1 | /* |
2 | * Copyright (c) 1997, 2018, 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. |
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. |
85 | 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 | |
105 | 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 | |
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 | // |
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------------------------------------------ |
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 | return NULL; |
256 | } |
257 | |
258 | //------------------------------sub-------------------------------------------- |
259 | // A subtract node differences it's two inputs. |
260 | const Type *SubINode::sub( const Type *t1, const Type *t2 ) const { |
261 | const TypeInt *r0 = t1->is_int(); // Handy access |
262 | const TypeInt *r1 = t2->is_int(); |
263 | int32_t lo = java_subtract(r0->_lo, r1->_hi); |
264 | int32_t hi = java_subtract(r0->_hi, r1->_lo); |
265 | |
266 | // We next check for 32-bit overflow. |
267 | // If that happens, we just assume all integers are possible. |
268 | if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR |
269 | ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND |
270 | (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR |
271 | ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs |
272 | return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen)); |
273 | else // Overflow; assume all integers |
274 | return TypeInt::INT; |
275 | } |
276 | |
277 | //============================================================================= |
278 | //------------------------------Ideal------------------------------------------ |
279 | Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
280 | Node *in1 = in(1); |
281 | Node *in2 = in(2); |
282 | uint op1 = in1->Opcode(); |
283 | uint op2 = in2->Opcode(); |
284 | |
285 | #ifdef ASSERT |
286 | // Check for dead loop |
287 | if( phase->eqv( in1, this ) || phase->eqv( in2, this ) || |
288 | ( ( op1 == Op_AddL || op1 == Op_SubL ) && |
289 | ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) || |
290 | phase->eqv( in1->in(1), in1 ) || phase->eqv( in1->in(2), in1 ) ) ) ) |
291 | assert(false, "dead loop in SubLNode::Ideal" ); |
292 | #endif |
293 | |
294 | if( phase->type( in2 ) == Type::TOP ) return NULL; |
295 | const TypeLong *i = phase->type( in2 )->isa_long(); |
296 | // Convert "x-c0" into "x+ -c0". |
297 | if( i && // Might be bottom or top... |
298 | i->is_con() ) |
299 | return new AddLNode(in1, phase->longcon(-i->get_con())); |
300 | |
301 | // Convert "(x+c0) - y" into (x-y) + c0" |
302 | // Do not collapse (x+c0)-y if "+" is a loop increment or |
303 | // if "y" is a loop induction variable. |
304 | if( op1 == Op_AddL && ok_to_convert(in1, in2) ) { |
305 | Node *in11 = in1->in(1); |
306 | const Type *tadd = phase->type( in1->in(2) ); |
307 | if( tadd->singleton() && tadd != Type::TOP ) { |
308 | Node *sub2 = phase->transform( new SubLNode( in11, in2 )); |
309 | return new AddLNode( sub2, in1->in(2) ); |
310 | } |
311 | } |
312 | |
313 | // Convert "x - (y+c0)" into "(x-y) - c0" |
314 | // Need the same check as in above optimization but reversed. |
315 | if (op2 == Op_AddL && ok_to_convert(in2, in1)) { |
316 | Node* in21 = in2->in(1); |
317 | Node* in22 = in2->in(2); |
318 | const TypeLong* tcon = phase->type(in22)->isa_long(); |
319 | if (tcon != NULL && tcon->is_con()) { |
320 | Node* sub2 = phase->transform( new SubLNode(in1, in21) ); |
321 | Node* neg_c0 = phase->longcon(- tcon->get_con()); |
322 | return new AddLNode(sub2, neg_c0); |
323 | } |
324 | } |
325 | |
326 | const Type *t1 = phase->type( in1 ); |
327 | if( t1 == Type::TOP ) return NULL; |
328 | |
329 | #ifdef ASSERT |
330 | // Check for dead loop |
331 | if( ( op2 == Op_AddL || op2 == Op_SubL ) && |
332 | ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) || |
333 | phase->eqv( in2->in(1), in2 ) || phase->eqv( in2->in(2), in2 ) ) ) |
334 | assert(false, "dead loop in SubLNode::Ideal" ); |
335 | #endif |
336 | |
337 | // Convert "x - (x+y)" into "-y" |
338 | if( op2 == Op_AddL && |
339 | phase->eqv( in1, in2->in(1) ) ) |
340 | return new SubLNode( phase->makecon(TypeLong::ZERO), in2->in(2)); |
341 | // Convert "x - (y+x)" into "-y" |
342 | if( op2 == Op_AddL && |
343 | phase->eqv( in1, in2->in(2) ) ) |
344 | return new SubLNode( phase->makecon(TypeLong::ZERO),in2->in(1)); |
345 | |
346 | // Convert "0 - (x-y)" into "y-x" |
347 | if( phase->type( in1 ) == TypeLong::ZERO && op2 == Op_SubL ) |
348 | return new SubLNode( in2->in(2), in2->in(1) ); |
349 | |
350 | // Convert "(X+A) - (X+B)" into "A - B" |
351 | if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) ) |
352 | return new SubLNode( in1->in(2), in2->in(2) ); |
353 | |
354 | // Convert "(A+X) - (B+X)" into "A - B" |
355 | if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) ) |
356 | return new SubLNode( in1->in(1), in2->in(1) ); |
357 | |
358 | // Convert "A-(B-C)" into (A+C)-B" |
359 | if( op2 == Op_SubL && in2->outcnt() == 1) { |
360 | Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) ); |
361 | return new SubLNode( add1, in2->in(1) ); |
362 | } |
363 | |
364 | return NULL; |
365 | } |
366 | |
367 | //------------------------------sub-------------------------------------------- |
368 | // A subtract node differences it's two inputs. |
369 | const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const { |
370 | const TypeLong *r0 = t1->is_long(); // Handy access |
371 | const TypeLong *r1 = t2->is_long(); |
372 | jlong lo = java_subtract(r0->_lo, r1->_hi); |
373 | jlong hi = java_subtract(r0->_hi, r1->_lo); |
374 | |
375 | // We next check for 32-bit overflow. |
376 | // If that happens, we just assume all integers are possible. |
377 | if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR |
378 | ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND |
379 | (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR |
380 | ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs |
381 | return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen)); |
382 | else // Overflow; assume all integers |
383 | return TypeLong::LONG; |
384 | } |
385 | |
386 | //============================================================================= |
387 | //------------------------------Value------------------------------------------ |
388 | // A subtract node differences its two inputs. |
389 | const Type* SubFPNode::Value(PhaseGVN* phase) const { |
390 | const Node* in1 = in(1); |
391 | const Node* in2 = in(2); |
392 | // Either input is TOP ==> the result is TOP |
393 | const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); |
394 | if( t1 == Type::TOP ) return Type::TOP; |
395 | const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); |
396 | if( t2 == Type::TOP ) return Type::TOP; |
397 | |
398 | // if both operands are infinity of same sign, the result is NaN; do |
399 | // not replace with zero |
400 | if( (t1->is_finite() && t2->is_finite()) ) { |
401 | if( phase->eqv(in1, in2) ) return add_id(); |
402 | } |
403 | |
404 | // Either input is BOTTOM ==> the result is the local BOTTOM |
405 | const Type *bot = bottom_type(); |
406 | if( (t1 == bot) || (t2 == bot) || |
407 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
408 | return bot; |
409 | |
410 | return sub(t1,t2); // Local flavor of type subtraction |
411 | } |
412 | |
413 | |
414 | //============================================================================= |
415 | //------------------------------Ideal------------------------------------------ |
416 | Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
417 | const Type *t2 = phase->type( in(2) ); |
418 | // Convert "x-c0" into "x+ -c0". |
419 | if( t2->base() == Type::FloatCon ) { // Might be bottom or top... |
420 | // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) ); |
421 | } |
422 | |
423 | // Not associative because of boundary conditions (infinity) |
424 | if( IdealizedNumerics && !phase->C->method()->is_strict() ) { |
425 | // Convert "x - (x+y)" into "-y" |
426 | if( in(2)->is_Add() && |
427 | phase->eqv(in(1),in(2)->in(1) ) ) |
428 | return new SubFNode( phase->makecon(TypeF::ZERO),in(2)->in(2)); |
429 | } |
430 | |
431 | // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes |
432 | // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0. |
433 | //if( phase->type(in(1)) == TypeF::ZERO ) |
434 | //return new (phase->C, 2) NegFNode(in(2)); |
435 | |
436 | return NULL; |
437 | } |
438 | |
439 | //------------------------------sub-------------------------------------------- |
440 | // A subtract node differences its two inputs. |
441 | const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const { |
442 | // no folding if one of operands is infinity or NaN, do not do constant folding |
443 | if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) { |
444 | return TypeF::make( t1->getf() - t2->getf() ); |
445 | } |
446 | else if( g_isnan(t1->getf()) ) { |
447 | return t1; |
448 | } |
449 | else if( g_isnan(t2->getf()) ) { |
450 | return t2; |
451 | } |
452 | else { |
453 | return Type::FLOAT; |
454 | } |
455 | } |
456 | |
457 | //============================================================================= |
458 | //------------------------------Ideal------------------------------------------ |
459 | Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
460 | const Type *t2 = phase->type( in(2) ); |
461 | // Convert "x-c0" into "x+ -c0". |
462 | if( t2->base() == Type::DoubleCon ) { // Might be bottom or top... |
463 | // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) ); |
464 | } |
465 | |
466 | // Not associative because of boundary conditions (infinity) |
467 | if( IdealizedNumerics && !phase->C->method()->is_strict() ) { |
468 | // Convert "x - (x+y)" into "-y" |
469 | if( in(2)->is_Add() && |
470 | phase->eqv(in(1),in(2)->in(1) ) ) |
471 | return new SubDNode( phase->makecon(TypeD::ZERO),in(2)->in(2)); |
472 | } |
473 | |
474 | // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes |
475 | // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0. |
476 | //if( phase->type(in(1)) == TypeD::ZERO ) |
477 | //return new (phase->C, 2) NegDNode(in(2)); |
478 | |
479 | return NULL; |
480 | } |
481 | |
482 | //------------------------------sub-------------------------------------------- |
483 | // A subtract node differences its two inputs. |
484 | const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const { |
485 | // no folding if one of operands is infinity or NaN, do not do constant folding |
486 | if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) { |
487 | return TypeD::make( t1->getd() - t2->getd() ); |
488 | } |
489 | else if( g_isnan(t1->getd()) ) { |
490 | return t1; |
491 | } |
492 | else if( g_isnan(t2->getd()) ) { |
493 | return t2; |
494 | } |
495 | else { |
496 | return Type::DOUBLE; |
497 | } |
498 | } |
499 | |
500 | //============================================================================= |
501 | //------------------------------Idealize--------------------------------------- |
502 | // Unlike SubNodes, compare must still flatten return value to the |
503 | // range -1, 0, 1. |
504 | // And optimizations like those for (X + Y) - X fail if overflow happens. |
505 | Node* CmpNode::Identity(PhaseGVN* phase) { |
506 | return this; |
507 | } |
508 | |
509 | #ifndef PRODUCT |
510 | //----------------------------related------------------------------------------ |
511 | // Related nodes of comparison nodes include all data inputs (until hitting a |
512 | // control boundary) as well as all outputs until and including control nodes |
513 | // as well as their projections. In compact mode, data inputs till depth 1 and |
514 | // all outputs till depth 1 are considered. |
515 | void CmpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
516 | if (compact) { |
517 | this->collect_nodes(in_rel, 1, false, true); |
518 | this->collect_nodes(out_rel, -1, false, false); |
519 | } else { |
520 | this->collect_nodes_in_all_data(in_rel, false); |
521 | this->collect_nodes_out_all_ctrl_boundary(out_rel); |
522 | // Now, find all control nodes in out_rel, and include their projections |
523 | // and projection targets (if any) in the result. |
524 | GrowableArray<Node*> proj(Compile::current()->unique()); |
525 | for (GrowableArrayIterator<Node*> it = out_rel->begin(); it != out_rel->end(); ++it) { |
526 | Node* n = *it; |
527 | if (n->is_CFG() && !n->is_Proj()) { |
528 | // Assume projections and projection targets are found at levels 1 and 2. |
529 | n->collect_nodes(&proj, -2, false, false); |
530 | for (GrowableArrayIterator<Node*> p = proj.begin(); p != proj.end(); ++p) { |
531 | out_rel->append_if_missing(*p); |
532 | } |
533 | proj.clear(); |
534 | } |
535 | } |
536 | } |
537 | } |
538 | #endif |
539 | |
540 | //============================================================================= |
541 | //------------------------------cmp-------------------------------------------- |
542 | // Simplify a CmpI (compare 2 integers) node, based on local information. |
543 | // If both inputs are constants, compare them. |
544 | const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const { |
545 | const TypeInt *r0 = t1->is_int(); // Handy access |
546 | const TypeInt *r1 = t2->is_int(); |
547 | |
548 | if( r0->_hi < r1->_lo ) // Range is always low? |
549 | return TypeInt::CC_LT; |
550 | else if( r0->_lo > r1->_hi ) // Range is always high? |
551 | return TypeInt::CC_GT; |
552 | |
553 | else if( r0->is_con() && r1->is_con() ) { // comparing constants? |
554 | assert(r0->get_con() == r1->get_con(), "must be equal" ); |
555 | return TypeInt::CC_EQ; // Equal results. |
556 | } else if( r0->_hi == r1->_lo ) // Range is never high? |
557 | return TypeInt::CC_LE; |
558 | else if( r0->_lo == r1->_hi ) // Range is never low? |
559 | return TypeInt::CC_GE; |
560 | return TypeInt::CC; // else use worst case results |
561 | } |
562 | |
563 | // Simplify a CmpU (compare 2 integers) node, based on local information. |
564 | // If both inputs are constants, compare them. |
565 | const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const { |
566 | assert(!t1->isa_ptr(), "obsolete usage of CmpU" ); |
567 | |
568 | // comparing two unsigned ints |
569 | const TypeInt *r0 = t1->is_int(); // Handy access |
570 | const TypeInt *r1 = t2->is_int(); |
571 | |
572 | // Current installed version |
573 | // Compare ranges for non-overlap |
574 | juint lo0 = r0->_lo; |
575 | juint hi0 = r0->_hi; |
576 | juint lo1 = r1->_lo; |
577 | juint hi1 = r1->_hi; |
578 | |
579 | // If either one has both negative and positive values, |
580 | // it therefore contains both 0 and -1, and since [0..-1] is the |
581 | // full unsigned range, the type must act as an unsigned bottom. |
582 | bool bot0 = ((jint)(lo0 ^ hi0) < 0); |
583 | bool bot1 = ((jint)(lo1 ^ hi1) < 0); |
584 | |
585 | if (bot0 || bot1) { |
586 | // All unsigned values are LE -1 and GE 0. |
587 | if (lo0 == 0 && hi0 == 0) { |
588 | return TypeInt::CC_LE; // 0 <= bot |
589 | } else if ((jint)lo0 == -1 && (jint)hi0 == -1) { |
590 | return TypeInt::CC_GE; // -1 >= bot |
591 | } else if (lo1 == 0 && hi1 == 0) { |
592 | return TypeInt::CC_GE; // bot >= 0 |
593 | } else if ((jint)lo1 == -1 && (jint)hi1 == -1) { |
594 | return TypeInt::CC_LE; // bot <= -1 |
595 | } |
596 | } else { |
597 | // We can use ranges of the form [lo..hi] if signs are the same. |
598 | assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid" ); |
599 | // results are reversed, '-' > '+' for unsigned compare |
600 | if (hi0 < lo1) { |
601 | return TypeInt::CC_LT; // smaller |
602 | } else if (lo0 > hi1) { |
603 | return TypeInt::CC_GT; // greater |
604 | } else if (hi0 == lo1 && lo0 == hi1) { |
605 | return TypeInt::CC_EQ; // Equal results |
606 | } else if (lo0 >= hi1) { |
607 | return TypeInt::CC_GE; |
608 | } else if (hi0 <= lo1) { |
609 | // Check for special case in Hashtable::get. (See below.) |
610 | if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) |
611 | return TypeInt::CC_LT; |
612 | return TypeInt::CC_LE; |
613 | } |
614 | } |
615 | // Check for special case in Hashtable::get - the hash index is |
616 | // mod'ed to the table size so the following range check is useless. |
617 | // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have |
618 | // to be positive. |
619 | // (This is a gross hack, since the sub method never |
620 | // looks at the structure of the node in any other case.) |
621 | if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) |
622 | return TypeInt::CC_LT; |
623 | return TypeInt::CC; // else use worst case results |
624 | } |
625 | |
626 | const Type* CmpUNode::Value(PhaseGVN* phase) const { |
627 | const Type* t = SubNode::Value_common(phase); |
628 | if (t != NULL) { |
629 | return t; |
630 | } |
631 | const Node* in1 = in(1); |
632 | const Node* in2 = in(2); |
633 | const Type* t1 = phase->type(in1); |
634 | const Type* t2 = phase->type(in2); |
635 | assert(t1->isa_int(), "CmpU has only Int type inputs" ); |
636 | if (t2 == TypeInt::INT) { // Compare to bottom? |
637 | return bottom_type(); |
638 | } |
639 | uint in1_op = in1->Opcode(); |
640 | if (in1_op == Op_AddI || in1_op == Op_SubI) { |
641 | // The problem rise when result of AddI(SubI) may overflow |
642 | // signed integer value. Let say the input type is |
643 | // [256, maxint] then +128 will create 2 ranges due to |
644 | // overflow: [minint, minint+127] and [384, maxint]. |
645 | // But C2 type system keep only 1 type range and as result |
646 | // it use general [minint, maxint] for this case which we |
647 | // can't optimize. |
648 | // |
649 | // Make 2 separate type ranges based on types of AddI(SubI) inputs |
650 | // and compare results of their compare. If results are the same |
651 | // CmpU node can be optimized. |
652 | const Node* in11 = in1->in(1); |
653 | const Node* in12 = in1->in(2); |
654 | const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11); |
655 | const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12); |
656 | // Skip cases when input types are top or bottom. |
657 | if ((t11 != Type::TOP) && (t11 != TypeInt::INT) && |
658 | (t12 != Type::TOP) && (t12 != TypeInt::INT)) { |
659 | const TypeInt *r0 = t11->is_int(); |
660 | const TypeInt *r1 = t12->is_int(); |
661 | jlong lo_r0 = r0->_lo; |
662 | jlong hi_r0 = r0->_hi; |
663 | jlong lo_r1 = r1->_lo; |
664 | jlong hi_r1 = r1->_hi; |
665 | if (in1_op == Op_SubI) { |
666 | jlong tmp = hi_r1; |
667 | hi_r1 = -lo_r1; |
668 | lo_r1 = -tmp; |
669 | // Note, for substructing [minint,x] type range |
670 | // long arithmetic provides correct overflow answer. |
671 | // The confusion come from the fact that in 32-bit |
672 | // -minint == minint but in 64-bit -minint == maxint+1. |
673 | } |
674 | jlong lo_long = lo_r0 + lo_r1; |
675 | jlong hi_long = hi_r0 + hi_r1; |
676 | int lo_tr1 = min_jint; |
677 | int hi_tr1 = (int)hi_long; |
678 | int lo_tr2 = (int)lo_long; |
679 | int hi_tr2 = max_jint; |
680 | bool underflow = lo_long != (jlong)lo_tr2; |
681 | bool overflow = hi_long != (jlong)hi_tr1; |
682 | // Use sub(t1, t2) when there is no overflow (one type range) |
683 | // or when both overflow and underflow (too complex). |
684 | if ((underflow != overflow) && (hi_tr1 < lo_tr2)) { |
685 | // Overflow only on one boundary, compare 2 separate type ranges. |
686 | int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here |
687 | const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w); |
688 | const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w); |
689 | const Type* cmp1 = sub(tr1, t2); |
690 | const Type* cmp2 = sub(tr2, t2); |
691 | if (cmp1 == cmp2) { |
692 | return cmp1; // Hit! |
693 | } |
694 | } |
695 | } |
696 | } |
697 | |
698 | return sub(t1, t2); // Local flavor of type subtraction |
699 | } |
700 | |
701 | bool CmpUNode::is_index_range_check() const { |
702 | // Check for the "(X ModI Y) CmpU Y" shape |
703 | return (in(1)->Opcode() == Op_ModI && |
704 | in(1)->in(2)->eqv_uncast(in(2))); |
705 | } |
706 | |
707 | //------------------------------Idealize--------------------------------------- |
708 | Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) { |
709 | if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) { |
710 | switch (in(1)->Opcode()) { |
711 | case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL |
712 | return new CmpLNode(in(1)->in(1),in(1)->in(2)); |
713 | case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF |
714 | return new CmpFNode(in(1)->in(1),in(1)->in(2)); |
715 | case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD |
716 | return new CmpDNode(in(1)->in(1),in(1)->in(2)); |
717 | //case Op_SubI: |
718 | // If (x - y) cannot overflow, then ((x - y) <?> 0) |
719 | // can be turned into (x <?> y). |
720 | // This is handled (with more general cases) by Ideal_sub_algebra. |
721 | } |
722 | } |
723 | return NULL; // No change |
724 | } |
725 | |
726 | |
727 | //============================================================================= |
728 | // Simplify a CmpL (compare 2 longs ) node, based on local information. |
729 | // If both inputs are constants, compare them. |
730 | const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const { |
731 | const TypeLong *r0 = t1->is_long(); // Handy access |
732 | const TypeLong *r1 = t2->is_long(); |
733 | |
734 | if( r0->_hi < r1->_lo ) // Range is always low? |
735 | return TypeInt::CC_LT; |
736 | else if( r0->_lo > r1->_hi ) // Range is always high? |
737 | return TypeInt::CC_GT; |
738 | |
739 | else if( r0->is_con() && r1->is_con() ) { // comparing constants? |
740 | assert(r0->get_con() == r1->get_con(), "must be equal" ); |
741 | return TypeInt::CC_EQ; // Equal results. |
742 | } else if( r0->_hi == r1->_lo ) // Range is never high? |
743 | return TypeInt::CC_LE; |
744 | else if( r0->_lo == r1->_hi ) // Range is never low? |
745 | return TypeInt::CC_GE; |
746 | return TypeInt::CC; // else use worst case results |
747 | } |
748 | |
749 | |
750 | // Simplify a CmpUL (compare 2 unsigned longs) node, based on local information. |
751 | // If both inputs are constants, compare them. |
752 | const Type* CmpULNode::sub(const Type* t1, const Type* t2) const { |
753 | assert(!t1->isa_ptr(), "obsolete usage of CmpUL" ); |
754 | |
755 | // comparing two unsigned longs |
756 | const TypeLong* r0 = t1->is_long(); // Handy access |
757 | const TypeLong* r1 = t2->is_long(); |
758 | |
759 | // Current installed version |
760 | // Compare ranges for non-overlap |
761 | julong lo0 = r0->_lo; |
762 | julong hi0 = r0->_hi; |
763 | julong lo1 = r1->_lo; |
764 | julong hi1 = r1->_hi; |
765 | |
766 | // If either one has both negative and positive values, |
767 | // it therefore contains both 0 and -1, and since [0..-1] is the |
768 | // full unsigned range, the type must act as an unsigned bottom. |
769 | bool bot0 = ((jlong)(lo0 ^ hi0) < 0); |
770 | bool bot1 = ((jlong)(lo1 ^ hi1) < 0); |
771 | |
772 | if (bot0 || bot1) { |
773 | // All unsigned values are LE -1 and GE 0. |
774 | if (lo0 == 0 && hi0 == 0) { |
775 | return TypeInt::CC_LE; // 0 <= bot |
776 | } else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) { |
777 | return TypeInt::CC_GE; // -1 >= bot |
778 | } else if (lo1 == 0 && hi1 == 0) { |
779 | return TypeInt::CC_GE; // bot >= 0 |
780 | } else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) { |
781 | return TypeInt::CC_LE; // bot <= -1 |
782 | } |
783 | } else { |
784 | // We can use ranges of the form [lo..hi] if signs are the same. |
785 | assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid" ); |
786 | // results are reversed, '-' > '+' for unsigned compare |
787 | if (hi0 < lo1) { |
788 | return TypeInt::CC_LT; // smaller |
789 | } else if (lo0 > hi1) { |
790 | return TypeInt::CC_GT; // greater |
791 | } else if (hi0 == lo1 && lo0 == hi1) { |
792 | return TypeInt::CC_EQ; // Equal results |
793 | } else if (lo0 >= hi1) { |
794 | return TypeInt::CC_GE; |
795 | } else if (hi0 <= lo1) { |
796 | return TypeInt::CC_LE; |
797 | } |
798 | } |
799 | |
800 | return TypeInt::CC; // else use worst case results |
801 | } |
802 | |
803 | //============================================================================= |
804 | //------------------------------sub-------------------------------------------- |
805 | // Simplify an CmpP (compare 2 pointers) node, based on local information. |
806 | // If both inputs are constants, compare them. |
807 | const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const { |
808 | const TypePtr *r0 = t1->is_ptr(); // Handy access |
809 | const TypePtr *r1 = t2->is_ptr(); |
810 | |
811 | // Undefined inputs makes for an undefined result |
812 | if( TypePtr::above_centerline(r0->_ptr) || |
813 | TypePtr::above_centerline(r1->_ptr) ) |
814 | return Type::TOP; |
815 | |
816 | if (r0 == r1 && r0->singleton()) { |
817 | // Equal pointer constants (klasses, nulls, etc.) |
818 | return TypeInt::CC_EQ; |
819 | } |
820 | |
821 | // See if it is 2 unrelated classes. |
822 | const TypeOopPtr* p0 = r0->isa_oopptr(); |
823 | const TypeOopPtr* p1 = r1->isa_oopptr(); |
824 | if (p0 && p1) { |
825 | Node* in1 = in(1)->uncast(); |
826 | Node* in2 = in(2)->uncast(); |
827 | AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL); |
828 | AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL); |
829 | if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) { |
830 | return TypeInt::CC_GT; // different pointers |
831 | } |
832 | ciKlass* klass0 = p0->klass(); |
833 | bool xklass0 = p0->klass_is_exact(); |
834 | ciKlass* klass1 = p1->klass(); |
835 | bool xklass1 = p1->klass_is_exact(); |
836 | int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0); |
837 | if (klass0 && klass1 && |
838 | kps != 1 && // both or neither are klass pointers |
839 | klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces |
840 | klass1->is_loaded() && !klass1->is_interface() && |
841 | (!klass0->is_obj_array_klass() || |
842 | !klass0->as_obj_array_klass()->base_element_klass()->is_interface()) && |
843 | (!klass1->is_obj_array_klass() || |
844 | !klass1->as_obj_array_klass()->base_element_klass()->is_interface())) { |
845 | bool unrelated_classes = false; |
846 | // See if neither subclasses the other, or if the class on top |
847 | // is precise. In either of these cases, the compare is known |
848 | // to fail if at least one of the pointers is provably not null. |
849 | if (klass0->equals(klass1)) { // if types are unequal but klasses are equal |
850 | // Do nothing; we know nothing for imprecise types |
851 | } else if (klass0->is_subtype_of(klass1)) { |
852 | // If klass1's type is PRECISE, then classes are unrelated. |
853 | unrelated_classes = xklass1; |
854 | } else if (klass1->is_subtype_of(klass0)) { |
855 | // If klass0's type is PRECISE, then classes are unrelated. |
856 | unrelated_classes = xklass0; |
857 | } else { // Neither subtypes the other |
858 | unrelated_classes = true; |
859 | } |
860 | if (unrelated_classes) { |
861 | // The oops classes are known to be unrelated. If the joined PTRs of |
862 | // two oops is not Null and not Bottom, then we are sure that one |
863 | // of the two oops is non-null, and the comparison will always fail. |
864 | TypePtr::PTR jp = r0->join_ptr(r1->_ptr); |
865 | if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { |
866 | return TypeInt::CC_GT; |
867 | } |
868 | } |
869 | } |
870 | } |
871 | |
872 | // Known constants can be compared exactly |
873 | // Null can be distinguished from any NotNull pointers |
874 | // Unknown inputs makes an unknown result |
875 | if( r0->singleton() ) { |
876 | intptr_t bits0 = r0->get_con(); |
877 | if( r1->singleton() ) |
878 | return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; |
879 | return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; |
880 | } else if( r1->singleton() ) { |
881 | intptr_t bits1 = r1->get_con(); |
882 | return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; |
883 | } else |
884 | return TypeInt::CC; |
885 | } |
886 | |
887 | static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) { |
888 | // Return the klass node for (indirect load from OopHandle) |
889 | // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) |
890 | // or NULL if not matching. |
891 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
892 | n = bs->step_over_gc_barrier(n); |
893 | |
894 | if (n->Opcode() != Op_LoadP) return NULL; |
895 | |
896 | const TypeInstPtr* tp = phase->type(n)->isa_instptr(); |
897 | if (!tp || tp->klass() != phase->C->env()->Class_klass()) return NULL; |
898 | |
899 | Node* adr = n->in(MemNode::Address); |
900 | // First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier. |
901 | if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return NULL; |
902 | adr = adr->in(MemNode::Address); |
903 | |
904 | intptr_t off = 0; |
905 | Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off); |
906 | if (k == NULL) return NULL; |
907 | const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr(); |
908 | if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return NULL; |
909 | |
910 | // We've found the klass node of a Java mirror load. |
911 | return k; |
912 | } |
913 | |
914 | static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) { |
915 | // for ConP(Foo.class) return ConP(Foo.klass) |
916 | // otherwise return NULL |
917 | if (!n->is_Con()) return NULL; |
918 | |
919 | const TypeInstPtr* tp = phase->type(n)->isa_instptr(); |
920 | if (!tp) return NULL; |
921 | |
922 | ciType* mirror_type = tp->java_mirror_type(); |
923 | // TypeInstPtr::java_mirror_type() returns non-NULL for compile- |
924 | // time Class constants only. |
925 | if (!mirror_type) return NULL; |
926 | |
927 | // x.getClass() == int.class can never be true (for all primitive types) |
928 | // Return a ConP(NULL) node for this case. |
929 | if (mirror_type->is_classless()) { |
930 | return phase->makecon(TypePtr::NULL_PTR); |
931 | } |
932 | |
933 | // return the ConP(Foo.klass) |
934 | assert(mirror_type->is_klass(), "mirror_type should represent a Klass*" ); |
935 | return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass())); |
936 | } |
937 | |
938 | //------------------------------Ideal------------------------------------------ |
939 | // Normalize comparisons between Java mirror loads to compare the klass instead. |
940 | // |
941 | // Also check for the case of comparing an unknown klass loaded from the primary |
942 | // super-type array vs a known klass with no subtypes. This amounts to |
943 | // checking to see an unknown klass subtypes a known klass with no subtypes; |
944 | // this only happens on an exact match. We can shorten this test by 1 load. |
945 | Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) { |
946 | // Normalize comparisons between Java mirrors into comparisons of the low- |
947 | // level klass, where a dependent load could be shortened. |
948 | // |
949 | // The new pattern has a nice effect of matching the same pattern used in the |
950 | // fast path of instanceof/checkcast/Class.isInstance(), which allows |
951 | // redundant exact type check be optimized away by GVN. |
952 | // For example, in |
953 | // if (x.getClass() == Foo.class) { |
954 | // Foo foo = (Foo) x; |
955 | // // ... use a ... |
956 | // } |
957 | // a CmpPNode could be shared between if_acmpne and checkcast |
958 | { |
959 | Node* k1 = isa_java_mirror_load(phase, in(1)); |
960 | Node* k2 = isa_java_mirror_load(phase, in(2)); |
961 | Node* conk2 = isa_const_java_mirror(phase, in(2)); |
962 | |
963 | if (k1 && (k2 || conk2)) { |
964 | Node* lhs = k1; |
965 | Node* rhs = (k2 != NULL) ? k2 : conk2; |
966 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
967 | if (igvn != NULL) { |
968 | set_req_X(1, lhs, igvn); |
969 | set_req_X(2, rhs, igvn); |
970 | } else { |
971 | set_req(1, lhs); |
972 | set_req(2, rhs); |
973 | } |
974 | return this; |
975 | } |
976 | } |
977 | |
978 | // Constant pointer on right? |
979 | const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr(); |
980 | if (t2 == NULL || !t2->klass_is_exact()) |
981 | return NULL; |
982 | // Get the constant klass we are comparing to. |
983 | ciKlass* superklass = t2->klass(); |
984 | |
985 | // Now check for LoadKlass on left. |
986 | Node* ldk1 = in(1); |
987 | if (ldk1->is_DecodeNKlass()) { |
988 | ldk1 = ldk1->in(1); |
989 | if (ldk1->Opcode() != Op_LoadNKlass ) |
990 | return NULL; |
991 | } else if (ldk1->Opcode() != Op_LoadKlass ) |
992 | return NULL; |
993 | // Take apart the address of the LoadKlass: |
994 | Node* adr1 = ldk1->in(MemNode::Address); |
995 | intptr_t con2 = 0; |
996 | Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2); |
997 | if (ldk2 == NULL) |
998 | return NULL; |
999 | if (con2 == oopDesc::klass_offset_in_bytes()) { |
1000 | // We are inspecting an object's concrete class. |
1001 | // Short-circuit the check if the query is abstract. |
1002 | if (superklass->is_interface() || |
1003 | superklass->is_abstract()) { |
1004 | // Make it come out always false: |
1005 | this->set_req(2, phase->makecon(TypePtr::NULL_PTR)); |
1006 | return this; |
1007 | } |
1008 | } |
1009 | |
1010 | // Check for a LoadKlass from primary supertype array. |
1011 | // Any nested loadklass from loadklass+con must be from the p.s. array. |
1012 | if (ldk2->is_DecodeNKlass()) { |
1013 | // Keep ldk2 as DecodeN since it could be used in CmpP below. |
1014 | if (ldk2->in(1)->Opcode() != Op_LoadNKlass ) |
1015 | return NULL; |
1016 | } else if (ldk2->Opcode() != Op_LoadKlass) |
1017 | return NULL; |
1018 | |
1019 | // Verify that we understand the situation |
1020 | if (con2 != (intptr_t) superklass->super_check_offset()) |
1021 | return NULL; // Might be element-klass loading from array klass |
1022 | |
1023 | // If 'superklass' has no subklasses and is not an interface, then we are |
1024 | // assured that the only input which will pass the type check is |
1025 | // 'superklass' itself. |
1026 | // |
1027 | // We could be more liberal here, and allow the optimization on interfaces |
1028 | // which have a single implementor. This would require us to increase the |
1029 | // expressiveness of the add_dependency() mechanism. |
1030 | // %%% Do this after we fix TypeOopPtr: Deps are expressive enough now. |
1031 | |
1032 | // Object arrays must have their base element have no subtypes |
1033 | while (superklass->is_obj_array_klass()) { |
1034 | ciType* elem = superklass->as_obj_array_klass()->element_type(); |
1035 | superklass = elem->as_klass(); |
1036 | } |
1037 | if (superklass->is_instance_klass()) { |
1038 | ciInstanceKlass* ik = superklass->as_instance_klass(); |
1039 | if (ik->has_subklass() || ik->is_interface()) return NULL; |
1040 | // Add a dependency if there is a chance that a subclass will be added later. |
1041 | if (!ik->is_final()) { |
1042 | phase->C->dependencies()->assert_leaf_type(ik); |
1043 | } |
1044 | } |
1045 | |
1046 | // Bypass the dependent load, and compare directly |
1047 | this->set_req(1,ldk2); |
1048 | |
1049 | return this; |
1050 | } |
1051 | |
1052 | //============================================================================= |
1053 | //------------------------------sub-------------------------------------------- |
1054 | // Simplify an CmpN (compare 2 pointers) node, based on local information. |
1055 | // If both inputs are constants, compare them. |
1056 | const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const { |
1057 | const TypePtr *r0 = t1->make_ptr(); // Handy access |
1058 | const TypePtr *r1 = t2->make_ptr(); |
1059 | |
1060 | // Undefined inputs makes for an undefined result |
1061 | if ((r0 == NULL) || (r1 == NULL) || |
1062 | TypePtr::above_centerline(r0->_ptr) || |
1063 | TypePtr::above_centerline(r1->_ptr)) { |
1064 | return Type::TOP; |
1065 | } |
1066 | if (r0 == r1 && r0->singleton()) { |
1067 | // Equal pointer constants (klasses, nulls, etc.) |
1068 | return TypeInt::CC_EQ; |
1069 | } |
1070 | |
1071 | // See if it is 2 unrelated classes. |
1072 | const TypeOopPtr* p0 = r0->isa_oopptr(); |
1073 | const TypeOopPtr* p1 = r1->isa_oopptr(); |
1074 | if (p0 && p1) { |
1075 | ciKlass* klass0 = p0->klass(); |
1076 | bool xklass0 = p0->klass_is_exact(); |
1077 | ciKlass* klass1 = p1->klass(); |
1078 | bool xklass1 = p1->klass_is_exact(); |
1079 | int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0); |
1080 | if (klass0 && klass1 && |
1081 | kps != 1 && // both or neither are klass pointers |
1082 | !klass0->is_interface() && // do not trust interfaces |
1083 | !klass1->is_interface()) { |
1084 | bool unrelated_classes = false; |
1085 | // See if neither subclasses the other, or if the class on top |
1086 | // is precise. In either of these cases, the compare is known |
1087 | // to fail if at least one of the pointers is provably not null. |
1088 | if (klass0->equals(klass1)) { // if types are unequal but klasses are equal |
1089 | // Do nothing; we know nothing for imprecise types |
1090 | } else if (klass0->is_subtype_of(klass1)) { |
1091 | // If klass1's type is PRECISE, then classes are unrelated. |
1092 | unrelated_classes = xklass1; |
1093 | } else if (klass1->is_subtype_of(klass0)) { |
1094 | // If klass0's type is PRECISE, then classes are unrelated. |
1095 | unrelated_classes = xklass0; |
1096 | } else { // Neither subtypes the other |
1097 | unrelated_classes = true; |
1098 | } |
1099 | if (unrelated_classes) { |
1100 | // The oops classes are known to be unrelated. If the joined PTRs of |
1101 | // two oops is not Null and not Bottom, then we are sure that one |
1102 | // of the two oops is non-null, and the comparison will always fail. |
1103 | TypePtr::PTR jp = r0->join_ptr(r1->_ptr); |
1104 | if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { |
1105 | return TypeInt::CC_GT; |
1106 | } |
1107 | } |
1108 | } |
1109 | } |
1110 | |
1111 | // Known constants can be compared exactly |
1112 | // Null can be distinguished from any NotNull pointers |
1113 | // Unknown inputs makes an unknown result |
1114 | if( r0->singleton() ) { |
1115 | intptr_t bits0 = r0->get_con(); |
1116 | if( r1->singleton() ) |
1117 | return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; |
1118 | return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; |
1119 | } else if( r1->singleton() ) { |
1120 | intptr_t bits1 = r1->get_con(); |
1121 | return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; |
1122 | } else |
1123 | return TypeInt::CC; |
1124 | } |
1125 | |
1126 | //------------------------------Ideal------------------------------------------ |
1127 | Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) { |
1128 | return NULL; |
1129 | } |
1130 | |
1131 | //============================================================================= |
1132 | //------------------------------Value------------------------------------------ |
1133 | // Simplify an CmpF (compare 2 floats ) node, based on local information. |
1134 | // If both inputs are constants, compare them. |
1135 | const Type* CmpFNode::Value(PhaseGVN* phase) const { |
1136 | const Node* in1 = in(1); |
1137 | const Node* in2 = in(2); |
1138 | // Either input is TOP ==> the result is TOP |
1139 | const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); |
1140 | if( t1 == Type::TOP ) return Type::TOP; |
1141 | const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); |
1142 | if( t2 == Type::TOP ) return Type::TOP; |
1143 | |
1144 | // Not constants? Don't know squat - even if they are the same |
1145 | // value! If they are NaN's they compare to LT instead of EQ. |
1146 | const TypeF *tf1 = t1->isa_float_constant(); |
1147 | const TypeF *tf2 = t2->isa_float_constant(); |
1148 | if( !tf1 || !tf2 ) return TypeInt::CC; |
1149 | |
1150 | // This implements the Java bytecode fcmpl, so unordered returns -1. |
1151 | if( tf1->is_nan() || tf2->is_nan() ) |
1152 | return TypeInt::CC_LT; |
1153 | |
1154 | if( tf1->_f < tf2->_f ) return TypeInt::CC_LT; |
1155 | if( tf1->_f > tf2->_f ) return TypeInt::CC_GT; |
1156 | assert( tf1->_f == tf2->_f, "do not understand FP behavior" ); |
1157 | return TypeInt::CC_EQ; |
1158 | } |
1159 | |
1160 | |
1161 | //============================================================================= |
1162 | //------------------------------Value------------------------------------------ |
1163 | // Simplify an CmpD (compare 2 doubles ) node, based on local information. |
1164 | // If both inputs are constants, compare them. |
1165 | const Type* CmpDNode::Value(PhaseGVN* phase) const { |
1166 | const Node* in1 = in(1); |
1167 | const Node* in2 = in(2); |
1168 | // Either input is TOP ==> the result is TOP |
1169 | const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); |
1170 | if( t1 == Type::TOP ) return Type::TOP; |
1171 | const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); |
1172 | if( t2 == Type::TOP ) return Type::TOP; |
1173 | |
1174 | // Not constants? Don't know squat - even if they are the same |
1175 | // value! If they are NaN's they compare to LT instead of EQ. |
1176 | const TypeD *td1 = t1->isa_double_constant(); |
1177 | const TypeD *td2 = t2->isa_double_constant(); |
1178 | if( !td1 || !td2 ) return TypeInt::CC; |
1179 | |
1180 | // This implements the Java bytecode dcmpl, so unordered returns -1. |
1181 | if( td1->is_nan() || td2->is_nan() ) |
1182 | return TypeInt::CC_LT; |
1183 | |
1184 | if( td1->_d < td2->_d ) return TypeInt::CC_LT; |
1185 | if( td1->_d > td2->_d ) return TypeInt::CC_GT; |
1186 | assert( td1->_d == td2->_d, "do not understand FP behavior" ); |
1187 | return TypeInt::CC_EQ; |
1188 | } |
1189 | |
1190 | //------------------------------Ideal------------------------------------------ |
1191 | Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
1192 | // Check if we can change this to a CmpF and remove a ConvD2F operation. |
1193 | // Change (CMPD (F2D (float)) (ConD value)) |
1194 | // To (CMPF (float) (ConF value)) |
1195 | // Valid when 'value' does not lose precision as a float. |
1196 | // Benefits: eliminates conversion, does not require 24-bit mode |
1197 | |
1198 | // NaNs prevent commuting operands. This transform works regardless of the |
1199 | // order of ConD and ConvF2D inputs by preserving the original order. |
1200 | int idx_f2d = 1; // ConvF2D on left side? |
1201 | if( in(idx_f2d)->Opcode() != Op_ConvF2D ) |
1202 | idx_f2d = 2; // No, swap to check for reversed args |
1203 | int idx_con = 3-idx_f2d; // Check for the constant on other input |
1204 | |
1205 | if( ConvertCmpD2CmpF && |
1206 | in(idx_f2d)->Opcode() == Op_ConvF2D && |
1207 | in(idx_con)->Opcode() == Op_ConD ) { |
1208 | const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant(); |
1209 | double t2_value_as_double = t2->_d; |
1210 | float t2_value_as_float = (float)t2_value_as_double; |
1211 | if( t2_value_as_double == (double)t2_value_as_float ) { |
1212 | // Test value can be represented as a float |
1213 | // Eliminate the conversion to double and create new comparison |
1214 | Node *new_in1 = in(idx_f2d)->in(1); |
1215 | Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) ); |
1216 | if( idx_f2d != 1 ) { // Must flip args to match original order |
1217 | Node *tmp = new_in1; |
1218 | new_in1 = new_in2; |
1219 | new_in2 = tmp; |
1220 | } |
1221 | CmpFNode *new_cmp = (Opcode() == Op_CmpD3) |
1222 | ? new CmpF3Node( new_in1, new_in2 ) |
1223 | : new CmpFNode ( new_in1, new_in2 ) ; |
1224 | return new_cmp; // Changed to CmpFNode |
1225 | } |
1226 | // Testing value required the precision of a double |
1227 | } |
1228 | return NULL; // No change |
1229 | } |
1230 | |
1231 | |
1232 | //============================================================================= |
1233 | //------------------------------cc2logical------------------------------------- |
1234 | // Convert a condition code type to a logical type |
1235 | const Type *BoolTest::cc2logical( const Type *CC ) const { |
1236 | if( CC == Type::TOP ) return Type::TOP; |
1237 | if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse |
1238 | const TypeInt *ti = CC->is_int(); |
1239 | if( ti->is_con() ) { // Only 1 kind of condition codes set? |
1240 | // Match low order 2 bits |
1241 | int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0; |
1242 | if( _test & 4 ) tmp = 1-tmp; // Optionally complement result |
1243 | return TypeInt::make(tmp); // Boolean result |
1244 | } |
1245 | |
1246 | if( CC == TypeInt::CC_GE ) { |
1247 | if( _test == ge ) return TypeInt::ONE; |
1248 | if( _test == lt ) return TypeInt::ZERO; |
1249 | } |
1250 | if( CC == TypeInt::CC_LE ) { |
1251 | if( _test == le ) return TypeInt::ONE; |
1252 | if( _test == gt ) return TypeInt::ZERO; |
1253 | } |
1254 | |
1255 | return TypeInt::BOOL; |
1256 | } |
1257 | |
1258 | //------------------------------dump_spec------------------------------------- |
1259 | // Print special per-node info |
1260 | void BoolTest::dump_on(outputStream *st) const { |
1261 | const char *msg[] = {"eq" ,"gt" ,"of" ,"lt" ,"ne" ,"le" ,"nof" ,"ge" }; |
1262 | st->print("%s" , msg[_test]); |
1263 | } |
1264 | |
1265 | // Returns the logical AND of two tests (or 'never' if both tests can never be true). |
1266 | // For example, a test for 'le' followed by a test for 'lt' is equivalent with 'lt'. |
1267 | BoolTest::mask BoolTest::merge(BoolTest other) const { |
1268 | const mask res[illegal+1][illegal+1] = { |
1269 | // eq, gt, of, lt, ne, le, nof, ge, never, illegal |
1270 | {eq, never, illegal, never, never, eq, illegal, eq, never, illegal}, // eq |
1271 | {never, gt, illegal, never, gt, never, illegal, gt, never, illegal}, // gt |
1272 | {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // of |
1273 | {never, never, illegal, lt, lt, lt, illegal, never, never, illegal}, // lt |
1274 | {never, gt, illegal, lt, ne, lt, illegal, gt, never, illegal}, // ne |
1275 | {eq, never, illegal, lt, lt, le, illegal, eq, never, illegal}, // le |
1276 | {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // nof |
1277 | {eq, gt, illegal, never, gt, eq, illegal, ge, never, illegal}, // ge |
1278 | {never, never, never, never, never, never, never, never, never, illegal}, // never |
1279 | {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal}}; // illegal |
1280 | return res[_test][other._test]; |
1281 | } |
1282 | |
1283 | //============================================================================= |
1284 | uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); } |
1285 | uint BoolNode::size_of() const { return sizeof(BoolNode); } |
1286 | |
1287 | //------------------------------operator==------------------------------------- |
1288 | bool BoolNode::cmp( const Node &n ) const { |
1289 | const BoolNode *b = (const BoolNode *)&n; // Cast up |
1290 | return (_test._test == b->_test._test); |
1291 | } |
1292 | |
1293 | //-------------------------------make_predicate-------------------------------- |
1294 | Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) { |
1295 | if (test_value->is_Con()) return test_value; |
1296 | if (test_value->is_Bool()) return test_value; |
1297 | if (test_value->is_CMove() && |
1298 | test_value->in(CMoveNode::Condition)->is_Bool()) { |
1299 | BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool(); |
1300 | const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse)); |
1301 | const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue)); |
1302 | if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) { |
1303 | return bol; |
1304 | } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) { |
1305 | return phase->transform( bol->negate(phase) ); |
1306 | } |
1307 | // Else fall through. The CMove gets in the way of the test. |
1308 | // It should be the case that make_predicate(bol->as_int_value()) == bol. |
1309 | } |
1310 | Node* cmp = new CmpINode(test_value, phase->intcon(0)); |
1311 | cmp = phase->transform(cmp); |
1312 | Node* bol = new BoolNode(cmp, BoolTest::ne); |
1313 | return phase->transform(bol); |
1314 | } |
1315 | |
1316 | //--------------------------------as_int_value--------------------------------- |
1317 | Node* BoolNode::as_int_value(PhaseGVN* phase) { |
1318 | // Inverse to make_predicate. The CMove probably boils down to a Conv2B. |
1319 | Node* cmov = CMoveNode::make(NULL, this, |
1320 | phase->intcon(0), phase->intcon(1), |
1321 | TypeInt::BOOL); |
1322 | return phase->transform(cmov); |
1323 | } |
1324 | |
1325 | //----------------------------------negate------------------------------------- |
1326 | BoolNode* BoolNode::negate(PhaseGVN* phase) { |
1327 | return new BoolNode(in(1), _test.negate()); |
1328 | } |
1329 | |
1330 | // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub |
1331 | // overflows and we can prove that C is not in the two resulting ranges. |
1332 | // This optimization is similar to the one performed by CmpUNode::Value(). |
1333 | Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op, |
1334 | int cmp1_op, const TypeInt* cmp2_type) { |
1335 | // Only optimize eq/ne integer comparison of add/sub |
1336 | if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && |
1337 | (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) { |
1338 | // Skip cases were inputs of add/sub are not integers or of bottom type |
1339 | const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int(); |
1340 | const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int(); |
1341 | if ((r0 != NULL) && (r0 != TypeInt::INT) && |
1342 | (r1 != NULL) && (r1 != TypeInt::INT) && |
1343 | (cmp2_type != TypeInt::INT)) { |
1344 | // Compute exact (long) type range of add/sub result |
1345 | jlong lo_long = r0->_lo; |
1346 | jlong hi_long = r0->_hi; |
1347 | if (cmp1_op == Op_AddI) { |
1348 | lo_long += r1->_lo; |
1349 | hi_long += r1->_hi; |
1350 | } else { |
1351 | lo_long -= r1->_hi; |
1352 | hi_long -= r1->_lo; |
1353 | } |
1354 | // Check for over-/underflow by casting to integer |
1355 | int lo_int = (int)lo_long; |
1356 | int hi_int = (int)hi_long; |
1357 | bool underflow = lo_long != (jlong)lo_int; |
1358 | bool overflow = hi_long != (jlong)hi_int; |
1359 | if ((underflow != overflow) && (hi_int < lo_int)) { |
1360 | // Overflow on one boundary, compute resulting type ranges: |
1361 | // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT] |
1362 | int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here |
1363 | const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w); |
1364 | const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w); |
1365 | // Compare second input of cmp to both type ranges |
1366 | const Type* sub_tr1 = cmp->sub(tr1, cmp2_type); |
1367 | const Type* sub_tr2 = cmp->sub(tr2, cmp2_type); |
1368 | if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) { |
1369 | // The result of the add/sub will never equal cmp2. Replace BoolNode |
1370 | // by false (0) if it tests for equality and by true (1) otherwise. |
1371 | return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1); |
1372 | } |
1373 | } |
1374 | } |
1375 | } |
1376 | return NULL; |
1377 | } |
1378 | |
1379 | static bool is_counted_loop_cmp(Node *cmp) { |
1380 | Node *n = cmp->in(1)->in(1); |
1381 | return n != NULL && |
1382 | n->is_Phi() && |
1383 | n->in(0) != NULL && |
1384 | n->in(0)->is_CountedLoop() && |
1385 | n->in(0)->as_CountedLoop()->phi() == n; |
1386 | } |
1387 | |
1388 | //------------------------------Ideal------------------------------------------ |
1389 | Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1390 | // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)". |
1391 | // This moves the constant to the right. Helps value-numbering. |
1392 | Node *cmp = in(1); |
1393 | if( !cmp->is_Sub() ) return NULL; |
1394 | int cop = cmp->Opcode(); |
1395 | if( cop == Op_FastLock || cop == Op_FastUnlock) return NULL; |
1396 | Node *cmp1 = cmp->in(1); |
1397 | Node *cmp2 = cmp->in(2); |
1398 | if( !cmp1 ) return NULL; |
1399 | |
1400 | if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) { |
1401 | return NULL; |
1402 | } |
1403 | |
1404 | // Constant on left? |
1405 | Node *con = cmp1; |
1406 | uint op2 = cmp2->Opcode(); |
1407 | // Move constants to the right of compare's to canonicalize. |
1408 | // Do not muck with Opaque1 nodes, as this indicates a loop |
1409 | // guard that cannot change shape. |
1410 | if( con->is_Con() && !cmp2->is_Con() && op2 != Op_Opaque1 && |
1411 | // Because of NaN's, CmpD and CmpF are not commutative |
1412 | cop != Op_CmpD && cop != Op_CmpF && |
1413 | // Protect against swapping inputs to a compare when it is used by a |
1414 | // counted loop exit, which requires maintaining the loop-limit as in(2) |
1415 | !is_counted_loop_exit_test() ) { |
1416 | // Ok, commute the constant to the right of the cmp node. |
1417 | // Clone the Node, getting a new Node of the same class |
1418 | cmp = cmp->clone(); |
1419 | // Swap inputs to the clone |
1420 | cmp->swap_edges(1, 2); |
1421 | cmp = phase->transform( cmp ); |
1422 | return new BoolNode( cmp, _test.commute() ); |
1423 | } |
1424 | |
1425 | // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)". |
1426 | // The XOR-1 is an idiom used to flip the sense of a bool. We flip the |
1427 | // test instead. |
1428 | int cmp1_op = cmp1->Opcode(); |
1429 | const TypeInt* cmp2_type = phase->type(cmp2)->isa_int(); |
1430 | if (cmp2_type == NULL) return NULL; |
1431 | Node* j_xor = cmp1; |
1432 | if( cmp2_type == TypeInt::ZERO && |
1433 | cmp1_op == Op_XorI && |
1434 | j_xor->in(1) != j_xor && // An xor of itself is dead |
1435 | phase->type( j_xor->in(1) ) == TypeInt::BOOL && |
1436 | phase->type( j_xor->in(2) ) == TypeInt::ONE && |
1437 | (_test._test == BoolTest::eq || |
1438 | _test._test == BoolTest::ne) ) { |
1439 | Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2)); |
1440 | return new BoolNode( ncmp, _test.negate() ); |
1441 | } |
1442 | |
1443 | // Change ((x & m) u<= m) or ((m & x) u<= m) to always true |
1444 | // Same with ((x & m) u< m+1) and ((m & x) u< m+1) |
1445 | if (cop == Op_CmpU && |
1446 | cmp1_op == Op_AndI) { |
1447 | Node* bound = NULL; |
1448 | if (_test._test == BoolTest::le) { |
1449 | bound = cmp2; |
1450 | } else if (_test._test == BoolTest::lt && |
1451 | cmp2->Opcode() == Op_AddI && |
1452 | cmp2->in(2)->find_int_con(0) == 1) { |
1453 | bound = cmp2->in(1); |
1454 | } |
1455 | if (cmp1->in(2) == bound || cmp1->in(1) == bound) { |
1456 | return ConINode::make(1); |
1457 | } |
1458 | } |
1459 | |
1460 | // Change ((x & (m - 1)) u< m) into (m > 0) |
1461 | // This is the off-by-one variant of the above |
1462 | if (cop == Op_CmpU && |
1463 | _test._test == BoolTest::lt && |
1464 | cmp1_op == Op_AndI) { |
1465 | Node* l = cmp1->in(1); |
1466 | Node* r = cmp1->in(2); |
1467 | for (int repeat = 0; repeat < 2; repeat++) { |
1468 | bool match = r->Opcode() == Op_AddI && r->in(2)->find_int_con(0) == -1 && |
1469 | r->in(1) == cmp2; |
1470 | if (match) { |
1471 | // arraylength known to be non-negative, so a (arraylength != 0) is sufficient, |
1472 | // but to be compatible with the array range check pattern, use (arraylength u> 0) |
1473 | Node* ncmp = cmp2->Opcode() == Op_LoadRange |
1474 | ? phase->transform(new CmpUNode(cmp2, phase->intcon(0))) |
1475 | : phase->transform(new CmpINode(cmp2, phase->intcon(0))); |
1476 | return new BoolNode(ncmp, BoolTest::gt); |
1477 | } else { |
1478 | // commute and try again |
1479 | l = cmp1->in(2); |
1480 | r = cmp1->in(1); |
1481 | } |
1482 | } |
1483 | } |
1484 | |
1485 | // Change x u< 1 or x u<= 0 to x == 0 |
1486 | if (cop == Op_CmpU && |
1487 | cmp1_op != Op_LoadRange && |
1488 | ((_test._test == BoolTest::lt && |
1489 | cmp2->find_int_con(-1) == 1) || |
1490 | (_test._test == BoolTest::le && |
1491 | cmp2->find_int_con(-1) == 0))) { |
1492 | Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0))); |
1493 | return new BoolNode(ncmp, BoolTest::eq); |
1494 | } |
1495 | |
1496 | // Change (arraylength <= 0) or (arraylength == 0) |
1497 | // into (arraylength u<= 0) |
1498 | // Also change (arraylength != 0) into (arraylength u> 0) |
1499 | // The latter version matches the code pattern generated for |
1500 | // array range checks, which will more likely be optimized later. |
1501 | if (cop == Op_CmpI && |
1502 | cmp1_op == Op_LoadRange && |
1503 | cmp2->find_int_con(-1) == 0) { |
1504 | if (_test._test == BoolTest::le || _test._test == BoolTest::eq) { |
1505 | Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); |
1506 | return new BoolNode(ncmp, BoolTest::le); |
1507 | } else if (_test._test == BoolTest::ne) { |
1508 | Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); |
1509 | return new BoolNode(ncmp, BoolTest::gt); |
1510 | } |
1511 | } |
1512 | |
1513 | // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)". |
1514 | // This is a standard idiom for branching on a boolean value. |
1515 | Node *c2b = cmp1; |
1516 | if( cmp2_type == TypeInt::ZERO && |
1517 | cmp1_op == Op_Conv2B && |
1518 | (_test._test == BoolTest::eq || |
1519 | _test._test == BoolTest::ne) ) { |
1520 | Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int() |
1521 | ? (Node*)new CmpINode(c2b->in(1),cmp2) |
1522 | : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR)) |
1523 | ); |
1524 | return new BoolNode( ncmp, _test._test ); |
1525 | } |
1526 | |
1527 | // Comparing a SubI against a zero is equal to comparing the SubI |
1528 | // arguments directly. This only works for eq and ne comparisons |
1529 | // due to possible integer overflow. |
1530 | if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && |
1531 | (cop == Op_CmpI) && |
1532 | (cmp1_op == Op_SubI) && |
1533 | ( cmp2_type == TypeInt::ZERO ) ) { |
1534 | Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2))); |
1535 | return new BoolNode( ncmp, _test._test ); |
1536 | } |
1537 | |
1538 | // Same as above but with and AddI of a constant |
1539 | if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && |
1540 | cop == Op_CmpI && |
1541 | cmp1_op == Op_AddI && |
1542 | cmp1->in(2) != NULL && |
1543 | phase->type(cmp1->in(2))->isa_int() && |
1544 | phase->type(cmp1->in(2))->is_int()->is_con() && |
1545 | cmp2_type == TypeInt::ZERO && |
1546 | !is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape |
1547 | ) { |
1548 | const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int(); |
1549 | Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi))); |
1550 | return new BoolNode( ncmp, _test._test ); |
1551 | } |
1552 | |
1553 | // Change "bool eq/ne (cmp (phi (X -X) 0))" into "bool eq/ne (cmp X 0)" |
1554 | // since zero check of conditional negation of an integer is equal to |
1555 | // zero check of the integer directly. |
1556 | if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && |
1557 | (cop == Op_CmpI) && |
1558 | (cmp2_type == TypeInt::ZERO) && |
1559 | (cmp1_op == Op_Phi)) { |
1560 | // There should be a diamond phi with true path at index 1 or 2 |
1561 | PhiNode *phi = cmp1->as_Phi(); |
1562 | int idx_true = phi->is_diamond_phi(); |
1563 | if (idx_true != 0) { |
1564 | // True input is in(idx_true) while false input is in(3 - idx_true) |
1565 | Node *tin = phi->in(idx_true); |
1566 | Node *fin = phi->in(3 - idx_true); |
1567 | if ((tin->Opcode() == Op_SubI) && |
1568 | (phase->type(tin->in(1)) == TypeInt::ZERO) && |
1569 | (tin->in(2) == fin)) { |
1570 | // Found conditional negation at true path, create a new CmpINode without that |
1571 | Node *ncmp = phase->transform(new CmpINode(fin, cmp2)); |
1572 | return new BoolNode(ncmp, _test._test); |
1573 | } |
1574 | if ((fin->Opcode() == Op_SubI) && |
1575 | (phase->type(fin->in(1)) == TypeInt::ZERO) && |
1576 | (fin->in(2) == tin)) { |
1577 | // Found conditional negation at false path, create a new CmpINode without that |
1578 | Node *ncmp = phase->transform(new CmpINode(tin, cmp2)); |
1579 | return new BoolNode(ncmp, _test._test); |
1580 | } |
1581 | } |
1582 | } |
1583 | |
1584 | // Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the |
1585 | // most general case because negating 0x80000000 does nothing. Needed for |
1586 | // the CmpF3/SubI/CmpI idiom. |
1587 | if( cop == Op_CmpI && |
1588 | cmp1_op == Op_SubI && |
1589 | cmp2_type == TypeInt::ZERO && |
1590 | phase->type( cmp1->in(1) ) == TypeInt::ZERO && |
1591 | phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) { |
1592 | Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2)); |
1593 | return new BoolNode( ncmp, _test.commute() ); |
1594 | } |
1595 | |
1596 | // Try to optimize signed integer comparison |
1597 | return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type); |
1598 | |
1599 | // The transformation below is not valid for either signed or unsigned |
1600 | // comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE. |
1601 | // This transformation can be resurrected when we are able to |
1602 | // make inferences about the range of values being subtracted from |
1603 | // (or added to) relative to the wraparound point. |
1604 | // |
1605 | // // Remove +/-1's if possible. |
1606 | // // "X <= Y-1" becomes "X < Y" |
1607 | // // "X+1 <= Y" becomes "X < Y" |
1608 | // // "X < Y+1" becomes "X <= Y" |
1609 | // // "X-1 < Y" becomes "X <= Y" |
1610 | // // Do not this to compares off of the counted-loop-end. These guys are |
1611 | // // checking the trip counter and they want to use the post-incremented |
1612 | // // counter. If they use the PRE-incremented counter, then the counter has |
1613 | // // to be incremented in a private block on a loop backedge. |
1614 | // if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd ) |
1615 | // return NULL; |
1616 | // #ifndef PRODUCT |
1617 | // // Do not do this in a wash GVN pass during verification. |
1618 | // // Gets triggered by too many simple optimizations to be bothered with |
1619 | // // re-trying it again and again. |
1620 | // if( !phase->allow_progress() ) return NULL; |
1621 | // #endif |
1622 | // // Not valid for unsigned compare because of corner cases in involving zero. |
1623 | // // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an |
1624 | // // exception in case X is 0 (because 0-1 turns into 4billion unsigned but |
1625 | // // "0 <=u Y" is always true). |
1626 | // if( cmp->Opcode() == Op_CmpU ) return NULL; |
1627 | // int cmp2_op = cmp2->Opcode(); |
1628 | // if( _test._test == BoolTest::le ) { |
1629 | // if( cmp1_op == Op_AddI && |
1630 | // phase->type( cmp1->in(2) ) == TypeInt::ONE ) |
1631 | // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt ); |
1632 | // else if( cmp2_op == Op_AddI && |
1633 | // phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 ) |
1634 | // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt ); |
1635 | // } else if( _test._test == BoolTest::lt ) { |
1636 | // if( cmp1_op == Op_AddI && |
1637 | // phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 ) |
1638 | // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le ); |
1639 | // else if( cmp2_op == Op_AddI && |
1640 | // phase->type( cmp2->in(2) ) == TypeInt::ONE ) |
1641 | // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le ); |
1642 | // } |
1643 | } |
1644 | |
1645 | //------------------------------Value------------------------------------------ |
1646 | // Simplify a Bool (convert condition codes to boolean (1 or 0)) node, |
1647 | // based on local information. If the input is constant, do it. |
1648 | const Type* BoolNode::Value(PhaseGVN* phase) const { |
1649 | return _test.cc2logical( phase->type( in(1) ) ); |
1650 | } |
1651 | |
1652 | #ifndef PRODUCT |
1653 | //------------------------------dump_spec-------------------------------------- |
1654 | // Dump special per-node info |
1655 | void BoolNode::dump_spec(outputStream *st) const { |
1656 | st->print("[" ); |
1657 | _test.dump_on(st); |
1658 | st->print("]" ); |
1659 | } |
1660 | |
1661 | //-------------------------------related--------------------------------------- |
1662 | // A BoolNode's related nodes are all of its data inputs, and all of its |
1663 | // outputs until control nodes are hit, which are included. In compact |
1664 | // representation, inputs till level 3 and immediate outputs are included. |
1665 | void BoolNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { |
1666 | if (compact) { |
1667 | this->collect_nodes(in_rel, 3, false, true); |
1668 | this->collect_nodes(out_rel, -1, false, false); |
1669 | } else { |
1670 | this->collect_nodes_in_all_data(in_rel, false); |
1671 | this->collect_nodes_out_all_ctrl_boundary(out_rel); |
1672 | } |
1673 | } |
1674 | #endif |
1675 | |
1676 | //----------------------is_counted_loop_exit_test------------------------------ |
1677 | // Returns true if node is used by a counted loop node. |
1678 | bool BoolNode::is_counted_loop_exit_test() { |
1679 | for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { |
1680 | Node* use = fast_out(i); |
1681 | if (use->is_CountedLoopEnd()) { |
1682 | return true; |
1683 | } |
1684 | } |
1685 | return false; |
1686 | } |
1687 | |
1688 | //============================================================================= |
1689 | //------------------------------Value------------------------------------------ |
1690 | // Compute sqrt |
1691 | const Type* SqrtDNode::Value(PhaseGVN* phase) const { |
1692 | const Type *t1 = phase->type( in(1) ); |
1693 | if( t1 == Type::TOP ) return Type::TOP; |
1694 | if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; |
1695 | double d = t1->getd(); |
1696 | if( d < 0.0 ) return Type::DOUBLE; |
1697 | return TypeD::make( sqrt( d ) ); |
1698 | } |
1699 | |
1700 | const Type* SqrtFNode::Value(PhaseGVN* phase) const { |
1701 | const Type *t1 = phase->type( in(1) ); |
1702 | if( t1 == Type::TOP ) return Type::TOP; |
1703 | if( t1->base() != Type::FloatCon ) return Type::FLOAT; |
1704 | float f = t1->getf(); |
1705 | if( f < 0.0f ) return Type::FLOAT; |
1706 | return TypeF::make( (float)sqrt( (double)f ) ); |
1707 | } |
1708 | |