1 /*
2 * Copyright (c) 1999, 2013, 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 "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mathexactnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "prims/nativeLookup.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "trace/traceMacros.hpp"
43
44 class LibraryIntrinsic : public InlineCallGenerator {
45 // Extend the set of intrinsics known to the runtime:
46 public:
47 private:
48 bool _is_virtual;
49 bool _is_predicted;
50 bool _does_virtual_dispatch;
51 vmIntrinsics::ID _intrinsic_id;
52
53 public:
54 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
55 : InlineCallGenerator(m),
56 _is_virtual(is_virtual),
57 _is_predicted(is_predicted),
58 _does_virtual_dispatch(does_virtual_dispatch),
59 _intrinsic_id(id)
60 {
61 }
62 virtual bool is_intrinsic() const { return true; }
63 virtual bool is_virtual() const { return _is_virtual; }
64 virtual bool is_predicted() const { return _is_predicted; }
65 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
66 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
67 virtual Node* generate_predicate(JVMState* jvms);
68 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
69 };
70
71
72 // Local helper class for LibraryIntrinsic:
73 class LibraryCallKit : public GraphKit {
74 private:
75 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
76 Node* _result; // the result node, if any
77 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
78
79 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
80
81 public:
82 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
83 : GraphKit(jvms),
84 _intrinsic(intrinsic),
85 _result(NULL)
86 {
87 // Check if this is a root compile. In that case we don't have a caller.
88 if (!jvms->has_method()) {
89 _reexecute_sp = sp();
90 } else {
91 // Find out how many arguments the interpreter needs when deoptimizing
92 // and save the stack pointer value so it can used by uncommon_trap.
93 // We find the argument count by looking at the declared signature.
94 bool ignored_will_link;
95 ciSignature* declared_signature = NULL;
96 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
97 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
98 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
99 }
100 }
101
102 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
103
104 ciMethod* caller() const { return jvms()->method(); }
105 int bci() const { return jvms()->bci(); }
106 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
107 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
108 ciMethod* callee() const { return _intrinsic->method(); }
109
110 bool try_to_inline();
111 Node* try_to_predicate();
112
113 void push_result() {
114 // Push the result onto the stack.
115 if (!stopped() && result() != NULL) {
116 BasicType bt = result()->bottom_type()->basic_type();
117 push_node(bt, result());
118 }
119 }
120
121 private:
122 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
123 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
124 }
125
126 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
127 void set_result(RegionNode* region, PhiNode* value);
128 Node* result() { return _result; }
129
130 virtual int reexecute_sp() { return _reexecute_sp; }
131
132 // Helper functions to inline natives
133 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
134 Node* generate_slow_guard(Node* test, RegionNode* region);
135 Node* generate_fair_guard(Node* test, RegionNode* region);
136 Node* generate_negative_guard(Node* index, RegionNode* region,
137 // resulting CastII of index:
138 Node* *pos_index = NULL);
139 Node* generate_nonpositive_guard(Node* index, bool never_negative,
140 // resulting CastII of index:
141 Node* *pos_index = NULL);
142 Node* generate_limit_guard(Node* offset, Node* subseq_length,
143 Node* array_length,
144 RegionNode* region);
145 Node* generate_current_thread(Node* &tls_output);
146 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
147 bool disjoint_bases, const char* &name, bool dest_uninitialized);
148 Node* load_mirror_from_klass(Node* klass);
149 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
150 RegionNode* region, int null_path,
151 int offset);
152 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
153 RegionNode* region, int null_path) {
154 int offset = java_lang_Class::klass_offset_in_bytes();
155 return load_klass_from_mirror_common(mirror, never_see_null,
156 region, null_path,
157 offset);
158 }
159 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::array_klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* generate_access_flags_guard(Node* kls,
167 int modifier_mask, int modifier_bits,
168 RegionNode* region);
169 Node* generate_interface_guard(Node* kls, RegionNode* region);
170 Node* generate_array_guard(Node* kls, RegionNode* region) {
171 return generate_array_guard_common(kls, region, false, false);
172 }
173 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
174 return generate_array_guard_common(kls, region, false, true);
175 }
176 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, true, false);
178 }
179 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, true, true);
181 }
182 Node* generate_array_guard_common(Node* kls, RegionNode* region,
183 bool obj_array, bool not_array);
184 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
185 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
186 bool is_virtual = false, bool is_static = false);
187 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
188 return generate_method_call(method_id, false, true);
189 }
190 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
191 return generate_method_call(method_id, true, false);
192 }
193 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
194
195 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
196 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
197 bool inline_string_compareTo();
198 bool inline_string_indexOf();
199 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
200 bool inline_string_equals();
201 Node* round_double_node(Node* n);
202 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
203 bool inline_math_native(vmIntrinsics::ID id);
204 bool inline_trig(vmIntrinsics::ID id);
205 bool inline_math(vmIntrinsics::ID id);
206 template <typename OverflowOp>
207 bool inline_math_overflow(Node* arg1, Node* arg2);
208 template <typename OverflowOp>
209 bool inline_math_overflow(bool is_unary);
210 void inline_math_mathExact(Node* math, Node* test);
211 bool inline_math_addExactI(bool is_increment);
212 bool inline_math_addExactL(bool is_increment);
213 bool inline_math_multiplyExactI();
214 bool inline_math_multiplyExactL();
215 bool inline_math_negateExactI();
216 bool inline_math_negateExactL();
217 bool inline_math_subtractExactI(bool is_decrement);
218 bool inline_math_subtractExactL(bool is_decrement);
219 bool inline_exp();
220 bool inline_pow();
221 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
222 bool inline_min_max(vmIntrinsics::ID id);
223 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
224 // This returns Type::AnyPtr, RawPtr, or OopPtr.
225 int classify_unsafe_addr(Node* &base, Node* &offset);
226 Node* make_unsafe_address(Node* base, Node* offset);
227 // Helper for inline_unsafe_access.
228 // Generates the guards that check whether the result of
229 // Unsafe.getObject should be recorded in an SATB log buffer.
230 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
231 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
232 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
233 static bool klass_needs_init_guard(Node* kls);
234 bool inline_unsafe_allocate();
235 bool inline_unsafe_copyMemory();
236 bool inline_native_currentThread();
237 #ifdef TRACE_HAVE_INTRINSICS
238 bool inline_native_classID();
239 bool inline_native_threadID();
240 #endif
241 bool inline_native_time_funcs(address method, const char* funcName);
242 bool inline_native_isInterrupted();
243 bool inline_native_Class_query(vmIntrinsics::ID id);
244 bool inline_native_subtype_check();
245
246 bool inline_native_newArray();
247 bool inline_native_getLength();
248 bool inline_array_copyOf(bool is_copyOfRange);
249 bool inline_array_equals();
250 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
251 bool inline_native_clone(bool is_virtual);
252 bool inline_native_Reflection_getCallerClass();
253 // Helper function for inlining native object hash method
254 bool inline_native_hashcode(bool is_virtual, bool is_static);
255 bool inline_native_getClass();
256
257 // Helper functions for inlining arraycopy
258 bool inline_arraycopy();
259 void generate_arraycopy(const TypePtr* adr_type,
260 BasicType basic_elem_type,
261 Node* src, Node* src_offset,
262 Node* dest, Node* dest_offset,
263 Node* copy_length,
264 bool disjoint_bases = false,
265 bool length_never_negative = false,
266 RegionNode* slow_region = NULL);
267 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
268 RegionNode* slow_region);
269 void generate_clear_array(const TypePtr* adr_type,
270 Node* dest,
271 BasicType basic_elem_type,
272 Node* slice_off,
273 Node* slice_len,
274 Node* slice_end);
275 bool generate_block_arraycopy(const TypePtr* adr_type,
276 BasicType basic_elem_type,
277 AllocateNode* alloc,
278 Node* src, Node* src_offset,
279 Node* dest, Node* dest_offset,
280 Node* dest_size, bool dest_uninitialized);
281 void generate_slow_arraycopy(const TypePtr* adr_type,
282 Node* src, Node* src_offset,
283 Node* dest, Node* dest_offset,
284 Node* copy_length, bool dest_uninitialized);
285 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
286 Node* dest_elem_klass,
287 Node* src, Node* src_offset,
288 Node* dest, Node* dest_offset,
289 Node* copy_length, bool dest_uninitialized);
290 Node* generate_generic_arraycopy(const TypePtr* adr_type,
291 Node* src, Node* src_offset,
292 Node* dest, Node* dest_offset,
293 Node* copy_length, bool dest_uninitialized);
294 void generate_unchecked_arraycopy(const TypePtr* adr_type,
295 BasicType basic_elem_type,
296 bool disjoint_bases,
297 Node* src, Node* src_offset,
298 Node* dest, Node* dest_offset,
299 Node* copy_length, bool dest_uninitialized);
300 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
301 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
302 bool inline_unsafe_ordered_store(BasicType type);
303 bool inline_unsafe_fence(vmIntrinsics::ID id);
304 bool inline_fp_conversions(vmIntrinsics::ID id);
305 bool inline_number_methods(vmIntrinsics::ID id);
306 bool inline_reference_get();
307 bool inline_aescrypt_Block(vmIntrinsics::ID id);
308 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
309 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
310 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
311 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
312 bool inline_encodeISOArray();
313 bool inline_updateCRC32();
314 bool inline_updateBytesCRC32();
315 bool inline_updateByteBufferCRC32();
316 };
317
318
319 //---------------------------make_vm_intrinsic----------------------------
320 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
321 vmIntrinsics::ID id = m->intrinsic_id();
322 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
323
324 if (DisableIntrinsic[0] != '\0'
325 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
326 // disabled by a user request on the command line:
327 // example: -XX:DisableIntrinsic=_hashCode,_getClass
328 return NULL;
329 }
330
331 if (!m->is_loaded()) {
332 // do not attempt to inline unloaded methods
333 return NULL;
334 }
335
336 // Only a few intrinsics implement a virtual dispatch.
337 // They are expensive calls which are also frequently overridden.
338 if (is_virtual) {
339 switch (id) {
340 case vmIntrinsics::_hashCode:
341 case vmIntrinsics::_clone:
342 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
343 break;
344 default:
345 return NULL;
346 }
347 }
348
349 // -XX:-InlineNatives disables nearly all intrinsics:
350 if (!InlineNatives) {
351 switch (id) {
352 case vmIntrinsics::_indexOf:
353 case vmIntrinsics::_compareTo:
354 case vmIntrinsics::_equals:
355 case vmIntrinsics::_equalsC:
356 case vmIntrinsics::_getAndAddInt:
357 case vmIntrinsics::_getAndAddLong:
358 case vmIntrinsics::_getAndSetInt:
359 case vmIntrinsics::_getAndSetLong:
360 case vmIntrinsics::_getAndSetObject:
361 case vmIntrinsics::_loadFence:
362 case vmIntrinsics::_storeFence:
363 case vmIntrinsics::_fullFence:
364 break; // InlineNatives does not control String.compareTo
365 case vmIntrinsics::_Reference_get:
366 break; // InlineNatives does not control Reference.get
367 default:
368 return NULL;
369 }
370 }
371
372 bool is_predicted = false;
373 bool does_virtual_dispatch = false;
374
375 switch (id) {
376 case vmIntrinsics::_compareTo:
377 if (!SpecialStringCompareTo) return NULL;
378 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
379 break;
380 case vmIntrinsics::_indexOf:
381 if (!SpecialStringIndexOf) return NULL;
382 break;
383 case vmIntrinsics::_equals:
384 if (!SpecialStringEquals) return NULL;
385 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
386 break;
387 case vmIntrinsics::_equalsC:
388 if (!SpecialArraysEquals) return NULL;
389 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
390 break;
391 case vmIntrinsics::_arraycopy:
392 if (!InlineArrayCopy) return NULL;
393 break;
394 case vmIntrinsics::_copyMemory:
395 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
396 if (!InlineArrayCopy) return NULL;
397 break;
398 case vmIntrinsics::_hashCode:
399 if (!InlineObjectHash) return NULL;
400 does_virtual_dispatch = true;
401 break;
402 case vmIntrinsics::_clone:
403 does_virtual_dispatch = true;
404 case vmIntrinsics::_copyOf:
405 case vmIntrinsics::_copyOfRange:
406 if (!InlineObjectCopy) return NULL;
407 // These also use the arraycopy intrinsic mechanism:
408 if (!InlineArrayCopy) return NULL;
409 break;
410 case vmIntrinsics::_encodeISOArray:
411 if (!SpecialEncodeISOArray) return NULL;
412 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
413 break;
414 case vmIntrinsics::_checkIndex:
415 // We do not intrinsify this. The optimizer does fine with it.
416 return NULL;
417
418 case vmIntrinsics::_getCallerClass:
419 if (!UseNewReflection) return NULL;
420 if (!InlineReflectionGetCallerClass) return NULL;
421 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
422 break;
423
424 case vmIntrinsics::_bitCount_i:
425 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
426 break;
427
428 case vmIntrinsics::_bitCount_l:
429 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
430 break;
431
432 case vmIntrinsics::_numberOfLeadingZeros_i:
433 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
434 break;
435
436 case vmIntrinsics::_numberOfLeadingZeros_l:
437 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
438 break;
439
440 case vmIntrinsics::_numberOfTrailingZeros_i:
441 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
442 break;
443
444 case vmIntrinsics::_numberOfTrailingZeros_l:
445 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
446 break;
447
448 case vmIntrinsics::_reverseBytes_c:
449 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
450 break;
451 case vmIntrinsics::_reverseBytes_s:
452 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
453 break;
454 case vmIntrinsics::_reverseBytes_i:
455 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
456 break;
457 case vmIntrinsics::_reverseBytes_l:
458 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
459 break;
460
461 case vmIntrinsics::_Reference_get:
462 // Use the intrinsic version of Reference.get() so that the value in
463 // the referent field can be registered by the G1 pre-barrier code.
464 // Also add memory barrier to prevent commoning reads from this field
465 // across safepoint since GC can change it value.
466 break;
467
468 case vmIntrinsics::_compareAndSwapObject:
469 #ifdef _LP64
470 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
471 #endif
472 break;
473
474 case vmIntrinsics::_compareAndSwapLong:
475 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
476 break;
477
478 case vmIntrinsics::_getAndAddInt:
479 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndAddLong:
483 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
484 break;
485
486 case vmIntrinsics::_getAndSetInt:
487 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
488 break;
489
490 case vmIntrinsics::_getAndSetLong:
491 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
492 break;
493
494 case vmIntrinsics::_getAndSetObject:
495 #ifdef _LP64
496 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
497 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
498 break;
499 #else
500 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
501 break;
502 #endif
503
504 case vmIntrinsics::_aescrypt_encryptBlock:
505 case vmIntrinsics::_aescrypt_decryptBlock:
506 if (!UseAESIntrinsics) return NULL;
507 break;
508
509 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
510 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
511 if (!UseAESIntrinsics) return NULL;
512 // these two require the predicated logic
513 is_predicted = true;
514 break;
515
516 case vmIntrinsics::_updateCRC32:
517 case vmIntrinsics::_updateBytesCRC32:
518 case vmIntrinsics::_updateByteBufferCRC32:
519 if (!UseCRC32Intrinsics) return NULL;
520 break;
521
522 case vmIntrinsics::_incrementExactI:
523 case vmIntrinsics::_addExactI:
524 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
525 break;
526 case vmIntrinsics::_incrementExactL:
527 case vmIntrinsics::_addExactL:
528 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
529 break;
530 case vmIntrinsics::_decrementExactI:
531 case vmIntrinsics::_subtractExactI:
532 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
533 break;
534 case vmIntrinsics::_decrementExactL:
535 case vmIntrinsics::_subtractExactL:
536 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
537 break;
538 case vmIntrinsics::_negateExactI:
539 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
540 break;
541 case vmIntrinsics::_negateExactL:
542 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
543 break;
544 case vmIntrinsics::_multiplyExactI:
545 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
546 break;
547 case vmIntrinsics::_multiplyExactL:
548 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
549 break;
550
551 default:
552 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
553 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
554 break;
555 }
556
557 // -XX:-InlineClassNatives disables natives from the Class class.
558 // The flag applies to all reflective calls, notably Array.newArray
559 // (visible to Java programmers as Array.newInstance).
560 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
561 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
562 if (!InlineClassNatives) return NULL;
563 }
564
565 // -XX:-InlineThreadNatives disables natives from the Thread class.
566 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
567 if (!InlineThreadNatives) return NULL;
568 }
569
570 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
571 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
572 m->holder()->name() == ciSymbol::java_lang_Float() ||
573 m->holder()->name() == ciSymbol::java_lang_Double()) {
574 if (!InlineMathNatives) return NULL;
575 }
576
577 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
578 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
579 if (!InlineUnsafeOps) return NULL;
580 }
581
582 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
583 }
584
585 //----------------------register_library_intrinsics-----------------------
586 // Initialize this file's data structures, for each Compile instance.
587 void Compile::register_library_intrinsics() {
588 // Nothing to do here.
589 }
590
591 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
592 LibraryCallKit kit(jvms, this);
593 Compile* C = kit.C;
594 int nodes = C->unique();
595 #ifndef PRODUCT
596 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
597 char buf[1000];
598 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
599 tty->print_cr("Intrinsic %s", str);
600 }
601 #endif
602 ciMethod* callee = kit.callee();
603 const int bci = kit.bci();
604
605 // Try to inline the intrinsic.
606 if (kit.try_to_inline()) {
607 if (C->print_intrinsics() || C->print_inlining()) {
608 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
609 }
610 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
611 if (C->log()) {
612 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
613 vmIntrinsics::name_at(intrinsic_id()),
614 (is_virtual() ? " virtual='1'" : ""),
615 C->unique() - nodes);
616 }
617 // Push the result from the inlined method onto the stack.
618 kit.push_result();
619 return kit.transfer_exceptions_into_jvms();
620 }
621
622 // The intrinsic bailed out
623 if (C->print_intrinsics() || C->print_inlining()) {
624 if (jvms->has_method()) {
625 // Not a root compile.
626 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
627 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
628 } else {
629 // Root compile
630 tty->print("Did not generate intrinsic %s%s at bci:%d in",
631 vmIntrinsics::name_at(intrinsic_id()),
632 (is_virtual() ? " (virtual)" : ""), bci);
633 }
634 }
635 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
636 return NULL;
637 }
638
639 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
640 LibraryCallKit kit(jvms, this);
641 Compile* C = kit.C;
642 int nodes = C->unique();
643 #ifndef PRODUCT
644 assert(is_predicted(), "sanity");
645 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
646 char buf[1000];
647 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
648 tty->print_cr("Predicate for intrinsic %s", str);
649 }
650 #endif
651 ciMethod* callee = kit.callee();
652 const int bci = kit.bci();
653
654 Node* slow_ctl = kit.try_to_predicate();
655 if (!kit.failing()) {
656 if (C->print_intrinsics() || C->print_inlining()) {
657 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
658 }
659 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
660 if (C->log()) {
661 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
662 vmIntrinsics::name_at(intrinsic_id()),
663 (is_virtual() ? " virtual='1'" : ""),
664 C->unique() - nodes);
665 }
666 return slow_ctl; // Could be NULL if the check folds.
667 }
668
669 // The intrinsic bailed out
670 if (C->print_intrinsics() || C->print_inlining()) {
671 if (jvms->has_method()) {
672 // Not a root compile.
673 const char* msg = "failed to generate predicate for intrinsic";
674 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
675 } else {
676 // Root compile
677 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
678 vmIntrinsics::name_at(intrinsic_id()),
679 (is_virtual() ? " (virtual)" : ""), bci);
680 }
681 }
682 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
683 return NULL;
684 }
685
686 bool LibraryCallKit::try_to_inline() {
687 // Handle symbolic names for otherwise undistinguished boolean switches:
688 const bool is_store = true;
689 const bool is_native_ptr = true;
690 const bool is_static = true;
691 const bool is_volatile = true;
692
693 if (!jvms()->has_method()) {
694 // Root JVMState has a null method.
695 assert(map()->memory()->Opcode() == Op_Parm, "");
696 // Insert the memory aliasing node
697 set_all_memory(reset_memory());
698 }
699 assert(merged_memory(), "");
700
701
702 switch (intrinsic_id()) {
703 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
704 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
705 case vmIntrinsics::_getClass: return inline_native_getClass();
706
707 case vmIntrinsics::_dsin:
708 case vmIntrinsics::_dcos:
709 case vmIntrinsics::_dtan:
710 case vmIntrinsics::_dabs:
711 case vmIntrinsics::_datan2:
712 case vmIntrinsics::_dsqrt:
713 case vmIntrinsics::_dexp:
714 case vmIntrinsics::_dlog:
715 case vmIntrinsics::_dlog10:
716 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
717
718 case vmIntrinsics::_min:
719 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
720
721 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
722 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
723 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
724 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
725 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
726 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
727 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
728 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
729 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
730 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
731 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
732 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
733
734 case vmIntrinsics::_arraycopy: return inline_arraycopy();
735
736 case vmIntrinsics::_compareTo: return inline_string_compareTo();
737 case vmIntrinsics::_indexOf: return inline_string_indexOf();
738 case vmIntrinsics::_equals: return inline_string_equals();
739
740 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
741 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
742 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
743 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
744 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
745 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
746 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
747 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
748 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
749
750 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
751 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
752 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
753 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
754 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
755 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
756 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
757 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
758 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
759
760 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
761 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
762 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
763 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
764 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
765 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
766 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
767 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
768
769 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
770 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
771 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
772 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
773 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
774 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
775 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
776 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
777
778 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
779 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
780 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
781 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
782 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
783 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
784 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
785 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
786 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
787
788 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
789 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
790 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
791 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
792 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
793 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
794 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
795 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
796 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
797
798 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
799 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
800 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
801 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
802
803 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
804 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
805 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
806
807 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
808 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
809 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
810
811 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
812 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
813 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
814 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
815 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
816
817 case vmIntrinsics::_loadFence:
818 case vmIntrinsics::_storeFence:
819 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
820
821 case vmIntrinsics::_currentThread: return inline_native_currentThread();
822 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
823
824 #ifdef TRACE_HAVE_INTRINSICS
825 case vmIntrinsics::_classID: return inline_native_classID();
826 case vmIntrinsics::_threadID: return inline_native_threadID();
827 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
828 #endif
829 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
830 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
831 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
832 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
833 case vmIntrinsics::_newArray: return inline_native_newArray();
834 case vmIntrinsics::_getLength: return inline_native_getLength();
835 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
836 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
837 case vmIntrinsics::_equalsC: return inline_array_equals();
838 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
839
840 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
841
842 case vmIntrinsics::_isInstance:
843 case vmIntrinsics::_getModifiers:
844 case vmIntrinsics::_isInterface:
845 case vmIntrinsics::_isArray:
846 case vmIntrinsics::_isPrimitive:
847 case vmIntrinsics::_getSuperclass:
848 case vmIntrinsics::_getComponentType:
849 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
850
851 case vmIntrinsics::_floatToRawIntBits:
852 case vmIntrinsics::_floatToIntBits:
853 case vmIntrinsics::_intBitsToFloat:
854 case vmIntrinsics::_doubleToRawLongBits:
855 case vmIntrinsics::_doubleToLongBits:
856 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
857
858 case vmIntrinsics::_numberOfLeadingZeros_i:
859 case vmIntrinsics::_numberOfLeadingZeros_l:
860 case vmIntrinsics::_numberOfTrailingZeros_i:
861 case vmIntrinsics::_numberOfTrailingZeros_l:
862 case vmIntrinsics::_bitCount_i:
863 case vmIntrinsics::_bitCount_l:
864 case vmIntrinsics::_reverseBytes_i:
865 case vmIntrinsics::_reverseBytes_l:
866 case vmIntrinsics::_reverseBytes_s:
867 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
868
869 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
870
871 case vmIntrinsics::_Reference_get: return inline_reference_get();
872
873 case vmIntrinsics::_aescrypt_encryptBlock:
874 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
875
876 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
877 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
878 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
879
880 case vmIntrinsics::_encodeISOArray:
881 return inline_encodeISOArray();
882
883 case vmIntrinsics::_updateCRC32:
884 return inline_updateCRC32();
885 case vmIntrinsics::_updateBytesCRC32:
886 return inline_updateBytesCRC32();
887 case vmIntrinsics::_updateByteBufferCRC32:
888 return inline_updateByteBufferCRC32();
889
890 default:
891 // If you get here, it may be that someone has added a new intrinsic
892 // to the list in vmSymbols.hpp without implementing it here.
893 #ifndef PRODUCT
894 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
895 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
896 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
897 }
898 #endif
899 return false;
900 }
901 }
902
903 Node* LibraryCallKit::try_to_predicate() {
904 if (!jvms()->has_method()) {
905 // Root JVMState has a null method.
906 assert(map()->memory()->Opcode() == Op_Parm, "");
907 // Insert the memory aliasing node
908 set_all_memory(reset_memory());
909 }
910 assert(merged_memory(), "");
911
912 switch (intrinsic_id()) {
913 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
914 return inline_cipherBlockChaining_AESCrypt_predicate(false);
915 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
916 return inline_cipherBlockChaining_AESCrypt_predicate(true);
917
918 default:
919 // If you get here, it may be that someone has added a new intrinsic
920 // to the list in vmSymbols.hpp without implementing it here.
921 #ifndef PRODUCT
922 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
923 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
924 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
925 }
926 #endif
927 Node* slow_ctl = control();
928 set_control(top()); // No fast path instrinsic
929 return slow_ctl;
930 }
931 }
932
933 //------------------------------set_result-------------------------------
934 // Helper function for finishing intrinsics.
935 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
936 record_for_igvn(region);
937 set_control(_gvn.transform(region));
938 set_result( _gvn.transform(value));
939 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
940 }
941
942 //------------------------------generate_guard---------------------------
943 // Helper function for generating guarded fast-slow graph structures.
944 // The given 'test', if true, guards a slow path. If the test fails
945 // then a fast path can be taken. (We generally hope it fails.)
946 // In all cases, GraphKit::control() is updated to the fast path.
947 // The returned value represents the control for the slow path.
948 // The return value is never 'top'; it is either a valid control
949 // or NULL if it is obvious that the slow path can never be taken.
950 // Also, if region and the slow control are not NULL, the slow edge
951 // is appended to the region.
952 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
953 if (stopped()) {
954 // Already short circuited.
955 return NULL;
956 }
957
958 // Build an if node and its projections.
959 // If test is true we take the slow path, which we assume is uncommon.
960 if (_gvn.type(test) == TypeInt::ZERO) {
961 // The slow branch is never taken. No need to build this guard.
962 return NULL;
963 }
964
965 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
966
967 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
968 if (if_slow == top()) {
969 // The slow branch is never taken. No need to build this guard.
970 return NULL;
971 }
972
973 if (region != NULL)
974 region->add_req(if_slow);
975
976 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
977 set_control(if_fast);
978
979 return if_slow;
980 }
981
982 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
983 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
984 }
985 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
986 return generate_guard(test, region, PROB_FAIR);
987 }
988
989 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
990 Node* *pos_index) {
991 if (stopped())
992 return NULL; // already stopped
993 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
994 return NULL; // index is already adequately typed
995 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
996 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
997 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
998 if (is_neg != NULL && pos_index != NULL) {
999 // Emulate effect of Parse::adjust_map_after_if.
1000 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1001 ccast->set_req(0, control());
1002 (*pos_index) = _gvn.transform(ccast);
1003 }
1004 return is_neg;
1005 }
1006
1007 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1008 Node* *pos_index) {
1009 if (stopped())
1010 return NULL; // already stopped
1011 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1012 return NULL; // index is already adequately typed
1013 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1014 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1015 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1016 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1017 if (is_notp != NULL && pos_index != NULL) {
1018 // Emulate effect of Parse::adjust_map_after_if.
1019 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1020 ccast->set_req(0, control());
1021 (*pos_index) = _gvn.transform(ccast);
1022 }
1023 return is_notp;
1024 }
1025
1026 // Make sure that 'position' is a valid limit index, in [0..length].
1027 // There are two equivalent plans for checking this:
1028 // A. (offset + copyLength) unsigned<= arrayLength
1029 // B. offset <= (arrayLength - copyLength)
1030 // We require that all of the values above, except for the sum and
1031 // difference, are already known to be non-negative.
1032 // Plan A is robust in the face of overflow, if offset and copyLength
1033 // are both hugely positive.
1034 //
1035 // Plan B is less direct and intuitive, but it does not overflow at
1036 // all, since the difference of two non-negatives is always
1037 // representable. Whenever Java methods must perform the equivalent
1038 // check they generally use Plan B instead of Plan A.
1039 // For the moment we use Plan A.
1040 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1041 Node* subseq_length,
1042 Node* array_length,
1043 RegionNode* region) {
1044 if (stopped())
1045 return NULL; // already stopped
1046 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1047 if (zero_offset && subseq_length->eqv_uncast(array_length))
1048 return NULL; // common case of whole-array copy
1049 Node* last = subseq_length;
1050 if (!zero_offset) // last += offset
1051 last = _gvn.transform(new (C) AddINode(last, offset));
1052 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1053 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1054 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1055 return is_over;
1056 }
1057
1058
1059 //--------------------------generate_current_thread--------------------
1060 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1061 ciKlass* thread_klass = env()->Thread_klass();
1062 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1063 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1064 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1065 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1066 tls_output = thread;
1067 return threadObj;
1068 }
1069
1070
1071 //------------------------------make_string_method_node------------------------
1072 // Helper method for String intrinsic functions. This version is called
1073 // with str1 and str2 pointing to String object nodes.
1074 //
1075 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1076 Node* no_ctrl = NULL;
1077
1078 // Get start addr of string
1079 Node* str1_value = load_String_value(no_ctrl, str1);
1080 Node* str1_offset = load_String_offset(no_ctrl, str1);
1081 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1082
1083 // Get length of string 1
1084 Node* str1_len = load_String_length(no_ctrl, str1);
1085
1086 Node* str2_value = load_String_value(no_ctrl, str2);
1087 Node* str2_offset = load_String_offset(no_ctrl, str2);
1088 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1089
1090 Node* str2_len = NULL;
1091 Node* result = NULL;
1092
1093 switch (opcode) {
1094 case Op_StrIndexOf:
1095 // Get length of string 2
1096 str2_len = load_String_length(no_ctrl, str2);
1097
1098 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1099 str1_start, str1_len, str2_start, str2_len);
1100 break;
1101 case Op_StrComp:
1102 // Get length of string 2
1103 str2_len = load_String_length(no_ctrl, str2);
1104
1105 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1106 str1_start, str1_len, str2_start, str2_len);
1107 break;
1108 case Op_StrEquals:
1109 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1110 str1_start, str2_start, str1_len);
1111 break;
1112 default:
1113 ShouldNotReachHere();
1114 return NULL;
1115 }
1116
1117 // All these intrinsics have checks.
1118 C->set_has_split_ifs(true); // Has chance for split-if optimization
1119
1120 return _gvn.transform(result);
1121 }
1122
1123 // Helper method for String intrinsic functions. This version is called
1124 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1125 // to Int nodes containing the lenghts of str1 and str2.
1126 //
1127 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1128 Node* result = NULL;
1129 switch (opcode) {
1130 case Op_StrIndexOf:
1131 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1132 str1_start, cnt1, str2_start, cnt2);
1133 break;
1134 case Op_StrComp:
1135 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1136 str1_start, cnt1, str2_start, cnt2);
1137 break;
1138 case Op_StrEquals:
1139 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1140 str1_start, str2_start, cnt1);
1141 break;
1142 default:
1143 ShouldNotReachHere();
1144 return NULL;
1145 }
1146
1147 // All these intrinsics have checks.
1148 C->set_has_split_ifs(true); // Has chance for split-if optimization
1149
1150 return _gvn.transform(result);
1151 }
1152
1153 //------------------------------inline_string_compareTo------------------------
1154 // public int java.lang.String.compareTo(String anotherString);
1155 bool LibraryCallKit::inline_string_compareTo() {
1156 Node* receiver = null_check(argument(0));
1157 Node* arg = null_check(argument(1));
1158 if (stopped()) {
1159 return true;
1160 }
1161 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1162 return true;
1163 }
1164
1165 //------------------------------inline_string_equals------------------------
1166 bool LibraryCallKit::inline_string_equals() {
1167 Node* receiver = null_check_receiver();
1168 // NOTE: Do not null check argument for String.equals() because spec
1169 // allows to specify NULL as argument.
1170 Node* argument = this->argument(1);
1171 if (stopped()) {
1172 return true;
1173 }
1174
1175 // paths (plus control) merge
1176 RegionNode* region = new (C) RegionNode(5);
1177 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1178
1179 // does source == target string?
1180 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1181 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1182
1183 Node* if_eq = generate_slow_guard(bol, NULL);
1184 if (if_eq != NULL) {
1185 // receiver == argument
1186 phi->init_req(2, intcon(1));
1187 region->init_req(2, if_eq);
1188 }
1189
1190 // get String klass for instanceOf
1191 ciInstanceKlass* klass = env()->String_klass();
1192
1193 if (!stopped()) {
1194 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1195 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1196 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1197
1198 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1199 //instanceOf == true, fallthrough
1200
1201 if (inst_false != NULL) {
1202 phi->init_req(3, intcon(0));
1203 region->init_req(3, inst_false);
1204 }
1205 }
1206
1207 if (!stopped()) {
1208 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1209
1210 // Properly cast the argument to String
1211 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1212 // This path is taken only when argument's type is String:NotNull.
1213 argument = cast_not_null(argument, false);
1214
1215 Node* no_ctrl = NULL;
1216
1217 // Get start addr of receiver
1218 Node* receiver_val = load_String_value(no_ctrl, receiver);
1219 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1220 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1221
1222 // Get length of receiver
1223 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1224
1225 // Get start addr of argument
1226 Node* argument_val = load_String_value(no_ctrl, argument);
1227 Node* argument_offset = load_String_offset(no_ctrl, argument);
1228 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1229
1230 // Get length of argument
1231 Node* argument_cnt = load_String_length(no_ctrl, argument);
1232
1233 // Check for receiver count != argument count
1234 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1235 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1236 Node* if_ne = generate_slow_guard(bol, NULL);
1237 if (if_ne != NULL) {
1238 phi->init_req(4, intcon(0));
1239 region->init_req(4, if_ne);
1240 }
1241
1242 // Check for count == 0 is done by assembler code for StrEquals.
1243
1244 if (!stopped()) {
1245 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1246 phi->init_req(1, equals);
1247 region->init_req(1, control());
1248 }
1249 }
1250
1251 // post merge
1252 set_control(_gvn.transform(region));
1253 record_for_igvn(region);
1254
1255 set_result(_gvn.transform(phi));
1256 return true;
1257 }
1258
1259 //------------------------------inline_array_equals----------------------------
1260 bool LibraryCallKit::inline_array_equals() {
1261 Node* arg1 = argument(0);
1262 Node* arg2 = argument(1);
1263 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1264 return true;
1265 }
1266
1267 // Java version of String.indexOf(constant string)
1268 // class StringDecl {
1269 // StringDecl(char[] ca) {
1270 // offset = 0;
1271 // count = ca.length;
1272 // value = ca;
1273 // }
1274 // int offset;
1275 // int count;
1276 // char[] value;
1277 // }
1278 //
1279 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1280 // int targetOffset, int cache_i, int md2) {
1281 // int cache = cache_i;
1282 // int sourceOffset = string_object.offset;
1283 // int sourceCount = string_object.count;
1284 // int targetCount = target_object.length;
1285 //
1286 // int targetCountLess1 = targetCount - 1;
1287 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1288 //
1289 // char[] source = string_object.value;
1290 // char[] target = target_object;
1291 // int lastChar = target[targetCountLess1];
1292 //
1293 // outer_loop:
1294 // for (int i = sourceOffset; i < sourceEnd; ) {
1295 // int src = source[i + targetCountLess1];
1296 // if (src == lastChar) {
1297 // // With random strings and a 4-character alphabet,
1298 // // reverse matching at this point sets up 0.8% fewer
1299 // // frames, but (paradoxically) makes 0.3% more probes.
1300 // // Since those probes are nearer the lastChar probe,
1301 // // there is may be a net D$ win with reverse matching.
1302 // // But, reversing loop inhibits unroll of inner loop
1303 // // for unknown reason. So, does running outer loop from
1304 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1305 // for (int j = 0; j < targetCountLess1; j++) {
1306 // if (target[targetOffset + j] != source[i+j]) {
1307 // if ((cache & (1 << source[i+j])) == 0) {
1308 // if (md2 < j+1) {
1309 // i += j+1;
1310 // continue outer_loop;
1311 // }
1312 // }
1313 // i += md2;
1314 // continue outer_loop;
1315 // }
1316 // }
1317 // return i - sourceOffset;
1318 // }
1319 // if ((cache & (1 << src)) == 0) {
1320 // i += targetCountLess1;
1321 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1322 // i++;
1323 // }
1324 // return -1;
1325 // }
1326
1327 //------------------------------string_indexOf------------------------
1328 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1329 jint cache_i, jint md2_i) {
1330
1331 Node* no_ctrl = NULL;
1332 float likely = PROB_LIKELY(0.9);
1333 float unlikely = PROB_UNLIKELY(0.9);
1334
1335 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1336
1337 Node* source = load_String_value(no_ctrl, string_object);
1338 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1339 Node* sourceCount = load_String_length(no_ctrl, string_object);
1340
1341 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1342 jint target_length = target_array->length();
1343 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1344 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1345
1346 // String.value field is known to be @Stable.
1347 if (UseImplicitStableValues) {
1348 target = cast_array_to_stable(target, target_type);
1349 }
1350
1351 IdealKit kit(this, false, true);
1352 #define __ kit.
1353 Node* zero = __ ConI(0);
1354 Node* one = __ ConI(1);
1355 Node* cache = __ ConI(cache_i);
1356 Node* md2 = __ ConI(md2_i);
1357 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1358 Node* targetCount = __ ConI(target_length);
1359 Node* targetCountLess1 = __ ConI(target_length - 1);
1360 Node* targetOffset = __ ConI(targetOffset_i);
1361 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1362
1363 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1364 Node* outer_loop = __ make_label(2 /* goto */);
1365 Node* return_ = __ make_label(1);
1366
1367 __ set(rtn,__ ConI(-1));
1368 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1369 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1370 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1371 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1372 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1373 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1374 Node* tpj = __ AddI(targetOffset, __ value(j));
1375 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1376 Node* ipj = __ AddI(__ value(i), __ value(j));
1377 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1378 __ if_then(targ, BoolTest::ne, src2); {
1379 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1380 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1381 __ increment(i, __ AddI(__ value(j), one));
1382 __ goto_(outer_loop);
1383 } __ end_if(); __ dead(j);
1384 }__ end_if(); __ dead(j);
1385 __ increment(i, md2);
1386 __ goto_(outer_loop);
1387 }__ end_if();
1388 __ increment(j, one);
1389 }__ end_loop(); __ dead(j);
1390 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1391 __ goto_(return_);
1392 }__ end_if();
1393 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1394 __ increment(i, targetCountLess1);
1395 }__ end_if();
1396 __ increment(i, one);
1397 __ bind(outer_loop);
1398 }__ end_loop(); __ dead(i);
1399 __ bind(return_);
1400
1401 // Final sync IdealKit and GraphKit.
1402 final_sync(kit);
1403 Node* result = __ value(rtn);
1404 #undef __
1405 C->set_has_loops(true);
1406 return result;
1407 }
1408
1409 //------------------------------inline_string_indexOf------------------------
1410 bool LibraryCallKit::inline_string_indexOf() {
1411 Node* receiver = argument(0);
1412 Node* arg = argument(1);
1413
1414 Node* result;
1415 // Disable the use of pcmpestri until it can be guaranteed that
1416 // the load doesn't cross into the uncommited space.
1417 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1418 UseSSE42Intrinsics) {
1419 // Generate SSE4.2 version of indexOf
1420 // We currently only have match rules that use SSE4.2
1421
1422 receiver = null_check(receiver);
1423 arg = null_check(arg);
1424 if (stopped()) {
1425 return true;
1426 }
1427
1428 ciInstanceKlass* str_klass = env()->String_klass();
1429 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1430
1431 // Make the merge point
1432 RegionNode* result_rgn = new (C) RegionNode(4);
1433 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1434 Node* no_ctrl = NULL;
1435
1436 // Get start addr of source string
1437 Node* source = load_String_value(no_ctrl, receiver);
1438 Node* source_offset = load_String_offset(no_ctrl, receiver);
1439 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1440
1441 // Get length of source string
1442 Node* source_cnt = load_String_length(no_ctrl, receiver);
1443
1444 // Get start addr of substring
1445 Node* substr = load_String_value(no_ctrl, arg);
1446 Node* substr_offset = load_String_offset(no_ctrl, arg);
1447 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1448
1449 // Get length of source string
1450 Node* substr_cnt = load_String_length(no_ctrl, arg);
1451
1452 // Check for substr count > string count
1453 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1454 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1455 Node* if_gt = generate_slow_guard(bol, NULL);
1456 if (if_gt != NULL) {
1457 result_phi->init_req(2, intcon(-1));
1458 result_rgn->init_req(2, if_gt);
1459 }
1460
1461 if (!stopped()) {
1462 // Check for substr count == 0
1463 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1464 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1465 Node* if_zero = generate_slow_guard(bol, NULL);
1466 if (if_zero != NULL) {
1467 result_phi->init_req(3, intcon(0));
1468 result_rgn->init_req(3, if_zero);
1469 }
1470 }
1471
1472 if (!stopped()) {
1473 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1474 result_phi->init_req(1, result);
1475 result_rgn->init_req(1, control());
1476 }
1477 set_control(_gvn.transform(result_rgn));
1478 record_for_igvn(result_rgn);
1479 result = _gvn.transform(result_phi);
1480
1481 } else { // Use LibraryCallKit::string_indexOf
1482 // don't intrinsify if argument isn't a constant string.
1483 if (!arg->is_Con()) {
1484 return false;
1485 }
1486 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1487 if (str_type == NULL) {
1488 return false;
1489 }
1490 ciInstanceKlass* klass = env()->String_klass();
1491 ciObject* str_const = str_type->const_oop();
1492 if (str_const == NULL || str_const->klass() != klass) {
1493 return false;
1494 }
1495 ciInstance* str = str_const->as_instance();
1496 assert(str != NULL, "must be instance");
1497
1498 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1499 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1500
1501 int o;
1502 int c;
1503 if (java_lang_String::has_offset_field()) {
1504 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1505 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1506 } else {
1507 o = 0;
1508 c = pat->length();
1509 }
1510
1511 // constant strings have no offset and count == length which
1512 // simplifies the resulting code somewhat so lets optimize for that.
1513 if (o != 0 || c != pat->length()) {
1514 return false;
1515 }
1516
1517 receiver = null_check(receiver, T_OBJECT);
1518 // NOTE: No null check on the argument is needed since it's a constant String oop.
1519 if (stopped()) {
1520 return true;
1521 }
1522
1523 // The null string as a pattern always returns 0 (match at beginning of string)
1524 if (c == 0) {
1525 set_result(intcon(0));
1526 return true;
1527 }
1528
1529 // Generate default indexOf
1530 jchar lastChar = pat->char_at(o + (c - 1));
1531 int cache = 0;
1532 int i;
1533 for (i = 0; i < c - 1; i++) {
1534 assert(i < pat->length(), "out of range");
1535 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1536 }
1537
1538 int md2 = c;
1539 for (i = 0; i < c - 1; i++) {
1540 assert(i < pat->length(), "out of range");
1541 if (pat->char_at(o + i) == lastChar) {
1542 md2 = (c - 1) - i;
1543 }
1544 }
1545
1546 result = string_indexOf(receiver, pat, o, cache, md2);
1547 }
1548 set_result(result);
1549 return true;
1550 }
1551
1552 //--------------------------round_double_node--------------------------------
1553 // Round a double node if necessary.
1554 Node* LibraryCallKit::round_double_node(Node* n) {
1555 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1556 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1557 return n;
1558 }
1559
1560 //------------------------------inline_math-----------------------------------
1561 // public static double Math.abs(double)
1562 // public static double Math.sqrt(double)
1563 // public static double Math.log(double)
1564 // public static double Math.log10(double)
1565 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1566 Node* arg = round_double_node(argument(0));
1567 Node* n;
1568 switch (id) {
1569 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1570 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1571 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1572 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1573 default: fatal_unexpected_iid(id); break;
1574 }
1575 set_result(_gvn.transform(n));
1576 return true;
1577 }
1578
1579 //------------------------------inline_trig----------------------------------
1580 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1581 // argument reduction which will turn into a fast/slow diamond.
1582 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1583 Node* arg = round_double_node(argument(0));
1584 Node* n = NULL;
1585
1586 switch (id) {
1587 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1588 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1589 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1590 default: fatal_unexpected_iid(id); break;
1591 }
1592 n = _gvn.transform(n);
1593
1594 // Rounding required? Check for argument reduction!
1595 if (Matcher::strict_fp_requires_explicit_rounding) {
1596 static const double pi_4 = 0.7853981633974483;
1597 static const double neg_pi_4 = -0.7853981633974483;
1598 // pi/2 in 80-bit extended precision
1599 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1600 // -pi/2 in 80-bit extended precision
1601 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1602 // Cutoff value for using this argument reduction technique
1603 //static const double pi_2_minus_epsilon = 1.564660403643354;
1604 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1605
1606 // Pseudocode for sin:
1607 // if (x <= Math.PI / 4.0) {
1608 // if (x >= -Math.PI / 4.0) return fsin(x);
1609 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1610 // } else {
1611 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1612 // }
1613 // return StrictMath.sin(x);
1614
1615 // Pseudocode for cos:
1616 // if (x <= Math.PI / 4.0) {
1617 // if (x >= -Math.PI / 4.0) return fcos(x);
1618 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1619 // } else {
1620 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1621 // }
1622 // return StrictMath.cos(x);
1623
1624 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1625 // requires a special machine instruction to load it. Instead we'll try
1626 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1627 // probably do the math inside the SIN encoding.
1628
1629 // Make the merge point
1630 RegionNode* r = new (C) RegionNode(3);
1631 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1632
1633 // Flatten arg so we need only 1 test
1634 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1635 // Node for PI/4 constant
1636 Node *pi4 = makecon(TypeD::make(pi_4));
1637 // Check PI/4 : abs(arg)
1638 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1639 // Check: If PI/4 < abs(arg) then go slow
1640 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1641 // Branch either way
1642 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1643 set_control(opt_iff(r,iff));
1644
1645 // Set fast path result
1646 phi->init_req(2, n);
1647
1648 // Slow path - non-blocking leaf call
1649 Node* call = NULL;
1650 switch (id) {
1651 case vmIntrinsics::_dsin:
1652 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1653 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1654 "Sin", NULL, arg, top());
1655 break;
1656 case vmIntrinsics::_dcos:
1657 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1658 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1659 "Cos", NULL, arg, top());
1660 break;
1661 case vmIntrinsics::_dtan:
1662 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1663 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1664 "Tan", NULL, arg, top());
1665 break;
1666 }
1667 assert(control()->in(0) == call, "");
1668 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1669 r->init_req(1, control());
1670 phi->init_req(1, slow_result);
1671
1672 // Post-merge
1673 set_control(_gvn.transform(r));
1674 record_for_igvn(r);
1675 n = _gvn.transform(phi);
1676
1677 C->set_has_split_ifs(true); // Has chance for split-if optimization
1678 }
1679 set_result(n);
1680 return true;
1681 }
1682
1683 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1684 //-------------------
1685 //result=(result.isNaN())? funcAddr():result;
1686 // Check: If isNaN() by checking result!=result? then either trap
1687 // or go to runtime
1688 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1689 // Build the boolean node
1690 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1691
1692 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1693 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1694 // The pow or exp intrinsic returned a NaN, which requires a call
1695 // to the runtime. Recompile with the runtime call.
1696 uncommon_trap(Deoptimization::Reason_intrinsic,
1697 Deoptimization::Action_make_not_entrant);
1698 }
1699 set_result(result);
1700 } else {
1701 // If this inlining ever returned NaN in the past, we compile a call
1702 // to the runtime to properly handle corner cases
1703
1704 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1705 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1706 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1707
1708 if (!if_slow->is_top()) {
1709 RegionNode* result_region = new (C) RegionNode(3);
1710 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1711
1712 result_region->init_req(1, if_fast);
1713 result_val->init_req(1, result);
1714
1715 set_control(if_slow);
1716
1717 const TypePtr* no_memory_effects = NULL;
1718 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1719 no_memory_effects,
1720 x, top(), y, y ? top() : NULL);
1721 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1722 #ifdef ASSERT
1723 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1724 assert(value_top == top(), "second value must be top");
1725 #endif
1726
1727 result_region->init_req(2, control());
1728 result_val->init_req(2, value);
1729 set_result(result_region, result_val);
1730 } else {
1731 set_result(result);
1732 }
1733 }
1734 }
1735
1736 //------------------------------inline_exp-------------------------------------
1737 // Inline exp instructions, if possible. The Intel hardware only misses
1738 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1739 bool LibraryCallKit::inline_exp() {
1740 Node* arg = round_double_node(argument(0));
1741 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1742
1743 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1744
1745 C->set_has_split_ifs(true); // Has chance for split-if optimization
1746 return true;
1747 }
1748
1749 //------------------------------inline_pow-------------------------------------
1750 // Inline power instructions, if possible.
1751 bool LibraryCallKit::inline_pow() {
1752 // Pseudocode for pow
1753 // if (x <= 0.0) {
1754 // long longy = (long)y;
1755 // if ((double)longy == y) { // if y is long
1756 // if (y + 1 == y) longy = 0; // huge number: even
1757 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1758 // } else {
1759 // result = NaN;
1760 // }
1761 // } else {
1762 // result = DPow(x,y);
1763 // }
1764 // if (result != result)? {
1765 // result = uncommon_trap() or runtime_call();
1766 // }
1767 // return result;
1768
1769 Node* x = round_double_node(argument(0));
1770 Node* y = round_double_node(argument(2));
1771
1772 Node* result = NULL;
1773
1774 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1775 // Short form: skip the fancy tests and just check for NaN result.
1776 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1777 } else {
1778 // If this inlining ever returned NaN in the past, include all
1779 // checks + call to the runtime.
1780
1781 // Set the merge point for If node with condition of (x <= 0.0)
1782 // There are four possible paths to region node and phi node
1783 RegionNode *r = new (C) RegionNode(4);
1784 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1785
1786 // Build the first if node: if (x <= 0.0)
1787 // Node for 0 constant
1788 Node *zeronode = makecon(TypeD::ZERO);
1789 // Check x:0
1790 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1791 // Check: If (x<=0) then go complex path
1792 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1793 // Branch either way
1794 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1795 // Fast path taken; set region slot 3
1796 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1797 r->init_req(3,fast_taken); // Capture fast-control
1798
1799 // Fast path not-taken, i.e. slow path
1800 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1801
1802 // Set fast path result
1803 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1804 phi->init_req(3, fast_result);
1805
1806 // Complex path
1807 // Build the second if node (if y is long)
1808 // Node for (long)y
1809 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1810 // Node for (double)((long) y)
1811 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1812 // Check (double)((long) y) : y
1813 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1814 // Check if (y isn't long) then go to slow path
1815
1816 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1817 // Branch either way
1818 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1819 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1820
1821 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1822
1823 // Calculate DPow(abs(x), y)*(1 & (long)y)
1824 // Node for constant 1
1825 Node *conone = longcon(1);
1826 // 1& (long)y
1827 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1828
1829 // A huge number is always even. Detect a huge number by checking
1830 // if y + 1 == y and set integer to be tested for parity to 0.
1831 // Required for corner case:
1832 // (long)9.223372036854776E18 = max_jlong
1833 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1834 // max_jlong is odd but 9.223372036854776E18 is even
1835 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1836 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1837 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1838 Node* correctedsign = NULL;
1839 if (ConditionalMoveLimit != 0) {
1840 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1841 } else {
1842 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1843 RegionNode *r = new (C) RegionNode(3);
1844 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1845 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1846 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1847 phi->init_req(1, signnode);
1848 phi->init_req(2, longcon(0));
1849 correctedsign = _gvn.transform(phi);
1850 ylong_path = _gvn.transform(r);
1851 record_for_igvn(r);
1852 }
1853
1854 // zero node
1855 Node *conzero = longcon(0);
1856 // Check (1&(long)y)==0?
1857 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1858 // Check if (1&(long)y)!=0?, if so the result is negative
1859 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1860 // abs(x)
1861 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1862 // abs(x)^y
1863 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1864 // -abs(x)^y
1865 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1866 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1867 Node *signresult = NULL;
1868 if (ConditionalMoveLimit != 0) {
1869 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1870 } else {
1871 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1872 RegionNode *r = new (C) RegionNode(3);
1873 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1874 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1875 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1876 phi->init_req(1, absxpowy);
1877 phi->init_req(2, negabsxpowy);
1878 signresult = _gvn.transform(phi);
1879 ylong_path = _gvn.transform(r);
1880 record_for_igvn(r);
1881 }
1882 // Set complex path fast result
1883 r->init_req(2, ylong_path);
1884 phi->init_req(2, signresult);
1885
1886 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1887 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1888 r->init_req(1,slow_path);
1889 phi->init_req(1,slow_result);
1890
1891 // Post merge
1892 set_control(_gvn.transform(r));
1893 record_for_igvn(r);
1894 result = _gvn.transform(phi);
1895 }
1896
1897 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1898
1899 C->set_has_split_ifs(true); // Has chance for split-if optimization
1900 return true;
1901 }
1902
1903 //------------------------------runtime_math-----------------------------
1904 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1905 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1906 "must be (DD)D or (D)D type");
1907
1908 // Inputs
1909 Node* a = round_double_node(argument(0));
1910 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1911
1912 const TypePtr* no_memory_effects = NULL;
1913 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1914 no_memory_effects,
1915 a, top(), b, b ? top() : NULL);
1916 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1917 #ifdef ASSERT
1918 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1919 assert(value_top == top(), "second value must be top");
1920 #endif
1921
1922 set_result(value);
1923 return true;
1924 }
1925
1926 //------------------------------inline_math_native-----------------------------
1927 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1928 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1929 switch (id) {
1930 // These intrinsics are not properly supported on all hardware
1931 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1932 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1933 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1934 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1935 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1936 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1937
1938 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1939 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1940 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1941 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1942
1943 // These intrinsics are supported on all hardware
1944 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1945 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1946
1947 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1948 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1949 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1950 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1951 #undef FN_PTR
1952
1953 // These intrinsics are not yet correctly implemented
1954 case vmIntrinsics::_datan2:
1955 return false;
1956
1957 default:
1958 fatal_unexpected_iid(id);
1959 return false;
1960 }
1961 }
1962
1963 static bool is_simple_name(Node* n) {
1964 return (n->req() == 1 // constant
1965 || (n->is_Type() && n->as_Type()->type()->singleton())
1966 || n->is_Proj() // parameter or return value
1967 || n->is_Phi() // local of some sort
1968 );
1969 }
1970
1971 //----------------------------inline_min_max-----------------------------------
1972 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1973 set_result(generate_min_max(id, argument(0), argument(1)));
1974 return true;
1975 }
1976
1977 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1978 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
1979 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1980 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1981 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1982
1983 {
1984 PreserveJVMState pjvms(this);
1985 PreserveReexecuteState preexecs(this);
1986 jvms()->set_should_reexecute(true);
1987
1988 set_control(slow_path);
1989 set_i_o(i_o());
1990
1991 uncommon_trap(Deoptimization::Reason_intrinsic,
1992 Deoptimization::Action_none);
1993 }
1994
1995 set_control(fast_path);
1996 set_result(math);
1997 }
1998
1999 template <typename OverflowOp>
2000 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2001 typedef typename OverflowOp::MathOp MathOp;
2002
2003 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2004 Node* operation = _gvn.transform( mathOp );
2005 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2006 inline_math_mathExact(operation, ofcheck);
2007 return true;
2008 }
2009
2010 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2011 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2012 }
2013
2014 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2015 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2016 }
2017
2018 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2019 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2020 }
2021
2022 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2023 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2024 }
2025
2026 bool LibraryCallKit::inline_math_negateExactI() {
2027 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2028 }
2029
2030 bool LibraryCallKit::inline_math_negateExactL() {
2031 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2032 }
2033
2034 bool LibraryCallKit::inline_math_multiplyExactI() {
2035 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2036 }
2037
2038 bool LibraryCallKit::inline_math_multiplyExactL() {
2039 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2040 }
2041
2042 Node*
2043 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2044 // These are the candidate return value:
2045 Node* xvalue = x0;
2046 Node* yvalue = y0;
2047
2048 if (xvalue == yvalue) {
2049 return xvalue;
2050 }
2051
2052 bool want_max = (id == vmIntrinsics::_max);
2053
2054 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2055 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2056 if (txvalue == NULL || tyvalue == NULL) return top();
2057 // This is not really necessary, but it is consistent with a
2058 // hypothetical MaxINode::Value method:
2059 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2060
2061 // %%% This folding logic should (ideally) be in a different place.
2062 // Some should be inside IfNode, and there to be a more reliable
2063 // transformation of ?: style patterns into cmoves. We also want
2064 // more powerful optimizations around cmove and min/max.
2065
2066 // Try to find a dominating comparison of these guys.
2067 // It can simplify the index computation for Arrays.copyOf
2068 // and similar uses of System.arraycopy.
2069 // First, compute the normalized version of CmpI(x, y).
2070 int cmp_op = Op_CmpI;
2071 Node* xkey = xvalue;
2072 Node* ykey = yvalue;
2073 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2074 if (ideal_cmpxy->is_Cmp()) {
2075 // E.g., if we have CmpI(length - offset, count),
2076 // it might idealize to CmpI(length, count + offset)
2077 cmp_op = ideal_cmpxy->Opcode();
2078 xkey = ideal_cmpxy->in(1);
2079 ykey = ideal_cmpxy->in(2);
2080 }
2081
2082 // Start by locating any relevant comparisons.
2083 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2084 Node* cmpxy = NULL;
2085 Node* cmpyx = NULL;
2086 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2087 Node* cmp = start_from->fast_out(k);
2088 if (cmp->outcnt() > 0 && // must have prior uses
2089 cmp->in(0) == NULL && // must be context-independent
2090 cmp->Opcode() == cmp_op) { // right kind of compare
2091 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2092 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2093 }
2094 }
2095
2096 const int NCMPS = 2;
2097 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2098 int cmpn;
2099 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2100 if (cmps[cmpn] != NULL) break; // find a result
2101 }
2102 if (cmpn < NCMPS) {
2103 // Look for a dominating test that tells us the min and max.
2104 int depth = 0; // Limit search depth for speed
2105 Node* dom = control();
2106 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2107 if (++depth >= 100) break;
2108 Node* ifproj = dom;
2109 if (!ifproj->is_Proj()) continue;
2110 Node* iff = ifproj->in(0);
2111 if (!iff->is_If()) continue;
2112 Node* bol = iff->in(1);
2113 if (!bol->is_Bool()) continue;
2114 Node* cmp = bol->in(1);
2115 if (cmp == NULL) continue;
2116 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2117 if (cmps[cmpn] == cmp) break;
2118 if (cmpn == NCMPS) continue;
2119 BoolTest::mask btest = bol->as_Bool()->_test._test;
2120 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2121 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2122 // At this point, we know that 'x btest y' is true.
2123 switch (btest) {
2124 case BoolTest::eq:
2125 // They are proven equal, so we can collapse the min/max.
2126 // Either value is the answer. Choose the simpler.
2127 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2128 return yvalue;
2129 return xvalue;
2130 case BoolTest::lt: // x < y
2131 case BoolTest::le: // x <= y
2132 return (want_max ? yvalue : xvalue);
2133 case BoolTest::gt: // x > y
2134 case BoolTest::ge: // x >= y
2135 return (want_max ? xvalue : yvalue);
2136 }
2137 }
2138 }
2139
2140 // We failed to find a dominating test.
2141 // Let's pick a test that might GVN with prior tests.
2142 Node* best_bol = NULL;
2143 BoolTest::mask best_btest = BoolTest::illegal;
2144 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2145 Node* cmp = cmps[cmpn];
2146 if (cmp == NULL) continue;
2147 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2148 Node* bol = cmp->fast_out(j);
2149 if (!bol->is_Bool()) continue;
2150 BoolTest::mask btest = bol->as_Bool()->_test._test;
2151 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2152 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2153 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2154 best_bol = bol->as_Bool();
2155 best_btest = btest;
2156 }
2157 }
2158 }
2159
2160 Node* answer_if_true = NULL;
2161 Node* answer_if_false = NULL;
2162 switch (best_btest) {
2163 default:
2164 if (cmpxy == NULL)
2165 cmpxy = ideal_cmpxy;
2166 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2167 // and fall through:
2168 case BoolTest::lt: // x < y
2169 case BoolTest::le: // x <= y
2170 answer_if_true = (want_max ? yvalue : xvalue);
2171 answer_if_false = (want_max ? xvalue : yvalue);
2172 break;
2173 case BoolTest::gt: // x > y
2174 case BoolTest::ge: // x >= y
2175 answer_if_true = (want_max ? xvalue : yvalue);
2176 answer_if_false = (want_max ? yvalue : xvalue);
2177 break;
2178 }
2179
2180 jint hi, lo;
2181 if (want_max) {
2182 // We can sharpen the minimum.
2183 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2184 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2185 } else {
2186 // We can sharpen the maximum.
2187 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2188 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2189 }
2190
2191 // Use a flow-free graph structure, to avoid creating excess control edges
2192 // which could hinder other optimizations.
2193 // Since Math.min/max is often used with arraycopy, we want
2194 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2195 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2196 answer_if_false, answer_if_true,
2197 TypeInt::make(lo, hi, widen));
2198
2199 return _gvn.transform(cmov);
2200
2201 /*
2202 // This is not as desirable as it may seem, since Min and Max
2203 // nodes do not have a full set of optimizations.
2204 // And they would interfere, anyway, with 'if' optimizations
2205 // and with CMoveI canonical forms.
2206 switch (id) {
2207 case vmIntrinsics::_min:
2208 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2209 case vmIntrinsics::_max:
2210 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2211 default:
2212 ShouldNotReachHere();
2213 }
2214 */
2215 }
2216
2217 inline int
2218 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2219 const TypePtr* base_type = TypePtr::NULL_PTR;
2220 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2221 if (base_type == NULL) {
2222 // Unknown type.
2223 return Type::AnyPtr;
2224 } else if (base_type == TypePtr::NULL_PTR) {
2225 // Since this is a NULL+long form, we have to switch to a rawptr.
2226 base = _gvn.transform(new (C) CastX2PNode(offset));
2227 offset = MakeConX(0);
2228 return Type::RawPtr;
2229 } else if (base_type->base() == Type::RawPtr) {
2230 return Type::RawPtr;
2231 } else if (base_type->isa_oopptr()) {
2232 // Base is never null => always a heap address.
2233 if (base_type->ptr() == TypePtr::NotNull) {
2234 return Type::OopPtr;
2235 }
2236 // Offset is small => always a heap address.
2237 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2238 if (offset_type != NULL &&
2239 base_type->offset() == 0 && // (should always be?)
2240 offset_type->_lo >= 0 &&
2241 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2242 return Type::OopPtr;
2243 }
2244 // Otherwise, it might either be oop+off or NULL+addr.
2245 return Type::AnyPtr;
2246 } else {
2247 // No information:
2248 return Type::AnyPtr;
2249 }
2250 }
2251
2252 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2253 int kind = classify_unsafe_addr(base, offset);
2254 if (kind == Type::RawPtr) {
2255 return basic_plus_adr(top(), base, offset);
2256 } else {
2257 return basic_plus_adr(base, offset);
2258 }
2259 }
2260
2261 //--------------------------inline_number_methods-----------------------------
2262 // inline int Integer.numberOfLeadingZeros(int)
2263 // inline int Long.numberOfLeadingZeros(long)
2264 //
2265 // inline int Integer.numberOfTrailingZeros(int)
2266 // inline int Long.numberOfTrailingZeros(long)
2267 //
2268 // inline int Integer.bitCount(int)
2269 // inline int Long.bitCount(long)
2270 //
2271 // inline char Character.reverseBytes(char)
2272 // inline short Short.reverseBytes(short)
2273 // inline int Integer.reverseBytes(int)
2274 // inline long Long.reverseBytes(long)
2275 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2276 Node* arg = argument(0);
2277 Node* n;
2278 switch (id) {
2279 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2280 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2281 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2282 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2283 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2284 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2285 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2286 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2287 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2288 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2289 default: fatal_unexpected_iid(id); break;
2290 }
2291 set_result(_gvn.transform(n));
2292 return true;
2293 }
2294
2295 //----------------------------inline_unsafe_access----------------------------
2296
2297 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2298
2299 // Helper that guards and inserts a pre-barrier.
2300 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2301 Node* pre_val, bool need_mem_bar) {
2302 // We could be accessing the referent field of a reference object. If so, when G1
2303 // is enabled, we need to log the value in the referent field in an SATB buffer.
2304 // This routine performs some compile time filters and generates suitable
2305 // runtime filters that guard the pre-barrier code.
2306 // Also add memory barrier for non volatile load from the referent field
2307 // to prevent commoning of loads across safepoint.
2308 if (!UseG1GC && !need_mem_bar)
2309 return;
2310
2311 // Some compile time checks.
2312
2313 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2314 const TypeX* otype = offset->find_intptr_t_type();
2315 if (otype != NULL && otype->is_con() &&
2316 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2317 // Constant offset but not the reference_offset so just return
2318 return;
2319 }
2320
2321 // We only need to generate the runtime guards for instances.
2322 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2323 if (btype != NULL) {
2324 if (btype->isa_aryptr()) {
2325 // Array type so nothing to do
2326 return;
2327 }
2328
2329 const TypeInstPtr* itype = btype->isa_instptr();
2330 if (itype != NULL) {
2331 // Can the klass of base_oop be statically determined to be
2332 // _not_ a sub-class of Reference and _not_ Object?
2333 ciKlass* klass = itype->klass();
2334 if ( klass->is_loaded() &&
2335 !klass->is_subtype_of(env()->Reference_klass()) &&
2336 !env()->Object_klass()->is_subtype_of(klass)) {
2337 return;
2338 }
2339 }
2340 }
2341
2342 // The compile time filters did not reject base_oop/offset so
2343 // we need to generate the following runtime filters
2344 //
2345 // if (offset == java_lang_ref_Reference::_reference_offset) {
2346 // if (instance_of(base, java.lang.ref.Reference)) {
2347 // pre_barrier(_, pre_val, ...);
2348 // }
2349 // }
2350
2351 float likely = PROB_LIKELY( 0.999);
2352 float unlikely = PROB_UNLIKELY(0.999);
2353
2354 IdealKit ideal(this);
2355 #define __ ideal.
2356
2357 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2358
2359 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2360 // Update graphKit memory and control from IdealKit.
2361 sync_kit(ideal);
2362
2363 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2364 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2365
2366 // Update IdealKit memory and control from graphKit.
2367 __ sync_kit(this);
2368
2369 Node* one = __ ConI(1);
2370 // is_instof == 0 if base_oop == NULL
2371 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2372
2373 // Update graphKit from IdeakKit.
2374 sync_kit(ideal);
2375
2376 // Use the pre-barrier to record the value in the referent field
2377 pre_barrier(false /* do_load */,
2378 __ ctrl(),
2379 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2380 pre_val /* pre_val */,
2381 T_OBJECT);
2382 if (need_mem_bar) {
2383 // Add memory barrier to prevent commoning reads from this field
2384 // across safepoint since GC can change its value.
2385 insert_mem_bar(Op_MemBarCPUOrder);
2386 }
2387 // Update IdealKit from graphKit.
2388 __ sync_kit(this);
2389
2390 } __ end_if(); // _ref_type != ref_none
2391 } __ end_if(); // offset == referent_offset
2392
2393 // Final sync IdealKit and GraphKit.
2394 final_sync(ideal);
2395 #undef __
2396 }
2397
2398
2399 // Interpret Unsafe.fieldOffset cookies correctly:
2400 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2401
2402 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2403 // Attempt to infer a sharper value type from the offset and base type.
2404 ciKlass* sharpened_klass = NULL;
2405
2406 // See if it is an instance field, with an object type.
2407 if (alias_type->field() != NULL) {
2408 assert(!is_native_ptr, "native pointer op cannot use a java address");
2409 if (alias_type->field()->type()->is_klass()) {
2410 sharpened_klass = alias_type->field()->type()->as_klass();
2411 }
2412 }
2413
2414 // See if it is a narrow oop array.
2415 if (adr_type->isa_aryptr()) {
2416 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2417 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2418 if (elem_type != NULL) {
2419 sharpened_klass = elem_type->klass();
2420 }
2421 }
2422 }
2423
2424 // The sharpened class might be unloaded if there is no class loader
2425 // contraint in place.
2426 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2427 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2428
2429 #ifndef PRODUCT
2430 if (C->print_intrinsics() || C->print_inlining()) {
2431 tty->print(" from base type: "); adr_type->dump();
2432 tty->print(" sharpened value: "); tjp->dump();
2433 }
2434 #endif
2435 // Sharpen the value type.
2436 return tjp;
2437 }
2438 return NULL;
2439 }
2440
2441 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2442 if (callee()->is_static()) return false; // caller must have the capability!
2443
2444 #ifndef PRODUCT
2445 {
2446 ResourceMark rm;
2447 // Check the signatures.
2448 ciSignature* sig = callee()->signature();
2449 #ifdef ASSERT
2450 if (!is_store) {
2451 // Object getObject(Object base, int/long offset), etc.
2452 BasicType rtype = sig->return_type()->basic_type();
2453 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2454 rtype = T_ADDRESS; // it is really a C void*
2455 assert(rtype == type, "getter must return the expected value");
2456 if (!is_native_ptr) {
2457 assert(sig->count() == 2, "oop getter has 2 arguments");
2458 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2459 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2460 } else {
2461 assert(sig->count() == 1, "native getter has 1 argument");
2462 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2463 }
2464 } else {
2465 // void putObject(Object base, int/long offset, Object x), etc.
2466 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2467 if (!is_native_ptr) {
2468 assert(sig->count() == 3, "oop putter has 3 arguments");
2469 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2470 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2471 } else {
2472 assert(sig->count() == 2, "native putter has 2 arguments");
2473 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2474 }
2475 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2476 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2477 vtype = T_ADDRESS; // it is really a C void*
2478 assert(vtype == type, "putter must accept the expected value");
2479 }
2480 #endif // ASSERT
2481 }
2482 #endif //PRODUCT
2483
2484 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2485
2486 Node* receiver = argument(0); // type: oop
2487
2488 // Build address expression. See the code in inline_unsafe_prefetch.
2489 Node* adr;
2490 Node* heap_base_oop = top();
2491 Node* offset = top();
2492 Node* val;
2493
2494 if (!is_native_ptr) {
2495 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2496 Node* base = argument(1); // type: oop
2497 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2498 offset = argument(2); // type: long
2499 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2500 // to be plain byte offsets, which are also the same as those accepted
2501 // by oopDesc::field_base.
2502 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2503 "fieldOffset must be byte-scaled");
2504 // 32-bit machines ignore the high half!
2505 offset = ConvL2X(offset);
2506 adr = make_unsafe_address(base, offset);
2507 heap_base_oop = base;
2508 val = is_store ? argument(4) : NULL;
2509 } else {
2510 Node* ptr = argument(1); // type: long
2511 ptr = ConvL2X(ptr); // adjust Java long to machine word
2512 adr = make_unsafe_address(NULL, ptr);
2513 val = is_store ? argument(3) : NULL;
2514 }
2515
2516 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2517
2518 // First guess at the value type.
2519 const Type *value_type = Type::get_const_basic_type(type);
2520
2521 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2522 // there was not enough information to nail it down.
2523 Compile::AliasType* alias_type = C->alias_type(adr_type);
2524 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2525
2526 // We will need memory barriers unless we can determine a unique
2527 // alias category for this reference. (Note: If for some reason
2528 // the barriers get omitted and the unsafe reference begins to "pollute"
2529 // the alias analysis of the rest of the graph, either Compile::can_alias
2530 // or Compile::must_alias will throw a diagnostic assert.)
2531 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2532
2533 // If we are reading the value of the referent field of a Reference
2534 // object (either by using Unsafe directly or through reflection)
2535 // then, if G1 is enabled, we need to record the referent in an
2536 // SATB log buffer using the pre-barrier mechanism.
2537 // Also we need to add memory barrier to prevent commoning reads
2538 // from this field across safepoint since GC can change its value.
2539 bool need_read_barrier = !is_native_ptr && !is_store &&
2540 offset != top() && heap_base_oop != top();
2541
2542 if (!is_store && type == T_OBJECT) {
2543 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2544 if (tjp != NULL) {
2545 value_type = tjp;
2546 }
2547 }
2548
2549 receiver = null_check(receiver);
2550 if (stopped()) {
2551 return true;
2552 }
2553 // Heap pointers get a null-check from the interpreter,
2554 // as a courtesy. However, this is not guaranteed by Unsafe,
2555 // and it is not possible to fully distinguish unintended nulls
2556 // from intended ones in this API.
2557
2558 if (is_volatile) {
2559 // We need to emit leading and trailing CPU membars (see below) in
2560 // addition to memory membars when is_volatile. This is a little
2561 // too strong, but avoids the need to insert per-alias-type
2562 // volatile membars (for stores; compare Parse::do_put_xxx), which
2563 // we cannot do effectively here because we probably only have a
2564 // rough approximation of type.
2565 need_mem_bar = true;
2566 // For Stores, place a memory ordering barrier now.
2567 if (is_store) {
2568 insert_mem_bar(Op_MemBarRelease);
2569 } else {
2570 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2571 insert_mem_bar(Op_MemBarVolatile);
2572 }
2573 }
2574 }
2575
2576 // Memory barrier to prevent normal and 'unsafe' accesses from
2577 // bypassing each other. Happens after null checks, so the
2578 // exception paths do not take memory state from the memory barrier,
2579 // so there's no problems making a strong assert about mixing users
2580 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2581 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2582 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2583
2584 if (!is_store) {
2585 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2586 // load value
2587 switch (type) {
2588 case T_BOOLEAN:
2589 case T_CHAR:
2590 case T_BYTE:
2591 case T_SHORT:
2592 case T_INT:
2593 case T_LONG:
2594 case T_FLOAT:
2595 case T_DOUBLE:
2596 break;
2597 case T_OBJECT:
2598 if (need_read_barrier) {
2599 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2600 }
2601 break;
2602 case T_ADDRESS:
2603 // Cast to an int type.
2604 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2605 p = ConvX2L(p);
2606 break;
2607 default:
2608 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2609 break;
2610 }
2611 // The load node has the control of the preceding MemBarCPUOrder. All
2612 // following nodes will have the control of the MemBarCPUOrder inserted at
2613 // the end of this method. So, pushing the load onto the stack at a later
2614 // point is fine.
2615 set_result(p);
2616 } else {
2617 // place effect of store into memory
2618 switch (type) {
2619 case T_DOUBLE:
2620 val = dstore_rounding(val);
2621 break;
2622 case T_ADDRESS:
2623 // Repackage the long as a pointer.
2624 val = ConvL2X(val);
2625 val = _gvn.transform(new (C) CastX2PNode(val));
2626 break;
2627 }
2628
2629 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2630 if (type != T_OBJECT ) {
2631 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2632 } else {
2633 // Possibly an oop being stored to Java heap or native memory
2634 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2635 // oop to Java heap.
2636 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2637 } else {
2638 // We can't tell at compile time if we are storing in the Java heap or outside
2639 // of it. So we need to emit code to conditionally do the proper type of
2640 // store.
2641
2642 IdealKit ideal(this);
2643 #define __ ideal.
2644 // QQQ who knows what probability is here??
2645 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2646 // Sync IdealKit and graphKit.
2647 sync_kit(ideal);
2648 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2649 // Update IdealKit memory.
2650 __ sync_kit(this);
2651 } __ else_(); {
2652 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2653 } __ end_if();
2654 // Final sync IdealKit and GraphKit.
2655 final_sync(ideal);
2656 #undef __
2657 }
2658 }
2659 }
2660
2661 if (is_volatile) {
2662 if (!is_store) {
2663 insert_mem_bar(Op_MemBarAcquire);
2664 } else {
2665 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2666 insert_mem_bar(Op_MemBarVolatile);
2667 }
2668 }
2669 }
2670
2671 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2672
2673 return true;
2674 }
2675
2676 //----------------------------inline_unsafe_prefetch----------------------------
2677
2678 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2679 #ifndef PRODUCT
2680 {
2681 ResourceMark rm;
2682 // Check the signatures.
2683 ciSignature* sig = callee()->signature();
2684 #ifdef ASSERT
2685 // Object getObject(Object base, int/long offset), etc.
2686 BasicType rtype = sig->return_type()->basic_type();
2687 if (!is_native_ptr) {
2688 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2689 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2690 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2691 } else {
2692 assert(sig->count() == 1, "native prefetch has 1 argument");
2693 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2694 }
2695 #endif // ASSERT
2696 }
2697 #endif // !PRODUCT
2698
2699 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2700
2701 const int idx = is_static ? 0 : 1;
2702 if (!is_static) {
2703 null_check_receiver();
2704 if (stopped()) {
2705 return true;
2706 }
2707 }
2708
2709 // Build address expression. See the code in inline_unsafe_access.
2710 Node *adr;
2711 if (!is_native_ptr) {
2712 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2713 Node* base = argument(idx + 0); // type: oop
2714 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2715 Node* offset = argument(idx + 1); // type: long
2716 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2717 // to be plain byte offsets, which are also the same as those accepted
2718 // by oopDesc::field_base.
2719 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2720 "fieldOffset must be byte-scaled");
2721 // 32-bit machines ignore the high half!
2722 offset = ConvL2X(offset);
2723 adr = make_unsafe_address(base, offset);
2724 } else {
2725 Node* ptr = argument(idx + 0); // type: long
2726 ptr = ConvL2X(ptr); // adjust Java long to machine word
2727 adr = make_unsafe_address(NULL, ptr);
2728 }
2729
2730 // Generate the read or write prefetch
2731 Node *prefetch;
2732 if (is_store) {
2733 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2734 } else {
2735 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2736 }
2737 prefetch->init_req(0, control());
2738 set_i_o(_gvn.transform(prefetch));
2739
2740 return true;
2741 }
2742
2743 //----------------------------inline_unsafe_load_store----------------------------
2744 // This method serves a couple of different customers (depending on LoadStoreKind):
2745 //
2746 // LS_cmpxchg:
2747 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2748 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2749 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2750 //
2751 // LS_xadd:
2752 // public int getAndAddInt( Object o, long offset, int delta)
2753 // public long getAndAddLong(Object o, long offset, long delta)
2754 //
2755 // LS_xchg:
2756 // int getAndSet(Object o, long offset, int newValue)
2757 // long getAndSet(Object o, long offset, long newValue)
2758 // Object getAndSet(Object o, long offset, Object newValue)
2759 //
2760 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2761 // This basic scheme here is the same as inline_unsafe_access, but
2762 // differs in enough details that combining them would make the code
2763 // overly confusing. (This is a true fact! I originally combined
2764 // them, but even I was confused by it!) As much code/comments as
2765 // possible are retained from inline_unsafe_access though to make
2766 // the correspondences clearer. - dl
2767
2768 if (callee()->is_static()) return false; // caller must have the capability!
2769
2770 #ifndef PRODUCT
2771 BasicType rtype;
2772 {
2773 ResourceMark rm;
2774 // Check the signatures.
2775 ciSignature* sig = callee()->signature();
2776 rtype = sig->return_type()->basic_type();
2777 if (kind == LS_xadd || kind == LS_xchg) {
2778 // Check the signatures.
2779 #ifdef ASSERT
2780 assert(rtype == type, "get and set must return the expected type");
2781 assert(sig->count() == 3, "get and set has 3 arguments");
2782 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2783 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2784 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2785 #endif // ASSERT
2786 } else if (kind == LS_cmpxchg) {
2787 // Check the signatures.
2788 #ifdef ASSERT
2789 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2790 assert(sig->count() == 4, "CAS has 4 arguments");
2791 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2792 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2793 #endif // ASSERT
2794 } else {
2795 ShouldNotReachHere();
2796 }
2797 }
2798 #endif //PRODUCT
2799
2800 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2801
2802 // Get arguments:
2803 Node* receiver = NULL;
2804 Node* base = NULL;
2805 Node* offset = NULL;
2806 Node* oldval = NULL;
2807 Node* newval = NULL;
2808 if (kind == LS_cmpxchg) {
2809 const bool two_slot_type = type2size[type] == 2;
2810 receiver = argument(0); // type: oop
2811 base = argument(1); // type: oop
2812 offset = argument(2); // type: long
2813 oldval = argument(4); // type: oop, int, or long
2814 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2815 } else if (kind == LS_xadd || kind == LS_xchg){
2816 receiver = argument(0); // type: oop
2817 base = argument(1); // type: oop
2818 offset = argument(2); // type: long
2819 oldval = NULL;
2820 newval = argument(4); // type: oop, int, or long
2821 }
2822
2823 // Null check receiver.
2824 receiver = null_check(receiver);
2825 if (stopped()) {
2826 return true;
2827 }
2828
2829 // Build field offset expression.
2830 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2831 // to be plain byte offsets, which are also the same as those accepted
2832 // by oopDesc::field_base.
2833 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2834 // 32-bit machines ignore the high half of long offsets
2835 offset = ConvL2X(offset);
2836 Node* adr = make_unsafe_address(base, offset);
2837 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2838
2839 // For CAS, unlike inline_unsafe_access, there seems no point in
2840 // trying to refine types. Just use the coarse types here.
2841 const Type *value_type = Type::get_const_basic_type(type);
2842 Compile::AliasType* alias_type = C->alias_type(adr_type);
2843 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2844
2845 if (kind == LS_xchg && type == T_OBJECT) {
2846 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2847 if (tjp != NULL) {
2848 value_type = tjp;
2849 }
2850 }
2851
2852 int alias_idx = C->get_alias_index(adr_type);
2853
2854 // Memory-model-wise, a LoadStore acts like a little synchronized
2855 // block, so needs barriers on each side. These don't translate
2856 // into actual barriers on most machines, but we still need rest of
2857 // compiler to respect ordering.
2858
2859 insert_mem_bar(Op_MemBarRelease);
2860 insert_mem_bar(Op_MemBarCPUOrder);
2861
2862 // 4984716: MemBars must be inserted before this
2863 // memory node in order to avoid a false
2864 // dependency which will confuse the scheduler.
2865 Node *mem = memory(alias_idx);
2866
2867 // For now, we handle only those cases that actually exist: ints,
2868 // longs, and Object. Adding others should be straightforward.
2869 Node* load_store;
2870 switch(type) {
2871 case T_INT:
2872 if (kind == LS_xadd) {
2873 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2874 } else if (kind == LS_xchg) {
2875 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2876 } else if (kind == LS_cmpxchg) {
2877 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2878 } else {
2879 ShouldNotReachHere();
2880 }
2881 break;
2882 case T_LONG:
2883 if (kind == LS_xadd) {
2884 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2885 } else if (kind == LS_xchg) {
2886 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2887 } else if (kind == LS_cmpxchg) {
2888 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2889 } else {
2890 ShouldNotReachHere();
2891 }
2892 break;
2893 case T_OBJECT:
2894 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2895 // could be delayed during Parse (for example, in adjust_map_after_if()).
2896 // Execute transformation here to avoid barrier generation in such case.
2897 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2898 newval = _gvn.makecon(TypePtr::NULL_PTR);
2899
2900 // Reference stores need a store barrier.
2901 if (kind == LS_xchg) {
2902 // If pre-barrier must execute before the oop store, old value will require do_load here.
2903 if (!can_move_pre_barrier()) {
2904 pre_barrier(true /* do_load*/,
2905 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2906 NULL /* pre_val*/,
2907 T_OBJECT);
2908 } // Else move pre_barrier to use load_store value, see below.
2909 } else if (kind == LS_cmpxchg) {
2910 // Same as for newval above:
2911 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2912 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2913 }
2914 // The only known value which might get overwritten is oldval.
2915 pre_barrier(false /* do_load */,
2916 control(), NULL, NULL, max_juint, NULL, NULL,
2917 oldval /* pre_val */,
2918 T_OBJECT);
2919 } else {
2920 ShouldNotReachHere();
2921 }
2922
2923 #ifdef _LP64
2924 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2925 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2926 if (kind == LS_xchg) {
2927 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2928 newval_enc, adr_type, value_type->make_narrowoop()));
2929 } else {
2930 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2931 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2932 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2933 newval_enc, oldval_enc));
2934 }
2935 } else
2936 #endif
2937 {
2938 if (kind == LS_xchg) {
2939 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2940 } else {
2941 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2942 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2943 }
2944 }
2945 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2946 break;
2947 default:
2948 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2949 break;
2950 }
2951
2952 // SCMemProjNodes represent the memory state of a LoadStore. Their
2953 // main role is to prevent LoadStore nodes from being optimized away
2954 // when their results aren't used.
2955 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2956 set_memory(proj, alias_idx);
2957
2958 if (type == T_OBJECT && kind == LS_xchg) {
2959 #ifdef _LP64
2960 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2961 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2962 }
2963 #endif
2964 if (can_move_pre_barrier()) {
2965 // Don't need to load pre_val. The old value is returned by load_store.
2966 // The pre_barrier can execute after the xchg as long as no safepoint
2967 // gets inserted between them.
2968 pre_barrier(false /* do_load */,
2969 control(), NULL, NULL, max_juint, NULL, NULL,
2970 load_store /* pre_val */,
2971 T_OBJECT);
2972 }
2973 }
2974
2975 // Add the trailing membar surrounding the access
2976 insert_mem_bar(Op_MemBarCPUOrder);
2977 insert_mem_bar(Op_MemBarAcquire);
2978
2979 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2980 set_result(load_store);
2981 return true;
2982 }
2983
2984 //----------------------------inline_unsafe_ordered_store----------------------
2985 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2986 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2987 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2988 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2989 // This is another variant of inline_unsafe_access, differing in
2990 // that it always issues store-store ("release") barrier and ensures
2991 // store-atomicity (which only matters for "long").
2992
2993 if (callee()->is_static()) return false; // caller must have the capability!
2994
2995 #ifndef PRODUCT
2996 {
2997 ResourceMark rm;
2998 // Check the signatures.
2999 ciSignature* sig = callee()->signature();
3000 #ifdef ASSERT
3001 BasicType rtype = sig->return_type()->basic_type();
3002 assert(rtype == T_VOID, "must return void");
3003 assert(sig->count() == 3, "has 3 arguments");
3004 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3005 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3006 #endif // ASSERT
3007 }
3008 #endif //PRODUCT
3009
3010 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3011
3012 // Get arguments:
3013 Node* receiver = argument(0); // type: oop
3014 Node* base = argument(1); // type: oop
3015 Node* offset = argument(2); // type: long
3016 Node* val = argument(4); // type: oop, int, or long
3017
3018 // Null check receiver.
3019 receiver = null_check(receiver);
3020 if (stopped()) {
3021 return true;
3022 }
3023
3024 // Build field offset expression.
3025 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3026 // 32-bit machines ignore the high half of long offsets
3027 offset = ConvL2X(offset);
3028 Node* adr = make_unsafe_address(base, offset);
3029 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3030 const Type *value_type = Type::get_const_basic_type(type);
3031 Compile::AliasType* alias_type = C->alias_type(adr_type);
3032
3033 insert_mem_bar(Op_MemBarRelease);
3034 insert_mem_bar(Op_MemBarCPUOrder);
3035 // Ensure that the store is atomic for longs:
3036 const bool require_atomic_access = true;
3037 Node* store;
3038 if (type == T_OBJECT) // reference stores need a store barrier.
3039 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3040 else {
3041 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3042 }
3043 insert_mem_bar(Op_MemBarCPUOrder);
3044 return true;
3045 }
3046
3047 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3048 // Regardless of form, don't allow previous ld/st to move down,
3049 // then issue acquire, release, or volatile mem_bar.
3050 insert_mem_bar(Op_MemBarCPUOrder);
3051 switch(id) {
3052 case vmIntrinsics::_loadFence:
3053 insert_mem_bar(Op_LoadFence);
3054 return true;
3055 case vmIntrinsics::_storeFence:
3056 insert_mem_bar(Op_StoreFence);
3057 return true;
3058 case vmIntrinsics::_fullFence:
3059 insert_mem_bar(Op_MemBarVolatile);
3060 return true;
3061 default:
3062 fatal_unexpected_iid(id);
3063 return false;
3064 }
3065 }
3066
3067 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3068 if (!kls->is_Con()) {
3069 return true;
3070 }
3071 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3072 if (klsptr == NULL) {
3073 return true;
3074 }
3075 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3076 // don't need a guard for a klass that is already initialized
3077 return !ik->is_initialized();
3078 }
3079
3080 //----------------------------inline_unsafe_allocate---------------------------
3081 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3082 bool LibraryCallKit::inline_unsafe_allocate() {
3083 if (callee()->is_static()) return false; // caller must have the capability!
3084
3085 null_check_receiver(); // null-check, then ignore
3086 Node* cls = null_check(argument(1));
3087 if (stopped()) return true;
3088
3089 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3090 kls = null_check(kls);
3091 if (stopped()) return true; // argument was like int.class
3092
3093 Node* test = NULL;
3094 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3095 // Note: The argument might still be an illegal value like
3096 // Serializable.class or Object[].class. The runtime will handle it.
3097 // But we must make an explicit check for initialization.
3098 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3099 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3100 // can generate code to load it as unsigned byte.
3101 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3102 Node* bits = intcon(InstanceKlass::fully_initialized);
3103 test = _gvn.transform(new (C) SubINode(inst, bits));
3104 // The 'test' is non-zero if we need to take a slow path.
3105 }
3106
3107 Node* obj = new_instance(kls, test);
3108 set_result(obj);
3109 return true;
3110 }
3111
3112 #ifdef TRACE_HAVE_INTRINSICS
3113 /*
3114 * oop -> myklass
3115 * myklass->trace_id |= USED
3116 * return myklass->trace_id & ~0x3
3117 */
3118 bool LibraryCallKit::inline_native_classID() {
3119 null_check_receiver(); // null-check, then ignore
3120 Node* cls = null_check(argument(1), T_OBJECT);
3121 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3122 kls = null_check(kls, T_OBJECT);
3123 ByteSize offset = TRACE_ID_OFFSET;
3124 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3125 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3126 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3127 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3128 Node* clsused = longcon(0x01l); // set the class bit
3129 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3130
3131 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3132 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3133 set_result(andl);
3134 return true;
3135 }
3136
3137 bool LibraryCallKit::inline_native_threadID() {
3138 Node* tls_ptr = NULL;
3139 Node* cur_thr = generate_current_thread(tls_ptr);
3140 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3141 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3142 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3143
3144 Node* threadid = NULL;
3145 size_t thread_id_size = OSThread::thread_id_size();
3146 if (thread_id_size == (size_t) BytesPerLong) {
3147 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3148 } else if (thread_id_size == (size_t) BytesPerInt) {
3149 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3150 } else {
3151 ShouldNotReachHere();
3152 }
3153 set_result(threadid);
3154 return true;
3155 }
3156 #endif
3157
3158 //------------------------inline_native_time_funcs--------------
3159 // inline code for System.currentTimeMillis() and System.nanoTime()
3160 // these have the same type and signature
3161 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3162 const TypeFunc* tf = OptoRuntime::void_long_Type();
3163 const TypePtr* no_memory_effects = NULL;
3164 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3165 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3166 #ifdef ASSERT
3167 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3168 assert(value_top == top(), "second value must be top");
3169 #endif
3170 set_result(value);
3171 return true;
3172 }
3173
3174 //------------------------inline_native_currentThread------------------
3175 bool LibraryCallKit::inline_native_currentThread() {
3176 Node* junk = NULL;
3177 set_result(generate_current_thread(junk));
3178 return true;
3179 }
3180
3181 //------------------------inline_native_isInterrupted------------------
3182 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3183 bool LibraryCallKit::inline_native_isInterrupted() {
3184 // Add a fast path to t.isInterrupted(clear_int):
3185 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3186 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3187 // So, in the common case that the interrupt bit is false,
3188 // we avoid making a call into the VM. Even if the interrupt bit
3189 // is true, if the clear_int argument is false, we avoid the VM call.
3190 // However, if the receiver is not currentThread, we must call the VM,
3191 // because there must be some locking done around the operation.
3192
3193 // We only go to the fast case code if we pass two guards.
3194 // Paths which do not pass are accumulated in the slow_region.
3195
3196 enum {
3197 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3198 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3199 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3200 PATH_LIMIT
3201 };
3202
3203 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3204 // out of the function.
3205 insert_mem_bar(Op_MemBarCPUOrder);
3206
3207 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3208 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3209
3210 RegionNode* slow_region = new (C) RegionNode(1);
3211 record_for_igvn(slow_region);
3212
3213 // (a) Receiving thread must be the current thread.
3214 Node* rec_thr = argument(0);
3215 Node* tls_ptr = NULL;
3216 Node* cur_thr = generate_current_thread(tls_ptr);
3217 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3218 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3219
3220 generate_slow_guard(bol_thr, slow_region);
3221
3222 // (b) Interrupt bit on TLS must be false.
3223 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3224 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3225 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3226
3227 // Set the control input on the field _interrupted read to prevent it floating up.
3228 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3229 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3230 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3231
3232 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3233
3234 // First fast path: if (!TLS._interrupted) return false;
3235 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3236 result_rgn->init_req(no_int_result_path, false_bit);
3237 result_val->init_req(no_int_result_path, intcon(0));
3238
3239 // drop through to next case
3240 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3241
3242 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3243 Node* clr_arg = argument(1);
3244 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3245 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3246 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3247
3248 // Second fast path: ... else if (!clear_int) return true;
3249 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3250 result_rgn->init_req(no_clear_result_path, false_arg);
3251 result_val->init_req(no_clear_result_path, intcon(1));
3252
3253 // drop through to next case
3254 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3255
3256 // (d) Otherwise, go to the slow path.
3257 slow_region->add_req(control());
3258 set_control( _gvn.transform(slow_region));
3259
3260 if (stopped()) {
3261 // There is no slow path.
3262 result_rgn->init_req(slow_result_path, top());
3263 result_val->init_req(slow_result_path, top());
3264 } else {
3265 // non-virtual because it is a private non-static
3266 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3267
3268 Node* slow_val = set_results_for_java_call(slow_call);
3269 // this->control() comes from set_results_for_java_call
3270
3271 Node* fast_io = slow_call->in(TypeFunc::I_O);
3272 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3273
3274 // These two phis are pre-filled with copies of of the fast IO and Memory
3275 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3276 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3277
3278 result_rgn->init_req(slow_result_path, control());
3279 result_io ->init_req(slow_result_path, i_o());
3280 result_mem->init_req(slow_result_path, reset_memory());
3281 result_val->init_req(slow_result_path, slow_val);
3282
3283 set_all_memory(_gvn.transform(result_mem));
3284 set_i_o( _gvn.transform(result_io));
3285 }
3286
3287 C->set_has_split_ifs(true); // Has chance for split-if optimization
3288 set_result(result_rgn, result_val);
3289 return true;
3290 }
3291
3292 //---------------------------load_mirror_from_klass----------------------------
3293 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3294 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3295 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3296 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3297 }
3298
3299 //-----------------------load_klass_from_mirror_common-------------------------
3300 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3301 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3302 // and branch to the given path on the region.
3303 // If never_see_null, take an uncommon trap on null, so we can optimistically
3304 // compile for the non-null case.
3305 // If the region is NULL, force never_see_null = true.
3306 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3307 bool never_see_null,
3308 RegionNode* region,
3309 int null_path,
3310 int offset) {
3311 if (region == NULL) never_see_null = true;
3312 Node* p = basic_plus_adr(mirror, offset);
3313 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3314 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3315 Node* null_ctl = top();
3316 kls = null_check_oop(kls, &null_ctl, never_see_null);
3317 if (region != NULL) {
3318 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3319 region->init_req(null_path, null_ctl);
3320 } else {
3321 assert(null_ctl == top(), "no loose ends");
3322 }
3323 return kls;
3324 }
3325
3326 //--------------------(inline_native_Class_query helpers)---------------------
3327 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3328 // Fall through if (mods & mask) == bits, take the guard otherwise.
3329 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3330 // Branch around if the given klass has the given modifier bit set.
3331 // Like generate_guard, adds a new path onto the region.
3332 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3333 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3334 Node* mask = intcon(modifier_mask);
3335 Node* bits = intcon(modifier_bits);
3336 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3337 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3338 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3339 return generate_fair_guard(bol, region);
3340 }
3341 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3342 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3343 }
3344
3345 //-------------------------inline_native_Class_query-------------------
3346 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3347 const Type* return_type = TypeInt::BOOL;
3348 Node* prim_return_value = top(); // what happens if it's a primitive class?
3349 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3350 bool expect_prim = false; // most of these guys expect to work on refs
3351
3352 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3353
3354 Node* mirror = argument(0);
3355 Node* obj = top();
3356
3357 switch (id) {
3358 case vmIntrinsics::_isInstance:
3359 // nothing is an instance of a primitive type
3360 prim_return_value = intcon(0);
3361 obj = argument(1);
3362 break;
3363 case vmIntrinsics::_getModifiers:
3364 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3365 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3366 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3367 break;
3368 case vmIntrinsics::_isInterface:
3369 prim_return_value = intcon(0);
3370 break;
3371 case vmIntrinsics::_isArray:
3372 prim_return_value = intcon(0);
3373 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3374 break;
3375 case vmIntrinsics::_isPrimitive:
3376 prim_return_value = intcon(1);
3377 expect_prim = true; // obviously
3378 break;
3379 case vmIntrinsics::_getSuperclass:
3380 prim_return_value = null();
3381 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3382 break;
3383 case vmIntrinsics::_getComponentType:
3384 prim_return_value = null();
3385 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3386 break;
3387 case vmIntrinsics::_getClassAccessFlags:
3388 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3389 return_type = TypeInt::INT; // not bool! 6297094
3390 break;
3391 default:
3392 fatal_unexpected_iid(id);
3393 break;
3394 }
3395
3396 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3397 if (mirror_con == NULL) return false; // cannot happen?
3398
3399 #ifndef PRODUCT
3400 if (C->print_intrinsics() || C->print_inlining()) {
3401 ciType* k = mirror_con->java_mirror_type();
3402 if (k) {
3403 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3404 k->print_name();
3405 tty->cr();
3406 }
3407 }
3408 #endif
3409
3410 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3411 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3412 record_for_igvn(region);
3413 PhiNode* phi = new (C) PhiNode(region, return_type);
3414
3415 // The mirror will never be null of Reflection.getClassAccessFlags, however
3416 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3417 // if it is. See bug 4774291.
3418
3419 // For Reflection.getClassAccessFlags(), the null check occurs in
3420 // the wrong place; see inline_unsafe_access(), above, for a similar
3421 // situation.
3422 mirror = null_check(mirror);
3423 // If mirror or obj is dead, only null-path is taken.
3424 if (stopped()) return true;
3425
3426 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3427
3428 // Now load the mirror's klass metaobject, and null-check it.
3429 // Side-effects region with the control path if the klass is null.
3430 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3431 // If kls is null, we have a primitive mirror.
3432 phi->init_req(_prim_path, prim_return_value);
3433 if (stopped()) { set_result(region, phi); return true; }
3434 bool safe_for_replace = (region->in(_prim_path) == top());
3435
3436 Node* p; // handy temp
3437 Node* null_ctl;
3438
3439 // Now that we have the non-null klass, we can perform the real query.
3440 // For constant classes, the query will constant-fold in LoadNode::Value.
3441 Node* query_value = top();
3442 switch (id) {
3443 case vmIntrinsics::_isInstance:
3444 // nothing is an instance of a primitive type
3445 query_value = gen_instanceof(obj, kls, safe_for_replace);
3446 break;
3447
3448 case vmIntrinsics::_getModifiers:
3449 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3450 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3451 break;
3452
3453 case vmIntrinsics::_isInterface:
3454 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3455 if (generate_interface_guard(kls, region) != NULL)
3456 // A guard was added. If the guard is taken, it was an interface.
3457 phi->add_req(intcon(1));
3458 // If we fall through, it's a plain class.
3459 query_value = intcon(0);
3460 break;
3461
3462 case vmIntrinsics::_isArray:
3463 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3464 if (generate_array_guard(kls, region) != NULL)
3465 // A guard was added. If the guard is taken, it was an array.
3466 phi->add_req(intcon(1));
3467 // If we fall through, it's a plain class.
3468 query_value = intcon(0);
3469 break;
3470
3471 case vmIntrinsics::_isPrimitive:
3472 query_value = intcon(0); // "normal" path produces false
3473 break;
3474
3475 case vmIntrinsics::_getSuperclass:
3476 // The rules here are somewhat unfortunate, but we can still do better
3477 // with random logic than with a JNI call.
3478 // Interfaces store null or Object as _super, but must report null.
3479 // Arrays store an intermediate super as _super, but must report Object.
3480 // Other types can report the actual _super.
3481 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3482 if (generate_interface_guard(kls, region) != NULL)
3483 // A guard was added. If the guard is taken, it was an interface.
3484 phi->add_req(null());
3485 if (generate_array_guard(kls, region) != NULL)
3486 // A guard was added. If the guard is taken, it was an array.
3487 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3488 // If we fall through, it's a plain class. Get its _super.
3489 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3490 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3491 null_ctl = top();
3492 kls = null_check_oop(kls, &null_ctl);
3493 if (null_ctl != top()) {
3494 // If the guard is taken, Object.superClass is null (both klass and mirror).
3495 region->add_req(null_ctl);
3496 phi ->add_req(null());
3497 }
3498 if (!stopped()) {
3499 query_value = load_mirror_from_klass(kls);
3500 }
3501 break;
3502
3503 case vmIntrinsics::_getComponentType:
3504 if (generate_array_guard(kls, region) != NULL) {
3505 // Be sure to pin the oop load to the guard edge just created:
3506 Node* is_array_ctrl = region->in(region->req()-1);
3507 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3508 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3509 phi->add_req(cmo);
3510 }
3511 query_value = null(); // non-array case is null
3512 break;
3513
3514 case vmIntrinsics::_getClassAccessFlags:
3515 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3516 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3517 break;
3518
3519 default:
3520 fatal_unexpected_iid(id);
3521 break;
3522 }
3523
3524 // Fall-through is the normal case of a query to a real class.
3525 phi->init_req(1, query_value);
3526 region->init_req(1, control());
3527
3528 C->set_has_split_ifs(true); // Has chance for split-if optimization
3529 set_result(region, phi);
3530 return true;
3531 }
3532
3533 //--------------------------inline_native_subtype_check------------------------
3534 // This intrinsic takes the JNI calls out of the heart of
3535 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3536 bool LibraryCallKit::inline_native_subtype_check() {
3537 // Pull both arguments off the stack.
3538 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3539 args[0] = argument(0);
3540 args[1] = argument(1);
3541 Node* klasses[2]; // corresponding Klasses: superk, subk
3542 klasses[0] = klasses[1] = top();
3543
3544 enum {
3545 // A full decision tree on {superc is prim, subc is prim}:
3546 _prim_0_path = 1, // {P,N} => false
3547 // {P,P} & superc!=subc => false
3548 _prim_same_path, // {P,P} & superc==subc => true
3549 _prim_1_path, // {N,P} => false
3550 _ref_subtype_path, // {N,N} & subtype check wins => true
3551 _both_ref_path, // {N,N} & subtype check loses => false
3552 PATH_LIMIT
3553 };
3554
3555 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3556 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3557 record_for_igvn(region);
3558
3559 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3560 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3561 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3562
3563 // First null-check both mirrors and load each mirror's klass metaobject.
3564 int which_arg;
3565 for (which_arg = 0; which_arg <= 1; which_arg++) {
3566 Node* arg = args[which_arg];
3567 arg = null_check(arg);
3568 if (stopped()) break;
3569 args[which_arg] = arg;
3570
3571 Node* p = basic_plus_adr(arg, class_klass_offset);
3572 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3573 klasses[which_arg] = _gvn.transform(kls);
3574 }
3575
3576 // Having loaded both klasses, test each for null.
3577 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3578 for (which_arg = 0; which_arg <= 1; which_arg++) {
3579 Node* kls = klasses[which_arg];
3580 Node* null_ctl = top();
3581 kls = null_check_oop(kls, &null_ctl, never_see_null);
3582 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3583 region->init_req(prim_path, null_ctl);
3584 if (stopped()) break;
3585 klasses[which_arg] = kls;
3586 }
3587
3588 if (!stopped()) {
3589 // now we have two reference types, in klasses[0..1]
3590 Node* subk = klasses[1]; // the argument to isAssignableFrom
3591 Node* superk = klasses[0]; // the receiver
3592 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3593 // now we have a successful reference subtype check
3594 region->set_req(_ref_subtype_path, control());
3595 }
3596
3597 // If both operands are primitive (both klasses null), then
3598 // we must return true when they are identical primitives.
3599 // It is convenient to test this after the first null klass check.
3600 set_control(region->in(_prim_0_path)); // go back to first null check
3601 if (!stopped()) {
3602 // Since superc is primitive, make a guard for the superc==subc case.
3603 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3604 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3605 generate_guard(bol_eq, region, PROB_FAIR);
3606 if (region->req() == PATH_LIMIT+1) {
3607 // A guard was added. If the added guard is taken, superc==subc.
3608 region->swap_edges(PATH_LIMIT, _prim_same_path);
3609 region->del_req(PATH_LIMIT);
3610 }
3611 region->set_req(_prim_0_path, control()); // Not equal after all.
3612 }
3613
3614 // these are the only paths that produce 'true':
3615 phi->set_req(_prim_same_path, intcon(1));
3616 phi->set_req(_ref_subtype_path, intcon(1));
3617
3618 // pull together the cases:
3619 assert(region->req() == PATH_LIMIT, "sane region");
3620 for (uint i = 1; i < region->req(); i++) {
3621 Node* ctl = region->in(i);
3622 if (ctl == NULL || ctl == top()) {
3623 region->set_req(i, top());
3624 phi ->set_req(i, top());
3625 } else if (phi->in(i) == NULL) {
3626 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3627 }
3628 }
3629
3630 set_control(_gvn.transform(region));
3631 set_result(_gvn.transform(phi));
3632 return true;
3633 }
3634
3635 //---------------------generate_array_guard_common------------------------
3636 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3637 bool obj_array, bool not_array) {
3638 // If obj_array/non_array==false/false:
3639 // Branch around if the given klass is in fact an array (either obj or prim).
3640 // If obj_array/non_array==false/true:
3641 // Branch around if the given klass is not an array klass of any kind.
3642 // If obj_array/non_array==true/true:
3643 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3644 // If obj_array/non_array==true/false:
3645 // Branch around if the kls is an oop array (Object[] or subtype)
3646 //
3647 // Like generate_guard, adds a new path onto the region.
3648 jint layout_con = 0;
3649 Node* layout_val = get_layout_helper(kls, layout_con);
3650 if (layout_val == NULL) {
3651 bool query = (obj_array
3652 ? Klass::layout_helper_is_objArray(layout_con)
3653 : Klass::layout_helper_is_array(layout_con));
3654 if (query == not_array) {
3655 return NULL; // never a branch
3656 } else { // always a branch
3657 Node* always_branch = control();
3658 if (region != NULL)
3659 region->add_req(always_branch);
3660 set_control(top());
3661 return always_branch;
3662 }
3663 }
3664 // Now test the correct condition.
3665 jint nval = (obj_array
3666 ? ((jint)Klass::_lh_array_tag_type_value
3667 << Klass::_lh_array_tag_shift)
3668 : Klass::_lh_neutral_value);
3669 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3670 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3671 // invert the test if we are looking for a non-array
3672 if (not_array) btest = BoolTest(btest).negate();
3673 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3674 return generate_fair_guard(bol, region);
3675 }
3676
3677
3678 //-----------------------inline_native_newArray--------------------------
3679 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3680 bool LibraryCallKit::inline_native_newArray() {
3681 Node* mirror = argument(0);
3682 Node* count_val = argument(1);
3683
3684 mirror = null_check(mirror);
3685 // If mirror or obj is dead, only null-path is taken.
3686 if (stopped()) return true;
3687
3688 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3689 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3690 PhiNode* result_val = new(C) PhiNode(result_reg,
3691 TypeInstPtr::NOTNULL);
3692 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3693 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3694 TypePtr::BOTTOM);
3695
3696 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3697 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3698 result_reg, _slow_path);
3699 Node* normal_ctl = control();
3700 Node* no_array_ctl = result_reg->in(_slow_path);
3701
3702 // Generate code for the slow case. We make a call to newArray().
3703 set_control(no_array_ctl);
3704 if (!stopped()) {
3705 // Either the input type is void.class, or else the
3706 // array klass has not yet been cached. Either the
3707 // ensuing call will throw an exception, or else it
3708 // will cache the array klass for next time.
3709 PreserveJVMState pjvms(this);
3710 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3711 Node* slow_result = set_results_for_java_call(slow_call);
3712 // this->control() comes from set_results_for_java_call
3713 result_reg->set_req(_slow_path, control());
3714 result_val->set_req(_slow_path, slow_result);
3715 result_io ->set_req(_slow_path, i_o());
3716 result_mem->set_req(_slow_path, reset_memory());
3717 }
3718
3719 set_control(normal_ctl);
3720 if (!stopped()) {
3721 // Normal case: The array type has been cached in the java.lang.Class.
3722 // The following call works fine even if the array type is polymorphic.
3723 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3724 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3725 result_reg->init_req(_normal_path, control());
3726 result_val->init_req(_normal_path, obj);
3727 result_io ->init_req(_normal_path, i_o());
3728 result_mem->init_req(_normal_path, reset_memory());
3729 }
3730
3731 // Return the combined state.
3732 set_i_o( _gvn.transform(result_io) );
3733 set_all_memory( _gvn.transform(result_mem));
3734
3735 C->set_has_split_ifs(true); // Has chance for split-if optimization
3736 set_result(result_reg, result_val);
3737 return true;
3738 }
3739
3740 //----------------------inline_native_getLength--------------------------
3741 // public static native int java.lang.reflect.Array.getLength(Object array);
3742 bool LibraryCallKit::inline_native_getLength() {
3743 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3744
3745 Node* array = null_check(argument(0));
3746 // If array is dead, only null-path is taken.
3747 if (stopped()) return true;
3748
3749 // Deoptimize if it is a non-array.
3750 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3751
3752 if (non_array != NULL) {
3753 PreserveJVMState pjvms(this);
3754 set_control(non_array);
3755 uncommon_trap(Deoptimization::Reason_intrinsic,
3756 Deoptimization::Action_maybe_recompile);
3757 }
3758
3759 // If control is dead, only non-array-path is taken.
3760 if (stopped()) return true;
3761
3762 // The works fine even if the array type is polymorphic.
3763 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3764 Node* result = load_array_length(array);
3765
3766 C->set_has_split_ifs(true); // Has chance for split-if optimization
3767 set_result(result);
3768 return true;
3769 }
3770
3771 //------------------------inline_array_copyOf----------------------------
3772 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3773 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3774 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3775 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3776
3777 // Get the arguments.
3778 Node* original = argument(0);
3779 Node* start = is_copyOfRange? argument(1): intcon(0);
3780 Node* end = is_copyOfRange? argument(2): argument(1);
3781 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3782
3783 Node* newcopy;
3784
3785 // Set the original stack and the reexecute bit for the interpreter to reexecute
3786 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3787 { PreserveReexecuteState preexecs(this);
3788 jvms()->set_should_reexecute(true);
3789
3790 array_type_mirror = null_check(array_type_mirror);
3791 original = null_check(original);
3792
3793 // Check if a null path was taken unconditionally.
3794 if (stopped()) return true;
3795
3796 Node* orig_length = load_array_length(original);
3797
3798 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3799 klass_node = null_check(klass_node);
3800
3801 RegionNode* bailout = new (C) RegionNode(1);
3802 record_for_igvn(bailout);
3803
3804 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3805 // Bail out if that is so.
3806 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3807 if (not_objArray != NULL) {
3808 // Improve the klass node's type from the new optimistic assumption:
3809 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3810 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3811 Node* cast = new (C) CastPPNode(klass_node, akls);
3812 cast->init_req(0, control());
3813 klass_node = _gvn.transform(cast);
3814 }
3815
3816 // Bail out if either start or end is negative.
3817 generate_negative_guard(start, bailout, &start);
3818 generate_negative_guard(end, bailout, &end);
3819
3820 Node* length = end;
3821 if (_gvn.type(start) != TypeInt::ZERO) {
3822 length = _gvn.transform(new (C) SubINode(end, start));
3823 }
3824
3825 // Bail out if length is negative.
3826 // Without this the new_array would throw
3827 // NegativeArraySizeException but IllegalArgumentException is what
3828 // should be thrown
3829 generate_negative_guard(length, bailout, &length);
3830
3831 if (bailout->req() > 1) {
3832 PreserveJVMState pjvms(this);
3833 set_control(_gvn.transform(bailout));
3834 uncommon_trap(Deoptimization::Reason_intrinsic,
3835 Deoptimization::Action_maybe_recompile);
3836 }
3837
3838 if (!stopped()) {
3839 // How many elements will we copy from the original?
3840 // The answer is MinI(orig_length - start, length).
3841 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3842 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3843
3844 newcopy = new_array(klass_node, length, 0); // no argments to push
3845
3846 // Generate a direct call to the right arraycopy function(s).
3847 // We know the copy is disjoint but we might not know if the
3848 // oop stores need checking.
3849 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3850 // This will fail a store-check if x contains any non-nulls.
3851 bool disjoint_bases = true;
3852 // if start > orig_length then the length of the copy may be
3853 // negative.
3854 bool length_never_negative = !is_copyOfRange;
3855 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3856 original, start, newcopy, intcon(0), moved,
3857 disjoint_bases, length_never_negative);
3858 }
3859 } // original reexecute is set back here
3860
3861 C->set_has_split_ifs(true); // Has chance for split-if optimization
3862 if (!stopped()) {
3863 set_result(newcopy);
3864 }
3865 return true;
3866 }
3867
3868
3869 //----------------------generate_virtual_guard---------------------------
3870 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3871 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3872 RegionNode* slow_region) {
3873 ciMethod* method = callee();
3874 int vtable_index = method->vtable_index();
3875 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3876 err_msg_res("bad index %d", vtable_index));
3877 // Get the Method* out of the appropriate vtable entry.
3878 int entry_offset = (InstanceKlass::vtable_start_offset() +
3879 vtable_index*vtableEntry::size()) * wordSize +
3880 vtableEntry::method_offset_in_bytes();
3881 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3882 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3883
3884 // Compare the target method with the expected method (e.g., Object.hashCode).
3885 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3886
3887 Node* native_call = makecon(native_call_addr);
3888 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3889 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3890
3891 return generate_slow_guard(test_native, slow_region);
3892 }
3893
3894 //-----------------------generate_method_call----------------------------
3895 // Use generate_method_call to make a slow-call to the real
3896 // method if the fast path fails. An alternative would be to
3897 // use a stub like OptoRuntime::slow_arraycopy_Java.
3898 // This only works for expanding the current library call,
3899 // not another intrinsic. (E.g., don't use this for making an
3900 // arraycopy call inside of the copyOf intrinsic.)
3901 CallJavaNode*
3902 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3903 // When compiling the intrinsic method itself, do not use this technique.
3904 guarantee(callee() != C->method(), "cannot make slow-call to self");
3905
3906 ciMethod* method = callee();
3907 // ensure the JVMS we have will be correct for this call
3908 guarantee(method_id == method->intrinsic_id(), "must match");
3909
3910 const TypeFunc* tf = TypeFunc::make(method);
3911 CallJavaNode* slow_call;
3912 if (is_static) {
3913 assert(!is_virtual, "");
3914 slow_call = new(C) CallStaticJavaNode(C, tf,
3915 SharedRuntime::get_resolve_static_call_stub(),
3916 method, bci());
3917 } else if (is_virtual) {
3918 null_check_receiver();
3919 int vtable_index = Method::invalid_vtable_index;
3920 if (UseInlineCaches) {
3921 // Suppress the vtable call
3922 } else {
3923 // hashCode and clone are not a miranda methods,
3924 // so the vtable index is fixed.
3925 // No need to use the linkResolver to get it.
3926 vtable_index = method->vtable_index();
3927 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3928 err_msg_res("bad index %d", vtable_index));
3929 }
3930 slow_call = new(C) CallDynamicJavaNode(tf,
3931 SharedRuntime::get_resolve_virtual_call_stub(),
3932 method, vtable_index, bci());
3933 } else { // neither virtual nor static: opt_virtual
3934 null_check_receiver();
3935 slow_call = new(C) CallStaticJavaNode(C, tf,
3936 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3937 method, bci());
3938 slow_call->set_optimized_virtual(true);
3939 }
3940 set_arguments_for_java_call(slow_call);
3941 set_edges_for_java_call(slow_call);
3942 return slow_call;
3943 }
3944
3945
3946 //------------------------------inline_native_hashcode--------------------
3947 // Build special case code for calls to hashCode on an object.
3948 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3949 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3950 assert(!(is_virtual && is_static), "either virtual, special, or static");
3951
3952 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3953
3954 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3955 PhiNode* result_val = new(C) PhiNode(result_reg,
3956 TypeInt::INT);
3957 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3958 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3959 TypePtr::BOTTOM);
3960 Node* obj = NULL;
3961 if (!is_static) {
3962 // Check for hashing null object
3963 obj = null_check_receiver();
3964 if (stopped()) return true; // unconditionally null
3965 result_reg->init_req(_null_path, top());
3966 result_val->init_req(_null_path, top());
3967 } else {
3968 // Do a null check, and return zero if null.
3969 // System.identityHashCode(null) == 0
3970 obj = argument(0);
3971 Node* null_ctl = top();
3972 obj = null_check_oop(obj, &null_ctl);
3973 result_reg->init_req(_null_path, null_ctl);
3974 result_val->init_req(_null_path, _gvn.intcon(0));
3975 }
3976
3977 // Unconditionally null? Then return right away.
3978 if (stopped()) {
3979 set_control( result_reg->in(_null_path));
3980 if (!stopped())
3981 set_result(result_val->in(_null_path));
3982 return true;
3983 }
3984
3985 // After null check, get the object's klass.
3986 Node* obj_klass = load_object_klass(obj);
3987
3988 // This call may be virtual (invokevirtual) or bound (invokespecial).
3989 // For each case we generate slightly different code.
3990
3991 // We only go to the fast case code if we pass a number of guards. The
3992 // paths which do not pass are accumulated in the slow_region.
3993 RegionNode* slow_region = new (C) RegionNode(1);
3994 record_for_igvn(slow_region);
3995
3996 // If this is a virtual call, we generate a funny guard. We pull out
3997 // the vtable entry corresponding to hashCode() from the target object.
3998 // If the target method which we are calling happens to be the native
3999 // Object hashCode() method, we pass the guard. We do not need this
4000 // guard for non-virtual calls -- the caller is known to be the native
4001 // Object hashCode().
4002 if (is_virtual) {
4003 generate_virtual_guard(obj_klass, slow_region);
4004 }
4005
4006 // Get the header out of the object, use LoadMarkNode when available
4007 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4008 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4009
4010 // Test the header to see if it is unlocked.
4011 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4012 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4013 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4014 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4015 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4016
4017 generate_slow_guard(test_unlocked, slow_region);
4018
4019 // Get the hash value and check to see that it has been properly assigned.
4020 // We depend on hash_mask being at most 32 bits and avoid the use of
4021 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4022 // vm: see markOop.hpp.
4023 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4024 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4025 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4026 // This hack lets the hash bits live anywhere in the mark object now, as long
4027 // as the shift drops the relevant bits into the low 32 bits. Note that
4028 // Java spec says that HashCode is an int so there's no point in capturing
4029 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4030 hshifted_header = ConvX2I(hshifted_header);
4031 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4032
4033 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4034 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4035 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4036
4037 generate_slow_guard(test_assigned, slow_region);
4038
4039 Node* init_mem = reset_memory();
4040 // fill in the rest of the null path:
4041 result_io ->init_req(_null_path, i_o());
4042 result_mem->init_req(_null_path, init_mem);
4043
4044 result_val->init_req(_fast_path, hash_val);
4045 result_reg->init_req(_fast_path, control());
4046 result_io ->init_req(_fast_path, i_o());
4047 result_mem->init_req(_fast_path, init_mem);
4048
4049 // Generate code for the slow case. We make a call to hashCode().
4050 set_control(_gvn.transform(slow_region));
4051 if (!stopped()) {
4052 // No need for PreserveJVMState, because we're using up the present state.
4053 set_all_memory(init_mem);
4054 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4055 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4056 Node* slow_result = set_results_for_java_call(slow_call);
4057 // this->control() comes from set_results_for_java_call
4058 result_reg->init_req(_slow_path, control());
4059 result_val->init_req(_slow_path, slow_result);
4060 result_io ->set_req(_slow_path, i_o());
4061 result_mem ->set_req(_slow_path, reset_memory());
4062 }
4063
4064 // Return the combined state.
4065 set_i_o( _gvn.transform(result_io) );
4066 set_all_memory( _gvn.transform(result_mem));
4067
4068 set_result(result_reg, result_val);
4069 return true;
4070 }
4071
4072 //---------------------------inline_native_getClass----------------------------
4073 // public final native Class<?> java.lang.Object.getClass();
4074 //
4075 // Build special case code for calls to getClass on an object.
4076 bool LibraryCallKit::inline_native_getClass() {
4077 Node* obj = null_check_receiver();
4078 if (stopped()) return true;
4079 set_result(load_mirror_from_klass(load_object_klass(obj)));
4080 return true;
4081 }
4082
4083 //-----------------inline_native_Reflection_getCallerClass---------------------
4084 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4085 //
4086 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4087 //
4088 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4089 // in that it must skip particular security frames and checks for
4090 // caller sensitive methods.
4091 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4092 #ifndef PRODUCT
4093 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4094 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4095 }
4096 #endif
4097
4098 if (!jvms()->has_method()) {
4099 #ifndef PRODUCT
4100 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4101 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4102 }
4103 #endif
4104 return false;
4105 }
4106
4107 // Walk back up the JVM state to find the caller at the required
4108 // depth.
4109 JVMState* caller_jvms = jvms();
4110
4111 // Cf. JVM_GetCallerClass
4112 // NOTE: Start the loop at depth 1 because the current JVM state does
4113 // not include the Reflection.getCallerClass() frame.
4114 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4115 ciMethod* m = caller_jvms->method();
4116 switch (n) {
4117 case 0:
4118 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4119 break;
4120 case 1:
4121 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4122 if (!m->caller_sensitive()) {
4123 #ifndef PRODUCT
4124 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4125 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4126 }
4127 #endif
4128 return false; // bail-out; let JVM_GetCallerClass do the work
4129 }
4130 break;
4131 default:
4132 if (!m->is_ignored_by_security_stack_walk()) {
4133 // We have reached the desired frame; return the holder class.
4134 // Acquire method holder as java.lang.Class and push as constant.
4135 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4136 ciInstance* caller_mirror = caller_klass->java_mirror();
4137 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4138
4139 #ifndef PRODUCT
4140 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4141 tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
4142 tty->print_cr(" JVM state at this point:");
4143 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4144 ciMethod* m = jvms()->of_depth(i)->method();
4145 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4146 }
4147 }
4148 #endif
4149 return true;
4150 }
4151 break;
4152 }
4153 }
4154
4155 #ifndef PRODUCT
4156 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4157 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4158 tty->print_cr(" JVM state at this point:");
4159 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4160 ciMethod* m = jvms()->of_depth(i)->method();
4161 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4162 }
4163 }
4164 #endif
4165
4166 return false; // bail-out; let JVM_GetCallerClass do the work
4167 }
4168
4169 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4170 Node* arg = argument(0);
4171 Node* result;
4172
4173 switch (id) {
4174 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4175 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4176 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4177 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4178
4179 case vmIntrinsics::_doubleToLongBits: {
4180 // two paths (plus control) merge in a wood
4181 RegionNode *r = new (C) RegionNode(3);
4182 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4183
4184 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4185 // Build the boolean node
4186 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4187
4188 // Branch either way.
4189 // NaN case is less traveled, which makes all the difference.
4190 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4191 Node *opt_isnan = _gvn.transform(ifisnan);
4192 assert( opt_isnan->is_If(), "Expect an IfNode");
4193 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4194 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4195
4196 set_control(iftrue);
4197
4198 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4199 Node *slow_result = longcon(nan_bits); // return NaN
4200 phi->init_req(1, _gvn.transform( slow_result ));
4201 r->init_req(1, iftrue);
4202
4203 // Else fall through
4204 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4205 set_control(iffalse);
4206
4207 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4208 r->init_req(2, iffalse);
4209
4210 // Post merge
4211 set_control(_gvn.transform(r));
4212 record_for_igvn(r);
4213
4214 C->set_has_split_ifs(true); // Has chance for split-if optimization
4215 result = phi;
4216 assert(result->bottom_type()->isa_long(), "must be");
4217 break;
4218 }
4219
4220 case vmIntrinsics::_floatToIntBits: {
4221 // two paths (plus control) merge in a wood
4222 RegionNode *r = new (C) RegionNode(3);
4223 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4224
4225 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4226 // Build the boolean node
4227 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4228
4229 // Branch either way.
4230 // NaN case is less traveled, which makes all the difference.
4231 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4232 Node *opt_isnan = _gvn.transform(ifisnan);
4233 assert( opt_isnan->is_If(), "Expect an IfNode");
4234 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4235 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4236
4237 set_control(iftrue);
4238
4239 static const jint nan_bits = 0x7fc00000;
4240 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4241 phi->init_req(1, _gvn.transform( slow_result ));
4242 r->init_req(1, iftrue);
4243
4244 // Else fall through
4245 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4246 set_control(iffalse);
4247
4248 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4249 r->init_req(2, iffalse);
4250
4251 // Post merge
4252 set_control(_gvn.transform(r));
4253 record_for_igvn(r);
4254
4255 C->set_has_split_ifs(true); // Has chance for split-if optimization
4256 result = phi;
4257 assert(result->bottom_type()->isa_int(), "must be");
4258 break;
4259 }
4260
4261 default:
4262 fatal_unexpected_iid(id);
4263 break;
4264 }
4265 set_result(_gvn.transform(result));
4266 return true;
4267 }
4268
4269 #ifdef _LP64
4270 #define XTOP ,top() /*additional argument*/
4271 #else //_LP64
4272 #define XTOP /*no additional argument*/
4273 #endif //_LP64
4274
4275 //----------------------inline_unsafe_copyMemory-------------------------
4276 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4277 bool LibraryCallKit::inline_unsafe_copyMemory() {
4278 if (callee()->is_static()) return false; // caller must have the capability!
4279 null_check_receiver(); // null-check receiver
4280 if (stopped()) return true;
4281
4282 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4283
4284 Node* src_ptr = argument(1); // type: oop
4285 Node* src_off = ConvL2X(argument(2)); // type: long
4286 Node* dst_ptr = argument(4); // type: oop
4287 Node* dst_off = ConvL2X(argument(5)); // type: long
4288 Node* size = ConvL2X(argument(7)); // type: long
4289
4290 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4291 "fieldOffset must be byte-scaled");
4292
4293 Node* src = make_unsafe_address(src_ptr, src_off);
4294 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4295
4296 // Conservatively insert a memory barrier on all memory slices.
4297 // Do not let writes of the copy source or destination float below the copy.
4298 insert_mem_bar(Op_MemBarCPUOrder);
4299
4300 // Call it. Note that the length argument is not scaled.
4301 make_runtime_call(RC_LEAF|RC_NO_FP,
4302 OptoRuntime::fast_arraycopy_Type(),
4303 StubRoutines::unsafe_arraycopy(),
4304 "unsafe_arraycopy",
4305 TypeRawPtr::BOTTOM,
4306 src, dst, size XTOP);
4307
4308 // Do not let reads of the copy destination float above the copy.
4309 insert_mem_bar(Op_MemBarCPUOrder);
4310
4311 return true;
4312 }
4313
4314 //------------------------clone_coping-----------------------------------
4315 // Helper function for inline_native_clone.
4316 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4317 assert(obj_size != NULL, "");
4318 Node* raw_obj = alloc_obj->in(1);
4319 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4320
4321 AllocateNode* alloc = NULL;
4322 if (ReduceBulkZeroing) {
4323 // We will be completely responsible for initializing this object -
4324 // mark Initialize node as complete.
4325 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4326 // The object was just allocated - there should be no any stores!
4327 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4328 // Mark as complete_with_arraycopy so that on AllocateNode
4329 // expansion, we know this AllocateNode is initialized by an array
4330 // copy and a StoreStore barrier exists after the array copy.
4331 alloc->initialization()->set_complete_with_arraycopy();
4332 }
4333
4334 // Copy the fastest available way.
4335 // TODO: generate fields copies for small objects instead.
4336 Node* src = obj;
4337 Node* dest = alloc_obj;
4338 Node* size = _gvn.transform(obj_size);
4339
4340 // Exclude the header but include array length to copy by 8 bytes words.
4341 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4342 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4343 instanceOopDesc::base_offset_in_bytes();
4344 // base_off:
4345 // 8 - 32-bit VM
4346 // 12 - 64-bit VM, compressed klass
4347 // 16 - 64-bit VM, normal klass
4348 if (base_off % BytesPerLong != 0) {
4349 assert(UseCompressedClassPointers, "");
4350 if (is_array) {
4351 // Exclude length to copy by 8 bytes words.
4352 base_off += sizeof(int);
4353 } else {
4354 // Include klass to copy by 8 bytes words.
4355 base_off = instanceOopDesc::klass_offset_in_bytes();
4356 }
4357 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4358 }
4359 src = basic_plus_adr(src, base_off);
4360 dest = basic_plus_adr(dest, base_off);
4361
4362 // Compute the length also, if needed:
4363 Node* countx = size;
4364 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4365 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4366
4367 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4368 bool disjoint_bases = true;
4369 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4370 src, NULL, dest, NULL, countx,
4371 /*dest_uninitialized*/true);
4372
4373 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4374 if (card_mark) {
4375 assert(!is_array, "");
4376 // Put in store barrier for any and all oops we are sticking
4377 // into this object. (We could avoid this if we could prove
4378 // that the object type contains no oop fields at all.)
4379 Node* no_particular_value = NULL;
4380 Node* no_particular_field = NULL;
4381 int raw_adr_idx = Compile::AliasIdxRaw;
4382 post_barrier(control(),
4383 memory(raw_adr_type),
4384 alloc_obj,
4385 no_particular_field,
4386 raw_adr_idx,
4387 no_particular_value,
4388 T_OBJECT,
4389 false);
4390 }
4391
4392 // Do not let reads from the cloned object float above the arraycopy.
4393 if (alloc != NULL) {
4394 // Do not let stores that initialize this object be reordered with
4395 // a subsequent store that would make this object accessible by
4396 // other threads.
4397 // Record what AllocateNode this StoreStore protects so that
4398 // escape analysis can go from the MemBarStoreStoreNode to the
4399 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4400 // based on the escape status of the AllocateNode.
4401 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4402 } else {
4403 insert_mem_bar(Op_MemBarCPUOrder);
4404 }
4405 }
4406
4407 //------------------------inline_native_clone----------------------------
4408 // protected native Object java.lang.Object.clone();
4409 //
4410 // Here are the simple edge cases:
4411 // null receiver => normal trap
4412 // virtual and clone was overridden => slow path to out-of-line clone
4413 // not cloneable or finalizer => slow path to out-of-line Object.clone
4414 //
4415 // The general case has two steps, allocation and copying.
4416 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4417 //
4418 // Copying also has two cases, oop arrays and everything else.
4419 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4420 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4421 //
4422 // These steps fold up nicely if and when the cloned object's klass
4423 // can be sharply typed as an object array, a type array, or an instance.
4424 //
4425 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4426 PhiNode* result_val;
4427
4428 // Set the reexecute bit for the interpreter to reexecute
4429 // the bytecode that invokes Object.clone if deoptimization happens.
4430 { PreserveReexecuteState preexecs(this);
4431 jvms()->set_should_reexecute(true);
4432
4433 Node* obj = null_check_receiver();
4434 if (stopped()) return true;
4435
4436 Node* obj_klass = load_object_klass(obj);
4437 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4438 const TypeOopPtr* toop = ((tklass != NULL)
4439 ? tklass->as_instance_type()
4440 : TypeInstPtr::NOTNULL);
4441
4442 // Conservatively insert a memory barrier on all memory slices.
4443 // Do not let writes into the original float below the clone.
4444 insert_mem_bar(Op_MemBarCPUOrder);
4445
4446 // paths into result_reg:
4447 enum {
4448 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4449 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4450 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4451 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4452 PATH_LIMIT
4453 };
4454 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4455 result_val = new(C) PhiNode(result_reg,
4456 TypeInstPtr::NOTNULL);
4457 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4458 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4459 TypePtr::BOTTOM);
4460 record_for_igvn(result_reg);
4461
4462 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4463 int raw_adr_idx = Compile::AliasIdxRaw;
4464
4465 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4466 if (array_ctl != NULL) {
4467 // It's an array.
4468 PreserveJVMState pjvms(this);
4469 set_control(array_ctl);
4470 Node* obj_length = load_array_length(obj);
4471 Node* obj_size = NULL;
4472 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4473
4474 if (!use_ReduceInitialCardMarks()) {
4475 // If it is an oop array, it requires very special treatment,
4476 // because card marking is required on each card of the array.
4477 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4478 if (is_obja != NULL) {
4479 PreserveJVMState pjvms2(this);
4480 set_control(is_obja);
4481 // Generate a direct call to the right arraycopy function(s).
4482 bool disjoint_bases = true;
4483 bool length_never_negative = true;
4484 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4485 obj, intcon(0), alloc_obj, intcon(0),
4486 obj_length,
4487 disjoint_bases, length_never_negative);
4488 result_reg->init_req(_objArray_path, control());
4489 result_val->init_req(_objArray_path, alloc_obj);
4490 result_i_o ->set_req(_objArray_path, i_o());
4491 result_mem ->set_req(_objArray_path, reset_memory());
4492 }
4493 }
4494 // Otherwise, there are no card marks to worry about.
4495 // (We can dispense with card marks if we know the allocation
4496 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4497 // causes the non-eden paths to take compensating steps to
4498 // simulate a fresh allocation, so that no further
4499 // card marks are required in compiled code to initialize
4500 // the object.)
4501
4502 if (!stopped()) {
4503 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4504
4505 // Present the results of the copy.
4506 result_reg->init_req(_array_path, control());
4507 result_val->init_req(_array_path, alloc_obj);
4508 result_i_o ->set_req(_array_path, i_o());
4509 result_mem ->set_req(_array_path, reset_memory());
4510 }
4511 }
4512
4513 // We only go to the instance fast case code if we pass a number of guards.
4514 // The paths which do not pass are accumulated in the slow_region.
4515 RegionNode* slow_region = new (C) RegionNode(1);
4516 record_for_igvn(slow_region);
4517 if (!stopped()) {
4518 // It's an instance (we did array above). Make the slow-path tests.
4519 // If this is a virtual call, we generate a funny guard. We grab
4520 // the vtable entry corresponding to clone() from the target object.
4521 // If the target method which we are calling happens to be the
4522 // Object clone() method, we pass the guard. We do not need this
4523 // guard for non-virtual calls; the caller is known to be the native
4524 // Object clone().
4525 if (is_virtual) {
4526 generate_virtual_guard(obj_klass, slow_region);
4527 }
4528
4529 // The object must be cloneable and must not have a finalizer.
4530 // Both of these conditions may be checked in a single test.
4531 // We could optimize the cloneable test further, but we don't care.
4532 generate_access_flags_guard(obj_klass,
4533 // Test both conditions:
4534 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4535 // Must be cloneable but not finalizer:
4536 JVM_ACC_IS_CLONEABLE,
4537 slow_region);
4538 }
4539
4540 if (!stopped()) {
4541 // It's an instance, and it passed the slow-path tests.
4542 PreserveJVMState pjvms(this);
4543 Node* obj_size = NULL;
4544 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4545
4546 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4547
4548 // Present the results of the slow call.
4549 result_reg->init_req(_instance_path, control());
4550 result_val->init_req(_instance_path, alloc_obj);
4551 result_i_o ->set_req(_instance_path, i_o());
4552 result_mem ->set_req(_instance_path, reset_memory());
4553 }
4554
4555 // Generate code for the slow case. We make a call to clone().
4556 set_control(_gvn.transform(slow_region));
4557 if (!stopped()) {
4558 PreserveJVMState pjvms(this);
4559 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4560 Node* slow_result = set_results_for_java_call(slow_call);
4561 // this->control() comes from set_results_for_java_call
4562 result_reg->init_req(_slow_path, control());
4563 result_val->init_req(_slow_path, slow_result);
4564 result_i_o ->set_req(_slow_path, i_o());
4565 result_mem ->set_req(_slow_path, reset_memory());
4566 }
4567
4568 // Return the combined state.
4569 set_control( _gvn.transform(result_reg));
4570 set_i_o( _gvn.transform(result_i_o));
4571 set_all_memory( _gvn.transform(result_mem));
4572 } // original reexecute is set back here
4573
4574 set_result(_gvn.transform(result_val));
4575 return true;
4576 }
4577
4578 //------------------------------basictype2arraycopy----------------------------
4579 address LibraryCallKit::basictype2arraycopy(BasicType t,
4580 Node* src_offset,
4581 Node* dest_offset,
4582 bool disjoint_bases,
4583 const char* &name,
4584 bool dest_uninitialized) {
4585 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4586 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4587
4588 bool aligned = false;
4589 bool disjoint = disjoint_bases;
4590
4591 // if the offsets are the same, we can treat the memory regions as
4592 // disjoint, because either the memory regions are in different arrays,
4593 // or they are identical (which we can treat as disjoint.) We can also
4594 // treat a copy with a destination index less that the source index
4595 // as disjoint since a low->high copy will work correctly in this case.
4596 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4597 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4598 // both indices are constants
4599 int s_offs = src_offset_inttype->get_con();
4600 int d_offs = dest_offset_inttype->get_con();
4601 int element_size = type2aelembytes(t);
4602 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4603 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4604 if (s_offs >= d_offs) disjoint = true;
4605 } else if (src_offset == dest_offset && src_offset != NULL) {
4606 // This can occur if the offsets are identical non-constants.
4607 disjoint = true;
4608 }
4609
4610 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4611 }
4612
4613
4614 //------------------------------inline_arraycopy-----------------------
4615 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4616 // Object dest, int destPos,
4617 // int length);
4618 bool LibraryCallKit::inline_arraycopy() {
4619 // Get the arguments.
4620 Node* src = argument(0); // type: oop
4621 Node* src_offset = argument(1); // type: int
4622 Node* dest = argument(2); // type: oop
4623 Node* dest_offset = argument(3); // type: int
4624 Node* length = argument(4); // type: int
4625
4626 // Compile time checks. If any of these checks cannot be verified at compile time,
4627 // we do not make a fast path for this call. Instead, we let the call remain as it
4628 // is. The checks we choose to mandate at compile time are:
4629 //
4630 // (1) src and dest are arrays.
4631 const Type* src_type = src->Value(&_gvn);
4632 const Type* dest_type = dest->Value(&_gvn);
4633 const TypeAryPtr* top_src = src_type->isa_aryptr();
4634 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4635
4636 // Do we have the type of src?
4637 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4638 // Do we have the type of dest?
4639 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4640 // Is the type for src from speculation?
4641 bool src_spec = false;
4642 // Is the type for dest from speculation?
4643 bool dest_spec = false;
4644
4645 if (!has_src || !has_dest) {
4646 // We don't have sufficient type information, let's see if
4647 // speculative types can help. We need to have types for both src
4648 // and dest so that it pays off.
4649
4650 // Do we already have or could we have type information for src
4651 bool could_have_src = has_src;
4652 // Do we already have or could we have type information for dest
4653 bool could_have_dest = has_dest;
4654
4655 ciKlass* src_k = NULL;
4656 if (!has_src) {
4657 src_k = src_type->speculative_type();
4658 if (src_k != NULL && src_k->is_array_klass()) {
4659 could_have_src = true;
4660 }
4661 }
4662
4663 ciKlass* dest_k = NULL;
4664 if (!has_dest) {
4665 dest_k = dest_type->speculative_type();
4666 if (dest_k != NULL && dest_k->is_array_klass()) {
4667 could_have_dest = true;
4668 }
4669 }
4670
4671 if (could_have_src && could_have_dest) {
4672 // This is going to pay off so emit the required guards
4673 if (!has_src) {
4674 src = maybe_cast_profiled_obj(src, src_k);
4675 src_type = _gvn.type(src);
4676 top_src = src_type->isa_aryptr();
4677 has_src = (top_src != NULL && top_src->klass() != NULL);
4678 src_spec = true;
4679 }
4680 if (!has_dest) {
4681 dest = maybe_cast_profiled_obj(dest, dest_k);
4682 dest_type = _gvn.type(dest);
4683 top_dest = dest_type->isa_aryptr();
4684 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4685 dest_spec = true;
4686 }
4687 }
4688 }
4689
4690 if (!has_src || !has_dest) {
4691 // Conservatively insert a memory barrier on all memory slices.
4692 // Do not let writes into the source float below the arraycopy.
4693 insert_mem_bar(Op_MemBarCPUOrder);
4694
4695 // Call StubRoutines::generic_arraycopy stub.
4696 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4697 src, src_offset, dest, dest_offset, length);
4698
4699 // Do not let reads from the destination float above the arraycopy.
4700 // Since we cannot type the arrays, we don't know which slices
4701 // might be affected. We could restrict this barrier only to those
4702 // memory slices which pertain to array elements--but don't bother.
4703 if (!InsertMemBarAfterArraycopy)
4704 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4705 insert_mem_bar(Op_MemBarCPUOrder);
4706 return true;
4707 }
4708
4709 // (2) src and dest arrays must have elements of the same BasicType
4710 // Figure out the size and type of the elements we will be copying.
4711 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4712 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4713 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4714 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4715
4716 if (src_elem != dest_elem || dest_elem == T_VOID) {
4717 // The component types are not the same or are not recognized. Punt.
4718 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4719 generate_slow_arraycopy(TypePtr::BOTTOM,
4720 src, src_offset, dest, dest_offset, length,
4721 /*dest_uninitialized*/false);
4722 return true;
4723 }
4724
4725 if (src_elem == T_OBJECT) {
4726 // If both arrays are object arrays then having the exact types
4727 // for both will remove the need for a subtype check at runtime
4728 // before the call and may make it possible to pick a faster copy
4729 // routine (without a subtype check on every element)
4730 // Do we have the exact type of src?
4731 bool could_have_src = src_spec;
4732 // Do we have the exact type of dest?
4733 bool could_have_dest = dest_spec;
4734 ciKlass* src_k = top_src->klass();
4735 ciKlass* dest_k = top_dest->klass();
4736 if (!src_spec) {
4737 src_k = src_type->speculative_type();
4738 if (src_k != NULL && src_k->is_array_klass()) {
4739 could_have_src = true;
4740 }
4741 }
4742 if (!dest_spec) {
4743 dest_k = dest_type->speculative_type();
4744 if (dest_k != NULL && dest_k->is_array_klass()) {
4745 could_have_dest = true;
4746 }
4747 }
4748 if (could_have_src && could_have_dest) {
4749 // If we can have both exact types, emit the missing guards
4750 if (could_have_src && !src_spec) {
4751 src = maybe_cast_profiled_obj(src, src_k);
4752 }
4753 if (could_have_dest && !dest_spec) {
4754 dest = maybe_cast_profiled_obj(dest, dest_k);
4755 }
4756 }
4757 }
4758
4759 //---------------------------------------------------------------------------
4760 // We will make a fast path for this call to arraycopy.
4761
4762 // We have the following tests left to perform:
4763 //
4764 // (3) src and dest must not be null.
4765 // (4) src_offset must not be negative.
4766 // (5) dest_offset must not be negative.
4767 // (6) length must not be negative.
4768 // (7) src_offset + length must not exceed length of src.
4769 // (8) dest_offset + length must not exceed length of dest.
4770 // (9) each element of an oop array must be assignable
4771
4772 RegionNode* slow_region = new (C) RegionNode(1);
4773 record_for_igvn(slow_region);
4774
4775 // (3) operands must not be null
4776 // We currently perform our null checks with the null_check routine.
4777 // This means that the null exceptions will be reported in the caller
4778 // rather than (correctly) reported inside of the native arraycopy call.
4779 // This should be corrected, given time. We do our null check with the
4780 // stack pointer restored.
4781 src = null_check(src, T_ARRAY);
4782 dest = null_check(dest, T_ARRAY);
4783
4784 // (4) src_offset must not be negative.
4785 generate_negative_guard(src_offset, slow_region);
4786
4787 // (5) dest_offset must not be negative.
4788 generate_negative_guard(dest_offset, slow_region);
4789
4790 // (6) length must not be negative (moved to generate_arraycopy()).
4791 // generate_negative_guard(length, slow_region);
4792
4793 // (7) src_offset + length must not exceed length of src.
4794 generate_limit_guard(src_offset, length,
4795 load_array_length(src),
4796 slow_region);
4797
4798 // (8) dest_offset + length must not exceed length of dest.
4799 generate_limit_guard(dest_offset, length,
4800 load_array_length(dest),
4801 slow_region);
4802
4803 // (9) each element of an oop array must be assignable
4804 // The generate_arraycopy subroutine checks this.
4805
4806 // This is where the memory effects are placed:
4807 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4808 generate_arraycopy(adr_type, dest_elem,
4809 src, src_offset, dest, dest_offset, length,
4810 false, false, slow_region);
4811
4812 return true;
4813 }
4814
4815 //-----------------------------generate_arraycopy----------------------
4816 // Generate an optimized call to arraycopy.
4817 // Caller must guard against non-arrays.
4818 // Caller must determine a common array basic-type for both arrays.
4819 // Caller must validate offsets against array bounds.
4820 // The slow_region has already collected guard failure paths
4821 // (such as out of bounds length or non-conformable array types).
4822 // The generated code has this shape, in general:
4823 //
4824 // if (length == 0) return // via zero_path
4825 // slowval = -1
4826 // if (types unknown) {
4827 // slowval = call generic copy loop
4828 // if (slowval == 0) return // via checked_path
4829 // } else if (indexes in bounds) {
4830 // if ((is object array) && !(array type check)) {
4831 // slowval = call checked copy loop
4832 // if (slowval == 0) return // via checked_path
4833 // } else {
4834 // call bulk copy loop
4835 // return // via fast_path
4836 // }
4837 // }
4838 // // adjust params for remaining work:
4839 // if (slowval != -1) {
4840 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4841 // }
4842 // slow_region:
4843 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4844 // return // via slow_call_path
4845 //
4846 // This routine is used from several intrinsics: System.arraycopy,
4847 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4848 //
4849 void
4850 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4851 BasicType basic_elem_type,
4852 Node* src, Node* src_offset,
4853 Node* dest, Node* dest_offset,
4854 Node* copy_length,
4855 bool disjoint_bases,
4856 bool length_never_negative,
4857 RegionNode* slow_region) {
4858
4859 if (slow_region == NULL) {
4860 slow_region = new(C) RegionNode(1);
4861 record_for_igvn(slow_region);
4862 }
4863
4864 Node* original_dest = dest;
4865 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4866 bool dest_uninitialized = false;
4867
4868 // See if this is the initialization of a newly-allocated array.
4869 // If so, we will take responsibility here for initializing it to zero.
4870 // (Note: Because tightly_coupled_allocation performs checks on the
4871 // out-edges of the dest, we need to avoid making derived pointers
4872 // from it until we have checked its uses.)
4873 if (ReduceBulkZeroing
4874 && !ZeroTLAB // pointless if already zeroed
4875 && basic_elem_type != T_CONFLICT // avoid corner case
4876 && !src->eqv_uncast(dest)
4877 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4878 != NULL)
4879 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4880 && alloc->maybe_set_complete(&_gvn)) {
4881 // "You break it, you buy it."
4882 InitializeNode* init = alloc->initialization();
4883 assert(init->is_complete(), "we just did this");
4884 init->set_complete_with_arraycopy();
4885 assert(dest->is_CheckCastPP(), "sanity");
4886 assert(dest->in(0)->in(0) == init, "dest pinned");
4887 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4888 // From this point on, every exit path is responsible for
4889 // initializing any non-copied parts of the object to zero.
4890 // Also, if this flag is set we make sure that arraycopy interacts properly
4891 // with G1, eliding pre-barriers. See CR 6627983.
4892 dest_uninitialized = true;
4893 } else {
4894 // No zeroing elimination here.
4895 alloc = NULL;
4896 //original_dest = dest;
4897 //dest_uninitialized = false;
4898 }
4899
4900 // Results are placed here:
4901 enum { fast_path = 1, // normal void-returning assembly stub
4902 checked_path = 2, // special assembly stub with cleanup
4903 slow_call_path = 3, // something went wrong; call the VM
4904 zero_path = 4, // bypass when length of copy is zero
4905 bcopy_path = 5, // copy primitive array by 64-bit blocks
4906 PATH_LIMIT = 6
4907 };
4908 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4909 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4910 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4911 record_for_igvn(result_region);
4912 _gvn.set_type_bottom(result_i_o);
4913 _gvn.set_type_bottom(result_memory);
4914 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4915
4916 // The slow_control path:
4917 Node* slow_control;
4918 Node* slow_i_o = i_o();
4919 Node* slow_mem = memory(adr_type);
4920 debug_only(slow_control = (Node*) badAddress);
4921
4922 // Checked control path:
4923 Node* checked_control = top();
4924 Node* checked_mem = NULL;
4925 Node* checked_i_o = NULL;
4926 Node* checked_value = NULL;
4927
4928 if (basic_elem_type == T_CONFLICT) {
4929 assert(!dest_uninitialized, "");
4930 Node* cv = generate_generic_arraycopy(adr_type,
4931 src, src_offset, dest, dest_offset,
4932 copy_length, dest_uninitialized);
4933 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4934 checked_control = control();
4935 checked_i_o = i_o();
4936 checked_mem = memory(adr_type);
4937 checked_value = cv;
4938 set_control(top()); // no fast path
4939 }
4940
4941 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4942 if (not_pos != NULL) {
4943 PreserveJVMState pjvms(this);
4944 set_control(not_pos);
4945
4946 // (6) length must not be negative.
4947 if (!length_never_negative) {
4948 generate_negative_guard(copy_length, slow_region);
4949 }
4950
4951 // copy_length is 0.
4952 if (!stopped() && dest_uninitialized) {
4953 Node* dest_length = alloc->in(AllocateNode::ALength);
4954 if (copy_length->eqv_uncast(dest_length)
4955 || _gvn.find_int_con(dest_length, 1) <= 0) {
4956 // There is no zeroing to do. No need for a secondary raw memory barrier.
4957 } else {
4958 // Clear the whole thing since there are no source elements to copy.
4959 generate_clear_array(adr_type, dest, basic_elem_type,
4960 intcon(0), NULL,
4961 alloc->in(AllocateNode::AllocSize));
4962 // Use a secondary InitializeNode as raw memory barrier.
4963 // Currently it is needed only on this path since other
4964 // paths have stub or runtime calls as raw memory barriers.
4965 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4966 Compile::AliasIdxRaw,
4967 top())->as_Initialize();
4968 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4969 }
4970 }
4971
4972 // Present the results of the fast call.
4973 result_region->init_req(zero_path, control());
4974 result_i_o ->init_req(zero_path, i_o());
4975 result_memory->init_req(zero_path, memory(adr_type));
4976 }
4977
4978 if (!stopped() && dest_uninitialized) {
4979 // We have to initialize the *uncopied* part of the array to zero.
4980 // The copy destination is the slice dest[off..off+len]. The other slices
4981 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4982 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4983 Node* dest_length = alloc->in(AllocateNode::ALength);
4984 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4985 copy_length));
4986
4987 // If there is a head section that needs zeroing, do it now.
4988 if (find_int_con(dest_offset, -1) != 0) {
4989 generate_clear_array(adr_type, dest, basic_elem_type,
4990 intcon(0), dest_offset,
4991 NULL);
4992 }
4993
4994 // Next, perform a dynamic check on the tail length.
4995 // It is often zero, and we can win big if we prove this.
4996 // There are two wins: Avoid generating the ClearArray
4997 // with its attendant messy index arithmetic, and upgrade
4998 // the copy to a more hardware-friendly word size of 64 bits.
4999 Node* tail_ctl = NULL;
5000 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5001 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5002 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5003 tail_ctl = generate_slow_guard(bol_lt, NULL);
5004 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5005 }
5006
5007 // At this point, let's assume there is no tail.
5008 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5009 // There is no tail. Try an upgrade to a 64-bit copy.
5010 bool didit = false;
5011 { PreserveJVMState pjvms(this);
5012 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5013 src, src_offset, dest, dest_offset,
5014 dest_size, dest_uninitialized);
5015 if (didit) {
5016 // Present the results of the block-copying fast call.
5017 result_region->init_req(bcopy_path, control());
5018 result_i_o ->init_req(bcopy_path, i_o());
5019 result_memory->init_req(bcopy_path, memory(adr_type));
5020 }
5021 }
5022 if (didit)
5023 set_control(top()); // no regular fast path
5024 }
5025
5026 // Clear the tail, if any.
5027 if (tail_ctl != NULL) {
5028 Node* notail_ctl = stopped() ? NULL : control();
5029 set_control(tail_ctl);
5030 if (notail_ctl == NULL) {
5031 generate_clear_array(adr_type, dest, basic_elem_type,
5032 dest_tail, NULL,
5033 dest_size);
5034 } else {
5035 // Make a local merge.
5036 Node* done_ctl = new(C) RegionNode(3);
5037 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5038 done_ctl->init_req(1, notail_ctl);
5039 done_mem->init_req(1, memory(adr_type));
5040 generate_clear_array(adr_type, dest, basic_elem_type,
5041 dest_tail, NULL,
5042 dest_size);
5043 done_ctl->init_req(2, control());
5044 done_mem->init_req(2, memory(adr_type));
5045 set_control( _gvn.transform(done_ctl));
5046 set_memory( _gvn.transform(done_mem), adr_type );
5047 }
5048 }
5049 }
5050
5051 BasicType copy_type = basic_elem_type;
5052 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5053 if (!stopped() && copy_type == T_OBJECT) {
5054 // If src and dest have compatible element types, we can copy bits.
5055 // Types S[] and D[] are compatible if D is a supertype of S.
5056 //
5057 // If they are not, we will use checked_oop_disjoint_arraycopy,
5058 // which performs a fast optimistic per-oop check, and backs off
5059 // further to JVM_ArrayCopy on the first per-oop check that fails.
5060 // (Actually, we don't move raw bits only; the GC requires card marks.)
5061
5062 // Get the Klass* for both src and dest
5063 Node* src_klass = load_object_klass(src);
5064 Node* dest_klass = load_object_klass(dest);
5065
5066 // Generate the subtype check.
5067 // This might fold up statically, or then again it might not.
5068 //
5069 // Non-static example: Copying List<String>.elements to a new String[].
5070 // The backing store for a List<String> is always an Object[],
5071 // but its elements are always type String, if the generic types
5072 // are correct at the source level.
5073 //
5074 // Test S[] against D[], not S against D, because (probably)
5075 // the secondary supertype cache is less busy for S[] than S.
5076 // This usually only matters when D is an interface.
5077 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5078 // Plug failing path into checked_oop_disjoint_arraycopy
5079 if (not_subtype_ctrl != top()) {
5080 PreserveJVMState pjvms(this);
5081 set_control(not_subtype_ctrl);
5082 // (At this point we can assume disjoint_bases, since types differ.)
5083 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5084 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5085 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5086 Node* dest_elem_klass = _gvn.transform(n1);
5087 Node* cv = generate_checkcast_arraycopy(adr_type,
5088 dest_elem_klass,
5089 src, src_offset, dest, dest_offset,
5090 ConvI2X(copy_length), dest_uninitialized);
5091 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5092 checked_control = control();
5093 checked_i_o = i_o();
5094 checked_mem = memory(adr_type);
5095 checked_value = cv;
5096 }
5097 // At this point we know we do not need type checks on oop stores.
5098
5099 // Let's see if we need card marks:
5100 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5101 // If we do not need card marks, copy using the jint or jlong stub.
5102 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5103 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5104 "sizes agree");
5105 }
5106 }
5107
5108 if (!stopped()) {
5109 // Generate the fast path, if possible.
5110 PreserveJVMState pjvms(this);
5111 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5112 src, src_offset, dest, dest_offset,
5113 ConvI2X(copy_length), dest_uninitialized);
5114
5115 // Present the results of the fast call.
5116 result_region->init_req(fast_path, control());
5117 result_i_o ->init_req(fast_path, i_o());
5118 result_memory->init_req(fast_path, memory(adr_type));
5119 }
5120
5121 // Here are all the slow paths up to this point, in one bundle:
5122 slow_control = top();
5123 if (slow_region != NULL)
5124 slow_control = _gvn.transform(slow_region);
5125 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5126
5127 set_control(checked_control);
5128 if (!stopped()) {
5129 // Clean up after the checked call.
5130 // The returned value is either 0 or -1^K,
5131 // where K = number of partially transferred array elements.
5132 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5133 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5134 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5135
5136 // If it is 0, we are done, so transfer to the end.
5137 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5138 result_region->init_req(checked_path, checks_done);
5139 result_i_o ->init_req(checked_path, checked_i_o);
5140 result_memory->init_req(checked_path, checked_mem);
5141
5142 // If it is not zero, merge into the slow call.
5143 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5144 RegionNode* slow_reg2 = new(C) RegionNode(3);
5145 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5146 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5147 record_for_igvn(slow_reg2);
5148 slow_reg2 ->init_req(1, slow_control);
5149 slow_i_o2 ->init_req(1, slow_i_o);
5150 slow_mem2 ->init_req(1, slow_mem);
5151 slow_reg2 ->init_req(2, control());
5152 slow_i_o2 ->init_req(2, checked_i_o);
5153 slow_mem2 ->init_req(2, checked_mem);
5154
5155 slow_control = _gvn.transform(slow_reg2);
5156 slow_i_o = _gvn.transform(slow_i_o2);
5157 slow_mem = _gvn.transform(slow_mem2);
5158
5159 if (alloc != NULL) {
5160 // We'll restart from the very beginning, after zeroing the whole thing.
5161 // This can cause double writes, but that's OK since dest is brand new.
5162 // So we ignore the low 31 bits of the value returned from the stub.
5163 } else {
5164 // We must continue the copy exactly where it failed, or else
5165 // another thread might see the wrong number of writes to dest.
5166 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5167 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5168 slow_offset->init_req(1, intcon(0));
5169 slow_offset->init_req(2, checked_offset);
5170 slow_offset = _gvn.transform(slow_offset);
5171
5172 // Adjust the arguments by the conditionally incoming offset.
5173 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5174 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5175 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5176
5177 // Tweak the node variables to adjust the code produced below:
5178 src_offset = src_off_plus;
5179 dest_offset = dest_off_plus;
5180 copy_length = length_minus;
5181 }
5182 }
5183
5184 set_control(slow_control);
5185 if (!stopped()) {
5186 // Generate the slow path, if needed.
5187 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5188
5189 set_memory(slow_mem, adr_type);
5190 set_i_o(slow_i_o);
5191
5192 if (dest_uninitialized) {
5193 generate_clear_array(adr_type, dest, basic_elem_type,
5194 intcon(0), NULL,
5195 alloc->in(AllocateNode::AllocSize));
5196 }
5197
5198 generate_slow_arraycopy(adr_type,
5199 src, src_offset, dest, dest_offset,
5200 copy_length, /*dest_uninitialized*/false);
5201
5202 result_region->init_req(slow_call_path, control());
5203 result_i_o ->init_req(slow_call_path, i_o());
5204 result_memory->init_req(slow_call_path, memory(adr_type));
5205 }
5206
5207 // Remove unused edges.
5208 for (uint i = 1; i < result_region->req(); i++) {
5209 if (result_region->in(i) == NULL)
5210 result_region->init_req(i, top());
5211 }
5212
5213 // Finished; return the combined state.
5214 set_control( _gvn.transform(result_region));
5215 set_i_o( _gvn.transform(result_i_o) );
5216 set_memory( _gvn.transform(result_memory), adr_type );
5217
5218 // The memory edges above are precise in order to model effects around
5219 // array copies accurately to allow value numbering of field loads around
5220 // arraycopy. Such field loads, both before and after, are common in Java
5221 // collections and similar classes involving header/array data structures.
5222 //
5223 // But with low number of register or when some registers are used or killed
5224 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5225 // The next memory barrier is added to avoid it. If the arraycopy can be
5226 // optimized away (which it can, sometimes) then we can manually remove
5227 // the membar also.
5228 //
5229 // Do not let reads from the cloned object float above the arraycopy.
5230 if (alloc != NULL) {
5231 // Do not let stores that initialize this object be reordered with
5232 // a subsequent store that would make this object accessible by
5233 // other threads.
5234 // Record what AllocateNode this StoreStore protects so that
5235 // escape analysis can go from the MemBarStoreStoreNode to the
5236 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5237 // based on the escape status of the AllocateNode.
5238 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5239 } else if (InsertMemBarAfterArraycopy)
5240 insert_mem_bar(Op_MemBarCPUOrder);
5241 }
5242
5243
5244 // Helper function which determines if an arraycopy immediately follows
5245 // an allocation, with no intervening tests or other escapes for the object.
5246 AllocateArrayNode*
5247 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5248 RegionNode* slow_region) {
5249 if (stopped()) return NULL; // no fast path
5250 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5251
5252 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5253 if (alloc == NULL) return NULL;
5254
5255 Node* rawmem = memory(Compile::AliasIdxRaw);
5256 // Is the allocation's memory state untouched?
5257 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5258 // Bail out if there have been raw-memory effects since the allocation.
5259 // (Example: There might have been a call or safepoint.)
5260 return NULL;
5261 }
5262 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5263 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5264 return NULL;
5265 }
5266
5267 // There must be no unexpected observers of this allocation.
5268 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5269 Node* obs = ptr->fast_out(i);
5270 if (obs != this->map()) {
5271 return NULL;
5272 }
5273 }
5274
5275 // This arraycopy must unconditionally follow the allocation of the ptr.
5276 Node* alloc_ctl = ptr->in(0);
5277 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5278
5279 Node* ctl = control();
5280 while (ctl != alloc_ctl) {
5281 // There may be guards which feed into the slow_region.
5282 // Any other control flow means that we might not get a chance
5283 // to finish initializing the allocated object.
5284 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5285 IfNode* iff = ctl->in(0)->as_If();
5286 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5287 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5288 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5289 ctl = iff->in(0); // This test feeds the known slow_region.
5290 continue;
5291 }
5292 // One more try: Various low-level checks bottom out in
5293 // uncommon traps. If the debug-info of the trap omits
5294 // any reference to the allocation, as we've already
5295 // observed, then there can be no objection to the trap.
5296 bool found_trap = false;
5297 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5298 Node* obs = not_ctl->fast_out(j);
5299 if (obs->in(0) == not_ctl && obs->is_Call() &&
5300 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5301 found_trap = true; break;
5302 }
5303 }
5304 if (found_trap) {
5305 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5306 continue;
5307 }
5308 }
5309 return NULL;
5310 }
5311
5312 // If we get this far, we have an allocation which immediately
5313 // precedes the arraycopy, and we can take over zeroing the new object.
5314 // The arraycopy will finish the initialization, and provide
5315 // a new control state to which we will anchor the destination pointer.
5316
5317 return alloc;
5318 }
5319
5320 // Helper for initialization of arrays, creating a ClearArray.
5321 // It writes zero bits in [start..end), within the body of an array object.
5322 // The memory effects are all chained onto the 'adr_type' alias category.
5323 //
5324 // Since the object is otherwise uninitialized, we are free
5325 // to put a little "slop" around the edges of the cleared area,
5326 // as long as it does not go back into the array's header,
5327 // or beyond the array end within the heap.
5328 //
5329 // The lower edge can be rounded down to the nearest jint and the
5330 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5331 //
5332 // Arguments:
5333 // adr_type memory slice where writes are generated
5334 // dest oop of the destination array
5335 // basic_elem_type element type of the destination
5336 // slice_idx array index of first element to store
5337 // slice_len number of elements to store (or NULL)
5338 // dest_size total size in bytes of the array object
5339 //
5340 // Exactly one of slice_len or dest_size must be non-NULL.
5341 // If dest_size is non-NULL, zeroing extends to the end of the object.
5342 // If slice_len is non-NULL, the slice_idx value must be a constant.
5343 void
5344 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5345 Node* dest,
5346 BasicType basic_elem_type,
5347 Node* slice_idx,
5348 Node* slice_len,
5349 Node* dest_size) {
5350 // one or the other but not both of slice_len and dest_size:
5351 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5352 if (slice_len == NULL) slice_len = top();
5353 if (dest_size == NULL) dest_size = top();
5354
5355 // operate on this memory slice:
5356 Node* mem = memory(adr_type); // memory slice to operate on
5357
5358 // scaling and rounding of indexes:
5359 int scale = exact_log2(type2aelembytes(basic_elem_type));
5360 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5361 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5362 int bump_bit = (-1 << scale) & BytesPerInt;
5363
5364 // determine constant starts and ends
5365 const intptr_t BIG_NEG = -128;
5366 assert(BIG_NEG + 2*abase < 0, "neg enough");
5367 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5368 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5369 if (slice_len_con == 0) {
5370 return; // nothing to do here
5371 }
5372 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5373 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5374 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5375 assert(end_con < 0, "not two cons");
5376 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5377 BytesPerLong);
5378 }
5379
5380 if (start_con >= 0 && end_con >= 0) {
5381 // Constant start and end. Simple.
5382 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5383 start_con, end_con, &_gvn);
5384 } else if (start_con >= 0 && dest_size != top()) {
5385 // Constant start, pre-rounded end after the tail of the array.
5386 Node* end = dest_size;
5387 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5388 start_con, end, &_gvn);
5389 } else if (start_con >= 0 && slice_len != top()) {
5390 // Constant start, non-constant end. End needs rounding up.
5391 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5392 intptr_t end_base = abase + (slice_idx_con << scale);
5393 int end_round = (-1 << scale) & (BytesPerLong - 1);
5394 Node* end = ConvI2X(slice_len);
5395 if (scale != 0)
5396 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5397 end_base += end_round;
5398 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5399 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5400 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5401 start_con, end, &_gvn);
5402 } else if (start_con < 0 && dest_size != top()) {
5403 // Non-constant start, pre-rounded end after the tail of the array.
5404 // This is almost certainly a "round-to-end" operation.
5405 Node* start = slice_idx;
5406 start = ConvI2X(start);
5407 if (scale != 0)
5408 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5409 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5410 if ((bump_bit | clear_low) != 0) {
5411 int to_clear = (bump_bit | clear_low);
5412 // Align up mod 8, then store a jint zero unconditionally
5413 // just before the mod-8 boundary.
5414 if (((abase + bump_bit) & ~to_clear) - bump_bit
5415 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5416 bump_bit = 0;
5417 assert((abase & to_clear) == 0, "array base must be long-aligned");
5418 } else {
5419 // Bump 'start' up to (or past) the next jint boundary:
5420 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5421 assert((abase & clear_low) == 0, "array base must be int-aligned");
5422 }
5423 // Round bumped 'start' down to jlong boundary in body of array.
5424 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5425 if (bump_bit != 0) {
5426 // Store a zero to the immediately preceding jint:
5427 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5428 Node* p1 = basic_plus_adr(dest, x1);
5429 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5430 mem = _gvn.transform(mem);
5431 }
5432 }
5433 Node* end = dest_size; // pre-rounded
5434 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5435 start, end, &_gvn);
5436 } else {
5437 // Non-constant start, unrounded non-constant end.
5438 // (Nobody zeroes a random midsection of an array using this routine.)
5439 ShouldNotReachHere(); // fix caller
5440 }
5441
5442 // Done.
5443 set_memory(mem, adr_type);
5444 }
5445
5446
5447 bool
5448 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5449 BasicType basic_elem_type,
5450 AllocateNode* alloc,
5451 Node* src, Node* src_offset,
5452 Node* dest, Node* dest_offset,
5453 Node* dest_size, bool dest_uninitialized) {
5454 // See if there is an advantage from block transfer.
5455 int scale = exact_log2(type2aelembytes(basic_elem_type));
5456 if (scale >= LogBytesPerLong)
5457 return false; // it is already a block transfer
5458
5459 // Look at the alignment of the starting offsets.
5460 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5461
5462 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5463 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5464 if (src_off_con < 0 || dest_off_con < 0)
5465 // At present, we can only understand constants.
5466 return false;
5467
5468 intptr_t src_off = abase + (src_off_con << scale);
5469 intptr_t dest_off = abase + (dest_off_con << scale);
5470
5471 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5472 // Non-aligned; too bad.
5473 // One more chance: Pick off an initial 32-bit word.
5474 // This is a common case, since abase can be odd mod 8.
5475 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5476 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5477 Node* sptr = basic_plus_adr(src, src_off);
5478 Node* dptr = basic_plus_adr(dest, dest_off);
5479 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5480 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5481 src_off += BytesPerInt;
5482 dest_off += BytesPerInt;
5483 } else {
5484 return false;
5485 }
5486 }
5487 assert(src_off % BytesPerLong == 0, "");
5488 assert(dest_off % BytesPerLong == 0, "");
5489
5490 // Do this copy by giant steps.
5491 Node* sptr = basic_plus_adr(src, src_off);
5492 Node* dptr = basic_plus_adr(dest, dest_off);
5493 Node* countx = dest_size;
5494 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5495 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5496
5497 bool disjoint_bases = true; // since alloc != NULL
5498 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5499 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5500
5501 return true;
5502 }
5503
5504
5505 // Helper function; generates code for the slow case.
5506 // We make a call to a runtime method which emulates the native method,
5507 // but without the native wrapper overhead.
5508 void
5509 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5510 Node* src, Node* src_offset,
5511 Node* dest, Node* dest_offset,
5512 Node* copy_length, bool dest_uninitialized) {
5513 assert(!dest_uninitialized, "Invariant");
5514 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5515 OptoRuntime::slow_arraycopy_Type(),
5516 OptoRuntime::slow_arraycopy_Java(),
5517 "slow_arraycopy", adr_type,
5518 src, src_offset, dest, dest_offset,
5519 copy_length);
5520
5521 // Handle exceptions thrown by this fellow:
5522 make_slow_call_ex(call, env()->Throwable_klass(), false);
5523 }
5524
5525 // Helper function; generates code for cases requiring runtime checks.
5526 Node*
5527 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5528 Node* dest_elem_klass,
5529 Node* src, Node* src_offset,
5530 Node* dest, Node* dest_offset,
5531 Node* copy_length, bool dest_uninitialized) {
5532 if (stopped()) return NULL;
5533
5534 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5535 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5536 return NULL;
5537 }
5538
5539 // Pick out the parameters required to perform a store-check
5540 // for the target array. This is an optimistic check. It will
5541 // look in each non-null element's class, at the desired klass's
5542 // super_check_offset, for the desired klass.
5543 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5544 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5545 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5546 Node* check_offset = ConvI2X(_gvn.transform(n3));
5547 Node* check_value = dest_elem_klass;
5548
5549 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5550 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5551
5552 // (We know the arrays are never conjoint, because their types differ.)
5553 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5554 OptoRuntime::checkcast_arraycopy_Type(),
5555 copyfunc_addr, "checkcast_arraycopy", adr_type,
5556 // five arguments, of which two are
5557 // intptr_t (jlong in LP64)
5558 src_start, dest_start,
5559 copy_length XTOP,
5560 check_offset XTOP,
5561 check_value);
5562
5563 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5564 }
5565
5566
5567 // Helper function; generates code for cases requiring runtime checks.
5568 Node*
5569 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5570 Node* src, Node* src_offset,
5571 Node* dest, Node* dest_offset,
5572 Node* copy_length, bool dest_uninitialized) {
5573 assert(!dest_uninitialized, "Invariant");
5574 if (stopped()) return NULL;
5575 address copyfunc_addr = StubRoutines::generic_arraycopy();
5576 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5577 return NULL;
5578 }
5579
5580 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5581 OptoRuntime::generic_arraycopy_Type(),
5582 copyfunc_addr, "generic_arraycopy", adr_type,
5583 src, src_offset, dest, dest_offset, copy_length);
5584
5585 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5586 }
5587
5588 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5589 void
5590 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5591 BasicType basic_elem_type,
5592 bool disjoint_bases,
5593 Node* src, Node* src_offset,
5594 Node* dest, Node* dest_offset,
5595 Node* copy_length, bool dest_uninitialized) {
5596 if (stopped()) return; // nothing to do
5597
5598 Node* src_start = src;
5599 Node* dest_start = dest;
5600 if (src_offset != NULL || dest_offset != NULL) {
5601 assert(src_offset != NULL && dest_offset != NULL, "");
5602 src_start = array_element_address(src, src_offset, basic_elem_type);
5603 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5604 }
5605
5606 // Figure out which arraycopy runtime method to call.
5607 const char* copyfunc_name = "arraycopy";
5608 address copyfunc_addr =
5609 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5610 disjoint_bases, copyfunc_name, dest_uninitialized);
5611
5612 // Call it. Note that the count_ix value is not scaled to a byte-size.
5613 make_runtime_call(RC_LEAF|RC_NO_FP,
5614 OptoRuntime::fast_arraycopy_Type(),
5615 copyfunc_addr, copyfunc_name, adr_type,
5616 src_start, dest_start, copy_length XTOP);
5617 }
5618
5619 //-------------inline_encodeISOArray-----------------------------------
5620 // encode char[] to byte[] in ISO_8859_1
5621 bool LibraryCallKit::inline_encodeISOArray() {
5622 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5623 // no receiver since it is static method
5624 Node *src = argument(0);
5625 Node *src_offset = argument(1);
5626 Node *dst = argument(2);
5627 Node *dst_offset = argument(3);
5628 Node *length = argument(4);
5629
5630 const Type* src_type = src->Value(&_gvn);
5631 const Type* dst_type = dst->Value(&_gvn);
5632 const TypeAryPtr* top_src = src_type->isa_aryptr();
5633 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5634 if (top_src == NULL || top_src->klass() == NULL ||
5635 top_dest == NULL || top_dest->klass() == NULL) {
5636 // failed array check
5637 return false;
5638 }
5639
5640 // Figure out the size and type of the elements we will be copying.
5641 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5642 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5643 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5644 return false;
5645 }
5646 Node* src_start = array_element_address(src, src_offset, src_elem);
5647 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5648 // 'src_start' points to src array + scaled offset
5649 // 'dst_start' points to dst array + scaled offset
5650
5651 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5652 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5653 enc = _gvn.transform(enc);
5654 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5655 set_memory(res_mem, mtype);
5656 set_result(enc);
5657 return true;
5658 }
5659
5660 /**
5661 * Calculate CRC32 for byte.
5662 * int java.util.zip.CRC32.update(int crc, int b)
5663 */
5664 bool LibraryCallKit::inline_updateCRC32() {
5665 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5666 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5667 // no receiver since it is static method
5668 Node* crc = argument(0); // type: int
5669 Node* b = argument(1); // type: int
5670
5671 /*
5672 * int c = ~ crc;
5673 * b = timesXtoThe32[(b ^ c) & 0xFF];
5674 * b = b ^ (c >>> 8);
5675 * crc = ~b;
5676 */
5677
5678 Node* M1 = intcon(-1);
5679 crc = _gvn.transform(new (C) XorINode(crc, M1));
5680 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5681 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5682
5683 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5684 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5685 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5686 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5687
5688 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5689 result = _gvn.transform(new (C) XorINode(crc, result));
5690 result = _gvn.transform(new (C) XorINode(result, M1));
5691 set_result(result);
5692 return true;
5693 }
5694
5695 /**
5696 * Calculate CRC32 for byte[] array.
5697 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5698 */
5699 bool LibraryCallKit::inline_updateBytesCRC32() {
5700 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5701 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5702 // no receiver since it is static method
5703 Node* crc = argument(0); // type: int
5704 Node* src = argument(1); // type: oop
5705 Node* offset = argument(2); // type: int
5706 Node* length = argument(3); // type: int
5707
5708 const Type* src_type = src->Value(&_gvn);
5709 const TypeAryPtr* top_src = src_type->isa_aryptr();
5710 if (top_src == NULL || top_src->klass() == NULL) {
5711 // failed array check
5712 return false;
5713 }
5714
5715 // Figure out the size and type of the elements we will be copying.
5716 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5717 if (src_elem != T_BYTE) {
5718 return false;
5719 }
5720
5721 // 'src_start' points to src array + scaled offset
5722 Node* src_start = array_element_address(src, offset, src_elem);
5723
5724 // We assume that range check is done by caller.
5725 // TODO: generate range check (offset+length < src.length) in debug VM.
5726
5727 // Call the stub.
5728 address stubAddr = StubRoutines::updateBytesCRC32();
5729 const char *stubName = "updateBytesCRC32";
5730
5731 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5732 stubAddr, stubName, TypePtr::BOTTOM,
5733 crc, src_start, length);
5734 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5735 set_result(result);
5736 return true;
5737 }
5738
5739 /**
5740 * Calculate CRC32 for ByteBuffer.
5741 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5742 */
5743 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5744 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5745 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5746 // no receiver since it is static method
5747 Node* crc = argument(0); // type: int
5748 Node* src = argument(1); // type: long
5749 Node* offset = argument(3); // type: int
5750 Node* length = argument(4); // type: int
5751
5752 src = ConvL2X(src); // adjust Java long to machine word
5753 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5754 offset = ConvI2X(offset);
5755
5756 // 'src_start' points to src array + scaled offset
5757 Node* src_start = basic_plus_adr(top(), base, offset);
5758
5759 // Call the stub.
5760 address stubAddr = StubRoutines::updateBytesCRC32();
5761 const char *stubName = "updateBytesCRC32";
5762
5763 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5764 stubAddr, stubName, TypePtr::BOTTOM,
5765 crc, src_start, length);
5766 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5767 set_result(result);
5768 return true;
5769 }
5770
5771 //----------------------------inline_reference_get----------------------------
5772 // public T java.lang.ref.Reference.get();
5773 bool LibraryCallKit::inline_reference_get() {
5774 const int referent_offset = java_lang_ref_Reference::referent_offset;
5775 guarantee(referent_offset > 0, "should have already been set");
5776
5777 // Get the argument:
5778 Node* reference_obj = null_check_receiver();
5779 if (stopped()) return true;
5780
5781 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5782
5783 ciInstanceKlass* klass = env()->Object_klass();
5784 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5785
5786 Node* no_ctrl = NULL;
5787 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5788
5789 // Use the pre-barrier to record the value in the referent field
5790 pre_barrier(false /* do_load */,
5791 control(),
5792 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5793 result /* pre_val */,
5794 T_OBJECT);
5795
5796 // Add memory barrier to prevent commoning reads from this field
5797 // across safepoint since GC can change its value.
5798 insert_mem_bar(Op_MemBarCPUOrder);
5799
5800 set_result(result);
5801 return true;
5802 }
5803
5804
5805 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5806 bool is_exact=true, bool is_static=false) {
5807
5808 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5809 assert(tinst != NULL, "obj is null");
5810 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5811 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5812
5813 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5814 ciSymbol::make(fieldTypeString),
5815 is_static);
5816 if (field == NULL) return (Node *) NULL;
5817 assert (field != NULL, "undefined field");
5818
5819 // Next code copied from Parse::do_get_xxx():
5820
5821 // Compute address and memory type.
5822 int offset = field->offset_in_bytes();
5823 bool is_vol = field->is_volatile();
5824 ciType* field_klass = field->type();
5825 assert(field_klass->is_loaded(), "should be loaded");
5826 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5827 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5828 BasicType bt = field->layout_type();
5829
5830 // Build the resultant type of the load
5831 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5832
5833 // Build the load.
5834 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5835 return loadedField;
5836 }
5837
5838
5839 //------------------------------inline_aescrypt_Block-----------------------
5840 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5841 address stubAddr;
5842 const char *stubName;
5843 assert(UseAES, "need AES instruction support");
5844
5845 switch(id) {
5846 case vmIntrinsics::_aescrypt_encryptBlock:
5847 stubAddr = StubRoutines::aescrypt_encryptBlock();
5848 stubName = "aescrypt_encryptBlock";
5849 break;
5850 case vmIntrinsics::_aescrypt_decryptBlock:
5851 stubAddr = StubRoutines::aescrypt_decryptBlock();
5852 stubName = "aescrypt_decryptBlock";
5853 break;
5854 }
5855 if (stubAddr == NULL) return false;
5856
5857 Node* aescrypt_object = argument(0);
5858 Node* src = argument(1);
5859 Node* src_offset = argument(2);
5860 Node* dest = argument(3);
5861 Node* dest_offset = argument(4);
5862
5863 // (1) src and dest are arrays.
5864 const Type* src_type = src->Value(&_gvn);
5865 const Type* dest_type = dest->Value(&_gvn);
5866 const TypeAryPtr* top_src = src_type->isa_aryptr();
5867 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5868 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5869
5870 // for the quick and dirty code we will skip all the checks.
5871 // we are just trying to get the call to be generated.
5872 Node* src_start = src;
5873 Node* dest_start = dest;
5874 if (src_offset != NULL || dest_offset != NULL) {
5875 assert(src_offset != NULL && dest_offset != NULL, "");
5876 src_start = array_element_address(src, src_offset, T_BYTE);
5877 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5878 }
5879
5880 // now need to get the start of its expanded key array
5881 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5882 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5883 if (k_start == NULL) return false;
5884
5885 if (Matcher::pass_original_key_for_aes()) {
5886 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5887 // compatibility issues between Java key expansion and SPARC crypto instructions
5888 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5889 if (original_k_start == NULL) return false;
5890
5891 // Call the stub.
5892 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5893 stubAddr, stubName, TypePtr::BOTTOM,
5894 src_start, dest_start, k_start, original_k_start);
5895 } else {
5896 // Call the stub.
5897 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5898 stubAddr, stubName, TypePtr::BOTTOM,
5899 src_start, dest_start, k_start);
5900 }
5901
5902 return true;
5903 }
5904
5905 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5906 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5907 address stubAddr;
5908 const char *stubName;
5909
5910 assert(UseAES, "need AES instruction support");
5911
5912 switch(id) {
5913 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5914 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5915 stubName = "cipherBlockChaining_encryptAESCrypt";
5916 break;
5917 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5918 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5919 stubName = "cipherBlockChaining_decryptAESCrypt";
5920 break;
5921 }
5922 if (stubAddr == NULL) return false;
5923
5924 Node* cipherBlockChaining_object = argument(0);
5925 Node* src = argument(1);
5926 Node* src_offset = argument(2);
5927 Node* len = argument(3);
5928 Node* dest = argument(4);
5929 Node* dest_offset = argument(5);
5930
5931 // (1) src and dest are arrays.
5932 const Type* src_type = src->Value(&_gvn);
5933 const Type* dest_type = dest->Value(&_gvn);
5934 const TypeAryPtr* top_src = src_type->isa_aryptr();
5935 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5936 assert (top_src != NULL && top_src->klass() != NULL
5937 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5938
5939 // checks are the responsibility of the caller
5940 Node* src_start = src;
5941 Node* dest_start = dest;
5942 if (src_offset != NULL || dest_offset != NULL) {
5943 assert(src_offset != NULL && dest_offset != NULL, "");
5944 src_start = array_element_address(src, src_offset, T_BYTE);
5945 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5946 }
5947
5948 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5949 // (because of the predicated logic executed earlier).
5950 // so we cast it here safely.
5951 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5952
5953 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5954 if (embeddedCipherObj == NULL) return false;
5955
5956 // cast it to what we know it will be at runtime
5957 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5958 assert(tinst != NULL, "CBC obj is null");
5959 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5960 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5961 if (!klass_AESCrypt->is_loaded()) return false;
5962
5963 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5964 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5965 const TypeOopPtr* xtype = aklass->as_instance_type();
5966 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5967 aescrypt_object = _gvn.transform(aescrypt_object);
5968
5969 // we need to get the start of the aescrypt_object's expanded key array
5970 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5971 if (k_start == NULL) return false;
5972
5973 // similarly, get the start address of the r vector
5974 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5975 if (objRvec == NULL) return false;
5976 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5977
5978 Node* cbcCrypt;
5979 if (Matcher::pass_original_key_for_aes()) {
5980 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5981 // compatibility issues between Java key expansion and SPARC crypto instructions
5982 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5983 if (original_k_start == NULL) return false;
5984
5985 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5986 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5987 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5988 stubAddr, stubName, TypePtr::BOTTOM,
5989 src_start, dest_start, k_start, r_start, len, original_k_start);
5990 } else {
5991 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5992 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5993 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5994 stubAddr, stubName, TypePtr::BOTTOM,
5995 src_start, dest_start, k_start, r_start, len);
5996 }
5997
5998 // return cipher length (int)
5999 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6000 set_result(retvalue);
6001 return true;
6002 }
6003
6004 //------------------------------get_key_start_from_aescrypt_object-----------------------
6005 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6006 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6007 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6008 if (objAESCryptKey == NULL) return (Node *) NULL;
6009
6010 // now have the array, need to get the start address of the K array
6011 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6012 return k_start;
6013 }
6014
6015 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6016 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6017 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6018 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6019 if (objAESCryptKey == NULL) return (Node *) NULL;
6020
6021 // now have the array, need to get the start address of the lastKey array
6022 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6023 return original_k_start;
6024 }
6025
6026 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6027 // Return node representing slow path of predicate check.
6028 // the pseudo code we want to emulate with this predicate is:
6029 // for encryption:
6030 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6031 // for decryption:
6032 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6033 // note cipher==plain is more conservative than the original java code but that's OK
6034 //
6035 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6036 // First, check receiver for NULL since it is virtual method.
6037 Node* objCBC = argument(0);
6038 objCBC = null_check(objCBC);
6039
6040 if (stopped()) return NULL; // Always NULL
6041
6042 // Load embeddedCipher field of CipherBlockChaining object.
6043 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6044
6045 // get AESCrypt klass for instanceOf check
6046 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6047 // will have same classloader as CipherBlockChaining object
6048 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6049 assert(tinst != NULL, "CBCobj is null");
6050 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6051
6052 // we want to do an instanceof comparison against the AESCrypt class
6053 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6054 if (!klass_AESCrypt->is_loaded()) {
6055 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6056 Node* ctrl = control();
6057 set_control(top()); // no regular fast path
6058 return ctrl;
6059 }
6060 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6061
6062 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6063 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6064 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6065
6066 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6067
6068 // for encryption, we are done
6069 if (!decrypting)
6070 return instof_false; // even if it is NULL
6071
6072 // for decryption, we need to add a further check to avoid
6073 // taking the intrinsic path when cipher and plain are the same
6074 // see the original java code for why.
6075 RegionNode* region = new(C) RegionNode(3);
6076 region->init_req(1, instof_false);
6077 Node* src = argument(1);
6078 Node* dest = argument(4);
6079 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6080 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6081 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6082 region->init_req(2, src_dest_conjoint);
6083
6084 record_for_igvn(region);
6085 return _gvn.transform(region);
6086 }