Mercurial > hg > graal-jvmci-8
annotate src/share/vm/opto/mulnode.cpp @ 418:72c5366e5d86
6743900: frequency based block layout
Summary: post-register allocation pass that drives block layout by edge frequencies
Reviewed-by: never, kvn
author | rasbold |
---|---|
date | Thu, 06 Nov 2008 14:59:10 -0800 |
parents | 78c058bc5cdc |
children | 3b5ac9e7e6ea |
rev | line source |
---|---|
0 | 1 /* |
196 | 2 * Copyright 1997-2008 Sun Microsystems, Inc. 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 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 // Portions of code courtesy of Clifford Click | |
26 | |
27 #include "incls/_precompiled.incl" | |
28 #include "incls/_mulnode.cpp.incl" | |
29 | |
30 | |
31 //============================================================================= | |
32 //------------------------------hash------------------------------------------- | |
33 // Hash function over MulNodes. Needs to be commutative; i.e., I swap | |
34 // (commute) inputs to MulNodes willy-nilly so the hash function must return | |
35 // the same value in the presence of edge swapping. | |
36 uint MulNode::hash() const { | |
37 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode(); | |
38 } | |
39 | |
40 //------------------------------Identity--------------------------------------- | |
41 // Multiplying a one preserves the other argument | |
42 Node *MulNode::Identity( PhaseTransform *phase ) { | |
43 register const Type *one = mul_id(); // The multiplicative identity | |
44 if( phase->type( in(1) )->higher_equal( one ) ) return in(2); | |
45 if( phase->type( in(2) )->higher_equal( one ) ) return in(1); | |
46 | |
47 return this; | |
48 } | |
49 | |
50 //------------------------------Ideal------------------------------------------ | |
51 // We also canonicalize the Node, moving constants to the right input, | |
52 // and flatten expressions (so that 1+x+2 becomes x+3). | |
53 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
54 const Type *t1 = phase->type( in(1) ); | |
55 const Type *t2 = phase->type( in(2) ); | |
56 Node *progress = NULL; // Progress flag | |
57 // We are OK if right is a constant, or right is a load and | |
58 // left is a non-constant. | |
59 if( !(t2->singleton() || | |
60 (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) { | |
61 if( t1->singleton() || // Left input is a constant? | |
62 // Otherwise, sort inputs (commutativity) to help value numbering. | |
63 (in(1)->_idx > in(2)->_idx) ) { | |
64 swap_edges(1, 2); | |
65 const Type *t = t1; | |
66 t1 = t2; | |
67 t2 = t; | |
68 progress = this; // Made progress | |
69 } | |
70 } | |
71 | |
72 // If the right input is a constant, and the left input is a product of a | |
73 // constant, flatten the expression tree. | |
74 uint op = Opcode(); | |
75 if( t2->singleton() && // Right input is a constant? | |
76 op != Op_MulF && // Float & double cannot reassociate | |
77 op != Op_MulD ) { | |
78 if( t2 == Type::TOP ) return NULL; | |
79 Node *mul1 = in(1); | |
80 #ifdef ASSERT | |
81 // Check for dead loop | |
82 int op1 = mul1->Opcode(); | |
83 if( phase->eqv( mul1, this ) || phase->eqv( in(2), this ) || | |
84 ( op1 == mul_opcode() || op1 == add_opcode() ) && | |
85 ( phase->eqv( mul1->in(1), this ) || phase->eqv( mul1->in(2), this ) || | |
86 phase->eqv( mul1->in(1), mul1 ) || phase->eqv( mul1->in(2), mul1 ) ) ) | |
87 assert(false, "dead loop in MulNode::Ideal"); | |
88 #endif | |
89 | |
90 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply? | |
91 // Mul of a constant? | |
92 const Type *t12 = phase->type( mul1->in(2) ); | |
93 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant? | |
94 // Compute new constant; check for overflow | |
95 const Type *tcon01 = mul1->as_Mul()->mul_ring(t2,t12); | |
96 if( tcon01->singleton() ) { | |
97 // The Mul of the flattened expression | |
98 set_req(1, mul1->in(1)); | |
99 set_req(2, phase->makecon( tcon01 )); | |
100 t2 = tcon01; | |
101 progress = this; // Made progress | |
102 } | |
103 } | |
104 } | |
105 // If the right input is a constant, and the left input is an add of a | |
106 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0 | |
107 const Node *add1 = in(1); | |
108 if( add1->Opcode() == add_opcode() ) { // Left input is an add? | |
109 // Add of a constant? | |
110 const Type *t12 = phase->type( add1->in(2) ); | |
111 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant? | |
112 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" ); | |
113 // Compute new constant; check for overflow | |
114 const Type *tcon01 = mul_ring(t2,t12); | |
115 if( tcon01->singleton() ) { | |
116 | |
117 // Convert (X+con1)*con0 into X*con0 | |
118 Node *mul = clone(); // mul = ()*con0 | |
119 mul->set_req(1,add1->in(1)); // mul = X*con0 | |
120 mul = phase->transform(mul); | |
121 | |
122 Node *add2 = add1->clone(); | |
123 add2->set_req(1, mul); // X*con0 + con0*con1 | |
124 add2->set_req(2, phase->makecon(tcon01) ); | |
125 progress = add2; | |
126 } | |
127 } | |
128 } // End of is left input an add | |
129 } // End of is right input a Mul | |
130 | |
131 return progress; | |
132 } | |
133 | |
134 //------------------------------Value----------------------------------------- | |
135 const Type *MulNode::Value( PhaseTransform *phase ) const { | |
136 const Type *t1 = phase->type( in(1) ); | |
137 const Type *t2 = phase->type( in(2) ); | |
138 // Either input is TOP ==> the result is TOP | |
139 if( t1 == Type::TOP ) return Type::TOP; | |
140 if( t2 == Type::TOP ) return Type::TOP; | |
141 | |
142 // Either input is ZERO ==> the result is ZERO. | |
143 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0 | |
144 int op = Opcode(); | |
145 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) { | |
146 const Type *zero = add_id(); // The multiplicative zero | |
147 if( t1->higher_equal( zero ) ) return zero; | |
148 if( t2->higher_equal( zero ) ) return zero; | |
149 } | |
150 | |
151 // Either input is BOTTOM ==> the result is the local BOTTOM | |
152 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) | |
153 return bottom_type(); | |
154 | |
404
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
155 #if defined(IA32) |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
156 // 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
|
157 // multiplication with round-to-zero on this platform. |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
158 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
|
159 return TypeD::DOUBLE; |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
160 } |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
161 #endif |
78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
rasbold
parents:
196
diff
changeset
|
162 |
0 | 163 return mul_ring(t1,t2); // Local flavor of type multiplication |
164 } | |
165 | |
166 | |
167 //============================================================================= | |
168 //------------------------------Ideal------------------------------------------ | |
169 // Check for power-of-2 multiply, then try the regular MulNode::Ideal | |
170 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
171 // Swap constant to right | |
172 jint con; | |
173 if ((con = in(1)->find_int_con(0)) != 0) { | |
174 swap_edges(1, 2); | |
175 // Finish rest of method to use info in 'con' | |
176 } else if ((con = in(2)->find_int_con(0)) == 0) { | |
177 return MulNode::Ideal(phase, can_reshape); | |
178 } | |
179 | |
180 // Now we have a constant Node on the right and the constant in con | |
181 if( con == 0 ) return NULL; // By zero is handled by Value call | |
182 if( con == 1 ) return NULL; // By one is handled by Identity call | |
183 | |
184 // Check for negative constant; if so negate the final result | |
185 bool sign_flip = false; | |
186 if( con < 0 ) { | |
187 con = -con; | |
188 sign_flip = true; | |
189 } | |
190 | |
191 // Get low bit; check for being the only bit | |
192 Node *res = NULL; | |
193 jint bit1 = con & -con; // Extract low bit | |
194 if( bit1 == con ) { // Found a power of 2? | |
195 res = new (phase->C, 3) LShiftINode( in(1), phase->intcon(log2_intptr(bit1)) ); | |
196 } else { | |
197 | |
198 // Check for constant with 2 bits set | |
199 jint bit2 = con-bit1; | |
200 bit2 = bit2 & -bit2; // Extract 2nd bit | |
201 if( bit2 + bit1 == con ) { // Found all bits in con? | |
202 Node *n1 = phase->transform( new (phase->C, 3) LShiftINode( in(1), phase->intcon(log2_intptr(bit1)) ) ); | |
203 Node *n2 = phase->transform( new (phase->C, 3) LShiftINode( in(1), phase->intcon(log2_intptr(bit2)) ) ); | |
204 res = new (phase->C, 3) AddINode( n2, n1 ); | |
205 | |
206 } else if (is_power_of_2(con+1)) { | |
207 // Sleezy: power-of-2 -1. Next time be generic. | |
208 jint temp = (jint) (con + 1); | |
209 Node *n1 = phase->transform( new (phase->C, 3) LShiftINode( in(1), phase->intcon(log2_intptr(temp)) ) ); | |
210 res = new (phase->C, 3) SubINode( n1, in(1) ); | |
211 } else { | |
212 return MulNode::Ideal(phase, can_reshape); | |
213 } | |
214 } | |
215 | |
216 if( sign_flip ) { // Need to negate result? | |
217 res = phase->transform(res);// Transform, before making the zero con | |
218 res = new (phase->C, 3) SubINode(phase->intcon(0),res); | |
219 } | |
220 | |
221 return res; // Return final result | |
222 } | |
223 | |
224 //------------------------------mul_ring--------------------------------------- | |
225 // Compute the product type of two integer ranges into this node. | |
226 const Type *MulINode::mul_ring(const Type *t0, const Type *t1) const { | |
227 const TypeInt *r0 = t0->is_int(); // Handy access | |
228 const TypeInt *r1 = t1->is_int(); | |
229 | |
230 // Fetch endpoints of all ranges | |
231 int32 lo0 = r0->_lo; | |
232 double a = (double)lo0; | |
233 int32 hi0 = r0->_hi; | |
234 double b = (double)hi0; | |
235 int32 lo1 = r1->_lo; | |
236 double c = (double)lo1; | |
237 int32 hi1 = r1->_hi; | |
238 double d = (double)hi1; | |
239 | |
240 // Compute all endpoints & check for overflow | |
241 int32 A = lo0*lo1; | |
242 if( (double)A != a*c ) return TypeInt::INT; // Overflow? | |
243 int32 B = lo0*hi1; | |
244 if( (double)B != a*d ) return TypeInt::INT; // Overflow? | |
245 int32 C = hi0*lo1; | |
246 if( (double)C != b*c ) return TypeInt::INT; // Overflow? | |
247 int32 D = hi0*hi1; | |
248 if( (double)D != b*d ) return TypeInt::INT; // Overflow? | |
249 | |
250 if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints | |
251 else { lo0 = B; hi0 = A; } | |
252 if( C < D ) { | |
253 if( C < lo0 ) lo0 = C; | |
254 if( D > hi0 ) hi0 = D; | |
255 } else { | |
256 if( D < lo0 ) lo0 = D; | |
257 if( C > hi0 ) hi0 = C; | |
258 } | |
259 return TypeInt::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); | |
260 } | |
261 | |
262 | |
263 //============================================================================= | |
264 //------------------------------Ideal------------------------------------------ | |
265 // Check for power-of-2 multiply, then try the regular MulNode::Ideal | |
266 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
267 // Swap constant to right | |
268 jlong con; | |
269 if ((con = in(1)->find_long_con(0)) != 0) { | |
270 swap_edges(1, 2); | |
271 // Finish rest of method to use info in 'con' | |
272 } else if ((con = in(2)->find_long_con(0)) == 0) { | |
273 return MulNode::Ideal(phase, can_reshape); | |
274 } | |
275 | |
276 // Now we have a constant Node on the right and the constant in con | |
277 if( con == CONST64(0) ) return NULL; // By zero is handled by Value call | |
278 if( con == CONST64(1) ) return NULL; // By one is handled by Identity call | |
279 | |
280 // Check for negative constant; if so negate the final result | |
281 bool sign_flip = false; | |
282 if( con < 0 ) { | |
283 con = -con; | |
284 sign_flip = true; | |
285 } | |
286 | |
287 // Get low bit; check for being the only bit | |
288 Node *res = NULL; | |
289 jlong bit1 = con & -con; // Extract low bit | |
290 if( bit1 == con ) { // Found a power of 2? | |
291 res = new (phase->C, 3) LShiftLNode( in(1), phase->intcon(log2_long(bit1)) ); | |
292 } else { | |
293 | |
294 // Check for constant with 2 bits set | |
295 jlong bit2 = con-bit1; | |
296 bit2 = bit2 & -bit2; // Extract 2nd bit | |
297 if( bit2 + bit1 == con ) { // Found all bits in con? | |
298 Node *n1 = phase->transform( new (phase->C, 3) LShiftLNode( in(1), phase->intcon(log2_long(bit1)) ) ); | |
299 Node *n2 = phase->transform( new (phase->C, 3) LShiftLNode( in(1), phase->intcon(log2_long(bit2)) ) ); | |
300 res = new (phase->C, 3) AddLNode( n2, n1 ); | |
301 | |
302 } else if (is_power_of_2_long(con+1)) { | |
303 // Sleezy: power-of-2 -1. Next time be generic. | |
304 jlong temp = (jlong) (con + 1); | |
305 Node *n1 = phase->transform( new (phase->C, 3) LShiftLNode( in(1), phase->intcon(log2_long(temp)) ) ); | |
306 res = new (phase->C, 3) SubLNode( n1, in(1) ); | |
307 } else { | |
308 return MulNode::Ideal(phase, can_reshape); | |
309 } | |
310 } | |
311 | |
312 if( sign_flip ) { // Need to negate result? | |
313 res = phase->transform(res);// Transform, before making the zero con | |
314 res = new (phase->C, 3) SubLNode(phase->longcon(0),res); | |
315 } | |
316 | |
317 return res; // Return final result | |
318 } | |
319 | |
320 //------------------------------mul_ring--------------------------------------- | |
321 // Compute the product type of two integer ranges into this node. | |
322 const Type *MulLNode::mul_ring(const Type *t0, const Type *t1) const { | |
323 const TypeLong *r0 = t0->is_long(); // Handy access | |
324 const TypeLong *r1 = t1->is_long(); | |
325 | |
326 // Fetch endpoints of all ranges | |
327 jlong lo0 = r0->_lo; | |
328 double a = (double)lo0; | |
329 jlong hi0 = r0->_hi; | |
330 double b = (double)hi0; | |
331 jlong lo1 = r1->_lo; | |
332 double c = (double)lo1; | |
333 jlong hi1 = r1->_hi; | |
334 double d = (double)hi1; | |
335 | |
336 // Compute all endpoints & check for overflow | |
337 jlong A = lo0*lo1; | |
338 if( (double)A != a*c ) return TypeLong::LONG; // Overflow? | |
339 jlong B = lo0*hi1; | |
340 if( (double)B != a*d ) return TypeLong::LONG; // Overflow? | |
341 jlong C = hi0*lo1; | |
342 if( (double)C != b*c ) return TypeLong::LONG; // Overflow? | |
343 jlong D = hi0*hi1; | |
344 if( (double)D != b*d ) return TypeLong::LONG; // Overflow? | |
345 | |
346 if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints | |
347 else { lo0 = B; hi0 = A; } | |
348 if( C < D ) { | |
349 if( C < lo0 ) lo0 = C; | |
350 if( D > hi0 ) hi0 = D; | |
351 } else { | |
352 if( D < lo0 ) lo0 = D; | |
353 if( C > hi0 ) hi0 = C; | |
354 } | |
355 return TypeLong::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); | |
356 } | |
357 | |
358 //============================================================================= | |
359 //------------------------------mul_ring--------------------------------------- | |
360 // Compute the product type of two double ranges into this node. | |
361 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const { | |
362 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT; | |
363 return TypeF::make( t0->getf() * t1->getf() ); | |
364 } | |
365 | |
366 //============================================================================= | |
367 //------------------------------mul_ring--------------------------------------- | |
368 // Compute the product type of two double ranges into this node. | |
369 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const { | |
370 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
|
371 // We must be multiplying 2 double constants. |
0 | 372 return TypeD::make( t0->getd() * t1->getd() ); |
373 } | |
374 | |
375 //============================================================================= | |
145 | 376 //------------------------------Value------------------------------------------ |
377 const Type *MulHiLNode::Value( PhaseTransform *phase ) const { | |
378 // Either input is TOP ==> the result is TOP | |
379 const Type *t1 = phase->type( in(1) ); | |
380 const Type *t2 = phase->type( in(2) ); | |
381 if( t1 == Type::TOP ) return Type::TOP; | |
382 if( t2 == Type::TOP ) return Type::TOP; | |
383 | |
384 // Either input is BOTTOM ==> the result is the local BOTTOM | |
385 const Type *bot = bottom_type(); | |
386 if( (t1 == bot) || (t2 == bot) || | |
387 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
388 return bot; | |
389 | |
390 // It is not worth trying to constant fold this stuff! | |
391 return TypeLong::LONG; | |
392 } | |
393 | |
394 //============================================================================= | |
0 | 395 //------------------------------mul_ring--------------------------------------- |
396 // Supplied function returns the product of the inputs IN THE CURRENT RING. | |
397 // For the logical operations the ring's MUL is really a logical AND function. | |
398 // This also type-checks the inputs for sanity. Guaranteed never to | |
399 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. | |
400 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const { | |
401 const TypeInt *r0 = t0->is_int(); // Handy access | |
402 const TypeInt *r1 = t1->is_int(); | |
403 int widen = MAX2(r0->_widen,r1->_widen); | |
404 | |
405 // If either input is a constant, might be able to trim cases | |
406 if( !r0->is_con() && !r1->is_con() ) | |
407 return TypeInt::INT; // No constants to be had | |
408 | |
409 // Both constants? Return bits | |
410 if( r0->is_con() && r1->is_con() ) | |
411 return TypeInt::make( r0->get_con() & r1->get_con() ); | |
412 | |
413 if( r0->is_con() && r0->get_con() > 0 ) | |
414 return TypeInt::make(0, r0->get_con(), widen); | |
415 | |
416 if( r1->is_con() && r1->get_con() > 0 ) | |
417 return TypeInt::make(0, r1->get_con(), widen); | |
418 | |
419 if( r0 == TypeInt::BOOL || r1 == TypeInt::BOOL ) { | |
420 return TypeInt::BOOL; | |
421 } | |
422 | |
423 return TypeInt::INT; // No constants to be had | |
424 } | |
425 | |
426 //------------------------------Identity--------------------------------------- | |
427 // Masking off the high bits of an unsigned load is not required | |
428 Node *AndINode::Identity( PhaseTransform *phase ) { | |
429 | |
430 // x & x => x | |
431 if (phase->eqv(in(1), in(2))) return in(1); | |
432 | |
433 Node *load = in(1); | |
434 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
435 if( t2 && t2->is_con() ) { | |
436 int con = t2->get_con(); | |
437 // Masking off high bits which are always zero is useless. | |
438 const TypeInt* t1 = phase->type( in(1) )->isa_int(); | |
439 if (t1 != NULL && t1->_lo >= 0) { | |
440 jint t1_support = ((jint)1 << (1 + log2_intptr(t1->_hi))) - 1; | |
441 if ((t1_support & con) == t1_support) | |
442 return load; | |
443 } | |
444 uint lop = load->Opcode(); | |
445 if( lop == Op_LoadC && | |
446 con == 0x0000FFFF ) // Already zero-extended | |
447 return load; | |
448 // Masking off the high bits of a unsigned-shift-right is not | |
449 // needed either. | |
450 if( lop == Op_URShiftI ) { | |
451 const TypeInt *t12 = phase->type( load->in(2) )->isa_int(); | |
452 if( t12 && t12->is_con() ) { | |
453 int shift_con = t12->get_con(); | |
454 int mask = max_juint >> shift_con; | |
455 if( (mask&con) == mask ) // If AND is useless, skip it | |
456 return load; | |
457 } | |
458 } | |
459 } | |
460 return MulNode::Identity(phase); | |
461 } | |
462 | |
463 //------------------------------Ideal------------------------------------------ | |
464 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
465 // Special case constant AND mask | |
466 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
467 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); | |
468 const int mask = t2->get_con(); | |
469 Node *load = in(1); | |
470 uint lop = load->Opcode(); | |
471 | |
472 // Masking bits off of a Character? Hi bits are already zero. | |
473 if( lop == Op_LoadC && | |
474 (mask & 0xFFFF0000) ) // Can we make a smaller mask? | |
475 return new (phase->C, 3) AndINode(load,phase->intcon(mask&0xFFFF)); | |
476 | |
477 // Masking bits off of a Short? Loading a Character does some masking | |
478 if( lop == Op_LoadS && | |
479 (mask & 0xFFFF0000) == 0 ) { | |
480 Node *ldc = new (phase->C, 3) LoadCNode(load->in(MemNode::Control), | |
481 load->in(MemNode::Memory), | |
482 load->in(MemNode::Address), | |
483 load->adr_type()); | |
484 ldc = phase->transform(ldc); | |
485 return new (phase->C, 3) AndINode(ldc,phase->intcon(mask&0xFFFF)); | |
486 } | |
487 | |
488 // Masking sign bits off of a Byte? Let the matcher use an unsigned load | |
489 if( lop == Op_LoadB && | |
490 (!in(0) && load->in(0)) && | |
491 (mask == 0x000000FF) ) { | |
492 // Associate this node with the LoadB, so the matcher can see them together. | |
493 // If we don't do this, it is common for the LoadB to have one control | |
494 // edge, and the store or call containing this AndI to have a different | |
495 // control edge. This will cause Label_Root to group the AndI with | |
496 // the encoding store or call, so the matcher has no chance to match | |
497 // this AndI together with the LoadB. Setting the control edge here | |
498 // prevents Label_Root from grouping the AndI with the store or call, | |
499 // if it has a control edge that is inconsistent with the LoadB. | |
500 set_req(0, load->in(0)); | |
501 return this; | |
502 } | |
503 | |
504 // Masking off sign bits? Dont make them! | |
505 if( lop == Op_RShiftI ) { | |
506 const TypeInt *t12 = phase->type(load->in(2))->isa_int(); | |
507 if( t12 && t12->is_con() ) { // Shift is by a constant | |
508 int shift = t12->get_con(); | |
509 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
510 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift); | |
511 // If the AND'ing of the 2 masks has no bits, then only original shifted | |
512 // bits survive. NO sign-extension bits survive the maskings. | |
513 if( (sign_bits_mask & mask) == 0 ) { | |
514 // Use zero-fill shift instead | |
515 Node *zshift = phase->transform(new (phase->C, 3) URShiftINode(load->in(1),load->in(2))); | |
516 return new (phase->C, 3) AndINode( zshift, in(2) ); | |
517 } | |
518 } | |
519 } | |
520 | |
521 // Check for 'negate/and-1', a pattern emitted when someone asks for | |
522 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement | |
523 // plus 1) and the mask is of the low order bit. Skip the negate. | |
524 if( lop == Op_SubI && mask == 1 && load->in(1) && | |
525 phase->type(load->in(1)) == TypeInt::ZERO ) | |
526 return new (phase->C, 3) AndINode( load->in(2), in(2) ); | |
527 | |
528 return MulNode::Ideal(phase, can_reshape); | |
529 } | |
530 | |
531 //============================================================================= | |
532 //------------------------------mul_ring--------------------------------------- | |
533 // Supplied function returns the product of the inputs IN THE CURRENT RING. | |
534 // For the logical operations the ring's MUL is really a logical AND function. | |
535 // This also type-checks the inputs for sanity. Guaranteed never to | |
536 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. | |
537 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const { | |
538 const TypeLong *r0 = t0->is_long(); // Handy access | |
539 const TypeLong *r1 = t1->is_long(); | |
540 int widen = MAX2(r0->_widen,r1->_widen); | |
541 | |
542 // If either input is a constant, might be able to trim cases | |
543 if( !r0->is_con() && !r1->is_con() ) | |
544 return TypeLong::LONG; // No constants to be had | |
545 | |
546 // Both constants? Return bits | |
547 if( r0->is_con() && r1->is_con() ) | |
548 return TypeLong::make( r0->get_con() & r1->get_con() ); | |
549 | |
550 if( r0->is_con() && r0->get_con() > 0 ) | |
551 return TypeLong::make(CONST64(0), r0->get_con(), widen); | |
552 | |
553 if( r1->is_con() && r1->get_con() > 0 ) | |
554 return TypeLong::make(CONST64(0), r1->get_con(), widen); | |
555 | |
556 return TypeLong::LONG; // No constants to be had | |
557 } | |
558 | |
559 //------------------------------Identity--------------------------------------- | |
560 // Masking off the high bits of an unsigned load is not required | |
561 Node *AndLNode::Identity( PhaseTransform *phase ) { | |
562 | |
563 // x & x => x | |
564 if (phase->eqv(in(1), in(2))) return in(1); | |
565 | |
566 Node *usr = in(1); | |
567 const TypeLong *t2 = phase->type( in(2) )->isa_long(); | |
568 if( t2 && t2->is_con() ) { | |
569 jlong con = t2->get_con(); | |
570 // Masking off high bits which are always zero is useless. | |
571 const TypeLong* t1 = phase->type( in(1) )->isa_long(); | |
572 if (t1 != NULL && t1->_lo >= 0) { | |
573 jlong t1_support = ((jlong)1 << (1 + log2_long(t1->_hi))) - 1; | |
574 if ((t1_support & con) == t1_support) | |
575 return usr; | |
576 } | |
577 uint lop = usr->Opcode(); | |
578 // Masking off the high bits of a unsigned-shift-right is not | |
579 // needed either. | |
580 if( lop == Op_URShiftL ) { | |
581 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int(); | |
582 if( t12 && t12->is_con() ) { | |
583 int shift_con = t12->get_con(); | |
584 jlong mask = max_julong >> shift_con; | |
585 if( (mask&con) == mask ) // If AND is useless, skip it | |
586 return usr; | |
587 } | |
588 } | |
589 } | |
590 return MulNode::Identity(phase); | |
591 } | |
592 | |
593 //------------------------------Ideal------------------------------------------ | |
594 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
595 // Special case constant AND mask | |
596 const TypeLong *t2 = phase->type( in(2) )->isa_long(); | |
597 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); | |
598 const jlong mask = t2->get_con(); | |
599 | |
600 Node *rsh = in(1); | |
601 uint rop = rsh->Opcode(); | |
602 | |
603 // Masking off sign bits? Dont make them! | |
604 if( rop == Op_RShiftL ) { | |
605 const TypeInt *t12 = phase->type(rsh->in(2))->isa_int(); | |
606 if( t12 && t12->is_con() ) { // Shift is by a constant | |
607 int shift = t12->get_con(); | |
608 shift &= (BitsPerJavaInteger*2)-1; // semantics of Java shifts | |
609 const jlong sign_bits_mask = ~(((jlong)CONST64(1) << (jlong)(BitsPerJavaInteger*2 - shift)) -1); | |
610 // If the AND'ing of the 2 masks has no bits, then only original shifted | |
611 // bits survive. NO sign-extension bits survive the maskings. | |
612 if( (sign_bits_mask & mask) == 0 ) { | |
613 // Use zero-fill shift instead | |
614 Node *zshift = phase->transform(new (phase->C, 3) URShiftLNode(rsh->in(1),rsh->in(2))); | |
615 return new (phase->C, 3) AndLNode( zshift, in(2) ); | |
616 } | |
617 } | |
618 } | |
619 | |
620 return MulNode::Ideal(phase, can_reshape); | |
621 } | |
622 | |
623 //============================================================================= | |
624 //------------------------------Identity--------------------------------------- | |
625 Node *LShiftINode::Identity( PhaseTransform *phase ) { | |
626 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
627 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) ? in(1) : this; | |
628 } | |
629 | |
630 //------------------------------Ideal------------------------------------------ | |
631 // If the right input is a constant, and the left input is an add of a | |
632 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 | |
633 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
634 const Type *t = phase->type( in(2) ); | |
635 if( t == Type::TOP ) return NULL; // Right input is dead | |
636 const TypeInt *t2 = t->isa_int(); | |
637 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
638 const int con = t2->get_con() & ( BitsPerInt - 1 ); // masked shift count | |
639 | |
640 if ( con == 0 ) return NULL; // let Identity() handle 0 shift count | |
641 | |
642 // Left input is an add of a constant? | |
643 Node *add1 = in(1); | |
644 int add1_op = add1->Opcode(); | |
645 if( add1_op == Op_AddI ) { // Left input is an add? | |
646 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" ); | |
647 const TypeInt *t12 = phase->type(add1->in(2))->isa_int(); | |
648 if( t12 && t12->is_con() ){ // Left input is an add of a con? | |
649 // Transform is legal, but check for profit. Avoid breaking 'i2s' | |
650 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'. | |
651 if( con < 16 ) { | |
652 // Compute X << con0 | |
653 Node *lsh = phase->transform( new (phase->C, 3) LShiftINode( add1->in(1), in(2) ) ); | |
654 // Compute X<<con0 + (con1<<con0) | |
655 return new (phase->C, 3) AddINode( lsh, phase->intcon(t12->get_con() << con)); | |
656 } | |
657 } | |
658 } | |
659 | |
660 // Check for "(x>>c0)<<c0" which just masks off low bits | |
661 if( (add1_op == Op_RShiftI || add1_op == Op_URShiftI ) && | |
662 add1->in(2) == in(2) ) | |
663 // Convert to "(x & -(1<<c0))" | |
664 return new (phase->C, 3) AndINode(add1->in(1),phase->intcon( -(1<<con))); | |
665 | |
666 // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits | |
667 if( add1_op == Op_AndI ) { | |
668 Node *add2 = add1->in(1); | |
669 int add2_op = add2->Opcode(); | |
670 if( (add2_op == Op_RShiftI || add2_op == Op_URShiftI ) && | |
671 add2->in(2) == in(2) ) { | |
672 // Convert to "(x & (Y<<c0))" | |
673 Node *y_sh = phase->transform( new (phase->C, 3) LShiftINode( add1->in(2), in(2) ) ); | |
674 return new (phase->C, 3) AndINode( add2->in(1), y_sh ); | |
675 } | |
676 } | |
677 | |
678 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits | |
679 // before shifting them away. | |
680 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con); | |
681 if( add1_op == Op_AndI && | |
682 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) ) | |
683 return new (phase->C, 3) LShiftINode( add1->in(1), in(2) ); | |
684 | |
685 return NULL; | |
686 } | |
687 | |
688 //------------------------------Value------------------------------------------ | |
689 // A LShiftINode shifts its input2 left by input1 amount. | |
690 const Type *LShiftINode::Value( PhaseTransform *phase ) const { | |
691 const Type *t1 = phase->type( in(1) ); | |
692 const Type *t2 = phase->type( in(2) ); | |
693 // Either input is TOP ==> the result is TOP | |
694 if( t1 == Type::TOP ) return Type::TOP; | |
695 if( t2 == Type::TOP ) return Type::TOP; | |
696 | |
697 // Left input is ZERO ==> the result is ZERO. | |
698 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
699 // Shift by zero does nothing | |
700 if( t2 == TypeInt::ZERO ) return t1; | |
701 | |
702 // Either input is BOTTOM ==> the result is BOTTOM | |
703 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) || | |
704 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
705 return TypeInt::INT; | |
706 | |
707 const TypeInt *r1 = t1->is_int(); // Handy access | |
708 const TypeInt *r2 = t2->is_int(); // Handy access | |
709 | |
710 if (!r2->is_con()) | |
711 return TypeInt::INT; | |
712 | |
713 uint shift = r2->get_con(); | |
714 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
715 // Shift by a multiple of 32 does nothing: | |
716 if (shift == 0) return t1; | |
717 | |
718 // If the shift is a constant, shift the bounds of the type, | |
719 // unless this could lead to an overflow. | |
720 if (!r1->is_con()) { | |
721 jint lo = r1->_lo, hi = r1->_hi; | |
722 if (((lo << shift) >> shift) == lo && | |
723 ((hi << shift) >> shift) == hi) { | |
724 // No overflow. The range shifts up cleanly. | |
725 return TypeInt::make((jint)lo << (jint)shift, | |
726 (jint)hi << (jint)shift, | |
727 MAX2(r1->_widen,r2->_widen)); | |
728 } | |
729 return TypeInt::INT; | |
730 } | |
731 | |
732 return TypeInt::make( (jint)r1->get_con() << (jint)shift ); | |
733 } | |
734 | |
735 //============================================================================= | |
736 //------------------------------Identity--------------------------------------- | |
737 Node *LShiftLNode::Identity( PhaseTransform *phase ) { | |
738 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
739 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
740 } | |
741 | |
742 //------------------------------Ideal------------------------------------------ | |
743 // If the right input is a constant, and the left input is an add of a | |
744 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 | |
745 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
746 const Type *t = phase->type( in(2) ); | |
747 if( t == Type::TOP ) return NULL; // Right input is dead | |
748 const TypeInt *t2 = t->isa_int(); | |
749 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
750 const int con = t2->get_con() & ( BitsPerLong - 1 ); // masked shift count | |
751 | |
752 if ( con == 0 ) return NULL; // let Identity() handle 0 shift count | |
753 | |
754 // Left input is an add of a constant? | |
755 Node *add1 = in(1); | |
756 int add1_op = add1->Opcode(); | |
757 if( add1_op == Op_AddL ) { // Left input is an add? | |
758 // Avoid dead data cycles from dead loops | |
759 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" ); | |
760 const TypeLong *t12 = phase->type(add1->in(2))->isa_long(); | |
761 if( t12 && t12->is_con() ){ // Left input is an add of a con? | |
762 // Compute X << con0 | |
763 Node *lsh = phase->transform( new (phase->C, 3) LShiftLNode( add1->in(1), in(2) ) ); | |
764 // Compute X<<con0 + (con1<<con0) | |
765 return new (phase->C, 3) AddLNode( lsh, phase->longcon(t12->get_con() << con)); | |
766 } | |
767 } | |
768 | |
769 // Check for "(x>>c0)<<c0" which just masks off low bits | |
770 if( (add1_op == Op_RShiftL || add1_op == Op_URShiftL ) && | |
771 add1->in(2) == in(2) ) | |
772 // Convert to "(x & -(1<<c0))" | |
773 return new (phase->C, 3) AndLNode(add1->in(1),phase->longcon( -(CONST64(1)<<con))); | |
774 | |
775 // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits | |
776 if( add1_op == Op_AndL ) { | |
777 Node *add2 = add1->in(1); | |
778 int add2_op = add2->Opcode(); | |
779 if( (add2_op == Op_RShiftL || add2_op == Op_URShiftL ) && | |
780 add2->in(2) == in(2) ) { | |
781 // Convert to "(x & (Y<<c0))" | |
782 Node *y_sh = phase->transform( new (phase->C, 3) LShiftLNode( add1->in(2), in(2) ) ); | |
783 return new (phase->C, 3) AndLNode( add2->in(1), y_sh ); | |
784 } | |
785 } | |
786 | |
787 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits | |
788 // before shifting them away. | |
789 const jlong bits_mask = ((jlong)CONST64(1) << (jlong)(BitsPerJavaInteger*2 - con)) - CONST64(1); | |
790 if( add1_op == Op_AndL && | |
791 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) ) | |
792 return new (phase->C, 3) LShiftLNode( add1->in(1), in(2) ); | |
793 | |
794 return NULL; | |
795 } | |
796 | |
797 //------------------------------Value------------------------------------------ | |
798 // A LShiftLNode shifts its input2 left by input1 amount. | |
799 const Type *LShiftLNode::Value( PhaseTransform *phase ) const { | |
800 const Type *t1 = phase->type( in(1) ); | |
801 const Type *t2 = phase->type( in(2) ); | |
802 // Either input is TOP ==> the result is TOP | |
803 if( t1 == Type::TOP ) return Type::TOP; | |
804 if( t2 == Type::TOP ) return Type::TOP; | |
805 | |
806 // Left input is ZERO ==> the result is ZERO. | |
807 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
808 // Shift by zero does nothing | |
809 if( t2 == TypeInt::ZERO ) return t1; | |
810 | |
811 // Either input is BOTTOM ==> the result is BOTTOM | |
812 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) || | |
813 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) | |
814 return TypeLong::LONG; | |
815 | |
816 const TypeLong *r1 = t1->is_long(); // Handy access | |
817 const TypeInt *r2 = t2->is_int(); // Handy access | |
818 | |
819 if (!r2->is_con()) | |
820 return TypeLong::LONG; | |
821 | |
822 uint shift = r2->get_con(); | |
823 shift &= (BitsPerJavaInteger*2)-1; // semantics of Java shifts | |
824 // Shift by a multiple of 64 does nothing: | |
825 if (shift == 0) return t1; | |
826 | |
827 // If the shift is a constant, shift the bounds of the type, | |
828 // unless this could lead to an overflow. | |
829 if (!r1->is_con()) { | |
830 jlong lo = r1->_lo, hi = r1->_hi; | |
831 if (((lo << shift) >> shift) == lo && | |
832 ((hi << shift) >> shift) == hi) { | |
833 // No overflow. The range shifts up cleanly. | |
834 return TypeLong::make((jlong)lo << (jint)shift, | |
835 (jlong)hi << (jint)shift, | |
836 MAX2(r1->_widen,r2->_widen)); | |
837 } | |
838 return TypeLong::LONG; | |
839 } | |
840 | |
841 return TypeLong::make( (jlong)r1->get_con() << (jint)shift ); | |
842 } | |
843 | |
844 //============================================================================= | |
845 //------------------------------Identity--------------------------------------- | |
846 Node *RShiftINode::Identity( PhaseTransform *phase ) { | |
847 const TypeInt *t2 = phase->type(in(2))->isa_int(); | |
848 if( !t2 ) return this; | |
849 if ( t2->is_con() && ( t2->get_con() & ( BitsPerInt - 1 ) ) == 0 ) | |
850 return in(1); | |
851 | |
852 // Check for useless sign-masking | |
853 if( in(1)->Opcode() == Op_LShiftI && | |
854 in(1)->req() == 3 && | |
855 in(1)->in(2) == in(2) && | |
856 t2->is_con() ) { | |
857 uint shift = t2->get_con(); | |
858 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
859 // Compute masks for which this shifting doesn't change | |
860 int lo = (-1 << (BitsPerJavaInteger - shift-1)); // FFFF8000 | |
861 int hi = ~lo; // 00007FFF | |
862 const TypeInt *t11 = phase->type(in(1)->in(1))->isa_int(); | |
863 if( !t11 ) return this; | |
864 // Does actual value fit inside of mask? | |
865 if( lo <= t11->_lo && t11->_hi <= hi ) | |
866 return in(1)->in(1); // Then shifting is a nop | |
867 } | |
868 | |
869 return this; | |
870 } | |
871 | |
872 //------------------------------Ideal------------------------------------------ | |
873 Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
874 // Inputs may be TOP if they are dead. | |
875 const TypeInt *t1 = phase->type( in(1) )->isa_int(); | |
876 if( !t1 ) return NULL; // Left input is an integer | |
877 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
878 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
879 const TypeInt *t3; // type of in(1).in(2) | |
880 int shift = t2->get_con(); | |
881 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
882 | |
883 if ( shift == 0 ) return NULL; // let Identity() handle 0 shift count | |
884 | |
885 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller. | |
886 // Such expressions arise normally from shift chains like (byte)(x >> 24). | |
887 const Node *mask = in(1); | |
888 if( mask->Opcode() == Op_AndI && | |
889 (t3 = phase->type(mask->in(2))->isa_int()) && | |
890 t3->is_con() ) { | |
891 Node *x = mask->in(1); | |
892 jint maskbits = t3->get_con(); | |
893 // Convert to "(x >> shift) & (mask >> shift)" | |
894 Node *shr_nomask = phase->transform( new (phase->C, 3) RShiftINode(mask->in(1), in(2)) ); | |
895 return new (phase->C, 3) AndINode(shr_nomask, phase->intcon( maskbits >> shift)); | |
896 } | |
897 | |
898 // Check for "(short[i] <<16)>>16" which simply sign-extends | |
899 const Node *shl = in(1); | |
900 if( shl->Opcode() != Op_LShiftI ) return NULL; | |
901 | |
902 if( shift == 16 && | |
903 (t3 = phase->type(shl->in(2))->isa_int()) && | |
904 t3->is_con(16) ) { | |
905 Node *ld = shl->in(1); | |
906 if( ld->Opcode() == Op_LoadS ) { | |
907 // Sign extension is just useless here. Return a RShiftI of zero instead | |
908 // returning 'ld' directly. We cannot return an old Node directly as | |
909 // that is the job of 'Identity' calls and Identity calls only work on | |
910 // direct inputs ('ld' is an extra Node removed from 'this'). The | |
911 // combined optimization requires Identity only return direct inputs. | |
912 set_req(1, ld); | |
913 set_req(2, phase->intcon(0)); | |
914 return this; | |
915 } | |
916 else if( ld->Opcode() == Op_LoadC ) | |
917 // Replace zero-extension-load with sign-extension-load | |
918 return new (phase->C, 3) LoadSNode( ld->in(MemNode::Control), | |
919 ld->in(MemNode::Memory), | |
920 ld->in(MemNode::Address), | |
921 ld->adr_type()); | |
922 } | |
923 | |
924 // Check for "(byte[i] <<24)>>24" which simply sign-extends | |
925 if( shift == 24 && | |
926 (t3 = phase->type(shl->in(2))->isa_int()) && | |
927 t3->is_con(24) ) { | |
928 Node *ld = shl->in(1); | |
929 if( ld->Opcode() == Op_LoadB ) { | |
930 // Sign extension is just useless here | |
931 set_req(1, ld); | |
932 set_req(2, phase->intcon(0)); | |
933 return this; | |
934 } | |
935 } | |
936 | |
937 return NULL; | |
938 } | |
939 | |
940 //------------------------------Value------------------------------------------ | |
941 // A RShiftINode shifts its input2 right by input1 amount. | |
942 const Type *RShiftINode::Value( PhaseTransform *phase ) const { | |
943 const Type *t1 = phase->type( in(1) ); | |
944 const Type *t2 = phase->type( in(2) ); | |
945 // Either input is TOP ==> the result is TOP | |
946 if( t1 == Type::TOP ) return Type::TOP; | |
947 if( t2 == Type::TOP ) return Type::TOP; | |
948 | |
949 // Left input is ZERO ==> the result is ZERO. | |
950 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
951 // Shift by zero does nothing | |
952 if( t2 == TypeInt::ZERO ) return t1; | |
953 | |
954 // Either input is BOTTOM ==> the result is BOTTOM | |
955 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
956 return TypeInt::INT; | |
957 | |
958 if (t2 == TypeInt::INT) | |
959 return TypeInt::INT; | |
960 | |
961 const TypeInt *r1 = t1->is_int(); // Handy access | |
962 const TypeInt *r2 = t2->is_int(); // Handy access | |
963 | |
964 // If the shift is a constant, just shift the bounds of the type. | |
965 // For example, if the shift is 31, we just propagate sign bits. | |
966 if (r2->is_con()) { | |
967 uint shift = r2->get_con(); | |
968 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
969 // Shift by a multiple of 32 does nothing: | |
970 if (shift == 0) return t1; | |
971 // Calculate reasonably aggressive bounds for the result. | |
972 // This is necessary if we are to correctly type things | |
973 // like (x<<24>>24) == ((byte)x). | |
974 jint lo = (jint)r1->_lo >> (jint)shift; | |
975 jint hi = (jint)r1->_hi >> (jint)shift; | |
976 assert(lo <= hi, "must have valid bounds"); | |
977 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
978 #ifdef ASSERT | |
979 // Make sure we get the sign-capture idiom correct. | |
980 if (shift == BitsPerJavaInteger-1) { | |
981 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0"); | |
982 if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1"); | |
983 } | |
984 #endif | |
985 return ti; | |
986 } | |
987 | |
988 if( !r1->is_con() || !r2->is_con() ) | |
989 return TypeInt::INT; | |
990 | |
991 // Signed shift right | |
992 return TypeInt::make( r1->get_con() >> (r2->get_con()&31) ); | |
993 } | |
994 | |
995 //============================================================================= | |
996 //------------------------------Identity--------------------------------------- | |
997 Node *RShiftLNode::Identity( PhaseTransform *phase ) { | |
998 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
999 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
1000 } | |
1001 | |
1002 //------------------------------Value------------------------------------------ | |
1003 // A RShiftLNode shifts its input2 right by input1 amount. | |
1004 const Type *RShiftLNode::Value( PhaseTransform *phase ) const { | |
1005 const Type *t1 = phase->type( in(1) ); | |
1006 const Type *t2 = phase->type( in(2) ); | |
1007 // Either input is TOP ==> the result is TOP | |
1008 if( t1 == Type::TOP ) return Type::TOP; | |
1009 if( t2 == Type::TOP ) return Type::TOP; | |
1010 | |
1011 // Left input is ZERO ==> the result is ZERO. | |
1012 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1013 // Shift by zero does nothing | |
1014 if( t2 == TypeInt::ZERO ) return t1; | |
1015 | |
1016 // Either input is BOTTOM ==> the result is BOTTOM | |
1017 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1018 return TypeLong::LONG; | |
1019 | |
1020 if (t2 == TypeInt::INT) | |
1021 return TypeLong::LONG; | |
1022 | |
1023 const TypeLong *r1 = t1->is_long(); // Handy access | |
1024 const TypeInt *r2 = t2->is_int (); // Handy access | |
1025 | |
1026 // If the shift is a constant, just shift the bounds of the type. | |
1027 // For example, if the shift is 63, we just propagate sign bits. | |
1028 if (r2->is_con()) { | |
1029 uint shift = r2->get_con(); | |
1030 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts | |
1031 // Shift by a multiple of 64 does nothing: | |
1032 if (shift == 0) return t1; | |
1033 // Calculate reasonably aggressive bounds for the result. | |
1034 // This is necessary if we are to correctly type things | |
1035 // like (x<<24>>24) == ((byte)x). | |
1036 jlong lo = (jlong)r1->_lo >> (jlong)shift; | |
1037 jlong hi = (jlong)r1->_hi >> (jlong)shift; | |
1038 assert(lo <= hi, "must have valid bounds"); | |
1039 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1040 #ifdef ASSERT | |
1041 // Make sure we get the sign-capture idiom correct. | |
1042 if (shift == (2*BitsPerJavaInteger)-1) { | |
1043 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0"); | |
1044 if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1"); | |
1045 } | |
1046 #endif | |
1047 return tl; | |
1048 } | |
1049 | |
1050 return TypeLong::LONG; // Give up | |
1051 } | |
1052 | |
1053 //============================================================================= | |
1054 //------------------------------Identity--------------------------------------- | |
1055 Node *URShiftINode::Identity( PhaseTransform *phase ) { | |
1056 const TypeInt *ti = phase->type( in(2) )->isa_int(); | |
1057 if ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) return in(1); | |
1058 | |
1059 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x". | |
1060 // Happens during new-array length computation. | |
1061 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)] | |
1062 Node *add = in(1); | |
1063 if( add->Opcode() == Op_AddI ) { | |
1064 const TypeInt *t2 = phase->type(add->in(2))->isa_int(); | |
1065 if( t2 && t2->is_con(wordSize - 1) && | |
1066 add->in(1)->Opcode() == Op_LShiftI ) { | |
1067 // Check that shift_counts are LogBytesPerWord | |
1068 Node *lshift_count = add->in(1)->in(2); | |
1069 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int(); | |
1070 if( t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) && | |
1071 t_lshift_count == phase->type(in(2)) ) { | |
1072 Node *x = add->in(1)->in(1); | |
1073 const TypeInt *t_x = phase->type(x)->isa_int(); | |
1074 if( t_x != NULL && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord) ) { | |
1075 return x; | |
1076 } | |
1077 } | |
1078 } | |
1079 } | |
1080 | |
1081 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this; | |
1082 } | |
1083 | |
1084 //------------------------------Ideal------------------------------------------ | |
1085 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1086 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
1087 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
1088 const int con = t2->get_con() & 31; // Shift count is always masked | |
1089 if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count | |
1090 // We'll be wanting the right-shift amount as a mask of that many bits | |
1091 const int mask = right_n_bits(BitsPerJavaInteger - con); | |
1092 | |
1093 int in1_op = in(1)->Opcode(); | |
1094 | |
1095 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32 | |
1096 if( in1_op == Op_URShiftI ) { | |
1097 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int(); | |
1098 if( t12 && t12->is_con() ) { // Right input is a constant | |
1099 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" ); | |
1100 const int con2 = t12->get_con() & 31; // Shift count is always masked | |
1101 const int con3 = con+con2; | |
1102 if( con3 < 32 ) // Only merge shifts if total is < 32 | |
1103 return new (phase->C, 3) URShiftINode( in(1)->in(1), phase->intcon(con3) ); | |
1104 } | |
1105 } | |
1106 | |
1107 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z | |
1108 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". | |
1109 // If Q is "X << z" the rounding is useless. Look for patterns like | |
1110 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. | |
1111 Node *add = in(1); | |
1112 if( in1_op == Op_AddI ) { | |
1113 Node *lshl = add->in(1); | |
1114 if( lshl->Opcode() == Op_LShiftI && | |
1115 phase->type(lshl->in(2)) == t2 ) { | |
1116 Node *y_z = phase->transform( new (phase->C, 3) URShiftINode(add->in(2),in(2)) ); | |
1117 Node *sum = phase->transform( new (phase->C, 3) AddINode( lshl->in(1), y_z ) ); | |
1118 return new (phase->C, 3) AndINode( sum, phase->intcon(mask) ); | |
1119 } | |
1120 } | |
1121 | |
1122 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) | |
1123 // This shortens the mask. Also, if we are extracting a high byte and | |
1124 // storing it to a buffer, the mask will be removed completely. | |
1125 Node *andi = in(1); | |
1126 if( in1_op == Op_AndI ) { | |
1127 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int(); | |
1128 if( t3 && t3->is_con() ) { // Right input is a constant | |
1129 jint mask2 = t3->get_con(); | |
1130 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) | |
1131 Node *newshr = phase->transform( new (phase->C, 3) URShiftINode(andi->in(1), in(2)) ); | |
1132 return new (phase->C, 3) AndINode(newshr, phase->intcon(mask2)); | |
1133 // The negative values are easier to materialize than positive ones. | |
1134 // A typical case from address arithmetic is ((x & ~15) >> 4). | |
1135 // It's better to change that to ((x >> 4) & ~0) versus | |
1136 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64. | |
1137 } | |
1138 } | |
1139 | |
1140 // Check for "(X << z ) >>> z" which simply zero-extends | |
1141 Node *shl = in(1); | |
1142 if( in1_op == Op_LShiftI && | |
1143 phase->type(shl->in(2)) == t2 ) | |
1144 return new (phase->C, 3) AndINode( shl->in(1), phase->intcon(mask) ); | |
1145 | |
1146 return NULL; | |
1147 } | |
1148 | |
1149 //------------------------------Value------------------------------------------ | |
1150 // A URShiftINode shifts its input2 right by input1 amount. | |
1151 const Type *URShiftINode::Value( PhaseTransform *phase ) const { | |
1152 // (This is a near clone of RShiftINode::Value.) | |
1153 const Type *t1 = phase->type( in(1) ); | |
1154 const Type *t2 = phase->type( in(2) ); | |
1155 // Either input is TOP ==> the result is TOP | |
1156 if( t1 == Type::TOP ) return Type::TOP; | |
1157 if( t2 == Type::TOP ) return Type::TOP; | |
1158 | |
1159 // Left input is ZERO ==> the result is ZERO. | |
1160 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; | |
1161 // Shift by zero does nothing | |
1162 if( t2 == TypeInt::ZERO ) return t1; | |
1163 | |
1164 // Either input is BOTTOM ==> the result is BOTTOM | |
1165 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1166 return TypeInt::INT; | |
1167 | |
1168 if (t2 == TypeInt::INT) | |
1169 return TypeInt::INT; | |
1170 | |
1171 const TypeInt *r1 = t1->is_int(); // Handy access | |
1172 const TypeInt *r2 = t2->is_int(); // Handy access | |
1173 | |
1174 if (r2->is_con()) { | |
1175 uint shift = r2->get_con(); | |
1176 shift &= BitsPerJavaInteger-1; // semantics of Java shifts | |
1177 // Shift by a multiple of 32 does nothing: | |
1178 if (shift == 0) return t1; | |
1179 // Calculate reasonably aggressive bounds for the result. | |
1180 jint lo = (juint)r1->_lo >> (juint)shift; | |
1181 jint hi = (juint)r1->_hi >> (juint)shift; | |
1182 if (r1->_hi >= 0 && r1->_lo < 0) { | |
1183 // If the type has both negative and positive values, | |
1184 // there are two separate sub-domains to worry about: | |
1185 // The positive half and the negative half. | |
1186 jint neg_lo = lo; | |
1187 jint neg_hi = (juint)-1 >> (juint)shift; | |
1188 jint pos_lo = (juint) 0 >> (juint)shift; | |
1189 jint pos_hi = hi; | |
1190 lo = MIN2(neg_lo, pos_lo); // == 0 | |
1191 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; | |
1192 } | |
1193 assert(lo <= hi, "must have valid bounds"); | |
1194 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1195 #ifdef ASSERT | |
1196 // Make sure we get the sign-capture idiom correct. | |
1197 if (shift == BitsPerJavaInteger-1) { | |
1198 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0"); | |
1199 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1"); | |
1200 } | |
1201 #endif | |
1202 return ti; | |
1203 } | |
1204 | |
1205 // | |
1206 // Do not support shifted oops in info for GC | |
1207 // | |
1208 // else if( t1->base() == Type::InstPtr ) { | |
1209 // | |
1210 // const TypeInstPtr *o = t1->is_instptr(); | |
1211 // if( t1->singleton() ) | |
1212 // return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift ); | |
1213 // } | |
1214 // else if( t1->base() == Type::KlassPtr ) { | |
1215 // const TypeKlassPtr *o = t1->is_klassptr(); | |
1216 // if( t1->singleton() ) | |
1217 // return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift ); | |
1218 // } | |
1219 | |
1220 return TypeInt::INT; | |
1221 } | |
1222 | |
1223 //============================================================================= | |
1224 //------------------------------Identity--------------------------------------- | |
1225 Node *URShiftLNode::Identity( PhaseTransform *phase ) { | |
1226 const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int | |
1227 return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this; | |
1228 } | |
1229 | |
1230 //------------------------------Ideal------------------------------------------ | |
1231 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1232 const TypeInt *t2 = phase->type( in(2) )->isa_int(); | |
1233 if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant | |
1234 const int con = t2->get_con() & ( BitsPerLong - 1 ); // Shift count is always masked | |
1235 if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count | |
1236 // note: mask computation below does not work for 0 shift count | |
1237 // We'll be wanting the right-shift amount as a mask of that many bits | |
1238 const jlong mask = (((jlong)CONST64(1) << (jlong)(BitsPerJavaInteger*2 - con)) -1); | |
1239 | |
1240 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z | |
1241 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". | |
1242 // If Q is "X << z" the rounding is useless. Look for patterns like | |
1243 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. | |
1244 Node *add = in(1); | |
1245 if( add->Opcode() == Op_AddL ) { | |
1246 Node *lshl = add->in(1); | |
1247 if( lshl->Opcode() == Op_LShiftL && | |
1248 phase->type(lshl->in(2)) == t2 ) { | |
1249 Node *y_z = phase->transform( new (phase->C, 3) URShiftLNode(add->in(2),in(2)) ); | |
1250 Node *sum = phase->transform( new (phase->C, 3) AddLNode( lshl->in(1), y_z ) ); | |
1251 return new (phase->C, 3) AndLNode( sum, phase->longcon(mask) ); | |
1252 } | |
1253 } | |
1254 | |
1255 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) | |
1256 // This shortens the mask. Also, if we are extracting a high byte and | |
1257 // storing it to a buffer, the mask will be removed completely. | |
1258 Node *andi = in(1); | |
1259 if( andi->Opcode() == Op_AndL ) { | |
1260 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long(); | |
1261 if( t3 && t3->is_con() ) { // Right input is a constant | |
1262 jlong mask2 = t3->get_con(); | |
1263 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) | |
1264 Node *newshr = phase->transform( new (phase->C, 3) URShiftLNode(andi->in(1), in(2)) ); | |
1265 return new (phase->C, 3) AndLNode(newshr, phase->longcon(mask2)); | |
1266 } | |
1267 } | |
1268 | |
1269 // Check for "(X << z ) >>> z" which simply zero-extends | |
1270 Node *shl = in(1); | |
1271 if( shl->Opcode() == Op_LShiftL && | |
1272 phase->type(shl->in(2)) == t2 ) | |
1273 return new (phase->C, 3) AndLNode( shl->in(1), phase->longcon(mask) ); | |
1274 | |
1275 return NULL; | |
1276 } | |
1277 | |
1278 //------------------------------Value------------------------------------------ | |
1279 // A URShiftINode shifts its input2 right by input1 amount. | |
1280 const Type *URShiftLNode::Value( PhaseTransform *phase ) const { | |
1281 // (This is a near clone of RShiftLNode::Value.) | |
1282 const Type *t1 = phase->type( in(1) ); | |
1283 const Type *t2 = phase->type( in(2) ); | |
1284 // Either input is TOP ==> the result is TOP | |
1285 if( t1 == Type::TOP ) return Type::TOP; | |
1286 if( t2 == Type::TOP ) return Type::TOP; | |
1287 | |
1288 // Left input is ZERO ==> the result is ZERO. | |
1289 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; | |
1290 // Shift by zero does nothing | |
1291 if( t2 == TypeInt::ZERO ) return t1; | |
1292 | |
1293 // Either input is BOTTOM ==> the result is BOTTOM | |
1294 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) | |
1295 return TypeLong::LONG; | |
1296 | |
1297 if (t2 == TypeInt::INT) | |
1298 return TypeLong::LONG; | |
1299 | |
1300 const TypeLong *r1 = t1->is_long(); // Handy access | |
1301 const TypeInt *r2 = t2->is_int (); // Handy access | |
1302 | |
1303 if (r2->is_con()) { | |
1304 uint shift = r2->get_con(); | |
1305 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts | |
1306 // Shift by a multiple of 64 does nothing: | |
1307 if (shift == 0) return t1; | |
1308 // Calculate reasonably aggressive bounds for the result. | |
1309 jlong lo = (julong)r1->_lo >> (juint)shift; | |
1310 jlong hi = (julong)r1->_hi >> (juint)shift; | |
1311 if (r1->_hi >= 0 && r1->_lo < 0) { | |
1312 // If the type has both negative and positive values, | |
1313 // there are two separate sub-domains to worry about: | |
1314 // The positive half and the negative half. | |
1315 jlong neg_lo = lo; | |
1316 jlong neg_hi = (julong)-1 >> (juint)shift; | |
1317 jlong pos_lo = (julong) 0 >> (juint)shift; | |
1318 jlong pos_hi = hi; | |
1319 //lo = MIN2(neg_lo, pos_lo); // == 0 | |
1320 lo = neg_lo < pos_lo ? neg_lo : pos_lo; | |
1321 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; | |
1322 hi = neg_hi > pos_hi ? neg_hi : pos_hi; | |
1323 } | |
1324 assert(lo <= hi, "must have valid bounds"); | |
1325 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); | |
1326 #ifdef ASSERT | |
1327 // Make sure we get the sign-capture idiom correct. | |
1328 if (shift == (2*BitsPerJavaInteger)-1) { | |
1329 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0"); | |
1330 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1"); | |
1331 } | |
1332 #endif | |
1333 return tl; | |
1334 } | |
1335 | |
1336 return TypeLong::LONG; // Give up | |
1337 } |