Mercurial > hg > truffle
annotate src/share/vm/opto/mulnode.cpp @ 14714:b602356a9cfc
additional canonicalizers for accesses and value nodes (improves number of implicit null checks)
author | Lukas Stadler <lukas.stadler@oracle.com> |
---|---|
date | Thu, 20 Mar 2014 17:15:36 +0100 |
parents | 2113136690bc |
children | 4ca6dc0799b6 |
rev | line source |
---|---|
0 | 1 /* |
6842
b9a9ed0f8eeb
7197424: update copyright year to match last edit in jdk8 hotspot repository
mikael
parents:
6804
diff
changeset
|
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
897
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
897
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
897
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "memory/allocation.inline.hpp" | |
27 #include "opto/addnode.hpp" | |
28 #include "opto/connode.hpp" | |
29 #include "opto/memnode.hpp" | |
30 #include "opto/mulnode.hpp" | |
31 #include "opto/phaseX.hpp" | |
32 #include "opto/subnode.hpp" | |
0 | 33 |
1972 | 34 // Portions of code courtesy of Clifford Click |
0 | 35 |
36 | |
37 //============================================================================= | |
38 //------------------------------hash------------------------------------------- | |
39 // Hash function over MulNodes. Needs to be commutative; i.e., I swap | |
40 // (commute) inputs to MulNodes willy-nilly so the hash function must return | |
41 // the same value in the presence of edge swapping. | |
42 uint MulNode::hash() const { | |
43 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode(); | |
44 } | |
45 | |
46 //------------------------------Identity--------------------------------------- | |
47 // Multiplying a one preserves the other argument | |
48 Node *MulNode::Identity( PhaseTransform *phase ) { | |
49 register const Type *one = mul_id(); // The multiplicative identity | |
50 if( phase->type( in(1) )->higher_equal( one ) ) return in(2); | |
51 if( phase->type( in(2) )->higher_equal( one ) ) return in(1); | |
52 | |
53 return this; | |
54 } | |
55 | |
56 //------------------------------Ideal------------------------------------------ | |
57 // We also canonicalize the Node, moving constants to the right input, | |
58 // and flatten expressions (so that 1+x+2 becomes x+3). | |
59 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
60 const Type *t1 = phase->type( in(1) ); | |
61 const Type *t2 = phase->type( in(2) ); | |
62 Node *progress = NULL; // Progress flag | |
63 // We are OK if right is a constant, or right is a load and | |
64 // left is a non-constant. | |
65 if( !(t2->singleton() || | |
66 (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) { | |
67 if( t1->singleton() || // Left input is a constant? | |
68 // Otherwise, sort inputs (commutativity) to help value numbering. | |
69 (in(1)->_idx > in(2)->_idx) ) { | |
70 swap_edges(1, 2); | |
71 const Type *t = t1; | |
72 t1 = t2; | |
73 t2 = t; | |
74 progress = this; // Made progress | |
75 } | |
76 } | |
77 | |
78 // If the right input is a constant, and the left input is a product of a | |
79 // constant, flatten the expression tree. | |
80 uint op = Opcode(); | |
81 if( t2->singleton() && // Right input is a constant? | |
82 op != Op_MulF && // Float & double cannot reassociate | |
83 op != Op_MulD ) { | |
84 if( t2 == Type::TOP ) return NULL; | |
85 Node *mul1 = in(1); | |
86 #ifdef ASSERT | |
87 // Check for dead loop | |
88 int op1 = mul1->Opcode(); | |
89 if( phase->eqv( mul1, this ) || phase->eqv( in(2), this ) || | |
90 ( op1 == mul_opcode() || op1 == add_opcode() ) && | |
91 ( phase->eqv( mul1->in(1), this ) || phase->eqv( mul1->in(2), this ) || | |
92 phase->eqv( mul1->in(1), mul1 ) || phase->eqv( mul1->in(2), mul1 ) ) ) | |
93 assert(false, "dead loop in MulNode::Ideal"); | |
94 #endif | |
95 | |
96 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply? | |
97 // Mul of a constant? | |
98 const Type *t12 = phase->type( mul1->in(2) ); | |
99 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant? | |
100 // Compute new constant; check for overflow | |
3842 | 101 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12); |
0 | 102 if( tcon01->singleton() ) { |
103 // The Mul of the flattened expression | |
104 set_req(1, mul1->in(1)); | |
105 set_req(2, phase->makecon( tcon01 )); | |
106 t2 = tcon01; | |
107 progress = this; // Made progress | |
108 } | |
109 } | |
110 } | |
111 // If the right input is a constant, and the left input is an add of a | |
112 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0 | |
113 const Node *add1 = in(1); | |
114 if( add1->Opcode() == add_opcode() ) { // Left input is an add? | |
115 // Add of a constant? | |
116 const Type *t12 = phase->type( add1->in(2) ); | |
117 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant? | |
118 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" ); | |
119 // Compute new constant; check for overflow | |
120 const Type *tcon01 = mul_ring(t2,t12); | |
121 if( tcon01->singleton() ) { | |
122 | |
123 // Convert (X+con1)*con0 into X*con0 | |
124 Node *mul = clone(); // mul = ()*con0 | |
125 mul->set_req(1,add1->in(1)); // mul = X*con0 | |
126 mul = phase->transform(mul); | |
127 | |
128 Node *add2 = add1->clone(); | |
129 add2->set_req(1, mul); // X*con0 + con0*con1 | |
130 add2->set_req(2, phase->makecon(tcon01) ); | |
131 progress = add2; | |
132 } | |
133 } | |
134 } // End of is left input an add | |
135 } // End of is right input a Mul | |
136 | |
137 return progress; | |
138 } | |
139 | |
140 //------------------------------Value----------------------------------------- | |
141 const Type *MulNode::Value( PhaseTransform *phase ) const { | |
142 const Type *t1 = phase->type( in(1) ); | |
143 const Type *t2 = phase->type( in(2) ); | |
144 // Either input is TOP ==> the result is TOP | |
145 if( t1 == Type::TOP ) return Type::TOP; | |
146 if( t2 == Type::TOP ) return Type::TOP; | |
147 | |
148 // Either input is ZERO ==> the result is ZERO. | |
149 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0 | |
150 int op = Opcode(); | |
151 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) { | |
152 const Type *zero = add_id(); // The multiplicative zero | |
153 if( t1->higher_equal( zero ) ) return zero; | |
154 if( t2->higher_equal( zero ) ) return zero; | |
155 } | |
156 | |
157 // Either input is BOTTOM ==> the result is the local BOTTOM | |
158 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) | |
159 return bottom_type(); | |
160 | |
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
161 #if defined(IA32) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
162 // Can't trust native compilers to properly fold strict double |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
163 // multiplication with round-to-zero on this platform. |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
164 if (op == Op_MulD && phase->C->method()->is_strict()) { |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
165 return TypeD::DOUBLE; |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
166 } |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
167 #endif |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
168 |
0 | 169 return mul_ring(t1,t2); // Local flavor of type multiplication |
170 } | |
171 | |
172 | |
173 //============================================================================= | |
174 //------------------------------Ideal------------------------------------------ | |
175 // Check for power-of-2 multiply, then try the regular MulNode::Ideal | |
176 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
177 // Swap constant to right | |
178 jint con; | |
179 if ((con = in(1)->find_int_con(0)) != 0) { | |
180 swap_edges(1, 2); | |
181 // Finish rest of method to use info in 'con' | |
182 } else if ((con = in(2)->find_int_con(0)) == 0) { | |
183 return MulNode::Ideal(phase, can_reshape); | |
184 } | |
185 | |
186 // Now we have a constant Node on the right and the constant in con | |
187 if( con == 0 ) return NULL; // By zero is handled by Value call | |
188 if( con == 1 ) return NULL; // By one is handled by Identity call | |
189 | |
190 // Check for negative constant; if so negate the final result | |
191 bool sign_flip = false; | |
192 if( con < 0 ) { | |
193 con = -con; | |
194 sign_flip = true; | |
195 } | |
196 | |
197 // Get low bit; check for being the only bit | |
198 Node *res = NULL; | |
199 jint bit1 = con & -con; // Extract low bit | |
200 if( bit1 == con ) { // Found a power of 2? | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
201 res = new (phase->C) LShiftINode( in(1), phase->intcon(log2_intptr(bit1)) ); |
0 | 202 } else { |
203 | |
204 // Check for constant with 2 bits set | |
205 jint bit2 = con-bit1; | |
206 bit2 = bit2 & -bit2; // Extract 2nd bit | |
207 if( bit2 + bit1 == con ) { // Found all bits in con? | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
208 Node *n1 = phase->transform( new (phase->C) LShiftINode( in(1), phase->intcon(log2_intptr(bit1)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
209 Node *n2 = phase->transform( new (phase->C) LShiftINode( in(1), phase->intcon(log2_intptr(bit2)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
210 res = new (phase->C) AddINode( n2, n1 ); |
0 | 211 |
212 } else if (is_power_of_2(con+1)) { | |
213 // Sleezy: power-of-2 -1. Next time be generic. | |
214 jint temp = (jint) (con + 1); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
215 Node *n1 = phase->transform( new (phase->C) LShiftINode( in(1), phase->intcon(log2_intptr(temp)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
216 res = new (phase->C) SubINode( n1, in(1) ); |
0 | 217 } else { |
218 return MulNode::Ideal(phase, can_reshape); | |
219 } | |
220 } | |
221 | |
222 if( sign_flip ) { // Need to negate result? | |
223 res = phase->transform(res);// Transform, before making the zero con | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
224 res = new (phase->C) SubINode(phase->intcon(0),res); |
0 | 225 } |
226 | |
227 return res; // Return final result | |
228 } | |
229 | |
230 //------------------------------mul_ring--------------------------------------- | |
231 // Compute the product type of two integer ranges into this node. | |
232 const Type *MulINode::mul_ring(const Type *t0, const Type *t1) const { | |
233 const TypeInt *r0 = t0->is_int(); // Handy access | |
234 const TypeInt *r1 = t1->is_int(); | |
235 | |
236 // Fetch endpoints of all ranges | |
237 int32 lo0 = r0->_lo; | |
238 double a = (double)lo0; | |
239 int32 hi0 = r0->_hi; | |
240 double b = (double)hi0; | |
241 int32 lo1 = r1->_lo; | |
242 double c = (double)lo1; | |
243 int32 hi1 = r1->_hi; | |
244 double d = (double)hi1; | |
245 | |
246 // Compute all endpoints & check for overflow | |
247 int32 A = lo0*lo1; | |
248 if( (double)A != a*c ) return TypeInt::INT; // Overflow? | |
249 int32 B = lo0*hi1; | |
250 if( (double)B != a*d ) return TypeInt::INT; // Overflow? | |
251 int32 C = hi0*lo1; | |
252 if( (double)C != b*c ) return TypeInt::INT; // Overflow? | |
253 int32 D = hi0*hi1; | |
254 if( (double)D != b*d ) return TypeInt::INT; // Overflow? | |
255 | |
256 if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints | |
257 else { lo0 = B; hi0 = A; } | |
258 if( C < D ) { | |
259 if( C < lo0 ) lo0 = C; | |
260 if( D > hi0 ) hi0 = D; | |
261 } else { | |
262 if( D < lo0 ) lo0 = D; | |
263 if( C > hi0 ) hi0 = C; | |
264 } | |
265 return TypeInt::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); | |
266 } | |
267 | |
268 | |
269 //============================================================================= | |
270 //------------------------------Ideal------------------------------------------ | |
271 // Check for power-of-2 multiply, then try the regular MulNode::Ideal | |
272 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
273 // Swap constant to right | |
274 jlong con; | |
275 if ((con = in(1)->find_long_con(0)) != 0) { | |
276 swap_edges(1, 2); | |
277 // Finish rest of method to use info in 'con' | |
278 } else if ((con = in(2)->find_long_con(0)) == 0) { | |
279 return MulNode::Ideal(phase, can_reshape); | |
280 } | |
281 | |
282 // Now we have a constant Node on the right and the constant in con | |
283 if( con == CONST64(0) ) return NULL; // By zero is handled by Value call | |
284 if( con == CONST64(1) ) return NULL; // By one is handled by Identity call | |
285 | |
286 // Check for negative constant; if so negate the final result | |
287 bool sign_flip = false; | |
288 if( con < 0 ) { | |
289 con = -con; | |
290 sign_flip = true; | |
291 } | |
292 | |
293 // Get low bit; check for being the only bit | |
294 Node *res = NULL; | |
295 jlong bit1 = con & -con; // Extract low bit | |
296 if( bit1 == con ) { // Found a power of 2? | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
297 res = new (phase->C) LShiftLNode( in(1), phase->intcon(log2_long(bit1)) ); |
0 | 298 } else { |
299 | |
300 // Check for constant with 2 bits set | |
301 jlong bit2 = con-bit1; | |
302 bit2 = bit2 & -bit2; // Extract 2nd bit | |
303 if( bit2 + bit1 == con ) { // Found all bits in con? | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
304 Node *n1 = phase->transform( new (phase->C) LShiftLNode( in(1), phase->intcon(log2_long(bit1)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
305 Node *n2 = phase->transform( new (phase->C) LShiftLNode( in(1), phase->intcon(log2_long(bit2)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
306 res = new (phase->C) AddLNode( n2, n1 ); |
0 | 307 |
308 } else if (is_power_of_2_long(con+1)) { | |
309 // Sleezy: power-of-2 -1. Next time be generic. | |
310 jlong temp = (jlong) (con + 1); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
311 Node *n1 = phase->transform( new (phase->C) LShiftLNode( in(1), phase->intcon(log2_long(temp)) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
312 res = new (phase->C) SubLNode( n1, in(1) ); |
0 | 313 } else { |
314 return MulNode::Ideal(phase, can_reshape); | |
315 } | |
316 } | |
317 | |
318 if( sign_flip ) { // Need to negate result? | |
319 res = phase->transform(res);// Transform, before making the zero con | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
320 res = new (phase->C) SubLNode(phase->longcon(0),res); |
0 | 321 } |
322 | |
323 return res; // Return final result | |
324 } | |
325 | |
326 //------------------------------mul_ring--------------------------------------- | |
327 // Compute the product type of two integer ranges into this node. | |
328 const Type *MulLNode::mul_ring(const Type *t0, const Type *t1) const { | |
329 const TypeLong *r0 = t0->is_long(); // Handy access | |
330 const TypeLong *r1 = t1->is_long(); | |
331 | |
332 // Fetch endpoints of all ranges | |
333 jlong lo0 = r0->_lo; | |
334 double a = (double)lo0; | |
335 jlong hi0 = r0->_hi; | |
336 double b = (double)hi0; | |
337 jlong lo1 = r1->_lo; | |
338 double c = (double)lo1; | |
339 jlong hi1 = r1->_hi; | |
340 double d = (double)hi1; | |
341 | |
342 // Compute all endpoints & check for overflow | |
343 jlong A = lo0*lo1; | |
344 if( (double)A != a*c ) return TypeLong::LONG; // Overflow? | |
345 jlong B = lo0*hi1; | |
346 if( (double)B != a*d ) return TypeLong::LONG; // Overflow? | |
347 jlong C = hi0*lo1; | |
348 if( (double)C != b*c ) return TypeLong::LONG; // Overflow? | |
349 jlong D = hi0*hi1; | |
350 if( (double)D != b*d ) return TypeLong::LONG; // Overflow? | |
351 | |
352 if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints | |
353 else { lo0 = B; hi0 = A; } | |
354 if( C < D ) { | |
355 if( C < lo0 ) lo0 = C; | |
356 if( D > hi0 ) hi0 = D; | |
357 } else { | |
358 if( D < lo0 ) lo0 = D; | |
359 if( C > hi0 ) hi0 = C; | |
360 } | |
361 return TypeLong::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); | |
362 } | |
363 | |
364 //============================================================================= | |
365 //------------------------------mul_ring--------------------------------------- | |
366 // Compute the product type of two double ranges into this node. | |
367 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const { | |
368 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT; | |
369 return TypeF::make( t0->getf() * t1->getf() ); | |
370 } | |
371 | |
372 //============================================================================= | |
373 //------------------------------mul_ring--------------------------------------- | |
374 // Compute the product type of two double ranges into this node. | |
375 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const { | |
376 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE; | |
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
377 // We must be multiplying 2 double constants. |
0 | 378 return TypeD::make( t0->getd() * t1->getd() ); |
379 } | |
380 | |
381 //============================================================================= | |
145 | 382 //------------------------------Value------------------------------------------ |
383 const Type *MulHiLNode::Value( PhaseTransform *phase ) const { | |
384 // Either input is TOP ==> the result is TOP | |
385 const Type *t1 = phase->type( in(1) ); | |
386 const Type *t2 = phase->type( in(2) ); | |
387 if( t1 == Type::TOP ) return Type::TOP; | |
388 if( t2 == Type::TOP ) return Type::TOP; | |
389 | |
390 // Either input is BOTTOM ==> the result is the local BOTTOM | |
391 const Type *bot = bottom_type(); | |
392 if( (t1 == bot) || (t2 == bot) || | |
393 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
394 return bot; | |
395 | |
396 // It is not worth trying to constant fold this stuff! | |
397 return TypeLong::LONG; | |
398 } | |
399 | |
400 //============================================================================= | |
0 | 401 //------------------------------mul_ring--------------------------------------- |
402 // Supplied function returns the product of the inputs IN THE CURRENT RING. | |
403 // For the logical operations the ring's MUL is really a logical AND function. | |
404 // This also type-checks the inputs for sanity. Guaranteed never to | |
405 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. | |
406 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const { | |
407 const TypeInt *r0 = t0->is_int(); // Handy access | |
408 const TypeInt *r1 = t1->is_int(); | |
409 int widen = MAX2(r0->_widen,r1->_widen); | |
410 | |
411 // If either input is a constant, might be able to trim cases | |
412 if( !r0->is_con() && !r1->is_con() ) | |
413 return TypeInt::INT; // No constants to be had | |
414 | |
415 // Both constants? Return bits | |
416 if( r0->is_con() && r1->is_con() ) | |
417 return TypeInt::make( r0->get_con() & r1->get_con() ); | |
418 | |
419 if( r0->is_con() && r0->get_con() > 0 ) | |
420 return TypeInt::make(0, r0->get_con(), widen); | |
421 | |
422 if( r1->is_con() && r1->get_con() > 0 ) | |
423 return TypeInt::make(0, r1->get_con(), widen); | |
424 | |
425 if( r0 == TypeInt::BOOL || r1 == TypeInt::BOOL ) { | |
426 return TypeInt::BOOL; | |
427 } | |
428 | |
429 return TypeInt::INT; // No constants to be had | |
430 } | |
431 | |
432 //------------------------------Identity--------------------------------------- | |
433 // Masking off the high bits of an unsigned load is not required | |
434 Node *AndINode::Identity( PhaseTransform *phase ) { | |
435 | |
436 // x & x => x | |
437 if (phase->eqv(in(1), in(2))) return in(1); | |
438 | |
824 | 439 Node* in1 = in(1); |
440 uint op = in1->Opcode(); | |
441 const TypeInt* t2 = phase->type(in(2))->isa_int(); | |
442 if (t2 && t2->is_con()) { | |
0 | 443 int con = t2->get_con(); |
444 // Masking off high bits which are always zero is useless. | |
445 const TypeInt* t1 = phase->type( in(1) )->isa_int(); | |
446 if (t1 != NULL && t1->_lo >= 0) { | |
824 | 447 jint t1_support = right_n_bits(1 + log2_intptr(t1->_hi)); |
0 | 448 if ((t1_support & con) == t1_support) |
824 | 449 return in1; |
0 | 450 } |
451 // Masking off the high bits of a unsigned-shift-right is not | |
452 // needed either. | |
824 | 453 if (op == Op_URShiftI) { |
454 const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); | |
455 if (t12 && t12->is_con()) { // Shift is by a constant | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
456 int shift = t12->get_con(); |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
457 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
458 int mask = max_juint >> shift; |
824 | 459 if ((mask & con) == mask) // If AND is useless, skip it |
460 return in1; | |
0 | 461 } |
462 } | |
463 } | |
464 return MulNode::Identity(phase); | |
465 } | |
466 | |
467 //------------------------------Ideal------------------------------------------ | |
468 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
469 // Special case constant AND mask | |
470 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
471 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); | |
472 const int mask = t2->get_con(); | |
473 Node *load = in(1); | |
474 uint lop = load->Opcode(); | |
475 | |
476 // Masking bits off of a Character? Hi bits are already zero. | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
404
diff
changeset
|
477 if( lop == Op_LoadUS && |
0 | 478 (mask & 0xFFFF0000) ) // Can we make a smaller mask? |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
479 return new (phase->C) AndINode(load,phase->intcon(mask&0xFFFF)); |
0 | 480 |
481 // Masking bits off of a Short? Loading a Character does some masking | |
6891 | 482 if (can_reshape && |
483 load->outcnt() == 1 && load->unique_out() == this) { | |
484 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) { | |
485 Node *ldus = new (phase->C) LoadUSNode(load->in(MemNode::Control), | |
486 load->in(MemNode::Memory), | |
487 load->in(MemNode::Address), | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
488 load->adr_type(), |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
489 TypeInt::CHAR, MemNode::unordered); |
6891 | 490 ldus = phase->transform(ldus); |
491 return new (phase->C) AndINode(ldus, phase->intcon(mask & 0xFFFF)); | |
492 } | |
0 | 493 |
6891 | 494 // Masking sign bits off of a Byte? Do an unsigned byte load plus |
495 // an and. | |
496 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) { | |
497 Node* ldub = new (phase->C) LoadUBNode(load->in(MemNode::Control), | |
498 load->in(MemNode::Memory), | |
499 load->in(MemNode::Address), | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
500 load->adr_type(), |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
501 TypeInt::UBYTE, MemNode::unordered); |
6891 | 502 ldub = phase->transform(ldub); |
503 return new (phase->C) AndINode(ldub, phase->intcon(mask)); | |
504 } | |
0 | 505 } |
506 | |
507 // Masking off sign bits? Dont make them! | |
508 if( lop == Op_RShiftI ) { | |
509 const TypeInt *t12 = phase->type(load->in(2))->isa_int(); | |
510 if( t12 && t12->is_con() ) { // Shift is by a constant | |
511 int shift = t12->get_con(); | |
512 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
513 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift); | |
514 // If the AND'ing of the 2 masks has no bits, then only original shifted | |
515 // bits survive. NO sign-extension bits survive the maskings. | |
516 if( (sign_bits_mask & mask) == 0 ) { | |
517 // Use zero-fill shift instead | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
518 Node *zshift = phase->transform(new (phase->C) URShiftINode(load->in(1),load->in(2))); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
519 return new (phase->C) AndINode( zshift, in(2) ); |
0 | 520 } |
521 } | |
522 } | |
523 | |
524 // Check for 'negate/and-1', a pattern emitted when someone asks for | |
525 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement | |
526 // plus 1) and the mask is of the low order bit. Skip the negate. | |
527 if( lop == Op_SubI && mask == 1 && load->in(1) && | |
528 phase->type(load->in(1)) == TypeInt::ZERO ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
529 return new (phase->C) AndINode( load->in(2), in(2) ); |
0 | 530 |
531 return MulNode::Ideal(phase, can_reshape); | |
532 } | |
533 | |
534 //============================================================================= | |
535 //------------------------------mul_ring--------------------------------------- | |
536 // Supplied function returns the product of the inputs IN THE CURRENT RING. | |
537 // For the logical operations the ring's MUL is really a logical AND function. | |
538 // This also type-checks the inputs for sanity. Guaranteed never to | |
539 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. | |
540 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const { | |
541 const TypeLong *r0 = t0->is_long(); // Handy access | |
542 const TypeLong *r1 = t1->is_long(); | |
543 int widen = MAX2(r0->_widen,r1->_widen); | |
544 | |
545 // If either input is a constant, might be able to trim cases | |
546 if( !r0->is_con() && !r1->is_con() ) | |
547 return TypeLong::LONG; // No constants to be had | |
548 | |
549 // Both constants? Return bits | |
550 if( r0->is_con() && r1->is_con() ) | |
551 return TypeLong::make( r0->get_con() & r1->get_con() ); | |
552 | |
553 if( r0->is_con() && r0->get_con() > 0 ) | |
554 return TypeLong::make(CONST64(0), r0->get_con(), widen); | |
555 | |
556 if( r1->is_con() && r1->get_con() > 0 ) | |
557 return TypeLong::make(CONST64(0), r1->get_con(), widen); | |
558 | |
559 return TypeLong::LONG; // No constants to be had | |
560 } | |
561 | |
562 //------------------------------Identity--------------------------------------- | |
563 // Masking off the high bits of an unsigned load is not required | |
564 Node *AndLNode::Identity( PhaseTransform *phase ) { | |
565 | |
566 // x & x => x | |
567 if (phase->eqv(in(1), in(2))) return in(1); | |
568 | |
569 Node *usr = in(1); | |
570 const TypeLong *t2 = phase->type( in(2) )->isa_long(); | |
571 if( t2 && t2->is_con() ) { | |
572 jlong con = t2->get_con(); | |
573 // Masking off high bits which are always zero is useless. | |
574 const TypeLong* t1 = phase->type( in(1) )->isa_long(); | |
575 if (t1 != NULL && t1->_lo >= 0) { | |
576 jlong t1_support = ((jlong)1 << (1 + log2_long(t1->_hi))) - 1; | |
577 if ((t1_support & con) == t1_support) | |
578 return usr; | |
579 } | |
580 uint lop = usr->Opcode(); | |
581 // Masking off the high bits of a unsigned-shift-right is not | |
582 // needed either. | |
583 if( lop == Op_URShiftL ) { | |
584 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int(); | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
585 if( t12 && t12->is_con() ) { // Shift is by a constant |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
586 int shift = t12->get_con(); |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
587 shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
588 jlong mask = max_julong >> shift; |
0 | 589 if( (mask&con) == mask ) // If AND is useless, skip it |
590 return usr; | |
591 } | |
592 } | |
593 } | |
594 return MulNode::Identity(phase); | |
595 } | |
596 | |
597 //------------------------------Ideal------------------------------------------ | |
598 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
599 // Special case constant AND mask | |
600 const TypeLong *t2 = phase->type( in(2) )->isa_long(); | |
601 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); | |
602 const jlong mask = t2->get_con(); | |
603 | |
624 | 604 Node* in1 = in(1); |
605 uint op = in1->Opcode(); | |
606 | |
824 | 607 // Are we masking a long that was converted from an int with a mask |
897
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
608 // that fits in 32-bits? Commute them and use an AndINode. Don't |
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
609 // convert masks which would cause a sign extension of the integer |
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
610 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which |
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
611 // would be optimized away later in Identity. |
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
612 if (op == Op_ConvI2L && (mask & CONST64(0xFFFFFFFF80000000)) == 0) { |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
613 Node* andi = new (phase->C) AndINode(in1->in(1), phase->intcon(mask)); |
897
52898b0c43e9
6863155: Server compiler generates incorrect code (x86, long, bitshift, bitmask)
twisti
parents:
824
diff
changeset
|
614 andi = phase->transform(andi); |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
615 return new (phase->C) ConvI2LNode(andi); |
824 | 616 } |
617 | |
0 | 618 // Masking off sign bits? Dont make them! |
624 | 619 if (op == Op_RShiftL) { |
824 | 620 const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); |
0 | 621 if( t12 && t12->is_con() ) { // Shift is by a constant |
622 int shift = t12->get_con(); | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
623 shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
624 const jlong sign_bits_mask = ~(((jlong)CONST64(1) << (jlong)(BitsPerJavaLong - shift)) -1); |
0 | 625 // If the AND'ing of the 2 masks has no bits, then only original shifted |
626 // bits survive. NO sign-extension bits survive the maskings. | |
627 if( (sign_bits_mask & mask) == 0 ) { | |
628 // Use zero-fill shift instead | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
629 Node *zshift = phase->transform(new (phase->C) URShiftLNode(in1->in(1), in1->in(2))); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
630 return new (phase->C) AndLNode(zshift, in(2)); |
0 | 631 } |
632 } | |
633 } | |
634 | |
635 return MulNode::Ideal(phase, can_reshape); | |
636 } | |
637 | |
638 //============================================================================= | |
639 //------------------------------Identity--------------------------------------- | |
640 Node *LShiftINode::Identity( PhaseTransform *phase ) { | |
641 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
642 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) ? in(1) : this; | |
643 } | |
644 | |
645 //------------------------------Ideal------------------------------------------ | |
646 // If the right input is a constant, and the left input is an add of a | |
647 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 | |
648 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
649 const Type *t = phase->type( in(2) ); | |
650 if( t == Type::TOP ) return NULL; // Right input is dead | |
651 const TypeInt *t2 = t->isa_int(); | |
652 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
653 const int con = t2->get_con() & ( BitsPerInt - 1 ); // masked shift count | |
654 | |
655 if ( con == 0 ) return NULL; // let Identity() handle 0 shift count | |
656 | |
657 // Left input is an add of a constant? | |
658 Node *add1 = in(1); | |
659 int add1_op = add1->Opcode(); | |
660 if( add1_op == Op_AddI ) { // Left input is an add? | |
661 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" ); | |
662 const TypeInt *t12 = phase->type(add1->in(2))->isa_int(); | |
663 if( t12 && t12->is_con() ){ // Left input is an add of a con? | |
664 // Transform is legal, but check for profit. Avoid breaking 'i2s' | |
665 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'. | |
666 if( con < 16 ) { | |
667 // Compute X << con0 | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
668 Node *lsh = phase->transform( new (phase->C) LShiftINode( add1->in(1), in(2) ) ); |
0 | 669 // Compute X<<con0 + (con1<<con0) |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
670 return new (phase->C) AddINode( lsh, phase->intcon(t12->get_con() << con)); |
0 | 671 } |
672 } | |
673 } | |
674 | |
675 // Check for "(x>>c0)<<c0" which just masks off low bits | |
676 if( (add1_op == Op_RShiftI || add1_op == Op_URShiftI ) && | |
677 add1->in(2) == in(2) ) | |
678 // Convert to "(x & -(1<<c0))" | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
679 return new (phase->C) AndINode(add1->in(1),phase->intcon( -(1<<con))); |
0 | 680 |
681 // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits | |
682 if( add1_op == Op_AndI ) { | |
683 Node *add2 = add1->in(1); | |
684 int add2_op = add2->Opcode(); | |
685 if( (add2_op == Op_RShiftI || add2_op == Op_URShiftI ) && | |
686 add2->in(2) == in(2) ) { | |
687 // Convert to "(x & (Y<<c0))" | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
688 Node *y_sh = phase->transform( new (phase->C) LShiftINode( add1->in(2), in(2) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
689 return new (phase->C) AndINode( add2->in(1), y_sh ); |
0 | 690 } |
691 } | |
692 | |
693 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits | |
694 // before shifting them away. | |
695 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con); | |
696 if( add1_op == Op_AndI && | |
697 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
698 return new (phase->C) LShiftINode( add1->in(1), in(2) ); |
0 | 699 |
700 return NULL; | |
701 } | |
702 | |
703 //------------------------------Value------------------------------------------ | |
704 // A LShiftINode shifts its input2 left by input1 amount. | |
705 const Type *LShiftINode::Value( PhaseTransform *phase ) const { | |
706 const Type *t1 = phase->type( in(1) ); | |
707 const Type *t2 = phase->type( in(2) ); | |
708 // Either input is TOP ==> the result is TOP | |
709 if( t1 == Type::TOP ) return Type::TOP; | |
710 if( t2 == Type::TOP ) return Type::TOP; | |
711 | |
712 // Left input is ZERO ==> the result is ZERO. | |
713 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
714 // Shift by zero does nothing | |
715 if( t2 == TypeInt::ZERO ) return t1; | |
716 | |
717 // Either input is BOTTOM ==> the result is BOTTOM | |
718 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) || | |
719 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
720 return TypeInt::INT; | |
721 | |
722 const TypeInt *r1 = t1->is_int(); // Handy access | |
723 const TypeInt *r2 = t2->is_int(); // Handy access | |
724 | |
725 if (!r2->is_con()) | |
726 return TypeInt::INT; | |
727 | |
728 uint shift = r2->get_con(); | |
729 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
730 // Shift by a multiple of 32 does nothing: | |
731 if (shift == 0) return t1; | |
732 | |
733 // If the shift is a constant, shift the bounds of the type, | |
734 // unless this could lead to an overflow. | |
735 if (!r1->is_con()) { | |
736 jint lo = r1->_lo, hi = r1->_hi; | |
737 if (((lo << shift) >> shift) == lo && | |
738 ((hi << shift) >> shift) == hi) { | |
739 // No overflow. The range shifts up cleanly. | |
740 return TypeInt::make((jint)lo << (jint)shift, | |
741 (jint)hi << (jint)shift, | |
742 MAX2(r1->_widen,r2->_widen)); | |
743 } | |
744 return TypeInt::INT; | |
745 } | |
746 | |
747 return TypeInt::make( (jint)r1->get_con() << (jint)shift ); | |
748 } | |
749 | |
750 //============================================================================= | |
751 //------------------------------Identity--------------------------------------- | |
752 Node *LShiftLNode::Identity( PhaseTransform *phase ) { | |
753 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
754 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
755 } | |
756 | |
757 //------------------------------Ideal------------------------------------------ | |
758 // If the right input is a constant, and the left input is an add of a | |
759 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 | |
760 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
761 const Type *t = phase->type( in(2) ); | |
762 if( t == Type::TOP ) return NULL; // Right input is dead | |
763 const TypeInt *t2 = t->isa_int(); | |
764 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
765 const int con = t2->get_con() & ( BitsPerLong - 1 ); // masked shift count | |
766 | |
767 if ( con == 0 ) return NULL; // let Identity() handle 0 shift count | |
768 | |
769 // Left input is an add of a constant? | |
770 Node *add1 = in(1); | |
771 int add1_op = add1->Opcode(); | |
772 if( add1_op == Op_AddL ) { // Left input is an add? | |
773 // Avoid dead data cycles from dead loops | |
774 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" ); | |
775 const TypeLong *t12 = phase->type(add1->in(2))->isa_long(); | |
776 if( t12 && t12->is_con() ){ // Left input is an add of a con? | |
777 // Compute X << con0 | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
778 Node *lsh = phase->transform( new (phase->C) LShiftLNode( add1->in(1), in(2) ) ); |
0 | 779 // Compute X<<con0 + (con1<<con0) |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
780 return new (phase->C) AddLNode( lsh, phase->longcon(t12->get_con() << con)); |
0 | 781 } |
782 } | |
783 | |
784 // Check for "(x>>c0)<<c0" which just masks off low bits | |
785 if( (add1_op == Op_RShiftL || add1_op == Op_URShiftL ) && | |
786 add1->in(2) == in(2) ) | |
787 // Convert to "(x & -(1<<c0))" | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
788 return new (phase->C) AndLNode(add1->in(1),phase->longcon( -(CONST64(1)<<con))); |
0 | 789 |
790 // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits | |
791 if( add1_op == Op_AndL ) { | |
792 Node *add2 = add1->in(1); | |
793 int add2_op = add2->Opcode(); | |
794 if( (add2_op == Op_RShiftL || add2_op == Op_URShiftL ) && | |
795 add2->in(2) == in(2) ) { | |
796 // Convert to "(x & (Y<<c0))" | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
797 Node *y_sh = phase->transform( new (phase->C) LShiftLNode( add1->in(2), in(2) ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
798 return new (phase->C) AndLNode( add2->in(1), y_sh ); |
0 | 799 } |
800 } | |
801 | |
802 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits | |
803 // before shifting them away. | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
804 const jlong bits_mask = ((jlong)CONST64(1) << (jlong)(BitsPerJavaLong - con)) - CONST64(1); |
0 | 805 if( add1_op == Op_AndL && |
806 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
807 return new (phase->C) LShiftLNode( add1->in(1), in(2) ); |
0 | 808 |
809 return NULL; | |
810 } | |
811 | |
812 //------------------------------Value------------------------------------------ | |
813 // A LShiftLNode shifts its input2 left by input1 amount. | |
814 const Type *LShiftLNode::Value( PhaseTransform *phase ) const { | |
815 const Type *t1 = phase->type( in(1) ); | |
816 const Type *t2 = phase->type( in(2) ); | |
817 // Either input is TOP ==> the result is TOP | |
818 if( t1 == Type::TOP ) return Type::TOP; | |
819 if( t2 == Type::TOP ) return Type::TOP; | |
820 | |
821 // Left input is ZERO ==> the result is ZERO. | |
822 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
823 // Shift by zero does nothing | |
824 if( t2 == TypeInt::ZERO ) return t1; | |
825 | |
826 // Either input is BOTTOM ==> the result is BOTTOM | |
827 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) || | |
828 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
829 return TypeLong::LONG; | |
830 | |
831 const TypeLong *r1 = t1->is_long(); // Handy access | |
832 const TypeInt *r2 = t2->is_int(); // Handy access | |
833 | |
834 if (!r2->is_con()) | |
835 return TypeLong::LONG; | |
836 | |
837 uint shift = r2->get_con(); | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
838 shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
0 | 839 // Shift by a multiple of 64 does nothing: |
840 if (shift == 0) return t1; | |
841 | |
842 // If the shift is a constant, shift the bounds of the type, | |
843 // unless this could lead to an overflow. | |
844 if (!r1->is_con()) { | |
845 jlong lo = r1->_lo, hi = r1->_hi; | |
846 if (((lo << shift) >> shift) == lo && | |
847 ((hi << shift) >> shift) == hi) { | |
848 // No overflow. The range shifts up cleanly. | |
849 return TypeLong::make((jlong)lo << (jint)shift, | |
850 (jlong)hi << (jint)shift, | |
851 MAX2(r1->_widen,r2->_widen)); | |
852 } | |
853 return TypeLong::LONG; | |
854 } | |
855 | |
856 return TypeLong::make( (jlong)r1->get_con() << (jint)shift ); | |
857 } | |
858 | |
859 //============================================================================= | |
860 //------------------------------Identity--------------------------------------- | |
861 Node *RShiftINode::Identity( PhaseTransform *phase ) { | |
862 const TypeInt *t2 = phase->type(in(2))->isa_int(); | |
863 if( !t2 ) return this; | |
864 if ( t2->is_con() && ( t2->get_con() & ( BitsPerInt - 1 ) ) == 0 ) | |
865 return in(1); | |
866 | |
867 // Check for useless sign-masking | |
868 if( in(1)->Opcode() == Op_LShiftI && | |
869 in(1)->req() == 3 && | |
870 in(1)->in(2) == in(2) && | |
871 t2->is_con() ) { | |
872 uint shift = t2->get_con(); | |
873 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
874 // Compute masks for which this shifting doesn't change | |
875 int lo = (-1 << (BitsPerJavaInteger - shift-1)); // FFFF8000 | |
876 int hi = ~lo; // 00007FFF | |
877 const TypeInt *t11 = phase->type(in(1)->in(1))->isa_int(); | |
878 if( !t11 ) return this; | |
879 // Does actual value fit inside of mask? | |
880 if( lo <= t11->_lo && t11->_hi <= hi ) | |
881 return in(1)->in(1); // Then shifting is a nop | |
882 } | |
883 | |
884 return this; | |
885 } | |
886 | |
887 //------------------------------Ideal------------------------------------------ | |
888 Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
889 // Inputs may be TOP if they are dead. | |
890 const TypeInt *t1 = phase->type( in(1) )->isa_int(); | |
891 if( !t1 ) return NULL; // Left input is an integer | |
892 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
893 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
894 const TypeInt *t3; // type of in(1).in(2) | |
895 int shift = t2->get_con(); | |
896 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
897 | |
898 if ( shift == 0 ) return NULL; // let Identity() handle 0 shift count | |
899 | |
900 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller. | |
901 // Such expressions arise normally from shift chains like (byte)(x >> 24). | |
902 const Node *mask = in(1); | |
903 if( mask->Opcode() == Op_AndI && | |
904 (t3 = phase->type(mask->in(2))->isa_int()) && | |
905 t3->is_con() ) { | |
906 Node *x = mask->in(1); | |
907 jint maskbits = t3->get_con(); | |
908 // Convert to "(x >> shift) & (mask >> shift)" | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
909 Node *shr_nomask = phase->transform( new (phase->C) RShiftINode(mask->in(1), in(2)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
910 return new (phase->C) AndINode(shr_nomask, phase->intcon( maskbits >> shift)); |
0 | 911 } |
912 | |
913 // Check for "(short[i] <<16)>>16" which simply sign-extends | |
914 const Node *shl = in(1); | |
915 if( shl->Opcode() != Op_LShiftI ) return NULL; | |
916 | |
917 if( shift == 16 && | |
918 (t3 = phase->type(shl->in(2))->isa_int()) && | |
919 t3->is_con(16) ) { | |
920 Node *ld = shl->in(1); | |
921 if( ld->Opcode() == Op_LoadS ) { | |
922 // Sign extension is just useless here. Return a RShiftI of zero instead | |
923 // returning 'ld' directly. We cannot return an old Node directly as | |
924 // that is the job of 'Identity' calls and Identity calls only work on | |
925 // direct inputs ('ld' is an extra Node removed from 'this'). The | |
926 // combined optimization requires Identity only return direct inputs. | |
927 set_req(1, ld); | |
928 set_req(2, phase->intcon(0)); | |
929 return this; | |
930 } | |
6891 | 931 else if( can_reshape && |
932 ld->Opcode() == Op_LoadUS && | |
933 ld->outcnt() == 1 && ld->unique_out() == shl) | |
0 | 934 // Replace zero-extension-load with sign-extension-load |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
935 return new (phase->C) LoadSNode( ld->in(MemNode::Control), |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
936 ld->in(MemNode::Memory), |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
937 ld->in(MemNode::Address), |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
938 ld->adr_type(), TypeInt::SHORT, |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
6891
diff
changeset
|
939 MemNode::unordered); |
0 | 940 } |
941 | |
942 // Check for "(byte[i] <<24)>>24" which simply sign-extends | |
943 if( shift == 24 && | |
944 (t3 = phase->type(shl->in(2))->isa_int()) && | |
945 t3->is_con(24) ) { | |
946 Node *ld = shl->in(1); | |
947 if( ld->Opcode() == Op_LoadB ) { | |
948 // Sign extension is just useless here | |
949 set_req(1, ld); | |
950 set_req(2, phase->intcon(0)); | |
951 return this; | |
952 } | |
953 } | |
954 | |
955 return NULL; | |
956 } | |
957 | |
958 //------------------------------Value------------------------------------------ | |
959 // A RShiftINode shifts its input2 right by input1 amount. | |
960 const Type *RShiftINode::Value( PhaseTransform *phase ) const { | |
961 const Type *t1 = phase->type( in(1) ); | |
962 const Type *t2 = phase->type( in(2) ); | |
963 // Either input is TOP ==> the result is TOP | |
964 if( t1 == Type::TOP ) return Type::TOP; | |
965 if( t2 == Type::TOP ) return Type::TOP; | |
966 | |
967 // Left input is ZERO ==> the result is ZERO. | |
968 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
969 // Shift by zero does nothing | |
970 if( t2 == TypeInt::ZERO ) return t1; | |
971 | |
972 // Either input is BOTTOM ==> the result is BOTTOM | |
973 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
974 return TypeInt::INT; | |
975 | |
976 if (t2 == TypeInt::INT) | |
977 return TypeInt::INT; | |
978 | |
979 const TypeInt *r1 = t1->is_int(); // Handy access | |
980 const TypeInt *r2 = t2->is_int(); // Handy access | |
981 | |
982 // If the shift is a constant, just shift the bounds of the type. | |
983 // For example, if the shift is 31, we just propagate sign bits. | |
984 if (r2->is_con()) { | |
985 uint shift = r2->get_con(); | |
986 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
987 // Shift by a multiple of 32 does nothing: | |
988 if (shift == 0) return t1; | |
989 // Calculate reasonably aggressive bounds for the result. | |
990 // This is necessary if we are to correctly type things | |
991 // like (x<<24>>24) == ((byte)x). | |
992 jint lo = (jint)r1->_lo >> (jint)shift; | |
993 jint hi = (jint)r1->_hi >> (jint)shift; | |
994 assert(lo <= hi, "must have valid bounds"); | |
995 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
996 #ifdef ASSERT | |
997 // Make sure we get the sign-capture idiom correct. | |
998 if (shift == BitsPerJavaInteger-1) { | |
999 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0"); | |
1000 if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1"); | |
1001 } | |
1002 #endif | |
1003 return ti; | |
1004 } | |
1005 | |
1006 if( !r1->is_con() || !r2->is_con() ) | |
1007 return TypeInt::INT; | |
1008 | |
1009 // Signed shift right | |
1010 return TypeInt::make( r1->get_con() >> (r2->get_con()&31) ); | |
1011 } | |
1012 | |
1013 //============================================================================= | |
1014 //------------------------------Identity--------------------------------------- | |
1015 Node *RShiftLNode::Identity( PhaseTransform *phase ) { | |
1016 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
1017 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
1018 } | |
1019 | |
1020 //------------------------------Value------------------------------------------ | |
1021 // A RShiftLNode shifts its input2 right by input1 amount. | |
1022 const Type *RShiftLNode::Value( PhaseTransform *phase ) const { | |
1023 const Type *t1 = phase->type( in(1) ); | |
1024 const Type *t2 = phase->type( in(2) ); | |
1025 // Either input is TOP ==> the result is TOP | |
1026 if( t1 == Type::TOP ) return Type::TOP; | |
1027 if( t2 == Type::TOP ) return Type::TOP; | |
1028 | |
1029 // Left input is ZERO ==> the result is ZERO. | |
1030 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1031 // Shift by zero does nothing | |
1032 if( t2 == TypeInt::ZERO ) return t1; | |
1033 | |
1034 // Either input is BOTTOM ==> the result is BOTTOM | |
1035 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1036 return TypeLong::LONG; | |
1037 | |
1038 if (t2 == TypeInt::INT) | |
1039 return TypeLong::LONG; | |
1040 | |
1041 const TypeLong *r1 = t1->is_long(); // Handy access | |
1042 const TypeInt *r2 = t2->is_int (); // Handy access | |
1043 | |
1044 // If the shift is a constant, just shift the bounds of the type. | |
1045 // For example, if the shift is 63, we just propagate sign bits. | |
1046 if (r2->is_con()) { | |
1047 uint shift = r2->get_con(); | |
1048 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts | |
1049 // Shift by a multiple of 64 does nothing: | |
1050 if (shift == 0) return t1; | |
1051 // Calculate reasonably aggressive bounds for the result. | |
1052 // This is necessary if we are to correctly type things | |
1053 // like (x<<24>>24) == ((byte)x). | |
1054 jlong lo = (jlong)r1->_lo >> (jlong)shift; | |
1055 jlong hi = (jlong)r1->_hi >> (jlong)shift; | |
1056 assert(lo <= hi, "must have valid bounds"); | |
1057 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1058 #ifdef ASSERT | |
1059 // Make sure we get the sign-capture idiom correct. | |
1060 if (shift == (2*BitsPerJavaInteger)-1) { | |
1061 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0"); | |
1062 if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1"); | |
1063 } | |
1064 #endif | |
1065 return tl; | |
1066 } | |
1067 | |
1068 return TypeLong::LONG; // Give up | |
1069 } | |
1070 | |
1071 //============================================================================= | |
1072 //------------------------------Identity--------------------------------------- | |
1073 Node *URShiftINode::Identity( PhaseTransform *phase ) { | |
1074 const TypeInt *ti = phase->type( in(2) )->isa_int(); | |
1075 if ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) return in(1); | |
1076 | |
1077 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x". | |
1078 // Happens during new-array length computation. | |
1079 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)] | |
1080 Node *add = in(1); | |
1081 if( add->Opcode() == Op_AddI ) { | |
1082 const TypeInt *t2 = phase->type(add->in(2))->isa_int(); | |
1083 if( t2 && t2->is_con(wordSize - 1) && | |
1084 add->in(1)->Opcode() == Op_LShiftI ) { | |
1085 // Check that shift_counts are LogBytesPerWord | |
1086 Node *lshift_count = add->in(1)->in(2); | |
1087 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int(); | |
1088 if( t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) && | |
1089 t_lshift_count == phase->type(in(2)) ) { | |
1090 Node *x = add->in(1)->in(1); | |
1091 const TypeInt *t_x = phase->type(x)->isa_int(); | |
1092 if( t_x != NULL && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord) ) { | |
1093 return x; | |
1094 } | |
1095 } | |
1096 } | |
1097 } | |
1098 | |
1099 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this; | |
1100 } | |
1101 | |
1102 //------------------------------Ideal------------------------------------------ | |
1103 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1104 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
1105 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
1106 const int con = t2->get_con() & 31; // Shift count is always masked | |
1107 if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count | |
1108 // We'll be wanting the right-shift amount as a mask of that many bits | |
1109 const int mask = right_n_bits(BitsPerJavaInteger - con); | |
1110 | |
1111 int in1_op = in(1)->Opcode(); | |
1112 | |
1113 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32 | |
1114 if( in1_op == Op_URShiftI ) { | |
1115 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int(); | |
1116 if( t12 && t12->is_con() ) { // Right input is a constant | |
1117 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" ); | |
1118 const int con2 = t12->get_con() & 31; // Shift count is always masked | |
1119 const int con3 = con+con2; | |
1120 if( con3 < 32 ) // Only merge shifts if total is < 32 | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1121 return new (phase->C) URShiftINode( in(1)->in(1), phase->intcon(con3) ); |
0 | 1122 } |
1123 } | |
1124 | |
1125 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z | |
1126 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". | |
1127 // If Q is "X << z" the rounding is useless. Look for patterns like | |
1128 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. | |
1129 Node *add = in(1); | |
1130 if( in1_op == Op_AddI ) { | |
1131 Node *lshl = add->in(1); | |
1132 if( lshl->Opcode() == Op_LShiftI && | |
1133 phase->type(lshl->in(2)) == t2 ) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1134 Node *y_z = phase->transform( new (phase->C) URShiftINode(add->in(2),in(2)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1135 Node *sum = phase->transform( new (phase->C) AddINode( lshl->in(1), y_z ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1136 return new (phase->C) AndINode( sum, phase->intcon(mask) ); |
0 | 1137 } |
1138 } | |
1139 | |
1140 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) | |
1141 // This shortens the mask. Also, if we are extracting a high byte and | |
1142 // storing it to a buffer, the mask will be removed completely. | |
1143 Node *andi = in(1); | |
1144 if( in1_op == Op_AndI ) { | |
1145 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int(); | |
1146 if( t3 && t3->is_con() ) { // Right input is a constant | |
1147 jint mask2 = t3->get_con(); | |
1148 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1149 Node *newshr = phase->transform( new (phase->C) URShiftINode(andi->in(1), in(2)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1150 return new (phase->C) AndINode(newshr, phase->intcon(mask2)); |
0 | 1151 // The negative values are easier to materialize than positive ones. |
1152 // A typical case from address arithmetic is ((x & ~15) >> 4). | |
1153 // It's better to change that to ((x >> 4) & ~0) versus | |
1154 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64. | |
1155 } | |
1156 } | |
1157 | |
1158 // Check for "(X << z ) >>> z" which simply zero-extends | |
1159 Node *shl = in(1); | |
1160 if( in1_op == Op_LShiftI && | |
1161 phase->type(shl->in(2)) == t2 ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1162 return new (phase->C) AndINode( shl->in(1), phase->intcon(mask) ); |
0 | 1163 |
1164 return NULL; | |
1165 } | |
1166 | |
1167 //------------------------------Value------------------------------------------ | |
1168 // A URShiftINode shifts its input2 right by input1 amount. | |
1169 const Type *URShiftINode::Value( PhaseTransform *phase ) const { | |
1170 // (This is a near clone of RShiftINode::Value.) | |
1171 const Type *t1 = phase->type( in(1) ); | |
1172 const Type *t2 = phase->type( in(2) ); | |
1173 // Either input is TOP ==> the result is TOP | |
1174 if( t1 == Type::TOP ) return Type::TOP; | |
1175 if( t2 == Type::TOP ) return Type::TOP; | |
1176 | |
1177 // Left input is ZERO ==> the result is ZERO. | |
1178 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
1179 // Shift by zero does nothing | |
1180 if( t2 == TypeInt::ZERO ) return t1; | |
1181 | |
1182 // Either input is BOTTOM ==> the result is BOTTOM | |
1183 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1184 return TypeInt::INT; | |
1185 | |
1186 if (t2 == TypeInt::INT) | |
1187 return TypeInt::INT; | |
1188 | |
1189 const TypeInt *r1 = t1->is_int(); // Handy access | |
1190 const TypeInt *r2 = t2->is_int(); // Handy access | |
1191 | |
1192 if (r2->is_con()) { | |
1193 uint shift = r2->get_con(); | |
1194 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
1195 // Shift by a multiple of 32 does nothing: | |
1196 if (shift == 0) return t1; | |
1197 // Calculate reasonably aggressive bounds for the result. | |
1198 jint lo = (juint)r1->_lo >> (juint)shift; | |
1199 jint hi = (juint)r1->_hi >> (juint)shift; | |
1200 if (r1->_hi >= 0 && r1->_lo < 0) { | |
1201 // If the type has both negative and positive values, | |
1202 // there are two separate sub-domains to worry about: | |
1203 // The positive half and the negative half. | |
1204 jint neg_lo = lo; | |
1205 jint neg_hi = (juint)-1 >> (juint)shift; | |
1206 jint pos_lo = (juint) 0 >> (juint)shift; | |
1207 jint pos_hi = hi; | |
1208 lo = MIN2(neg_lo, pos_lo); // == 0 | |
1209 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; | |
1210 } | |
1211 assert(lo <= hi, "must have valid bounds"); | |
1212 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1213 #ifdef ASSERT | |
1214 // Make sure we get the sign-capture idiom correct. | |
1215 if (shift == BitsPerJavaInteger-1) { | |
1216 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0"); | |
1217 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1"); | |
1218 } | |
1219 #endif | |
1220 return ti; | |
1221 } | |
1222 | |
1223 // | |
1224 // Do not support shifted oops in info for GC | |
1225 // | |
1226 // else if( t1->base() == Type::InstPtr ) { | |
1227 // | |
1228 // const TypeInstPtr *o = t1->is_instptr(); | |
1229 // if( t1->singleton() ) | |
1230 // return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift ); | |
1231 // } | |
1232 // else if( t1->base() == Type::KlassPtr ) { | |
1233 // const TypeKlassPtr *o = t1->is_klassptr(); | |
1234 // if( t1->singleton() ) | |
1235 // return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift ); | |
1236 // } | |
1237 | |
1238 return TypeInt::INT; | |
1239 } | |
1240 | |
1241 //============================================================================= | |
1242 //------------------------------Identity--------------------------------------- | |
1243 Node *URShiftLNode::Identity( PhaseTransform *phase ) { | |
1244 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
1245 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
1246 } | |
1247 | |
1248 //------------------------------Ideal------------------------------------------ | |
1249 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1250 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
1251 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
1252 const int con = t2->get_con() & ( BitsPerLong - 1 ); // Shift count is always masked | |
1253 if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count | |
1254 // note: mask computation below does not work for 0 shift count | |
1255 // We'll be wanting the right-shift amount as a mask of that many bits | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
1256 const jlong mask = (((jlong)CONST64(1) << (jlong)(BitsPerJavaLong - con)) -1); |
0 | 1257 |
1258 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z | |
1259 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". | |
1260 // If Q is "X << z" the rounding is useless. Look for patterns like | |
1261 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. | |
1262 Node *add = in(1); | |
1263 if( add->Opcode() == Op_AddL ) { | |
1264 Node *lshl = add->in(1); | |
1265 if( lshl->Opcode() == Op_LShiftL && | |
1266 phase->type(lshl->in(2)) == t2 ) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1267 Node *y_z = phase->transform( new (phase->C) URShiftLNode(add->in(2),in(2)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1268 Node *sum = phase->transform( new (phase->C) AddLNode( lshl->in(1), y_z ) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1269 return new (phase->C) AndLNode( sum, phase->longcon(mask) ); |
0 | 1270 } |
1271 } | |
1272 | |
1273 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) | |
1274 // This shortens the mask. Also, if we are extracting a high byte and | |
1275 // storing it to a buffer, the mask will be removed completely. | |
1276 Node *andi = in(1); | |
1277 if( andi->Opcode() == Op_AndL ) { | |
1278 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long(); | |
1279 if( t3 && t3->is_con() ) { // Right input is a constant | |
1280 jlong mask2 = t3->get_con(); | |
1281 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1282 Node *newshr = phase->transform( new (phase->C) URShiftLNode(andi->in(1), in(2)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1283 return new (phase->C) AndLNode(newshr, phase->longcon(mask2)); |
0 | 1284 } |
1285 } | |
1286 | |
1287 // Check for "(X << z ) >>> z" which simply zero-extends | |
1288 Node *shl = in(1); | |
1289 if( shl->Opcode() == Op_LShiftL && | |
1290 phase->type(shl->in(2)) == t2 ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
3842
diff
changeset
|
1291 return new (phase->C) AndLNode( shl->in(1), phase->longcon(mask) ); |
0 | 1292 |
1293 return NULL; | |
1294 } | |
1295 | |
1296 //------------------------------Value------------------------------------------ | |
1297 // A URShiftINode shifts its input2 right by input1 amount. | |
1298 const Type *URShiftLNode::Value( PhaseTransform *phase ) const { | |
1299 // (This is a near clone of RShiftLNode::Value.) | |
1300 const Type *t1 = phase->type( in(1) ); | |
1301 const Type *t2 = phase->type( in(2) ); | |
1302 // Either input is TOP ==> the result is TOP | |
1303 if( t1 == Type::TOP ) return Type::TOP; | |
1304 if( t2 == Type::TOP ) return Type::TOP; | |
1305 | |
1306 // Left input is ZERO ==> the result is ZERO. | |
1307 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1308 // Shift by zero does nothing | |
1309 if( t2 == TypeInt::ZERO ) return t1; | |
1310 | |
1311 // Either input is BOTTOM ==> the result is BOTTOM | |
1312 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1313 return TypeLong::LONG; | |
1314 | |
1315 if (t2 == TypeInt::INT) | |
1316 return TypeLong::LONG; | |
1317 | |
1318 const TypeLong *r1 = t1->is_long(); // Handy access | |
1319 const TypeInt *r2 = t2->is_int (); // Handy access | |
1320 | |
1321 if (r2->is_con()) { | |
1322 uint shift = r2->get_con(); | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
1323 shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
0 | 1324 // Shift by a multiple of 64 does nothing: |
1325 if (shift == 0) return t1; | |
1326 // Calculate reasonably aggressive bounds for the result. | |
1327 jlong lo = (julong)r1->_lo >> (juint)shift; | |
1328 jlong hi = (julong)r1->_hi >> (juint)shift; | |
1329 if (r1->_hi >= 0 && r1->_lo < 0) { | |
1330 // If the type has both negative and positive values, | |
1331 // there are two separate sub-domains to worry about: | |
1332 // The positive half and the negative half. | |
1333 jlong neg_lo = lo; | |
1334 jlong neg_hi = (julong)-1 >> (juint)shift; | |
1335 jlong pos_lo = (julong) 0 >> (juint)shift; | |
1336 jlong pos_hi = hi; | |
1337 //lo = MIN2(neg_lo, pos_lo); // == 0 | |
1338 lo = neg_lo < pos_lo ? neg_lo : pos_lo; | |
1339 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; | |
1340 hi = neg_hi > pos_hi ? neg_hi : pos_hi; | |
1341 } | |
1342 assert(lo <= hi, "must have valid bounds"); | |
1343 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1344 #ifdef ASSERT | |
1345 // Make sure we get the sign-capture idiom correct. | |
559
7628781568e1
6795362: 32bit server compiler leads to wrong results on solaris-x86
twisti
parents:
558
diff
changeset
|
1346 if (shift == BitsPerJavaLong - 1) { |
0 | 1347 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0"); |
1348 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1"); | |
1349 } | |
1350 #endif | |
1351 return tl; | |
1352 } | |
1353 | |
1354 return TypeLong::LONG; // Give up | |
1355 } |