->method()->is_strict() ) {
449 // Convert "x - (x+y)" into "-y"
450 if( in(2)->is_Add() &&
451 phase->eqv(in(1),in(2)->in(1) ) )
452 return new SubFNode( phase->makecon(TypeF::ZERO),in(2)->in(2));
453 }
454
455 // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes
456 // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0.
457 //if( phase->type(in(1)) == TypeF::ZERO )
458 //return new (phase->C, 2) NegFNode(in(2));
459
460 return NULL;
461 }
462
463 //------------------------------sub--------------------------------------------
464 // A subtract node differences its two inputs.
sub(const Type * t1,const Type * t2) const465 const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const {
466 // no folding if one of operands is infinity or NaN, do not do constant folding
467 if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) {
468 return TypeF::make( t1->getf() - t2->getf() );
469 }
470 else if( g_isnan(t1->getf()) ) {
471 return t1;
472 }
473 else if( g_isnan(t2->getf()) ) {
474 return t2;
475 }
476 else {
477 return Type::FLOAT;
478 }
479 }
480
481 //=============================================================================
482 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)483 Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){
484 const Type *t2 = phase->type( in(2) );
485 // Convert "x-c0" into "x+ -c0".
486 if( t2->base() == Type::DoubleCon ) { // Might be bottom or top...
487 // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) );
488 }
489
490 // Not associative because of boundary conditions (infinity)
491 if( IdealizedNumerics && !phase->C->method()->is_strict() ) {
492 // Convert "x - (x+y)" into "-y"
493 if( in(2)->is_Add() &&
494 phase->eqv(in(1),in(2)->in(1) ) )
495 return new SubDNode( phase->makecon(TypeD::ZERO),in(2)->in(2));
496 }
497
498 // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes
499 // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0.
500 //if( phase->type(in(1)) == TypeD::ZERO )
501 //return new (phase->C, 2) NegDNode(in(2));
502
503 return NULL;
504 }
505
506 //------------------------------sub--------------------------------------------
507 // A subtract node differences its two inputs.
sub(const Type * t1,const Type * t2) const508 const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const {
509 // no folding if one of operands is infinity or NaN, do not do constant folding
510 if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) {
511 return TypeD::make( t1->getd() - t2->getd() );
512 }
513 else if( g_isnan(t1->getd()) ) {
514 return t1;
515 }
516 else if( g_isnan(t2->getd()) ) {
517 return t2;
518 }
519 else {
520 return Type::DOUBLE;
521 }
522 }
523
524 //=============================================================================
525 //------------------------------Idealize---------------------------------------
526 // Unlike SubNodes, compare must still flatten return value to the
527 // range -1, 0, 1.
528 // And optimizations like those for (X + Y) - X fail if overflow happens.
Identity(PhaseGVN * phase)529 Node* CmpNode::Identity(PhaseGVN* phase) {
530 return this;
531 }
532
533 #ifndef PRODUCT
534 //----------------------------related------------------------------------------
535 // Related nodes of comparison nodes include all data inputs (until hitting a
536 // control boundary) as well as all outputs until and including control nodes
537 // as well as their projections. In compact mode, data inputs till depth 1 and
538 // all outputs till depth 1 are considered.
related(GrowableArray<Node * > * in_rel,GrowableArray<Node * > * out_rel,bool compact) const539 void CmpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
540 if (compact) {
541 this->collect_nodes(in_rel, 1, false, true);
542 this->collect_nodes(out_rel, -1, false, false);
543 } else {
544 this->collect_nodes_in_all_data(in_rel, false);
545 this->collect_nodes_out_all_ctrl_boundary(out_rel);
546 // Now, find all control nodes in out_rel, and include their projections
547 // and projection targets (if any) in the result.
548 GrowableArray<Node*> proj(Compile::current()->unique());
549 for (GrowableArrayIterator<Node*> it = out_rel->begin(); it != out_rel->end(); ++it) {
550 Node* n = *it;
551 if (n->is_CFG() && !n->is_Proj()) {
552 // Assume projections and projection targets are found at levels 1 and 2.
553 n->collect_nodes(&proj, -2, false, false);
554 for (GrowableArrayIterator<Node*> p = proj.begin(); p != proj.end(); ++p) {
555 out_rel->append_if_missing(*p);
556 }
557 proj.clear();
558 }
559 }
560 }
561 }
562 #endif
563
564 //=============================================================================
565 //------------------------------cmp--------------------------------------------
566 // Simplify a CmpI (compare 2 integers) node, based on local information.
567 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const568 const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const {
569 const TypeInt *r0 = t1->is_int(); // Handy access
570 const TypeInt *r1 = t2->is_int();
571
572 if( r0->_hi < r1->_lo ) // Range is always low?
573 return TypeInt::CC_LT;
574 else if( r0->_lo > r1->_hi ) // Range is always high?
575 return TypeInt::CC_GT;
576
577 else if( r0->is_con() && r1->is_con() ) { // comparing constants?
578 assert(r0->get_con() == r1->get_con(), "must be equal");
579 return TypeInt::CC_EQ; // Equal results.
580 } else if( r0->_hi == r1->_lo ) // Range is never high?
581 return TypeInt::CC_LE;
582 else if( r0->_lo == r1->_hi ) // Range is never low?
583 return TypeInt::CC_GE;
584 return TypeInt::CC; // else use worst case results
585 }
586
587 // Simplify a CmpU (compare 2 integers) node, based on local information.
588 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const589 const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const {
590 assert(!t1->isa_ptr(), "obsolete usage of CmpU");
591
592 // comparing two unsigned ints
593 const TypeInt *r0 = t1->is_int(); // Handy access
594 const TypeInt *r1 = t2->is_int();
595
596 // Current installed version
597 // Compare ranges for non-overlap
598 juint lo0 = r0->_lo;
599 juint hi0 = r0->_hi;
600 juint lo1 = r1->_lo;
601 juint hi1 = r1->_hi;
602
603 // If either one has both negative and positive values,
604 // it therefore contains both 0 and -1, and since [0..-1] is the
605 // full unsigned range, the type must act as an unsigned bottom.
606 bool bot0 = ((jint)(lo0 ^ hi0) < 0);
607 bool bot1 = ((jint)(lo1 ^ hi1) < 0);
608
609 if (bot0 || bot1) {
610 // All unsigned values are LE -1 and GE 0.
611 if (lo0 == 0 && hi0 == 0) {
612 return TypeInt::CC_LE; // 0 <= bot
613 } else if ((jint)lo0 == -1 && (jint)hi0 == -1) {
614 return TypeInt::CC_GE; // -1 >= bot
615 } else if (lo1 == 0 && hi1 == 0) {
616 return TypeInt::CC_GE; // bot >= 0
617 } else if ((jint)lo1 == -1 && (jint)hi1 == -1) {
618 return TypeInt::CC_LE; // bot <= -1
619 }
620 } else {
621 // We can use ranges of the form [lo..hi] if signs are the same.
622 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
623 // results are reversed, '-' > '+' for unsigned compare
624 if (hi0 < lo1) {
625 return TypeInt::CC_LT; // smaller
626 } else if (lo0 > hi1) {
627 return TypeInt::CC_GT; // greater
628 } else if (hi0 == lo1 && lo0 == hi1) {
629 return TypeInt::CC_EQ; // Equal results
630 } else if (lo0 >= hi1) {
631 return TypeInt::CC_GE;
632 } else if (hi0 <= lo1) {
633 // Check for special case in Hashtable::get. (See below.)
634 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
635 return TypeInt::CC_LT;
636 return TypeInt::CC_LE;
637 }
638 }
639 // Check for special case in Hashtable::get - the hash index is
640 // mod'ed to the table size so the following range check is useless.
641 // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have
642 // to be positive.
643 // (This is a gross hack, since the sub method never
644 // looks at the structure of the node in any other case.)
645 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
646 return TypeInt::CC_LT;
647 return TypeInt::CC; // else use worst case results
648 }
649
Value(PhaseGVN * phase) const650 const Type* CmpUNode::Value(PhaseGVN* phase) const {
651 const Type* t = SubNode::Value_common(phase);
652 if (t != NULL) {
653 return t;
654 }
655 const Node* in1 = in(1);
656 const Node* in2 = in(2);
657 const Type* t1 = phase->type(in1);
658 const Type* t2 = phase->type(in2);
659 assert(t1->isa_int(), "CmpU has only Int type inputs");
660 if (t2 == TypeInt::INT) { // Compare to bottom?
661 return bottom_type();
662 }
663 uint in1_op = in1->Opcode();
664 if (in1_op == Op_AddI || in1_op == Op_SubI) {
665 // The problem rise when result of AddI(SubI) may overflow
666 // signed integer value. Let say the input type is
667 // [256, maxint] then +128 will create 2 ranges due to
668 // overflow: [minint, minint+127] and [384, maxint].
669 // But C2 type system keep only 1 type range and as result
670 // it use general [minint, maxint] for this case which we
671 // can't optimize.
672 //
673 // Make 2 separate type ranges based on types of AddI(SubI) inputs
674 // and compare results of their compare. If results are the same
675 // CmpU node can be optimized.
676 const Node* in11 = in1->in(1);
677 const Node* in12 = in1->in(2);
678 const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11);
679 const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12);
680 // Skip cases when input types are top or bottom.
681 if ((t11 != Type::TOP) && (t11 != TypeInt::INT) &&
682 (t12 != Type::TOP) && (t12 != TypeInt::INT)) {
683 const TypeInt *r0 = t11->is_int();
684 const TypeInt *r1 = t12->is_int();
685 jlong lo_r0 = r0->_lo;
686 jlong hi_r0 = r0->_hi;
687 jlong lo_r1 = r1->_lo;
688 jlong hi_r1 = r1->_hi;
689 if (in1_op == Op_SubI) {
690 jlong tmp = hi_r1;
691 hi_r1 = -lo_r1;
692 lo_r1 = -tmp;
693 // Note, for substructing [minint,x] type range
694 // long arithmetic provides correct overflow answer.
695 // The confusion come from the fact that in 32-bit
696 // -minint == minint but in 64-bit -minint == maxint+1.
697 }
698 jlong lo_long = lo_r0 + lo_r1;
699 jlong hi_long = hi_r0 + hi_r1;
700 int lo_tr1 = min_jint;
701 int hi_tr1 = (int)hi_long;
702 int lo_tr2 = (int)lo_long;
703 int hi_tr2 = max_jint;
704 bool underflow = lo_long != (jlong)lo_tr2;
705 bool overflow = hi_long != (jlong)hi_tr1;
706 // Use sub(t1, t2) when there is no overflow (one type range)
707 // or when both overflow and underflow (too complex).
708 if ((underflow != overflow) && (hi_tr1 < lo_tr2)) {
709 // Overflow only on one boundary, compare 2 separate type ranges.
710 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
711 const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w);
712 const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w);
713 const Type* cmp1 = sub(tr1, t2);
714 const Type* cmp2 = sub(tr2, t2);
715 if (cmp1 == cmp2) {
716 return cmp1; // Hit!
717 }
718 }
719 }
720 }
721
722 return sub(t1, t2); // Local flavor of type subtraction
723 }
724
is_index_range_check() const725 bool CmpUNode::is_index_range_check() const {
726 // Check for the "(X ModI Y) CmpU Y" shape
727 return (in(1)->Opcode() == Op_ModI &&
728 in(1)->in(2)->eqv_uncast(in(2)));
729 }
730
731 //------------------------------Idealize---------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)732 Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) {
733 if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) {
734 switch (in(1)->Opcode()) {
735 case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL
736 return new CmpLNode(in(1)->in(1),in(1)->in(2));
737 case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF
738 return new CmpFNode(in(1)->in(1),in(1)->in(2));
739 case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD
740 return new CmpDNode(in(1)->in(1),in(1)->in(2));
741 //case Op_SubI:
742 // If (x - y) cannot overflow, then ((x - y) <?> 0)
743 // can be turned into (x <?> y).
744 // This is handled (with more general cases) by Ideal_sub_algebra.
745 }
746 }
747 return NULL; // No change
748 }
749
750
751 //=============================================================================
752 // Simplify a CmpL (compare 2 longs ) node, based on local information.
753 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const754 const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const {
755 const TypeLong *r0 = t1->is_long(); // Handy access
756 const TypeLong *r1 = t2->is_long();
757
758 if( r0->_hi < r1->_lo ) // Range is always low?
759 return TypeInt::CC_LT;
760 else if( r0->_lo > r1->_hi ) // Range is always high?
761 return TypeInt::CC_GT;
762
763 else if( r0->is_con() && r1->is_con() ) { // comparing constants?
764 assert(r0->get_con() == r1->get_con(), "must be equal");
765 return TypeInt::CC_EQ; // Equal results.
766 } else if( r0->_hi == r1->_lo ) // Range is never high?
767 return TypeInt::CC_LE;
768 else if( r0->_lo == r1->_hi ) // Range is never low?
769 return TypeInt::CC_GE;
770 return TypeInt::CC; // else use worst case results
771 }
772
773
774 // Simplify a CmpUL (compare 2 unsigned longs) node, based on local information.
775 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const776 const Type* CmpULNode::sub(const Type* t1, const Type* t2) const {
777 assert(!t1->isa_ptr(), "obsolete usage of CmpUL");
778
779 // comparing two unsigned longs
780 const TypeLong* r0 = t1->is_long(); // Handy access
781 const TypeLong* r1 = t2->is_long();
782
783 // Current installed version
784 // Compare ranges for non-overlap
785 julong lo0 = r0->_lo;
786 julong hi0 = r0->_hi;
787 julong lo1 = r1->_lo;
788 julong hi1 = r1->_hi;
789
790 // If either one has both negative and positive values,
791 // it therefore contains both 0 and -1, and since [0..-1] is the
792 // full unsigned range, the type must act as an unsigned bottom.
793 bool bot0 = ((jlong)(lo0 ^ hi0) < 0);
794 bool bot1 = ((jlong)(lo1 ^ hi1) < 0);
795
796 if (bot0 || bot1) {
797 // All unsigned values are LE -1 and GE 0.
798 if (lo0 == 0 && hi0 == 0) {
799 return TypeInt::CC_LE; // 0 <= bot
800 } else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) {
801 return TypeInt::CC_GE; // -1 >= bot
802 } else if (lo1 == 0 && hi1 == 0) {
803 return TypeInt::CC_GE; // bot >= 0
804 } else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) {
805 return TypeInt::CC_LE; // bot <= -1
806 }
807 } else {
808 // We can use ranges of the form [lo..hi] if signs are the same.
809 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
810 // results are reversed, '-' > '+' for unsigned compare
811 if (hi0 < lo1) {
812 return TypeInt::CC_LT; // smaller
813 } else if (lo0 > hi1) {
814 return TypeInt::CC_GT; // greater
815 } else if (hi0 == lo1 && lo0 == hi1) {
816 return TypeInt::CC_EQ; // Equal results
817 } else if (lo0 >= hi1) {
818 return TypeInt::CC_GE;
819 } else if (hi0 <= lo1) {
820 return TypeInt::CC_LE;
821 }
822 }
823
824 return TypeInt::CC; // else use worst case results
825 }
826
827 //=============================================================================
828 //------------------------------sub--------------------------------------------
829 // Simplify an CmpP (compare 2 pointers) node, based on local information.
830 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const831 const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const {
832 const TypePtr *r0 = t1->is_ptr(); // Handy access
833 const TypePtr *r1 = t2->is_ptr();
834
835 // Undefined inputs makes for an undefined result
836 if( TypePtr::above_centerline(r0->_ptr) ||
837 TypePtr::above_centerline(r1->_ptr) )
838 return Type::TOP;
839
840 if (r0 == r1 && r0->singleton()) {
841 // Equal pointer constants (klasses, nulls, etc.)
842 return TypeInt::CC_EQ;
843 }
844
845 // See if it is 2 unrelated classes.
846 const TypeOopPtr* oop_p0 = r0->isa_oopptr();
847 const TypeOopPtr* oop_p1 = r1->isa_oopptr();
848 bool both_oop_ptr = oop_p0 && oop_p1;
849
850 if (both_oop_ptr) {
851 Node* in1 = in(1)->uncast();
852 Node* in2 = in(2)->uncast();
853 AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL);
854 AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL);
855 if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) {
856 return TypeInt::CC_GT; // different pointers
857 }
858 }
859
860 const TypeKlassPtr* klass_p0 = r0->isa_klassptr();
861 const TypeKlassPtr* klass_p1 = r1->isa_klassptr();
862
863 if (both_oop_ptr || (klass_p0 && klass_p1)) { // both or neither are klass pointers
864 ciKlass* klass0 = NULL;
865 bool xklass0 = false;
866 ciKlass* klass1 = NULL;
867 bool xklass1 = false;
868
869 if (oop_p0) {
870 klass0 = oop_p0->klass();
871 xklass0 = oop_p0->klass_is_exact();
872 } else {
873 assert(klass_p0, "must be non-null if oop_p0 is null");
874 klass0 = klass_p0->klass();
875 xklass0 = klass_p0->klass_is_exact();
876 }
877
878 if (oop_p1) {
879 klass1 = oop_p1->klass();
880 xklass1 = oop_p1->klass_is_exact();
881 } else {
882 assert(klass_p1, "must be non-null if oop_p1 is null");
883 klass1 = klass_p1->klass();
884 xklass1 = klass_p1->klass_is_exact();
885 }
886
887 if (klass0 && klass1 &&
888 klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces
889 klass1->is_loaded() && !klass1->is_interface() &&
890 (!klass0->is_obj_array_klass() ||
891 !klass0->as_obj_array_klass()->base_element_klass()->is_interface()) &&
892 (!klass1->is_obj_array_klass() ||
893 !klass1->as_obj_array_klass()->base_element_klass()->is_interface())) {
894 bool unrelated_classes = false;
895 // See if neither subclasses the other, or if the class on top
896 // is precise. In either of these cases, the compare is known
897 // to fail if at least one of the pointers is provably not null.
898 if (klass0->equals(klass1)) { // if types are unequal but klasses are equal
899 // Do nothing; we know nothing for imprecise types
900 } else if (klass0->is_subtype_of(klass1)) {
901 // If klass1's type is PRECISE, then classes are unrelated.
902 unrelated_classes = xklass1;
903 } else if (klass1->is_subtype_of(klass0)) {
904 // If klass0's type is PRECISE, then classes are unrelated.
905 unrelated_classes = xklass0;
906 } else { // Neither subtypes the other
907 unrelated_classes = true;
908 }
909 if (unrelated_classes) {
910 // The oops classes are known to be unrelated. If the joined PTRs of
911 // two oops is not Null and not Bottom, then we are sure that one
912 // of the two oops is non-null, and the comparison will always fail.
913 TypePtr::PTR jp = r0->join_ptr(r1->_ptr);
914 if (jp != TypePtr::Null && jp != TypePtr::BotPTR) {
915 return TypeInt::CC_GT;
916 }
917 }
918 }
919 }
920
921 // Known constants can be compared exactly
922 // Null can be distinguished from any NotNull pointers
923 // Unknown inputs makes an unknown result
924 if( r0->singleton() ) {
925 intptr_t bits0 = r0->get_con();
926 if( r1->singleton() )
927 return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT;
928 return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC;
929 } else if( r1->singleton() ) {
930 intptr_t bits1 = r1->get_con();
931 return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC;
932 } else
933 return TypeInt::CC;
934 }
935
isa_java_mirror_load(PhaseGVN * phase,Node * n)936 static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) {
937 // Return the klass node for (indirect load from OopHandle)
938 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
939 // or NULL if not matching.
940 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
941 n = bs->step_over_gc_barrier(n);
942
943 if (n->Opcode() != Op_LoadP) return NULL;
944
945 const TypeInstPtr* tp = phase->type(n)->isa_instptr();
946 if (!tp || tp->klass() != phase->C->env()->Class_klass()) return NULL;
947
948 Node* adr = n->in(MemNode::Address);
949 // First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier.
950 if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return NULL;
951 adr = adr->in(MemNode::Address);
952
953 intptr_t off = 0;
954 Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off);
955 if (k == NULL) return NULL;
956 const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr();
957 if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return NULL;
958
959 // We've found the klass node of a Java mirror load.
960 return k;
961 }
962
isa_const_java_mirror(PhaseGVN * phase,Node * n)963 static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) {
964 // for ConP(Foo.class) return ConP(Foo.klass)
965 // otherwise return NULL
966 if (!n->is_Con()) return NULL;
967
968 const TypeInstPtr* tp = phase->type(n)->isa_instptr();
969 if (!tp) return NULL;
970
971 ciType* mirror_type = tp->java_mirror_type();
972 // TypeInstPtr::java_mirror_type() returns non-NULL for compile-
973 // time Class constants only.
974 if (!mirror_type) return NULL;
975
976 // x.getClass() == int.class can never be true (for all primitive types)
977 // Return a ConP(NULL) node for this case.
978 if (mirror_type->is_classless()) {
979 return phase->makecon(TypePtr::NULL_PTR);
980 }
981
982 // return the ConP(Foo.klass)
983 assert(mirror_type->is_klass(), "mirror_type should represent a Klass*");
984 return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass()));
985 }
986
987 //------------------------------Ideal------------------------------------------
988 // Normalize comparisons between Java mirror loads to compare the klass instead.
989 //
990 // Also check for the case of comparing an unknown klass loaded from the primary
991 // super-type array vs a known klass with no subtypes. This amounts to
992 // checking to see an unknown klass subtypes a known klass with no subtypes;
993 // this only happens on an exact match. We can shorten this test by 1 load.
Ideal(PhaseGVN * phase,bool can_reshape)994 Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
995 // Normalize comparisons between Java mirrors into comparisons of the low-
996 // level klass, where a dependent load could be shortened.
997 //
998 // The new pattern has a nice effect of matching the same pattern used in the
999 // fast path of instanceof/checkcast/Class.isInstance(), which allows
1000 // redundant exact type check be optimized away by GVN.
1001 // For example, in
1002 // if (x.getClass() == Foo.class) {
1003 // Foo foo = (Foo) x;
1004 // // ... use a ...
1005 // }
1006 // a CmpPNode could be shared between if_acmpne and checkcast
1007 {
1008 Node* k1 = isa_java_mirror_load(phase, in(1));
1009 Node* k2 = isa_java_mirror_load(phase, in(2));
1010 Node* conk2 = isa_const_java_mirror(phase, in(2));
1011
1012 if (k1 && (k2 || conk2)) {
1013 Node* lhs = k1;
1014 Node* rhs = (k2 != NULL) ? k2 : conk2;
1015 PhaseIterGVN* igvn = phase->is_IterGVN();
1016 if (igvn != NULL) {
1017 set_req_X(1, lhs, igvn);
1018 set_req_X(2, rhs, igvn);
1019 } else {
1020 set_req(1, lhs);
1021 set_req(2, rhs);
1022 }
1023 return this;
1024 }
1025 }
1026
1027 // Constant pointer on right?
1028 const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr();
1029 if (t2 == NULL || !t2->klass_is_exact())
1030 return NULL;
1031 // Get the constant klass we are comparing to.
1032 ciKlass* superklass = t2->klass();
1033
1034 // Now check for LoadKlass on left.
1035 Node* ldk1 = in(1);
1036 if (ldk1->is_DecodeNKlass()) {
1037 ldk1 = ldk1->in(1);
1038 if (ldk1->Opcode() != Op_LoadNKlass )
1039 return NULL;
1040 } else if (ldk1->Opcode() != Op_LoadKlass )
1041 return NULL;
1042 // Take apart the address of the LoadKlass:
1043 Node* adr1 = ldk1->in(MemNode::Address);
1044 intptr_t con2 = 0;
1045 Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2);
1046 if (ldk2 == NULL)
1047 return NULL;
1048 if (con2 == oopDesc::klass_offset_in_bytes()) {
1049 // We are inspecting an object's concrete class.
1050 // Short-circuit the check if the query is abstract.
1051 if (superklass->is_interface() ||
1052 superklass->is_abstract()) {
1053 // Make it come out always false:
1054 this->set_req(2, phase->makecon(TypePtr::NULL_PTR));
1055 return this;
1056 }
1057 }
1058
1059 // Check for a LoadKlass from primary supertype array.
1060 // Any nested loadklass from loadklass+con must be from the p.s. array.
1061 if (ldk2->is_DecodeNKlass()) {
1062 // Keep ldk2 as DecodeN since it could be used in CmpP below.
1063 if (ldk2->in(1)->Opcode() != Op_LoadNKlass )
1064 return NULL;
1065 } else if (ldk2->Opcode() != Op_LoadKlass)
1066 return NULL;
1067
1068 // Verify that we understand the situation
1069 if (con2 != (intptr_t) superklass->super_check_offset())
1070 return NULL; // Might be element-klass loading from array klass
1071
1072 // If 'superklass' has no subklasses and is not an interface, then we are
1073 // assured that the only input which will pass the type check is
1074 // 'superklass' itself.
1075 //
1076 // We could be more liberal here, and allow the optimization on interfaces
1077 // which have a single implementor. This would require us to increase the
1078 // expressiveness of the add_dependency() mechanism.
1079 // %%% Do this after we fix TypeOopPtr: Deps are expressive enough now.
1080
1081 // Object arrays must have their base element have no subtypes
1082 while (superklass->is_obj_array_klass()) {
1083 ciType* elem = superklass->as_obj_array_klass()->element_type();
1084 superklass = elem->as_klass();
1085 }
1086 if (superklass->is_instance_klass()) {
1087 ciInstanceKlass* ik = superklass->as_instance_klass();
1088 if (ik->has_subklass() || ik->is_interface()) return NULL;
1089 // Add a dependency if there is a chance that a subclass will be added later.
1090 if (!ik->is_final()) {
1091 phase->C->dependencies()->assert_leaf_type(ik);
1092 }
1093 }
1094
1095 // Bypass the dependent load, and compare directly
1096 this->set_req(1,ldk2);
1097
1098 return this;
1099 }
1100
1101 //=============================================================================
1102 //------------------------------sub--------------------------------------------
1103 // Simplify an CmpN (compare 2 pointers) node, based on local information.
1104 // If both inputs are constants, compare them.
sub(const Type * t1,const Type * t2) const1105 const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const {
1106 ShouldNotReachHere();
1107 return bottom_type();
1108 }
1109
1110 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)1111 Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
1112 return NULL;
1113 }
1114
1115 //=============================================================================
1116 //------------------------------Value------------------------------------------
1117 // Simplify an CmpF (compare 2 floats ) node, based on local information.
1118 // If both inputs are constants, compare them.
Value(PhaseGVN * phase) const1119 const Type* CmpFNode::Value(PhaseGVN* phase) const {
1120 const Node* in1 = in(1);
1121 const Node* in2 = in(2);
1122 // Either input is TOP ==> the result is TOP
1123 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
1124 if( t1 == Type::TOP ) return Type::TOP;
1125 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
1126 if( t2 == Type::TOP ) return Type::TOP;
1127
1128 // Not constants? Don't know squat - even if they are the same
1129 // value! If they are NaN's they compare to LT instead of EQ.
1130 const TypeF *tf1 = t1->isa_float_constant();
1131 const TypeF *tf2 = t2->isa_float_constant();
1132 if( !tf1 || !tf2 ) return TypeInt::CC;
1133
1134 // This implements the Java bytecode fcmpl, so unordered returns -1.
1135 if( tf1->is_nan() || tf2->is_nan() )
1136 return TypeInt::CC_LT;
1137
1138 if( tf1->_f < tf2->_f ) return TypeInt::CC_LT;
1139 if( tf1->_f > tf2->_f ) return TypeInt::CC_GT;
1140 assert( tf1->_f == tf2->_f, "do not understand FP behavior" );
1141 return TypeInt::CC_EQ;
1142 }
1143
1144
1145 //=============================================================================
1146 //------------------------------Value------------------------------------------
1147 // Simplify an CmpD (compare 2 doubles ) node, based on local information.
1148 // If both inputs are constants, compare them.
Value(PhaseGVN * phase) const1149 const Type* CmpDNode::Value(PhaseGVN* phase) const {
1150 const Node* in1 = in(1);
1151 const Node* in2 = in(2);
1152 // Either input is TOP ==> the result is TOP
1153 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
1154 if( t1 == Type::TOP ) return Type::TOP;
1155 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
1156 if( t2 == Type::TOP ) return Type::TOP;
1157
1158 // Not constants? Don't know squat - even if they are the same
1159 // value! If they are NaN's they compare to LT instead of EQ.
1160 const TypeD *td1 = t1->isa_double_constant();
1161 const TypeD *td2 = t2->isa_double_constant();
1162 if( !td1 || !td2 ) return TypeInt::CC;
1163
1164 // This implements the Java bytecode dcmpl, so unordered returns -1.
1165 if( td1->is_nan() || td2->is_nan() )
1166 return TypeInt::CC_LT;
1167
1168 if( td1->_d < td2->_d ) return TypeInt::CC_LT;
1169 if( td1->_d > td2->_d ) return TypeInt::CC_GT;
1170 assert( td1->_d == td2->_d, "do not understand FP behavior" );
1171 return TypeInt::CC_EQ;
1172 }
1173
1174 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)1175 Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){
1176 // Check if we can change this to a CmpF and remove a ConvD2F operation.
1177 // Change (CMPD (F2D (float)) (ConD value))
1178 // To (CMPF (float) (ConF value))
1179 // Valid when 'value' does not lose precision as a float.
1180 // Benefits: eliminates conversion, does not require 24-bit mode
1181
1182 // NaNs prevent commuting operands. This transform works regardless of the
1183 // order of ConD and ConvF2D inputs by preserving the original order.
1184 int idx_f2d = 1; // ConvF2D on left side?
1185 if( in(idx_f2d)->Opcode() != Op_ConvF2D )
1186 idx_f2d = 2; // No, swap to check for reversed args
1187 int idx_con = 3-idx_f2d; // Check for the constant on other input
1188
1189 if( ConvertCmpD2CmpF &&
1190 in(idx_f2d)->Opcode() == Op_ConvF2D &&
1191 in(idx_con)->Opcode() == Op_ConD ) {
1192 const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant();
1193 double t2_value_as_double = t2->_d;
1194 float t2_value_as_float = (float)t2_value_as_double;
1195 if( t2_value_as_double == (double)t2_value_as_float ) {
1196 // Test value can be represented as a float
1197 // Eliminate the conversion to double and create new comparison
1198 Node *new_in1 = in(idx_f2d)->in(1);
1199 Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) );
1200 if( idx_f2d != 1 ) { // Must flip args to match original order
1201 Node *tmp = new_in1;
1202 new_in1 = new_in2;
1203 new_in2 = tmp;
1204 }
1205 CmpFNode *new_cmp = (Opcode() == Op_CmpD3)
1206 ? new CmpF3Node( new_in1, new_in2 )
1207 : new CmpFNode ( new_in1, new_in2 ) ;
1208 return new_cmp; // Changed to CmpFNode
1209 }
1210 // Testing value required the precision of a double
1211 }
1212 return NULL; // No change
1213 }
1214
1215
1216 //=============================================================================
1217 //------------------------------cc2logical-------------------------------------
1218 // Convert a condition code type to a logical type
cc2logical(const Type * CC) const1219 const Type *BoolTest::cc2logical( const Type *CC ) const {
1220 if( CC == Type::TOP ) return Type::TOP;
1221 if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse
1222 const TypeInt *ti = CC->is_int();
1223 if( ti->is_con() ) { // Only 1 kind of condition codes set?
1224 // Match low order 2 bits
1225 int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0;
1226 if( _test & 4 ) tmp = 1-tmp; // Optionally complement result
1227 return TypeInt::make(tmp); // Boolean result
1228 }
1229
1230 if( CC == TypeInt::CC_GE ) {
1231 if( _test == ge ) return TypeInt::ONE;
1232 if( _test == lt ) return TypeInt::ZERO;
1233 }
1234 if( CC == TypeInt::CC_LE ) {
1235 if( _test == le ) return TypeInt::ONE;
1236 if( _test == gt ) return TypeInt::ZERO;
1237 }
1238
1239 return TypeInt::BOOL;
1240 }
1241
1242 //------------------------------dump_spec-------------------------------------
1243 // Print special per-node info
dump_on(outputStream * st) const1244 void BoolTest::dump_on(outputStream *st) const {
1245 const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"};
1246 st->print("%s", msg[_test]);
1247 }
1248
1249 // Returns the logical AND of two tests (or 'never' if both tests can never be true).
1250 // For example, a test for 'le' followed by a test for 'lt' is equivalent with 'lt'.
merge(BoolTest other) const1251 BoolTest::mask BoolTest::merge(BoolTest other) const {
1252 const mask res[illegal+1][illegal+1] = {
1253 // eq, gt, of, lt, ne, le, nof, ge, never, illegal
1254 {eq, never, illegal, never, never, eq, illegal, eq, never, illegal}, // eq
1255 {never, gt, illegal, never, gt, never, illegal, gt, never, illegal}, // gt
1256 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // of
1257 {never, never, illegal, lt, lt, lt, illegal, never, never, illegal}, // lt
1258 {never, gt, illegal, lt, ne, lt, illegal, gt, never, illegal}, // ne
1259 {eq, never, illegal, lt, lt, le, illegal, eq, never, illegal}, // le
1260 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // nof
1261 {eq, gt, illegal, never, gt, eq, illegal, ge, never, illegal}, // ge
1262 {never, never, never, never, never, never, never, never, never, illegal}, // never
1263 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal}}; // illegal
1264 return res[_test][other._test];
1265 }
1266
1267 //=============================================================================
hash() const1268 uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); }
size_of() const1269 uint BoolNode::size_of() const { return sizeof(BoolNode); }
1270
1271 //------------------------------operator==-------------------------------------
cmp(const Node & n) const1272 bool BoolNode::cmp( const Node &n ) const {
1273 const BoolNode *b = (const BoolNode *)&n; // Cast up
1274 return (_test._test == b->_test._test);
1275 }
1276
1277 //-------------------------------make_predicate--------------------------------
make_predicate(Node * test_value,PhaseGVN * phase)1278 Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) {
1279 if (test_value->is_Con()) return test_value;
1280 if (test_value->is_Bool()) return test_value;
1281 if (test_value->is_CMove() &&
1282 test_value->in(CMoveNode::Condition)->is_Bool()) {
1283 BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool();
1284 const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse));
1285 const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue));
1286 if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) {
1287 return bol;
1288 } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) {
1289 return phase->transform( bol->negate(phase) );
1290 }
1291 // Else fall through. The CMove gets in the way of the test.
1292 // It should be the case that make_predicate(bol->as_int_value()) == bol.
1293 }
1294 Node* cmp = new CmpINode(test_value, phase->intcon(0));
1295 cmp = phase->transform(cmp);
1296 Node* bol = new BoolNode(cmp, BoolTest::ne);
1297 return phase->transform(bol);
1298 }
1299
1300 //--------------------------------as_int_value---------------------------------
as_int_value(PhaseGVN * phase)1301 Node* BoolNode::as_int_value(PhaseGVN* phase) {
1302 // Inverse to make_predicate. The CMove probably boils down to a Conv2B.
1303 Node* cmov = CMoveNode::make(NULL, this,
1304 phase->intcon(0), phase->intcon(1),
1305 TypeInt::BOOL);
1306 return phase->transform(cmov);
1307 }
1308
1309 //----------------------------------negate-------------------------------------
negate(PhaseGVN * phase)1310 BoolNode* BoolNode::negate(PhaseGVN* phase) {
1311 return new BoolNode(in(1), _test.negate());
1312 }
1313
1314 // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub
1315 // overflows and we can prove that C is not in the two resulting ranges.
1316 // This optimization is similar to the one performed by CmpUNode::Value().
fold_cmpI(PhaseGVN * phase,SubNode * cmp,Node * cmp1,int cmp_op,int cmp1_op,const TypeInt * cmp2_type)1317 Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op,
1318 int cmp1_op, const TypeInt* cmp2_type) {
1319 // Only optimize eq/ne integer comparison of add/sub
1320 if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1321 (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) {
1322 // Skip cases were inputs of add/sub are not integers or of bottom type
1323 const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int();
1324 const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int();
1325 if ((r0 != NULL) && (r0 != TypeInt::INT) &&
1326 (r1 != NULL) && (r1 != TypeInt::INT) &&
1327 (cmp2_type != TypeInt::INT)) {
1328 // Compute exact (long) type range of add/sub result
1329 jlong lo_long = r0->_lo;
1330 jlong hi_long = r0->_hi;
1331 if (cmp1_op == Op_AddI) {
1332 lo_long += r1->_lo;
1333 hi_long += r1->_hi;
1334 } else {
1335 lo_long -= r1->_hi;
1336 hi_long -= r1->_lo;
1337 }
1338 // Check for over-/underflow by casting to integer
1339 int lo_int = (int)lo_long;
1340 int hi_int = (int)hi_long;
1341 bool underflow = lo_long != (jlong)lo_int;
1342 bool overflow = hi_long != (jlong)hi_int;
1343 if ((underflow != overflow) && (hi_int < lo_int)) {
1344 // Overflow on one boundary, compute resulting type ranges:
1345 // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT]
1346 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
1347 const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w);
1348 const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w);
1349 // Compare second input of cmp to both type ranges
1350 const Type* sub_tr1 = cmp->sub(tr1, cmp2_type);
1351 const Type* sub_tr2 = cmp->sub(tr2, cmp2_type);
1352 if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) {
1353 // The result of the add/sub will never equal cmp2. Replace BoolNode
1354 // by false (0) if it tests for equality and by true (1) otherwise.
1355 return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1);
1356 }
1357 }
1358 }
1359 }
1360 return NULL;
1361 }
1362
is_counted_loop_cmp(Node * cmp)1363 static bool is_counted_loop_cmp(Node *cmp) {
1364 Node *n = cmp->in(1)->in(1);
1365 return n != NULL &&
1366 n->is_Phi() &&
1367 n->in(0) != NULL &&
1368 n->in(0)->is_CountedLoop() &&
1369 n->in(0)->as_CountedLoop()->phi() == n;
1370 }
1371
1372 //------------------------------Ideal------------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)1373 Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1374 // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)".
1375 // This moves the constant to the right. Helps value-numbering.
1376 Node *cmp = in(1);
1377 if( !cmp->is_Sub() ) return NULL;
1378 int cop = cmp->Opcode();
1379 if( cop == Op_FastLock || cop == Op_FastUnlock || cmp->is_SubTypeCheck()) return NULL;
1380 Node *cmp1 = cmp->in(1);
1381 Node *cmp2 = cmp->in(2);
1382 if( !cmp1 ) return NULL;
1383
1384 if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) {
1385 return NULL;
1386 }
1387
1388 // Constant on left?
1389 Node *con = cmp1;
1390 uint op2 = cmp2->Opcode();
1391 // Move constants to the right of compare's to canonicalize.
1392 // Do not muck with Opaque1 nodes, as this indicates a loop
1393 // guard that cannot change shape.
1394 if( con->is_Con() && !cmp2->is_Con() && op2 != Op_Opaque1 &&
1395 // Because of NaN's, CmpD and CmpF are not commutative
1396 cop != Op_CmpD && cop != Op_CmpF &&
1397 // Protect against swapping inputs to a compare when it is used by a
1398 // counted loop exit, which requires maintaining the loop-limit as in(2)
1399 !is_counted_loop_exit_test() ) {
1400 // Ok, commute the constant to the right of the cmp node.
1401 // Clone the Node, getting a new Node of the same class
1402 cmp = cmp->clone();
1403 // Swap inputs to the clone
1404 cmp->swap_edges(1, 2);
1405 cmp = phase->transform( cmp );
1406 return new BoolNode( cmp, _test.commute() );
1407 }
1408
1409 // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)".
1410 // The XOR-1 is an idiom used to flip the sense of a bool. We flip the
1411 // test instead.
1412 int cmp1_op = cmp1->Opcode();
1413 const TypeInt* cmp2_type = phase->type(cmp2)->isa_int();
1414 if (cmp2_type == NULL) return NULL;
1415 Node* j_xor = cmp1;
1416 if( cmp2_type == TypeInt::ZERO &&
1417 cmp1_op == Op_XorI &&
1418 j_xor->in(1) != j_xor && // An xor of itself is dead
1419 phase->type( j_xor->in(1) ) == TypeInt::BOOL &&
1420 phase->type( j_xor->in(2) ) == TypeInt::ONE &&
1421 (_test._test == BoolTest::eq ||
1422 _test._test == BoolTest::ne) ) {
1423 Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2));
1424 return new BoolNode( ncmp, _test.negate() );
1425 }
1426
1427 // Change ((x & m) u<= m) or ((m & x) u<= m) to always true
1428 // Same with ((x & m) u< m+1) and ((m & x) u< m+1)
1429 if (cop == Op_CmpU &&
1430 cmp1_op == Op_AndI) {
1431 Node* bound = NULL;
1432 if (_test._test == BoolTest::le) {
1433 bound = cmp2;
1434 } else if (_test._test == BoolTest::lt &&
1435 cmp2->Opcode() == Op_AddI &&
1436 cmp2->in(2)->find_int_con(0) == 1) {
1437 bound = cmp2->in(1);
1438 }
1439 if (cmp1->in(2) == bound || cmp1->in(1) == bound) {
1440 return ConINode::make(1);
1441 }
1442 }
1443
1444 // Change ((x & (m - 1)) u< m) into (m > 0)
1445 // This is the off-by-one variant of the above
1446 if (cop == Op_CmpU &&
1447 _test._test == BoolTest::lt &&
1448 cmp1_op == Op_AndI) {
1449 Node* l = cmp1->in(1);
1450 Node* r = cmp1->in(2);
1451 for (int repeat = 0; repeat < 2; repeat++) {
1452 bool match = r->Opcode() == Op_AddI && r->in(2)->find_int_con(0) == -1 &&
1453 r->in(1) == cmp2;
1454 if (match) {
1455 // arraylength known to be non-negative, so a (arraylength != 0) is sufficient,
1456 // but to be compatible with the array range check pattern, use (arraylength u> 0)
1457 Node* ncmp = cmp2->Opcode() == Op_LoadRange
1458 ? phase->transform(new CmpUNode(cmp2, phase->intcon(0)))
1459 : phase->transform(new CmpINode(cmp2, phase->intcon(0)));
1460 return new BoolNode(ncmp, BoolTest::gt);
1461 } else {
1462 // commute and try again
1463 l = cmp1->in(2);
1464 r = cmp1->in(1);
1465 }
1466 }
1467 }
1468
1469 // Change x u< 1 or x u<= 0 to x == 0
1470 if (cop == Op_CmpU &&
1471 cmp1_op != Op_LoadRange &&
1472 ((_test._test == BoolTest::lt &&
1473 cmp2->find_int_con(-1) == 1) ||
1474 (_test._test == BoolTest::le &&
1475 cmp2->find_int_con(-1) == 0))) {
1476 Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0)));
1477 return new BoolNode(ncmp, BoolTest::eq);
1478 }
1479
1480 // Change (arraylength <= 0) or (arraylength == 0)
1481 // into (arraylength u<= 0)
1482 // Also change (arraylength != 0) into (arraylength u> 0)
1483 // The latter version matches the code pattern generated for
1484 // array range checks, which will more likely be optimized later.
1485 if (cop == Op_CmpI &&
1486 cmp1_op == Op_LoadRange &&
1487 cmp2->find_int_con(-1) == 0) {
1488 if (_test._test == BoolTest::le || _test._test == BoolTest::eq) {
1489 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
1490 return new BoolNode(ncmp, BoolTest::le);
1491 } else if (_test._test == BoolTest::ne) {
1492 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
1493 return new BoolNode(ncmp, BoolTest::gt);
1494 }
1495 }
1496
1497 // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)".
1498 // This is a standard idiom for branching on a boolean value.
1499 Node *c2b = cmp1;
1500 if( cmp2_type == TypeInt::ZERO &&
1501 cmp1_op == Op_Conv2B &&
1502 (_test._test == BoolTest::eq ||
1503 _test._test == BoolTest::ne) ) {
1504 Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int()
1505 ? (Node*)new CmpINode(c2b->in(1),cmp2)
1506 : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR))
1507 );
1508 return new BoolNode( ncmp, _test._test );
1509 }
1510
1511 // Comparing a SubI against a zero is equal to comparing the SubI
1512 // arguments directly. This only works for eq and ne comparisons
1513 // due to possible integer overflow.
1514 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1515 (cop == Op_CmpI) &&
1516 (cmp1_op == Op_SubI) &&
1517 ( cmp2_type == TypeInt::ZERO ) ) {
1518 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2)));
1519 return new BoolNode( ncmp, _test._test );
1520 }
1521
1522 // Same as above but with and AddI of a constant
1523 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1524 cop == Op_CmpI &&
1525 cmp1_op == Op_AddI &&
1526 cmp1->in(2) != NULL &&
1527 phase->type(cmp1->in(2))->isa_int() &&
1528 phase->type(cmp1->in(2))->is_int()->is_con() &&
1529 cmp2_type == TypeInt::ZERO &&
1530 !is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape
1531 ) {
1532 const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int();
1533 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi)));
1534 return new BoolNode( ncmp, _test._test );
1535 }
1536
1537 // Change "bool eq/ne (cmp (phi (X -X) 0))" into "bool eq/ne (cmp X 0)"
1538 // since zero check of conditional negation of an integer is equal to
1539 // zero check of the integer directly.
1540 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1541 (cop == Op_CmpI) &&
1542 (cmp2_type == TypeInt::ZERO) &&
1543 (cmp1_op == Op_Phi)) {
1544 // There should be a diamond phi with true path at index 1 or 2
1545 PhiNode *phi = cmp1->as_Phi();
1546 int idx_true = phi->is_diamond_phi();
1547 if (idx_true != 0) {
1548 // True input is in(idx_true) while false input is in(3 - idx_true)
1549 Node *tin = phi->in(idx_true);
1550 Node *fin = phi->in(3 - idx_true);
1551 if ((tin->Opcode() == Op_SubI) &&
1552 (phase->type(tin->in(1)) == TypeInt::ZERO) &&
1553 (tin->in(2) == fin)) {
1554 // Found conditional negation at true path, create a new CmpINode without that
1555 Node *ncmp = phase->transform(new CmpINode(fin, cmp2));
1556 return new BoolNode(ncmp, _test._test);
1557 }
1558 if ((fin->Opcode() == Op_SubI) &&
1559 (phase->type(fin->in(1)) == TypeInt::ZERO) &&
1560 (fin->in(2) == tin)) {
1561 // Found conditional negation at false path, create a new CmpINode without that
1562 Node *ncmp = phase->transform(new CmpINode(tin, cmp2));
1563 return new BoolNode(ncmp, _test._test);
1564 }
1565 }
1566 }
1567
1568 // Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the
1569 // most general case because negating 0x80000000 does nothing. Needed for
1570 // the CmpF3/SubI/CmpI idiom.
1571 if( cop == Op_CmpI &&
1572 cmp1_op == Op_SubI &&
1573 cmp2_type == TypeInt::ZERO &&
1574 phase->type( cmp1->in(1) ) == TypeInt::ZERO &&
1575 phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) {
1576 Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2));
1577 return new BoolNode( ncmp, _test.commute() );
1578 }
1579
1580 // Try to optimize signed integer comparison
1581 return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type);
1582
1583 // The transformation below is not valid for either signed or unsigned
1584 // comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE.
1585 // This transformation can be resurrected when we are able to
1586 // make inferences about the range of values being subtracted from
1587 // (or added to) relative to the wraparound point.
1588 //
1589 // // Remove +/-1's if possible.
1590 // // "X <= Y-1" becomes "X < Y"
1591 // // "X+1 <= Y" becomes "X < Y"
1592 // // "X < Y+1" becomes "X <= Y"
1593 // // "X-1 < Y" becomes "X <= Y"
1594 // // Do not this to compares off of the counted-loop-end. These guys are
1595 // // checking the trip counter and they want to use the post-incremented
1596 // // counter. If they use the PRE-incremented counter, then the counter has
1597 // // to be incremented in a private block on a loop backedge.
1598 // if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd )
1599 // return NULL;
1600 // #ifndef PRODUCT
1601 // // Do not do this in a wash GVN pass during verification.
1602 // // Gets triggered by too many simple optimizations to be bothered with
1603 // // re-trying it again and again.
1604 // if( !phase->allow_progress() ) return NULL;
1605 // #endif
1606 // // Not valid for unsigned compare because of corner cases in involving zero.
1607 // // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an
1608 // // exception in case X is 0 (because 0-1 turns into 4billion unsigned but
1609 // // "0 <=u Y" is always true).
1610 // if( cmp->Opcode() == Op_CmpU ) return NULL;
1611 // int cmp2_op = cmp2->Opcode();
1612 // if( _test._test == BoolTest::le ) {
1613 // if( cmp1_op == Op_AddI &&
1614 // phase->type( cmp1->in(2) ) == TypeInt::ONE )
1615 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt );
1616 // else if( cmp2_op == Op_AddI &&
1617 // phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 )
1618 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt );
1619 // } else if( _test._test == BoolTest::lt ) {
1620 // if( cmp1_op == Op_AddI &&
1621 // phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 )
1622 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le );
1623 // else if( cmp2_op == Op_AddI &&
1624 // phase->type( cmp2->in(2) ) == TypeInt::ONE )
1625 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le );
1626 // }
1627 }
1628
1629 //------------------------------Value------------------------------------------
1630 // Simplify a Bool (convert condition codes to boolean (1 or 0)) node,
1631 // based on local information. If the input is constant, do it.
Value(PhaseGVN * phase) const1632 const Type* BoolNode::Value(PhaseGVN* phase) const {
1633 return _test.cc2logical( phase->type( in(1) ) );
1634 }
1635
1636 #ifndef PRODUCT
1637 //------------------------------dump_spec--------------------------------------
1638 // Dump special per-node info
dump_spec(outputStream * st) const1639 void BoolNode::dump_spec(outputStream *st) const {
1640 st->print("[");
1641 _test.dump_on(st);
1642 st->print("]");
1643 }
1644
1645 //-------------------------------related---------------------------------------
1646 // A BoolNode's related nodes are all of its data inputs, and all of its
1647 // outputs until control nodes are hit, which are included. In compact
1648 // representation, inputs till level 3 and immediate outputs are included.
related(GrowableArray<Node * > * in_rel,GrowableArray<Node * > * out_rel,bool compact) const1649 void BoolNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1650 if (compact) {
1651 this->collect_nodes(in_rel, 3, false, true);
1652 this->collect_nodes(out_rel, -1, false, false);
1653 } else {
1654 this->collect_nodes_in_all_data(in_rel, false);
1655 this->collect_nodes_out_all_ctrl_boundary(out_rel);
1656 }
1657 }
1658 #endif
1659
1660 //----------------------is_counted_loop_exit_test------------------------------
1661 // Returns true if node is used by a counted loop node.
is_counted_loop_exit_test()1662 bool BoolNode::is_counted_loop_exit_test() {
1663 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
1664 Node* use = fast_out(i);
1665 if (use->is_CountedLoopEnd()) {
1666 return true;
1667 }
1668 }
1669 return false;
1670 }
1671
1672 //=============================================================================
1673 //------------------------------Value------------------------------------------
1674 // Compute sqrt
Value(PhaseGVN * phase) const1675 const Type* SqrtDNode::Value(PhaseGVN* phase) const {
1676 const Type *t1 = phase->type( in(1) );
1677 if( t1 == Type::TOP ) return Type::TOP;
1678 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE;
1679 double d = t1->getd();
1680 if( d < 0.0 ) return Type::DOUBLE;
1681 return TypeD::make( sqrt( d ) );
1682 }
1683
Value(PhaseGVN * phase) const1684 const Type* SqrtFNode::Value(PhaseGVN* phase) const {
1685 const Type *t1 = phase->type( in(1) );
1686 if( t1 == Type::TOP ) return Type::TOP;
1687 if( t1->base() != Type::FloatCon ) return Type::FLOAT;
1688 float f = t1->getf();
1689 if( f < 0.0f ) return Type::FLOAT;
1690 return TypeF::make( (float)sqrt( (double)f ) );
1691 }
1692