1 | /* |
2 | * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "memory/allocation.inline.hpp" |
27 | #include "opto/addnode.hpp" |
28 | #include "opto/connode.hpp" |
29 | #include "opto/convertnode.hpp" |
30 | #include "opto/memnode.hpp" |
31 | #include "opto/mulnode.hpp" |
32 | #include "opto/phaseX.hpp" |
33 | #include "opto/subnode.hpp" |
34 | |
35 | // Portions of code courtesy of Clifford Click |
36 | |
37 | |
38 | //============================================================================= |
39 | //------------------------------hash------------------------------------------- |
40 | // Hash function over MulNodes. Needs to be commutative; i.e., I swap |
41 | // (commute) inputs to MulNodes willy-nilly so the hash function must return |
42 | // the same value in the presence of edge swapping. |
43 | uint MulNode::hash() const { |
44 | return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode(); |
45 | } |
46 | |
47 | //------------------------------Identity--------------------------------------- |
48 | // Multiplying a one preserves the other argument |
49 | Node* MulNode::Identity(PhaseGVN* phase) { |
50 | const Type *one = mul_id(); // The multiplicative identity |
51 | if( phase->type( in(1) )->higher_equal( one ) ) return in(2); |
52 | if( phase->type( in(2) )->higher_equal( one ) ) return in(1); |
53 | |
54 | return this; |
55 | } |
56 | |
57 | //------------------------------Ideal------------------------------------------ |
58 | // We also canonicalize the Node, moving constants to the right input, |
59 | // and flatten expressions (so that 1+x+2 becomes x+3). |
60 | Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
61 | const Type *t1 = phase->type( in(1) ); |
62 | const Type *t2 = phase->type( in(2) ); |
63 | Node *progress = NULL; // Progress flag |
64 | // We are OK if right is a constant, or right is a load and |
65 | // left is a non-constant. |
66 | if( !(t2->singleton() || |
67 | (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) { |
68 | if( t1->singleton() || // Left input is a constant? |
69 | // Otherwise, sort inputs (commutativity) to help value numbering. |
70 | (in(1)->_idx > in(2)->_idx) ) { |
71 | swap_edges(1, 2); |
72 | const Type *t = t1; |
73 | t1 = t2; |
74 | t2 = t; |
75 | progress = this; // Made progress |
76 | } |
77 | } |
78 | |
79 | // If the right input is a constant, and the left input is a product of a |
80 | // constant, flatten the expression tree. |
81 | uint op = Opcode(); |
82 | if( t2->singleton() && // Right input is a constant? |
83 | op != Op_MulF && // Float & double cannot reassociate |
84 | op != Op_MulD ) { |
85 | if( t2 == Type::TOP ) return NULL; |
86 | Node *mul1 = in(1); |
87 | #ifdef ASSERT |
88 | // Check for dead loop |
89 | int op1 = mul1->Opcode(); |
90 | if( phase->eqv( mul1, this ) || phase->eqv( in(2), this ) || |
91 | ( ( op1 == mul_opcode() || op1 == add_opcode() ) && |
92 | ( phase->eqv( mul1->in(1), this ) || phase->eqv( mul1->in(2), this ) || |
93 | phase->eqv( mul1->in(1), mul1 ) || phase->eqv( mul1->in(2), mul1 ) ) ) ) |
94 | assert(false, "dead loop in MulNode::Ideal" ); |
95 | #endif |
96 | |
97 | if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply? |
98 | // Mul of a constant? |
99 | const Type *t12 = phase->type( mul1->in(2) ); |
100 | if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant? |
101 | // Compute new constant; check for overflow |
102 | const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12); |
103 | if( tcon01->singleton() ) { |
104 | // The Mul of the flattened expression |
105 | set_req(1, mul1->in(1)); |
106 | set_req(2, phase->makecon( tcon01 )); |
107 | t2 = tcon01; |
108 | progress = this; // Made progress |
109 | } |
110 | } |
111 | } |
112 | // If the right input is a constant, and the left input is an add of a |
113 | // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0 |
114 | const Node *add1 = in(1); |
115 | if( add1->Opcode() == add_opcode() ) { // Left input is an add? |
116 | // Add of a constant? |
117 | const Type *t12 = phase->type( add1->in(2) ); |
118 | if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant? |
119 | assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" ); |
120 | // Compute new constant; check for overflow |
121 | const Type *tcon01 = mul_ring(t2,t12); |
122 | if( tcon01->singleton() ) { |
123 | |
124 | // Convert (X+con1)*con0 into X*con0 |
125 | Node *mul = clone(); // mul = ()*con0 |
126 | mul->set_req(1,add1->in(1)); // mul = X*con0 |
127 | mul = phase->transform(mul); |
128 | |
129 | Node *add2 = add1->clone(); |
130 | add2->set_req(1, mul); // X*con0 + con0*con1 |
131 | add2->set_req(2, phase->makecon(tcon01) ); |
132 | progress = add2; |
133 | } |
134 | } |
135 | } // End of is left input an add |
136 | } // End of is right input a Mul |
137 | |
138 | return progress; |
139 | } |
140 | |
141 | //------------------------------Value----------------------------------------- |
142 | const Type* MulNode::Value(PhaseGVN* phase) const { |
143 | const Type *t1 = phase->type( in(1) ); |
144 | const Type *t2 = phase->type( in(2) ); |
145 | // Either input is TOP ==> the result is TOP |
146 | if( t1 == Type::TOP ) return Type::TOP; |
147 | if( t2 == Type::TOP ) return Type::TOP; |
148 | |
149 | // Either input is ZERO ==> the result is ZERO. |
150 | // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0 |
151 | int op = Opcode(); |
152 | if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) { |
153 | const Type *zero = add_id(); // The multiplicative zero |
154 | if( t1->higher_equal( zero ) ) return zero; |
155 | if( t2->higher_equal( zero ) ) return zero; |
156 | } |
157 | |
158 | // Either input is BOTTOM ==> the result is the local BOTTOM |
159 | if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) |
160 | return bottom_type(); |
161 | |
162 | #if defined(IA32) |
163 | // Can't trust native compilers to properly fold strict double |
164 | // multiplication with round-to-zero on this platform. |
165 | if (op == Op_MulD && phase->C->method()->is_strict()) { |
166 | return TypeD::DOUBLE; |
167 | } |
168 | #endif |
169 | |
170 | return mul_ring(t1,t2); // Local flavor of type multiplication |
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 | |
193 | unsigned int abs_con = uabs(con); |
194 | if (abs_con != (unsigned int)con) { |
195 | sign_flip = true; |
196 | } |
197 | |
198 | // Get low bit; check for being the only bit |
199 | Node *res = NULL; |
200 | unsigned int bit1 = abs_con & (0-abs_con); // Extract low bit |
201 | if (bit1 == abs_con) { // Found a power of 2? |
202 | res = new LShiftINode(in(1), phase->intcon(log2_uint(bit1))); |
203 | } else { |
204 | |
205 | // Check for constant with 2 bits set |
206 | unsigned int bit2 = abs_con-bit1; |
207 | bit2 = bit2 & (0-bit2); // Extract 2nd bit |
208 | if (bit2 + bit1 == abs_con) { // Found all bits in con? |
209 | Node *n1 = phase->transform( new LShiftINode(in(1), phase->intcon(log2_uint(bit1)))); |
210 | Node *n2 = phase->transform( new LShiftINode(in(1), phase->intcon(log2_uint(bit2)))); |
211 | res = new AddINode(n2, n1); |
212 | |
213 | } else if (is_power_of_2(abs_con+1)) { |
214 | // Sleezy: power-of-2 -1. Next time be generic. |
215 | unsigned int temp = abs_con + 1; |
216 | Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2_uint(temp)))); |
217 | res = new SubINode(n1, in(1)); |
218 | } else { |
219 | return MulNode::Ideal(phase, can_reshape); |
220 | } |
221 | } |
222 | |
223 | if (sign_flip) { // Need to negate result? |
224 | res = phase->transform(res);// Transform, before making the zero con |
225 | res = new SubINode(phase->intcon(0),res); |
226 | } |
227 | |
228 | return res; // Return final result |
229 | } |
230 | |
231 | //------------------------------mul_ring--------------------------------------- |
232 | // Compute the product type of two integer ranges into this node. |
233 | const Type *MulINode::mul_ring(const Type *t0, const Type *t1) const { |
234 | const TypeInt *r0 = t0->is_int(); // Handy access |
235 | const TypeInt *r1 = t1->is_int(); |
236 | |
237 | // Fetch endpoints of all ranges |
238 | jint lo0 = r0->_lo; |
239 | double a = (double)lo0; |
240 | jint hi0 = r0->_hi; |
241 | double b = (double)hi0; |
242 | jint lo1 = r1->_lo; |
243 | double c = (double)lo1; |
244 | jint hi1 = r1->_hi; |
245 | double d = (double)hi1; |
246 | |
247 | // Compute all endpoints & check for overflow |
248 | int32_t A = java_multiply(lo0, lo1); |
249 | if( (double)A != a*c ) return TypeInt::INT; // Overflow? |
250 | int32_t B = java_multiply(lo0, hi1); |
251 | if( (double)B != a*d ) return TypeInt::INT; // Overflow? |
252 | int32_t C = java_multiply(hi0, lo1); |
253 | if( (double)C != b*c ) return TypeInt::INT; // Overflow? |
254 | int32_t D = java_multiply(hi0, hi1); |
255 | if( (double)D != b*d ) return TypeInt::INT; // Overflow? |
256 | |
257 | if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints |
258 | else { lo0 = B; hi0 = A; } |
259 | if( C < D ) { |
260 | if( C < lo0 ) lo0 = C; |
261 | if( D > hi0 ) hi0 = D; |
262 | } else { |
263 | if( D < lo0 ) lo0 = D; |
264 | if( C > hi0 ) hi0 = C; |
265 | } |
266 | return TypeInt::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); |
267 | } |
268 | |
269 | |
270 | //============================================================================= |
271 | //------------------------------Ideal------------------------------------------ |
272 | // Check for power-of-2 multiply, then try the regular MulNode::Ideal |
273 | Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
274 | // Swap constant to right |
275 | jlong con; |
276 | if ((con = in(1)->find_long_con(0)) != 0) { |
277 | swap_edges(1, 2); |
278 | // Finish rest of method to use info in 'con' |
279 | } else if ((con = in(2)->find_long_con(0)) == 0) { |
280 | return MulNode::Ideal(phase, can_reshape); |
281 | } |
282 | |
283 | // Now we have a constant Node on the right and the constant in con |
284 | if (con == CONST64(0)) return NULL; // By zero is handled by Value call |
285 | if (con == CONST64(1)) return NULL; // By one is handled by Identity call |
286 | |
287 | // Check for negative constant; if so negate the final result |
288 | bool sign_flip = false; |
289 | julong abs_con = uabs(con); |
290 | if (abs_con != (julong)con) { |
291 | sign_flip = true; |
292 | } |
293 | |
294 | // Get low bit; check for being the only bit |
295 | Node *res = NULL; |
296 | julong bit1 = abs_con & (0-abs_con); // Extract low bit |
297 | if (bit1 == abs_con) { // Found a power of 2? |
298 | res = new LShiftLNode(in(1), phase->intcon(log2_long(bit1))); |
299 | } else { |
300 | |
301 | // Check for constant with 2 bits set |
302 | julong bit2 = abs_con-bit1; |
303 | bit2 = bit2 & (0-bit2); // Extract 2nd bit |
304 | if (bit2 + bit1 == abs_con) { // Found all bits in con? |
305 | Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2_long(bit1)))); |
306 | Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2_long(bit2)))); |
307 | res = new AddLNode(n2, n1); |
308 | |
309 | } else if (is_power_of_2_long(abs_con+1)) { |
310 | // Sleezy: power-of-2 -1. Next time be generic. |
311 | julong temp = abs_con + 1; |
312 | Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2_long(temp)))); |
313 | res = new SubLNode(n1, in(1)); |
314 | } else { |
315 | return MulNode::Ideal(phase, can_reshape); |
316 | } |
317 | } |
318 | |
319 | if (sign_flip) { // Need to negate result? |
320 | res = phase->transform(res);// Transform, before making the zero con |
321 | res = new SubLNode(phase->longcon(0),res); |
322 | } |
323 | |
324 | return res; // Return final result |
325 | } |
326 | |
327 | //------------------------------mul_ring--------------------------------------- |
328 | // Compute the product type of two integer ranges into this node. |
329 | const Type *MulLNode::mul_ring(const Type *t0, const Type *t1) const { |
330 | const TypeLong *r0 = t0->is_long(); // Handy access |
331 | const TypeLong *r1 = t1->is_long(); |
332 | |
333 | // Fetch endpoints of all ranges |
334 | jlong lo0 = r0->_lo; |
335 | double a = (double)lo0; |
336 | jlong hi0 = r0->_hi; |
337 | double b = (double)hi0; |
338 | jlong lo1 = r1->_lo; |
339 | double c = (double)lo1; |
340 | jlong hi1 = r1->_hi; |
341 | double d = (double)hi1; |
342 | |
343 | // Compute all endpoints & check for overflow |
344 | jlong A = java_multiply(lo0, lo1); |
345 | if( (double)A != a*c ) return TypeLong::LONG; // Overflow? |
346 | jlong B = java_multiply(lo0, hi1); |
347 | if( (double)B != a*d ) return TypeLong::LONG; // Overflow? |
348 | jlong C = java_multiply(hi0, lo1); |
349 | if( (double)C != b*c ) return TypeLong::LONG; // Overflow? |
350 | jlong D = java_multiply(hi0, hi1); |
351 | if( (double)D != b*d ) return TypeLong::LONG; // Overflow? |
352 | |
353 | if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints |
354 | else { lo0 = B; hi0 = A; } |
355 | if( C < D ) { |
356 | if( C < lo0 ) lo0 = C; |
357 | if( D > hi0 ) hi0 = D; |
358 | } else { |
359 | if( D < lo0 ) lo0 = D; |
360 | if( C > hi0 ) hi0 = C; |
361 | } |
362 | return TypeLong::make(lo0, hi0, MAX2(r0->_widen,r1->_widen)); |
363 | } |
364 | |
365 | //============================================================================= |
366 | //------------------------------mul_ring--------------------------------------- |
367 | // Compute the product type of two double ranges into this node. |
368 | const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const { |
369 | if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT; |
370 | return TypeF::make( t0->getf() * t1->getf() ); |
371 | } |
372 | |
373 | //============================================================================= |
374 | //------------------------------mul_ring--------------------------------------- |
375 | // Compute the product type of two double ranges into this node. |
376 | const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const { |
377 | if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE; |
378 | // We must be multiplying 2 double constants. |
379 | return TypeD::make( t0->getd() * t1->getd() ); |
380 | } |
381 | |
382 | //============================================================================= |
383 | //------------------------------Value------------------------------------------ |
384 | const Type* MulHiLNode::Value(PhaseGVN* phase) const { |
385 | // Either input is TOP ==> the result is TOP |
386 | const Type *t1 = phase->type( in(1) ); |
387 | const Type *t2 = phase->type( in(2) ); |
388 | if( t1 == Type::TOP ) return Type::TOP; |
389 | if( t2 == Type::TOP ) return Type::TOP; |
390 | |
391 | // Either input is BOTTOM ==> the result is the local BOTTOM |
392 | const Type *bot = bottom_type(); |
393 | if( (t1 == bot) || (t2 == bot) || |
394 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
395 | return bot; |
396 | |
397 | // It is not worth trying to constant fold this stuff! |
398 | return TypeLong::LONG; |
399 | } |
400 | |
401 | //============================================================================= |
402 | //------------------------------mul_ring--------------------------------------- |
403 | // Supplied function returns the product of the inputs IN THE CURRENT RING. |
404 | // For the logical operations the ring's MUL is really a logical AND function. |
405 | // This also type-checks the inputs for sanity. Guaranteed never to |
406 | // be passed a TOP or BOTTOM type, these are filtered out by pre-check. |
407 | const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const { |
408 | const TypeInt *r0 = t0->is_int(); // Handy access |
409 | const TypeInt *r1 = t1->is_int(); |
410 | int widen = MAX2(r0->_widen,r1->_widen); |
411 | |
412 | // If either input is a constant, might be able to trim cases |
413 | if( !r0->is_con() && !r1->is_con() ) |
414 | return TypeInt::INT; // No constants to be had |
415 | |
416 | // Both constants? Return bits |
417 | if( r0->is_con() && r1->is_con() ) |
418 | return TypeInt::make( r0->get_con() & r1->get_con() ); |
419 | |
420 | if( r0->is_con() && r0->get_con() > 0 ) |
421 | return TypeInt::make(0, r0->get_con(), widen); |
422 | |
423 | if( r1->is_con() && r1->get_con() > 0 ) |
424 | return TypeInt::make(0, r1->get_con(), widen); |
425 | |
426 | if( r0 == TypeInt::BOOL || r1 == TypeInt::BOOL ) { |
427 | return TypeInt::BOOL; |
428 | } |
429 | |
430 | return TypeInt::INT; // No constants to be had |
431 | } |
432 | |
433 | //------------------------------Identity--------------------------------------- |
434 | // Masking off the high bits of an unsigned load is not required |
435 | Node* AndINode::Identity(PhaseGVN* phase) { |
436 | |
437 | // x & x => x |
438 | if (phase->eqv(in(1), in(2))) return in(1); |
439 | |
440 | Node* in1 = in(1); |
441 | uint op = in1->Opcode(); |
442 | const TypeInt* t2 = phase->type(in(2))->isa_int(); |
443 | if (t2 && t2->is_con()) { |
444 | int con = t2->get_con(); |
445 | // Masking off high bits which are always zero is useless. |
446 | const TypeInt* t1 = phase->type( in(1) )->isa_int(); |
447 | if (t1 != NULL && t1->_lo >= 0) { |
448 | jint t1_support = right_n_bits(1 + log2_jint(t1->_hi)); |
449 | if ((t1_support & con) == t1_support) |
450 | return in1; |
451 | } |
452 | // Masking off the high bits of a unsigned-shift-right is not |
453 | // needed either. |
454 | if (op == Op_URShiftI) { |
455 | const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); |
456 | if (t12 && t12->is_con()) { // Shift is by a constant |
457 | int shift = t12->get_con(); |
458 | shift &= BitsPerJavaInteger - 1; // semantics of Java shifts |
459 | int mask = max_juint >> shift; |
460 | if ((mask & con) == mask) // If AND is useless, skip it |
461 | return in1; |
462 | } |
463 | } |
464 | } |
465 | return MulNode::Identity(phase); |
466 | } |
467 | |
468 | //------------------------------Ideal------------------------------------------ |
469 | Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
470 | // Special case constant AND mask |
471 | const TypeInt *t2 = phase->type( in(2) )->isa_int(); |
472 | if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); |
473 | const int mask = t2->get_con(); |
474 | Node *load = in(1); |
475 | uint lop = load->Opcode(); |
476 | |
477 | // Masking bits off of a Character? Hi bits are already zero. |
478 | if( lop == Op_LoadUS && |
479 | (mask & 0xFFFF0000) ) // Can we make a smaller mask? |
480 | return new AndINode(load,phase->intcon(mask&0xFFFF)); |
481 | |
482 | // Masking bits off of a Short? Loading a Character does some masking |
483 | if (can_reshape && |
484 | load->outcnt() == 1 && load->unique_out() == this) { |
485 | if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) { |
486 | Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase); |
487 | ldus = phase->transform(ldus); |
488 | return new AndINode(ldus, phase->intcon(mask & 0xFFFF)); |
489 | } |
490 | |
491 | // Masking sign bits off of a Byte? Do an unsigned byte load plus |
492 | // an and. |
493 | if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) { |
494 | Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase); |
495 | ldub = phase->transform(ldub); |
496 | return new AndINode(ldub, phase->intcon(mask)); |
497 | } |
498 | } |
499 | |
500 | // Masking off sign bits? Dont make them! |
501 | if( lop == Op_RShiftI ) { |
502 | const TypeInt *t12 = phase->type(load->in(2))->isa_int(); |
503 | if( t12 && t12->is_con() ) { // Shift is by a constant |
504 | int shift = t12->get_con(); |
505 | shift &= BitsPerJavaInteger-1; // semantics of Java shifts |
506 | const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift); |
507 | // If the AND'ing of the 2 masks has no bits, then only original shifted |
508 | // bits survive. NO sign-extension bits survive the maskings. |
509 | if( (sign_bits_mask & mask) == 0 ) { |
510 | // Use zero-fill shift instead |
511 | Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2))); |
512 | return new AndINode( zshift, in(2) ); |
513 | } |
514 | } |
515 | } |
516 | |
517 | // Check for 'negate/and-1', a pattern emitted when someone asks for |
518 | // 'mod 2'. Negate leaves the low order bit unchanged (think: complement |
519 | // plus 1) and the mask is of the low order bit. Skip the negate. |
520 | if( lop == Op_SubI && mask == 1 && load->in(1) && |
521 | phase->type(load->in(1)) == TypeInt::ZERO ) |
522 | return new AndINode( load->in(2), in(2) ); |
523 | |
524 | return MulNode::Ideal(phase, can_reshape); |
525 | } |
526 | |
527 | //============================================================================= |
528 | //------------------------------mul_ring--------------------------------------- |
529 | // Supplied function returns the product of the inputs IN THE CURRENT RING. |
530 | // For the logical operations the ring's MUL is really a logical AND function. |
531 | // This also type-checks the inputs for sanity. Guaranteed never to |
532 | // be passed a TOP or BOTTOM type, these are filtered out by pre-check. |
533 | const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const { |
534 | const TypeLong *r0 = t0->is_long(); // Handy access |
535 | const TypeLong *r1 = t1->is_long(); |
536 | int widen = MAX2(r0->_widen,r1->_widen); |
537 | |
538 | // If either input is a constant, might be able to trim cases |
539 | if( !r0->is_con() && !r1->is_con() ) |
540 | return TypeLong::LONG; // No constants to be had |
541 | |
542 | // Both constants? Return bits |
543 | if( r0->is_con() && r1->is_con() ) |
544 | return TypeLong::make( r0->get_con() & r1->get_con() ); |
545 | |
546 | if( r0->is_con() && r0->get_con() > 0 ) |
547 | return TypeLong::make(CONST64(0), r0->get_con(), widen); |
548 | |
549 | if( r1->is_con() && r1->get_con() > 0 ) |
550 | return TypeLong::make(CONST64(0), r1->get_con(), widen); |
551 | |
552 | return TypeLong::LONG; // No constants to be had |
553 | } |
554 | |
555 | //------------------------------Identity--------------------------------------- |
556 | // Masking off the high bits of an unsigned load is not required |
557 | Node* AndLNode::Identity(PhaseGVN* phase) { |
558 | |
559 | // x & x => x |
560 | if (phase->eqv(in(1), in(2))) return in(1); |
561 | |
562 | Node *usr = in(1); |
563 | const TypeLong *t2 = phase->type( in(2) )->isa_long(); |
564 | if( t2 && t2->is_con() ) { |
565 | jlong con = t2->get_con(); |
566 | // Masking off high bits which are always zero is useless. |
567 | const TypeLong* t1 = phase->type( in(1) )->isa_long(); |
568 | if (t1 != NULL && t1->_lo >= 0) { |
569 | int bit_count = log2_long(t1->_hi) + 1; |
570 | jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count)); |
571 | if ((t1_support & con) == t1_support) |
572 | return usr; |
573 | } |
574 | uint lop = usr->Opcode(); |
575 | // Masking off the high bits of a unsigned-shift-right is not |
576 | // needed either. |
577 | if( lop == Op_URShiftL ) { |
578 | const TypeInt *t12 = phase->type( usr->in(2) )->isa_int(); |
579 | if( t12 && t12->is_con() ) { // Shift is by a constant |
580 | int shift = t12->get_con(); |
581 | shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
582 | jlong mask = max_julong >> shift; |
583 | if( (mask&con) == mask ) // If AND is useless, skip it |
584 | return usr; |
585 | } |
586 | } |
587 | } |
588 | return MulNode::Identity(phase); |
589 | } |
590 | |
591 | //------------------------------Ideal------------------------------------------ |
592 | Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
593 | // Special case constant AND mask |
594 | const TypeLong *t2 = phase->type( in(2) )->isa_long(); |
595 | if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); |
596 | const jlong mask = t2->get_con(); |
597 | |
598 | Node* in1 = in(1); |
599 | uint op = in1->Opcode(); |
600 | |
601 | // Are we masking a long that was converted from an int with a mask |
602 | // that fits in 32-bits? Commute them and use an AndINode. Don't |
603 | // convert masks which would cause a sign extension of the integer |
604 | // value. This check includes UI2L masks (0x00000000FFFFFFFF) which |
605 | // would be optimized away later in Identity. |
606 | if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) { |
607 | Node* andi = new AndINode(in1->in(1), phase->intcon(mask)); |
608 | andi = phase->transform(andi); |
609 | return new ConvI2LNode(andi); |
610 | } |
611 | |
612 | // Masking off sign bits? Dont make them! |
613 | if (op == Op_RShiftL) { |
614 | const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); |
615 | if( t12 && t12->is_con() ) { // Shift is by a constant |
616 | int shift = t12->get_con(); |
617 | shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
618 | const jlong sign_bits_mask = ~(((jlong)CONST64(1) << (jlong)(BitsPerJavaLong - shift)) -1); |
619 | // If the AND'ing of the 2 masks has no bits, then only original shifted |
620 | // bits survive. NO sign-extension bits survive the maskings. |
621 | if( (sign_bits_mask & mask) == 0 ) { |
622 | // Use zero-fill shift instead |
623 | Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2))); |
624 | return new AndLNode(zshift, in(2)); |
625 | } |
626 | } |
627 | } |
628 | |
629 | return MulNode::Ideal(phase, can_reshape); |
630 | } |
631 | |
632 | //============================================================================= |
633 | |
634 | static int getShiftCon(PhaseGVN *phase, Node *shiftNode, int retVal) { |
635 | const Type *t = phase->type(shiftNode->in(2)); |
636 | if (t == Type::TOP) return retVal; // Right input is dead. |
637 | const TypeInt *t2 = t->isa_int(); |
638 | if (!t2 || !t2->is_con()) return retVal; // Right input is a constant. |
639 | |
640 | return t2->get_con(); |
641 | } |
642 | |
643 | static int maskShiftAmount(PhaseGVN *phase, Node *shiftNode, int nBits) { |
644 | int shift = getShiftCon(phase, shiftNode, 0); |
645 | int maskedShift = shift & (nBits - 1); |
646 | |
647 | if (maskedShift == 0) return 0; // Let Identity() handle 0 shift count. |
648 | |
649 | if (shift != maskedShift) { |
650 | shiftNode->set_req(2, phase->intcon(maskedShift)); // Replace shift count with masked value. |
651 | phase->igvn_rehash_node_delayed(shiftNode); |
652 | } |
653 | |
654 | return maskedShift; |
655 | } |
656 | |
657 | //------------------------------Identity--------------------------------------- |
658 | Node* LShiftINode::Identity(PhaseGVN* phase) { |
659 | return ((getShiftCon(phase, this, -1) & (BitsPerJavaInteger - 1)) == 0) ? in(1) : this; |
660 | } |
661 | |
662 | //------------------------------Ideal------------------------------------------ |
663 | // If the right input is a constant, and the left input is an add of a |
664 | // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 |
665 | Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
666 | int con = maskShiftAmount(phase, this, BitsPerJavaInteger); |
667 | if (con == 0) { |
668 | return NULL; |
669 | } |
670 | |
671 | // Left input is an add of a constant? |
672 | Node *add1 = in(1); |
673 | int add1_op = add1->Opcode(); |
674 | if( add1_op == Op_AddI ) { // Left input is an add? |
675 | assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" ); |
676 | const TypeInt *t12 = phase->type(add1->in(2))->isa_int(); |
677 | if( t12 && t12->is_con() ){ // Left input is an add of a con? |
678 | // Transform is legal, but check for profit. Avoid breaking 'i2s' |
679 | // and 'i2b' patterns which typically fold into 'StoreC/StoreB'. |
680 | if( con < 16 ) { |
681 | // Compute X << con0 |
682 | Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) ); |
683 | // Compute X<<con0 + (con1<<con0) |
684 | return new AddINode( lsh, phase->intcon(t12->get_con() << con)); |
685 | } |
686 | } |
687 | } |
688 | |
689 | // Check for "(x>>c0)<<c0" which just masks off low bits |
690 | if( (add1_op == Op_RShiftI || add1_op == Op_URShiftI ) && |
691 | add1->in(2) == in(2) ) |
692 | // Convert to "(x & -(1<<c0))" |
693 | return new AndINode(add1->in(1),phase->intcon( -(1<<con))); |
694 | |
695 | // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits |
696 | if( add1_op == Op_AndI ) { |
697 | Node *add2 = add1->in(1); |
698 | int add2_op = add2->Opcode(); |
699 | if( (add2_op == Op_RShiftI || add2_op == Op_URShiftI ) && |
700 | add2->in(2) == in(2) ) { |
701 | // Convert to "(x & (Y<<c0))" |
702 | Node *y_sh = phase->transform( new LShiftINode( add1->in(2), in(2) ) ); |
703 | return new AndINode( add2->in(1), y_sh ); |
704 | } |
705 | } |
706 | |
707 | // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits |
708 | // before shifting them away. |
709 | const jint bits_mask = right_n_bits(BitsPerJavaInteger-con); |
710 | if( add1_op == Op_AndI && |
711 | phase->type(add1->in(2)) == TypeInt::make( bits_mask ) ) |
712 | return new LShiftINode( add1->in(1), in(2) ); |
713 | |
714 | return NULL; |
715 | } |
716 | |
717 | //------------------------------Value------------------------------------------ |
718 | // A LShiftINode shifts its input2 left by input1 amount. |
719 | const Type* LShiftINode::Value(PhaseGVN* phase) const { |
720 | const Type *t1 = phase->type( in(1) ); |
721 | const Type *t2 = phase->type( in(2) ); |
722 | // Either input is TOP ==> the result is TOP |
723 | if( t1 == Type::TOP ) return Type::TOP; |
724 | if( t2 == Type::TOP ) return Type::TOP; |
725 | |
726 | // Left input is ZERO ==> the result is ZERO. |
727 | if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; |
728 | // Shift by zero does nothing |
729 | if( t2 == TypeInt::ZERO ) return t1; |
730 | |
731 | // Either input is BOTTOM ==> the result is BOTTOM |
732 | if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) || |
733 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
734 | return TypeInt::INT; |
735 | |
736 | const TypeInt *r1 = t1->is_int(); // Handy access |
737 | const TypeInt *r2 = t2->is_int(); // Handy access |
738 | |
739 | if (!r2->is_con()) |
740 | return TypeInt::INT; |
741 | |
742 | uint shift = r2->get_con(); |
743 | shift &= BitsPerJavaInteger-1; // semantics of Java shifts |
744 | // Shift by a multiple of 32 does nothing: |
745 | if (shift == 0) return t1; |
746 | |
747 | // If the shift is a constant, shift the bounds of the type, |
748 | // unless this could lead to an overflow. |
749 | if (!r1->is_con()) { |
750 | jint lo = r1->_lo, hi = r1->_hi; |
751 | if (((lo << shift) >> shift) == lo && |
752 | ((hi << shift) >> shift) == hi) { |
753 | // No overflow. The range shifts up cleanly. |
754 | return TypeInt::make((jint)lo << (jint)shift, |
755 | (jint)hi << (jint)shift, |
756 | MAX2(r1->_widen,r2->_widen)); |
757 | } |
758 | return TypeInt::INT; |
759 | } |
760 | |
761 | return TypeInt::make( (jint)r1->get_con() << (jint)shift ); |
762 | } |
763 | |
764 | //============================================================================= |
765 | //------------------------------Identity--------------------------------------- |
766 | Node* LShiftLNode::Identity(PhaseGVN* phase) { |
767 | return ((getShiftCon(phase, this, -1) & (BitsPerJavaLong - 1)) == 0) ? in(1) : this; |
768 | } |
769 | |
770 | //------------------------------Ideal------------------------------------------ |
771 | // If the right input is a constant, and the left input is an add of a |
772 | // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 |
773 | Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
774 | int con = maskShiftAmount(phase, this, BitsPerJavaLong); |
775 | if (con == 0) { |
776 | return NULL; |
777 | } |
778 | |
779 | // Left input is an add of a constant? |
780 | Node *add1 = in(1); |
781 | int add1_op = add1->Opcode(); |
782 | if( add1_op == Op_AddL ) { // Left input is an add? |
783 | // Avoid dead data cycles from dead loops |
784 | assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" ); |
785 | const TypeLong *t12 = phase->type(add1->in(2))->isa_long(); |
786 | if( t12 && t12->is_con() ){ // Left input is an add of a con? |
787 | // Compute X << con0 |
788 | Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) ); |
789 | // Compute X<<con0 + (con1<<con0) |
790 | return new AddLNode( lsh, phase->longcon(t12->get_con() << con)); |
791 | } |
792 | } |
793 | |
794 | // Check for "(x>>c0)<<c0" which just masks off low bits |
795 | if( (add1_op == Op_RShiftL || add1_op == Op_URShiftL ) && |
796 | add1->in(2) == in(2) ) |
797 | // Convert to "(x & -(1<<c0))" |
798 | return new AndLNode(add1->in(1),phase->longcon( -(CONST64(1)<<con))); |
799 | |
800 | // Check for "((x>>c0) & Y)<<c0" which just masks off more low bits |
801 | if( add1_op == Op_AndL ) { |
802 | Node *add2 = add1->in(1); |
803 | int add2_op = add2->Opcode(); |
804 | if( (add2_op == Op_RShiftL || add2_op == Op_URShiftL ) && |
805 | add2->in(2) == in(2) ) { |
806 | // Convert to "(x & (Y<<c0))" |
807 | Node *y_sh = phase->transform( new LShiftLNode( add1->in(2), in(2) ) ); |
808 | return new AndLNode( add2->in(1), y_sh ); |
809 | } |
810 | } |
811 | |
812 | // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits |
813 | // before shifting them away. |
814 | const jlong bits_mask = jlong(max_julong >> con); |
815 | if( add1_op == Op_AndL && |
816 | phase->type(add1->in(2)) == TypeLong::make( bits_mask ) ) |
817 | return new LShiftLNode( add1->in(1), in(2) ); |
818 | |
819 | return NULL; |
820 | } |
821 | |
822 | //------------------------------Value------------------------------------------ |
823 | // A LShiftLNode shifts its input2 left by input1 amount. |
824 | const Type* LShiftLNode::Value(PhaseGVN* phase) const { |
825 | const Type *t1 = phase->type( in(1) ); |
826 | const Type *t2 = phase->type( in(2) ); |
827 | // Either input is TOP ==> the result is TOP |
828 | if( t1 == Type::TOP ) return Type::TOP; |
829 | if( t2 == Type::TOP ) return Type::TOP; |
830 | |
831 | // Left input is ZERO ==> the result is ZERO. |
832 | if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; |
833 | // Shift by zero does nothing |
834 | if( t2 == TypeInt::ZERO ) return t1; |
835 | |
836 | // Either input is BOTTOM ==> the result is BOTTOM |
837 | if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) || |
838 | (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) |
839 | return TypeLong::LONG; |
840 | |
841 | const TypeLong *r1 = t1->is_long(); // Handy access |
842 | const TypeInt *r2 = t2->is_int(); // Handy access |
843 | |
844 | if (!r2->is_con()) |
845 | return TypeLong::LONG; |
846 | |
847 | uint shift = r2->get_con(); |
848 | shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
849 | // Shift by a multiple of 64 does nothing: |
850 | if (shift == 0) return t1; |
851 | |
852 | // If the shift is a constant, shift the bounds of the type, |
853 | // unless this could lead to an overflow. |
854 | if (!r1->is_con()) { |
855 | jlong lo = r1->_lo, hi = r1->_hi; |
856 | if (((lo << shift) >> shift) == lo && |
857 | ((hi << shift) >> shift) == hi) { |
858 | // No overflow. The range shifts up cleanly. |
859 | return TypeLong::make((jlong)lo << (jint)shift, |
860 | (jlong)hi << (jint)shift, |
861 | MAX2(r1->_widen,r2->_widen)); |
862 | } |
863 | return TypeLong::LONG; |
864 | } |
865 | |
866 | return TypeLong::make( (jlong)r1->get_con() << (jint)shift ); |
867 | } |
868 | |
869 | //============================================================================= |
870 | //------------------------------Identity--------------------------------------- |
871 | Node* RShiftINode::Identity(PhaseGVN* phase) { |
872 | int shift = getShiftCon(phase, this, -1); |
873 | if (shift == -1) return this; |
874 | if ((shift & (BitsPerJavaInteger - 1)) == 0) return in(1); |
875 | |
876 | // Check for useless sign-masking |
877 | if (in(1)->Opcode() == Op_LShiftI && |
878 | in(1)->req() == 3 && |
879 | in(1)->in(2) == in(2)) { |
880 | shift &= BitsPerJavaInteger-1; // semantics of Java shifts |
881 | // Compute masks for which this shifting doesn't change |
882 | int lo = (-1 << (BitsPerJavaInteger - ((uint)shift)-1)); // FFFF8000 |
883 | int hi = ~lo; // 00007FFF |
884 | const TypeInt *t11 = phase->type(in(1)->in(1))->isa_int(); |
885 | if (!t11) return this; |
886 | // Does actual value fit inside of mask? |
887 | if (lo <= t11->_lo && t11->_hi <= hi) { |
888 | return in(1)->in(1); // Then shifting is a nop |
889 | } |
890 | } |
891 | |
892 | return this; |
893 | } |
894 | |
895 | //------------------------------Ideal------------------------------------------ |
896 | Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
897 | // Inputs may be TOP if they are dead. |
898 | const TypeInt *t1 = phase->type(in(1))->isa_int(); |
899 | if (!t1) return NULL; // Left input is an integer |
900 | const TypeInt *t3; // type of in(1).in(2) |
901 | int shift = maskShiftAmount(phase, this, BitsPerJavaInteger); |
902 | if (shift == 0) { |
903 | return NULL; |
904 | } |
905 | |
906 | // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller. |
907 | // Such expressions arise normally from shift chains like (byte)(x >> 24). |
908 | const Node *mask = in(1); |
909 | if( mask->Opcode() == Op_AndI && |
910 | (t3 = phase->type(mask->in(2))->isa_int()) && |
911 | t3->is_con() ) { |
912 | Node *x = mask->in(1); |
913 | jint maskbits = t3->get_con(); |
914 | // Convert to "(x >> shift) & (mask >> shift)" |
915 | Node *shr_nomask = phase->transform( new RShiftINode(mask->in(1), in(2)) ); |
916 | return new AndINode(shr_nomask, phase->intcon( maskbits >> shift)); |
917 | } |
918 | |
919 | // Check for "(short[i] <<16)>>16" which simply sign-extends |
920 | const Node *shl = in(1); |
921 | if( shl->Opcode() != Op_LShiftI ) return NULL; |
922 | |
923 | if( shift == 16 && |
924 | (t3 = phase->type(shl->in(2))->isa_int()) && |
925 | t3->is_con(16) ) { |
926 | Node *ld = shl->in(1); |
927 | if( ld->Opcode() == Op_LoadS ) { |
928 | // Sign extension is just useless here. Return a RShiftI of zero instead |
929 | // returning 'ld' directly. We cannot return an old Node directly as |
930 | // that is the job of 'Identity' calls and Identity calls only work on |
931 | // direct inputs ('ld' is an extra Node removed from 'this'). The |
932 | // combined optimization requires Identity only return direct inputs. |
933 | set_req(1, ld); |
934 | set_req(2, phase->intcon(0)); |
935 | return this; |
936 | } |
937 | else if( can_reshape && |
938 | ld->Opcode() == Op_LoadUS && |
939 | ld->outcnt() == 1 && ld->unique_out() == shl) |
940 | // Replace zero-extension-load with sign-extension-load |
941 | return ld->as_Load()->convert_to_signed_load(*phase); |
942 | } |
943 | |
944 | // Check for "(byte[i] <<24)>>24" which simply sign-extends |
945 | if( shift == 24 && |
946 | (t3 = phase->type(shl->in(2))->isa_int()) && |
947 | t3->is_con(24) ) { |
948 | Node *ld = shl->in(1); |
949 | if( ld->Opcode() == Op_LoadB ) { |
950 | // Sign extension is just useless here |
951 | set_req(1, ld); |
952 | set_req(2, phase->intcon(0)); |
953 | return this; |
954 | } |
955 | } |
956 | |
957 | return NULL; |
958 | } |
959 | |
960 | //------------------------------Value------------------------------------------ |
961 | // A RShiftINode shifts its input2 right by input1 amount. |
962 | const Type* RShiftINode::Value(PhaseGVN* phase) const { |
963 | const Type *t1 = phase->type( in(1) ); |
964 | const Type *t2 = phase->type( in(2) ); |
965 | // Either input is TOP ==> the result is TOP |
966 | if( t1 == Type::TOP ) return Type::TOP; |
967 | if( t2 == Type::TOP ) return Type::TOP; |
968 | |
969 | // Left input is ZERO ==> the result is ZERO. |
970 | if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; |
971 | // Shift by zero does nothing |
972 | if( t2 == TypeInt::ZERO ) return t1; |
973 | |
974 | // Either input is BOTTOM ==> the result is BOTTOM |
975 | if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) |
976 | return TypeInt::INT; |
977 | |
978 | if (t2 == TypeInt::INT) |
979 | return TypeInt::INT; |
980 | |
981 | const TypeInt *r1 = t1->is_int(); // Handy access |
982 | const TypeInt *r2 = t2->is_int(); // Handy access |
983 | |
984 | // If the shift is a constant, just shift the bounds of the type. |
985 | // For example, if the shift is 31, we just propagate sign bits. |
986 | if (r2->is_con()) { |
987 | uint shift = r2->get_con(); |
988 | shift &= BitsPerJavaInteger-1; // semantics of Java shifts |
989 | // Shift by a multiple of 32 does nothing: |
990 | if (shift == 0) return t1; |
991 | // Calculate reasonably aggressive bounds for the result. |
992 | // This is necessary if we are to correctly type things |
993 | // like (x<<24>>24) == ((byte)x). |
994 | jint lo = (jint)r1->_lo >> (jint)shift; |
995 | jint hi = (jint)r1->_hi >> (jint)shift; |
996 | assert(lo <= hi, "must have valid bounds" ); |
997 | const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); |
998 | #ifdef ASSERT |
999 | // Make sure we get the sign-capture idiom correct. |
1000 | if (shift == BitsPerJavaInteger-1) { |
1001 | if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0" ); |
1002 | if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1" ); |
1003 | } |
1004 | #endif |
1005 | return ti; |
1006 | } |
1007 | |
1008 | if( !r1->is_con() || !r2->is_con() ) |
1009 | return TypeInt::INT; |
1010 | |
1011 | // Signed shift right |
1012 | return TypeInt::make( r1->get_con() >> (r2->get_con()&31) ); |
1013 | } |
1014 | |
1015 | //============================================================================= |
1016 | //------------------------------Identity--------------------------------------- |
1017 | Node* RShiftLNode::Identity(PhaseGVN* phase) { |
1018 | const TypeInt *ti = phase->type(in(2))->isa_int(); // Shift count is an int. |
1019 | return (ti && ti->is_con() && (ti->get_con() & (BitsPerJavaLong - 1)) == 0) ? in(1) : this; |
1020 | } |
1021 | |
1022 | //------------------------------Value------------------------------------------ |
1023 | // A RShiftLNode shifts its input2 right by input1 amount. |
1024 | const Type* RShiftLNode::Value(PhaseGVN* phase) const { |
1025 | const Type *t1 = phase->type( in(1) ); |
1026 | const Type *t2 = phase->type( in(2) ); |
1027 | // Either input is TOP ==> the result is TOP |
1028 | if( t1 == Type::TOP ) return Type::TOP; |
1029 | if( t2 == Type::TOP ) return Type::TOP; |
1030 | |
1031 | // Left input is ZERO ==> the result is ZERO. |
1032 | if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; |
1033 | // Shift by zero does nothing |
1034 | if( t2 == TypeInt::ZERO ) return t1; |
1035 | |
1036 | // Either input is BOTTOM ==> the result is BOTTOM |
1037 | if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) |
1038 | return TypeLong::LONG; |
1039 | |
1040 | if (t2 == TypeInt::INT) |
1041 | return TypeLong::LONG; |
1042 | |
1043 | const TypeLong *r1 = t1->is_long(); // Handy access |
1044 | const TypeInt *r2 = t2->is_int (); // Handy access |
1045 | |
1046 | // If the shift is a constant, just shift the bounds of the type. |
1047 | // For example, if the shift is 63, we just propagate sign bits. |
1048 | if (r2->is_con()) { |
1049 | uint shift = r2->get_con(); |
1050 | shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts |
1051 | // Shift by a multiple of 64 does nothing: |
1052 | if (shift == 0) return t1; |
1053 | // Calculate reasonably aggressive bounds for the result. |
1054 | // This is necessary if we are to correctly type things |
1055 | // like (x<<24>>24) == ((byte)x). |
1056 | jlong lo = (jlong)r1->_lo >> (jlong)shift; |
1057 | jlong hi = (jlong)r1->_hi >> (jlong)shift; |
1058 | assert(lo <= hi, "must have valid bounds" ); |
1059 | const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); |
1060 | #ifdef ASSERT |
1061 | // Make sure we get the sign-capture idiom correct. |
1062 | if (shift == (2*BitsPerJavaInteger)-1) { |
1063 | if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0" ); |
1064 | if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1" ); |
1065 | } |
1066 | #endif |
1067 | return tl; |
1068 | } |
1069 | |
1070 | return TypeLong::LONG; // Give up |
1071 | } |
1072 | |
1073 | //============================================================================= |
1074 | //------------------------------Identity--------------------------------------- |
1075 | Node* URShiftINode::Identity(PhaseGVN* phase) { |
1076 | int shift = getShiftCon(phase, this, -1); |
1077 | if ((shift & (BitsPerJavaInteger - 1)) == 0) return in(1); |
1078 | |
1079 | // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x". |
1080 | // Happens during new-array length computation. |
1081 | // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)] |
1082 | Node *add = in(1); |
1083 | if (add->Opcode() == Op_AddI) { |
1084 | const TypeInt *t2 = phase->type(add->in(2))->isa_int(); |
1085 | if (t2 && t2->is_con(wordSize - 1) && |
1086 | add->in(1)->Opcode() == Op_LShiftI) { |
1087 | // Check that shift_counts are LogBytesPerWord. |
1088 | Node *lshift_count = add->in(1)->in(2); |
1089 | const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int(); |
1090 | if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) && |
1091 | t_lshift_count == phase->type(in(2))) { |
1092 | Node *x = add->in(1)->in(1); |
1093 | const TypeInt *t_x = phase->type(x)->isa_int(); |
1094 | if (t_x != NULL && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) { |
1095 | return x; |
1096 | } |
1097 | } |
1098 | } |
1099 | } |
1100 | |
1101 | return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this; |
1102 | } |
1103 | |
1104 | //------------------------------Ideal------------------------------------------ |
1105 | Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1106 | int con = maskShiftAmount(phase, this, BitsPerJavaInteger); |
1107 | if (con == 0) { |
1108 | return NULL; |
1109 | } |
1110 | |
1111 | // We'll be wanting the right-shift amount as a mask of that many bits |
1112 | const int mask = right_n_bits(BitsPerJavaInteger - con); |
1113 | |
1114 | int in1_op = in(1)->Opcode(); |
1115 | |
1116 | // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32 |
1117 | if( in1_op == Op_URShiftI ) { |
1118 | const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int(); |
1119 | if( t12 && t12->is_con() ) { // Right input is a constant |
1120 | assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" ); |
1121 | const int con2 = t12->get_con() & 31; // Shift count is always masked |
1122 | const int con3 = con+con2; |
1123 | if( con3 < 32 ) // Only merge shifts if total is < 32 |
1124 | return new URShiftINode( in(1)->in(1), phase->intcon(con3) ); |
1125 | } |
1126 | } |
1127 | |
1128 | // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z |
1129 | // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". |
1130 | // If Q is "X << z" the rounding is useless. Look for patterns like |
1131 | // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. |
1132 | Node *add = in(1); |
1133 | const TypeInt *t2 = phase->type(in(2))->isa_int(); |
1134 | if (in1_op == Op_AddI) { |
1135 | Node *lshl = add->in(1); |
1136 | if( lshl->Opcode() == Op_LShiftI && |
1137 | phase->type(lshl->in(2)) == t2 ) { |
1138 | Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) ); |
1139 | Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) ); |
1140 | return new AndINode( sum, phase->intcon(mask) ); |
1141 | } |
1142 | } |
1143 | |
1144 | // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) |
1145 | // This shortens the mask. Also, if we are extracting a high byte and |
1146 | // storing it to a buffer, the mask will be removed completely. |
1147 | Node *andi = in(1); |
1148 | if( in1_op == Op_AndI ) { |
1149 | const TypeInt *t3 = phase->type( andi->in(2) )->isa_int(); |
1150 | if( t3 && t3->is_con() ) { // Right input is a constant |
1151 | jint mask2 = t3->get_con(); |
1152 | mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) |
1153 | Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) ); |
1154 | return new AndINode(newshr, phase->intcon(mask2)); |
1155 | // The negative values are easier to materialize than positive ones. |
1156 | // A typical case from address arithmetic is ((x & ~15) >> 4). |
1157 | // It's better to change that to ((x >> 4) & ~0) versus |
1158 | // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64. |
1159 | } |
1160 | } |
1161 | |
1162 | // Check for "(X << z ) >>> z" which simply zero-extends |
1163 | Node *shl = in(1); |
1164 | if( in1_op == Op_LShiftI && |
1165 | phase->type(shl->in(2)) == t2 ) |
1166 | return new AndINode( shl->in(1), phase->intcon(mask) ); |
1167 | |
1168 | return NULL; |
1169 | } |
1170 | |
1171 | //------------------------------Value------------------------------------------ |
1172 | // A URShiftINode shifts its input2 right by input1 amount. |
1173 | const Type* URShiftINode::Value(PhaseGVN* phase) const { |
1174 | // (This is a near clone of RShiftINode::Value.) |
1175 | const Type *t1 = phase->type( in(1) ); |
1176 | const Type *t2 = phase->type( in(2) ); |
1177 | // Either input is TOP ==> the result is TOP |
1178 | if( t1 == Type::TOP ) return Type::TOP; |
1179 | if( t2 == Type::TOP ) return Type::TOP; |
1180 | |
1181 | // Left input is ZERO ==> the result is ZERO. |
1182 | if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; |
1183 | // Shift by zero does nothing |
1184 | if( t2 == TypeInt::ZERO ) return t1; |
1185 | |
1186 | // Either input is BOTTOM ==> the result is BOTTOM |
1187 | if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) |
1188 | return TypeInt::INT; |
1189 | |
1190 | if (t2 == TypeInt::INT) |
1191 | return TypeInt::INT; |
1192 | |
1193 | const TypeInt *r1 = t1->is_int(); // Handy access |
1194 | const TypeInt *r2 = t2->is_int(); // Handy access |
1195 | |
1196 | if (r2->is_con()) { |
1197 | uint shift = r2->get_con(); |
1198 | shift &= BitsPerJavaInteger-1; // semantics of Java shifts |
1199 | // Shift by a multiple of 32 does nothing: |
1200 | if (shift == 0) return t1; |
1201 | // Calculate reasonably aggressive bounds for the result. |
1202 | jint lo = (juint)r1->_lo >> (juint)shift; |
1203 | jint hi = (juint)r1->_hi >> (juint)shift; |
1204 | if (r1->_hi >= 0 && r1->_lo < 0) { |
1205 | // If the type has both negative and positive values, |
1206 | // there are two separate sub-domains to worry about: |
1207 | // The positive half and the negative half. |
1208 | jint neg_lo = lo; |
1209 | jint neg_hi = (juint)-1 >> (juint)shift; |
1210 | jint pos_lo = (juint) 0 >> (juint)shift; |
1211 | jint pos_hi = hi; |
1212 | lo = MIN2(neg_lo, pos_lo); // == 0 |
1213 | hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; |
1214 | } |
1215 | assert(lo <= hi, "must have valid bounds" ); |
1216 | const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); |
1217 | #ifdef ASSERT |
1218 | // Make sure we get the sign-capture idiom correct. |
1219 | if (shift == BitsPerJavaInteger-1) { |
1220 | if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0" ); |
1221 | if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1" ); |
1222 | } |
1223 | #endif |
1224 | return ti; |
1225 | } |
1226 | |
1227 | // |
1228 | // Do not support shifted oops in info for GC |
1229 | // |
1230 | // else if( t1->base() == Type::InstPtr ) { |
1231 | // |
1232 | // const TypeInstPtr *o = t1->is_instptr(); |
1233 | // if( t1->singleton() ) |
1234 | // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift ); |
1235 | // } |
1236 | // else if( t1->base() == Type::KlassPtr ) { |
1237 | // const TypeKlassPtr *o = t1->is_klassptr(); |
1238 | // if( t1->singleton() ) |
1239 | // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift ); |
1240 | // } |
1241 | |
1242 | return TypeInt::INT; |
1243 | } |
1244 | |
1245 | //============================================================================= |
1246 | //------------------------------Identity--------------------------------------- |
1247 | Node* URShiftLNode::Identity(PhaseGVN* phase) { |
1248 | return ((getShiftCon(phase, this, -1) & (BitsPerJavaLong - 1)) == 0) ? in(1) : this; |
1249 | } |
1250 | |
1251 | //------------------------------Ideal------------------------------------------ |
1252 | Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
1253 | int con = maskShiftAmount(phase, this, BitsPerJavaLong); |
1254 | if (con == 0) { |
1255 | return NULL; |
1256 | } |
1257 | |
1258 | // We'll be wanting the right-shift amount as a mask of that many bits |
1259 | const jlong mask = jlong(max_julong >> con); |
1260 | |
1261 | // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z |
1262 | // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". |
1263 | // If Q is "X << z" the rounding is useless. Look for patterns like |
1264 | // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. |
1265 | Node *add = in(1); |
1266 | const TypeInt *t2 = phase->type(in(2))->isa_int(); |
1267 | if (add->Opcode() == Op_AddL) { |
1268 | Node *lshl = add->in(1); |
1269 | if( lshl->Opcode() == Op_LShiftL && |
1270 | phase->type(lshl->in(2)) == t2 ) { |
1271 | Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) ); |
1272 | Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) ); |
1273 | return new AndLNode( sum, phase->longcon(mask) ); |
1274 | } |
1275 | } |
1276 | |
1277 | // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) |
1278 | // This shortens the mask. Also, if we are extracting a high byte and |
1279 | // storing it to a buffer, the mask will be removed completely. |
1280 | Node *andi = in(1); |
1281 | if( andi->Opcode() == Op_AndL ) { |
1282 | const TypeLong *t3 = phase->type( andi->in(2) )->isa_long(); |
1283 | if( t3 && t3->is_con() ) { // Right input is a constant |
1284 | jlong mask2 = t3->get_con(); |
1285 | mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) |
1286 | Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) ); |
1287 | return new AndLNode(newshr, phase->longcon(mask2)); |
1288 | } |
1289 | } |
1290 | |
1291 | // Check for "(X << z ) >>> z" which simply zero-extends |
1292 | Node *shl = in(1); |
1293 | if( shl->Opcode() == Op_LShiftL && |
1294 | phase->type(shl->in(2)) == t2 ) |
1295 | return new AndLNode( shl->in(1), phase->longcon(mask) ); |
1296 | |
1297 | return NULL; |
1298 | } |
1299 | |
1300 | //------------------------------Value------------------------------------------ |
1301 | // A URShiftINode shifts its input2 right by input1 amount. |
1302 | const Type* URShiftLNode::Value(PhaseGVN* phase) const { |
1303 | // (This is a near clone of RShiftLNode::Value.) |
1304 | const Type *t1 = phase->type( in(1) ); |
1305 | const Type *t2 = phase->type( in(2) ); |
1306 | // Either input is TOP ==> the result is TOP |
1307 | if( t1 == Type::TOP ) return Type::TOP; |
1308 | if( t2 == Type::TOP ) return Type::TOP; |
1309 | |
1310 | // Left input is ZERO ==> the result is ZERO. |
1311 | if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; |
1312 | // Shift by zero does nothing |
1313 | if( t2 == TypeInt::ZERO ) return t1; |
1314 | |
1315 | // Either input is BOTTOM ==> the result is BOTTOM |
1316 | if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) |
1317 | return TypeLong::LONG; |
1318 | |
1319 | if (t2 == TypeInt::INT) |
1320 | return TypeLong::LONG; |
1321 | |
1322 | const TypeLong *r1 = t1->is_long(); // Handy access |
1323 | const TypeInt *r2 = t2->is_int (); // Handy access |
1324 | |
1325 | if (r2->is_con()) { |
1326 | uint shift = r2->get_con(); |
1327 | shift &= BitsPerJavaLong - 1; // semantics of Java shifts |
1328 | // Shift by a multiple of 64 does nothing: |
1329 | if (shift == 0) return t1; |
1330 | // Calculate reasonably aggressive bounds for the result. |
1331 | jlong lo = (julong)r1->_lo >> (juint)shift; |
1332 | jlong hi = (julong)r1->_hi >> (juint)shift; |
1333 | if (r1->_hi >= 0 && r1->_lo < 0) { |
1334 | // If the type has both negative and positive values, |
1335 | // there are two separate sub-domains to worry about: |
1336 | // The positive half and the negative half. |
1337 | jlong neg_lo = lo; |
1338 | jlong neg_hi = (julong)-1 >> (juint)shift; |
1339 | jlong pos_lo = (julong) 0 >> (juint)shift; |
1340 | jlong pos_hi = hi; |
1341 | //lo = MIN2(neg_lo, pos_lo); // == 0 |
1342 | lo = neg_lo < pos_lo ? neg_lo : pos_lo; |
1343 | //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; |
1344 | hi = neg_hi > pos_hi ? neg_hi : pos_hi; |
1345 | } |
1346 | assert(lo <= hi, "must have valid bounds" ); |
1347 | const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); |
1348 | #ifdef ASSERT |
1349 | // Make sure we get the sign-capture idiom correct. |
1350 | if (shift == BitsPerJavaLong - 1) { |
1351 | if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0" ); |
1352 | if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1" ); |
1353 | } |
1354 | #endif |
1355 | return tl; |
1356 | } |
1357 | |
1358 | return TypeLong::LONG; // Give up |
1359 | } |
1360 | |
1361 | //============================================================================= |
1362 | //------------------------------Value------------------------------------------ |
1363 | const Type* FmaDNode::Value(PhaseGVN* phase) const { |
1364 | const Type *t1 = phase->type(in(1)); |
1365 | if (t1 == Type::TOP) return Type::TOP; |
1366 | if (t1->base() != Type::DoubleCon) return Type::DOUBLE; |
1367 | const Type *t2 = phase->type(in(2)); |
1368 | if (t2 == Type::TOP) return Type::TOP; |
1369 | if (t2->base() != Type::DoubleCon) return Type::DOUBLE; |
1370 | const Type *t3 = phase->type(in(3)); |
1371 | if (t3 == Type::TOP) return Type::TOP; |
1372 | if (t3->base() != Type::DoubleCon) return Type::DOUBLE; |
1373 | #ifndef __STDC_IEC_559__ |
1374 | return Type::DOUBLE; |
1375 | #else |
1376 | double d1 = t1->getd(); |
1377 | double d2 = t2->getd(); |
1378 | double d3 = t3->getd(); |
1379 | return TypeD::make(fma(d1, d2, d3)); |
1380 | #endif |
1381 | } |
1382 | |
1383 | //============================================================================= |
1384 | //------------------------------Value------------------------------------------ |
1385 | const Type* FmaFNode::Value(PhaseGVN* phase) const { |
1386 | const Type *t1 = phase->type(in(1)); |
1387 | if (t1 == Type::TOP) return Type::TOP; |
1388 | if (t1->base() != Type::FloatCon) return Type::FLOAT; |
1389 | const Type *t2 = phase->type(in(2)); |
1390 | if (t2 == Type::TOP) return Type::TOP; |
1391 | if (t2->base() != Type::FloatCon) return Type::FLOAT; |
1392 | const Type *t3 = phase->type(in(3)); |
1393 | if (t3 == Type::TOP) return Type::TOP; |
1394 | if (t3->base() != Type::FloatCon) return Type::FLOAT; |
1395 | #ifndef __STDC_IEC_559__ |
1396 | return Type::FLOAT; |
1397 | #else |
1398 | float f1 = t1->getf(); |
1399 | float f2 = t2->getf(); |
1400 | float f3 = t3->getf(); |
1401 | return TypeF::make(fma(f1, f2, f3)); |
1402 | #endif |
1403 | } |
1404 | |
1405 | //============================================================================= |
1406 | //------------------------------hash------------------------------------------- |
1407 | // Hash function for MulAddS2INode. Operation is commutative with commutative pairs. |
1408 | // The hash function must return the same value when edge swapping is performed. |
1409 | uint MulAddS2INode::hash() const { |
1410 | return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode(); |
1411 | } |
1412 | |
1413 | |