1 /*
2 * Copyright (c) 1999, 2014, 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 "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callGenerator.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/countbitsnode.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/mathexactnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/mulnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subnode.hpp"
47 #include "prims/nativeLookup.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "trace/traceMacros.hpp"
50
51 class LibraryIntrinsic : public InlineCallGenerator {
52 // Extend the set of intrinsics known to the runtime:
53 public:
54 private:
55 bool _is_virtual;
56 bool _does_virtual_dispatch;
57 int8_t _predicates_count; // Intrinsic is predicated by several conditions
58 int8_t _last_predicate; // Last generated predicate
59 vmIntrinsics::ID _intrinsic_id;
60
61 public:
62 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
63 : InlineCallGenerator(m),
64 _is_virtual(is_virtual),
65 _does_virtual_dispatch(does_virtual_dispatch),
66 _predicates_count((int8_t)predicates_count),
67 _last_predicate((int8_t)-1),
68 _intrinsic_id(id)
69 {
70 }
71 virtual bool is_intrinsic() const { return true; }
72 virtual bool is_virtual() const { return _is_virtual; }
73 virtual bool is_predicated() const { return _predicates_count > 0; }
74 virtual int predicates_count() const { return _predicates_count; }
75 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
76 virtual JVMState* generate(JVMState* jvms);
77 virtual Node* generate_predicate(JVMState* jvms, int predicate);
78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
79 };
80
81
82 // Local helper class for LibraryIntrinsic:
83 class LibraryCallKit : public GraphKit {
84 private:
85 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
86 Node* _result; // the result node, if any
87 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
88
89 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
90
91 public:
92 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
93 : GraphKit(jvms),
94 _intrinsic(intrinsic),
95 _result(NULL)
96 {
97 // Check if this is a root compile. In that case we don't have a caller.
98 if (!jvms->has_method()) {
99 _reexecute_sp = sp();
100 } else {
101 // Find out how many arguments the interpreter needs when deoptimizing
102 // and save the stack pointer value so it can used by uncommon_trap.
103 // We find the argument count by looking at the declared signature.
104 bool ignored_will_link;
105 ciSignature* declared_signature = NULL;
106 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
107 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
108 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
109 }
110 }
111
112 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
113
114 ciMethod* caller() const { return jvms()->method(); }
115 int bci() const { return jvms()->bci(); }
116 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
117 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
118 ciMethod* callee() const { return _intrinsic->method(); }
119
120 bool try_to_inline(int predicate);
121 Node* try_to_predicate(int predicate);
122
123 void push_result() {
124 // Push the result onto the stack.
125 if (!stopped() && result() != NULL) {
126 BasicType bt = result()->bottom_type()->basic_type();
127 push_node(bt, result());
128 }
129 }
130
131 private:
132 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
134 }
135
136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
137 void set_result(RegionNode* region, PhiNode* value);
138 Node* result() { return _result; }
139
140 virtual int reexecute_sp() { return _reexecute_sp; }
141
142 // Helper functions to inline natives
143 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
144 Node* generate_slow_guard(Node* test, RegionNode* region);
145 Node* generate_fair_guard(Node* test, RegionNode* region);
146 Node* generate_negative_guard(Node* index, RegionNode* region,
147 // resulting CastII of index:
148 Node* *pos_index = NULL);
149 Node* generate_limit_guard(Node* offset, Node* subseq_length,
150 Node* array_length,
151 RegionNode* region);
152 Node* generate_current_thread(Node* &tls_output);
153 Node* load_mirror_from_klass(Node* klass);
154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path,
156 int offset);
157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path) {
159 int offset = java_lang_Class::klass_offset_in_bytes();
160 return load_klass_from_mirror_common(mirror, never_see_null,
161 region, null_path,
162 offset);
163 }
164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path) {
166 int offset = java_lang_Class::array_klass_offset_in_bytes();
167 return load_klass_from_mirror_common(mirror, never_see_null,
168 region, null_path,
169 offset);
170 }
171 Node* generate_access_flags_guard(Node* kls,
172 int modifier_mask, int modifier_bits,
173 RegionNode* region);
174 Node* generate_interface_guard(Node* kls, RegionNode* region);
175 Node* generate_array_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, false, false);
177 }
178 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, true);
180 }
181 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, true, false);
183 }
184 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, true);
186 }
187 Node* generate_array_guard_common(Node* kls, RegionNode* region,
188 bool obj_array, bool not_array);
189 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
190 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
191 bool is_virtual = false, bool is_static = false);
192 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
193 return generate_method_call(method_id, false, true);
194 }
195 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, true, false);
197 }
198 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
199
200 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
201 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
202 bool inline_string_compareTo();
203 bool inline_string_indexOf();
204 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
205 bool inline_string_equals();
206 Node* round_double_node(Node* n);
207 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
208 bool inline_math_native(vmIntrinsics::ID id);
209 bool inline_trig(vmIntrinsics::ID id);
210 bool inline_math(vmIntrinsics::ID id);
211 template <typename OverflowOp>
212 bool inline_math_overflow(Node* arg1, Node* arg2);
213 void inline_math_mathExact(Node* math, Node* test);
214 bool inline_math_addExactI(bool is_increment);
215 bool inline_math_addExactL(bool is_increment);
216 bool inline_math_multiplyExactI();
217 bool inline_math_multiplyExactL();
218 bool inline_math_negateExactI();
219 bool inline_math_negateExactL();
220 bool inline_math_subtractExactI(bool is_decrement);
221 bool inline_math_subtractExactL(bool is_decrement);
222 bool inline_exp();
223 bool inline_pow();
224 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_min_max(vmIntrinsics::ID id);
226 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
227 // This returns Type::AnyPtr, RawPtr, or OopPtr.
228 int classify_unsafe_addr(Node* &base, Node* &offset);
229 Node* make_unsafe_address(Node* base, Node* offset);
230 // Helper for inline_unsafe_access.
231 // Generates the guards that check whether the result of
232 // Unsafe.getObject should be recorded in an SATB log buffer.
233 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
234 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
235 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
236 static bool klass_needs_init_guard(Node* kls);
237 bool inline_unsafe_allocate();
238 bool inline_unsafe_copyMemory();
239 bool inline_native_currentThread();
240 #ifdef TRACE_HAVE_INTRINSICS
241 bool inline_native_classID();
242 bool inline_native_threadID();
243 #endif
244 bool inline_native_time_funcs(address method, const char* funcName);
245 bool inline_native_isInterrupted();
246 bool inline_native_Class_query(vmIntrinsics::ID id);
247 bool inline_native_subtype_check();
248
249 bool inline_native_newArray();
250 bool inline_native_getLength();
251 bool inline_array_copyOf(bool is_copyOfRange);
252 bool inline_array_equals();
253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
254 bool inline_native_clone(bool is_virtual);
255 bool inline_native_Reflection_getCallerClass();
256 // Helper function for inlining native object hash method
257 bool inline_native_hashcode(bool is_virtual, bool is_static);
258 bool inline_native_getClass();
259
260 // Helper functions for inlining arraycopy
261 bool inline_arraycopy();
262 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
263 RegionNode* slow_region);
264 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
265 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
266 bool inline_unsafe_ordered_store(BasicType type);
267 bool inline_unsafe_fence(vmIntrinsics::ID id);
268 bool inline_fp_conversions(vmIntrinsics::ID id);
269 bool inline_number_methods(vmIntrinsics::ID id);
270 bool inline_reference_get();
271 bool inline_aescrypt_Block(vmIntrinsics::ID id);
272 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
273 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
274 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
275 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
276 bool inline_sha_implCompress(vmIntrinsics::ID id);
277 bool inline_digestBase_implCompressMB(int predicate);
278 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
279 bool long_state, address stubAddr, const char *stubName,
280 Node* src_start, Node* ofs, Node* limit);
281 Node* get_state_from_sha_object(Node *sha_object);
282 Node* get_state_from_sha5_object(Node *sha_object);
283 Node* inline_digestBase_implCompressMB_predicate(int predicate);
284 bool inline_encodeISOArray();
285 bool inline_updateCRC32();
286 bool inline_updateBytesCRC32();
287 bool inline_updateByteBufferCRC32();
288 bool inline_multiplyToLen();
289 };
290
291
292 //---------------------------make_vm_intrinsic----------------------------
293 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
294 vmIntrinsics::ID id = m->intrinsic_id();
295 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
296
297 ccstr disable_intr = NULL;
298
299 if ((DisableIntrinsic[0] != '\0'
300 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
301 (method_has_option_value("DisableIntrinsic", disable_intr)
302 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
303 // disabled by a user request on the command line:
304 // example: -XX:DisableIntrinsic=_hashCode,_getClass
305 return NULL;
306 }
307
308 if (!m->is_loaded()) {
309 // do not attempt to inline unloaded methods
310 return NULL;
311 }
312
313 // Only a few intrinsics implement a virtual dispatch.
314 // They are expensive calls which are also frequently overridden.
315 if (is_virtual) {
316 switch (id) {
317 case vmIntrinsics::_hashCode:
318 case vmIntrinsics::_clone:
319 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
320 break;
321 default:
322 return NULL;
323 }
324 }
325
326 // -XX:-InlineNatives disables nearly all intrinsics:
327 if (!InlineNatives) {
328 switch (id) {
329 case vmIntrinsics::_indexOf:
330 case vmIntrinsics::_compareTo:
331 case vmIntrinsics::_equals:
332 case vmIntrinsics::_equalsC:
333 case vmIntrinsics::_getAndAddInt:
334 case vmIntrinsics::_getAndAddLong:
335 case vmIntrinsics::_getAndSetInt:
336 case vmIntrinsics::_getAndSetLong:
337 case vmIntrinsics::_getAndSetObject:
338 case vmIntrinsics::_loadFence:
339 case vmIntrinsics::_storeFence:
340 case vmIntrinsics::_fullFence:
341 break; // InlineNatives does not control String.compareTo
342 case vmIntrinsics::_Reference_get:
343 break; // InlineNatives does not control Reference.get
344 default:
345 return NULL;
346 }
347 }
348
349 int predicates = 0;
350 bool does_virtual_dispatch = false;
351
352 switch (id) {
353 case vmIntrinsics::_compareTo:
354 if (!SpecialStringCompareTo) return NULL;
355 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
356 break;
357 case vmIntrinsics::_indexOf:
358 if (!SpecialStringIndexOf) return NULL;
359 break;
360 case vmIntrinsics::_equals:
361 if (!SpecialStringEquals) return NULL;
362 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
363 break;
364 case vmIntrinsics::_equalsC:
365 if (!SpecialArraysEquals) return NULL;
366 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
367 break;
368 case vmIntrinsics::_arraycopy:
369 if (!InlineArrayCopy) return NULL;
370 break;
371 case vmIntrinsics::_copyMemory:
372 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
373 if (!InlineArrayCopy) return NULL;
374 break;
375 case vmIntrinsics::_hashCode:
376 if (!InlineObjectHash) return NULL;
377 does_virtual_dispatch = true;
378 break;
379 case vmIntrinsics::_clone:
380 does_virtual_dispatch = true;
381 case vmIntrinsics::_copyOf:
382 case vmIntrinsics::_copyOfRange:
383 if (!InlineObjectCopy) return NULL;
384 // These also use the arraycopy intrinsic mechanism:
385 if (!InlineArrayCopy) return NULL;
386 break;
387 case vmIntrinsics::_encodeISOArray:
388 if (!SpecialEncodeISOArray) return NULL;
389 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
390 break;
391 case vmIntrinsics::_checkIndex:
392 // We do not intrinsify this. The optimizer does fine with it.
393 return NULL;
394
395 case vmIntrinsics::_getCallerClass:
396 if (!InlineReflectionGetCallerClass) return NULL;
397 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
398 break;
399
400 case vmIntrinsics::_bitCount_i:
401 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
402 break;
403
404 case vmIntrinsics::_bitCount_l:
405 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
406 break;
407
408 case vmIntrinsics::_numberOfLeadingZeros_i:
409 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
410 break;
411
412 case vmIntrinsics::_numberOfLeadingZeros_l:
413 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
414 break;
415
416 case vmIntrinsics::_numberOfTrailingZeros_i:
417 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
418 break;
419
420 case vmIntrinsics::_numberOfTrailingZeros_l:
421 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
422 break;
423
424 case vmIntrinsics::_reverseBytes_c:
425 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
426 break;
427 case vmIntrinsics::_reverseBytes_s:
428 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
429 break;
430 case vmIntrinsics::_reverseBytes_i:
431 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
432 break;
433 case vmIntrinsics::_reverseBytes_l:
434 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
435 break;
436
437 case vmIntrinsics::_Reference_get:
438 // Use the intrinsic version of Reference.get() so that the value in
439 // the referent field can be registered by the G1 pre-barrier code.
440 // Also add memory barrier to prevent commoning reads from this field
441 // across safepoint since GC can change it value.
442 break;
443
444 case vmIntrinsics::_compareAndSwapObject:
445 #ifdef _LP64
446 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
447 #endif
448 break;
449
450 case vmIntrinsics::_compareAndSwapLong:
451 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
452 break;
453
454 case vmIntrinsics::_getAndAddInt:
455 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
456 break;
457
458 case vmIntrinsics::_getAndAddLong:
459 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
460 break;
461
462 case vmIntrinsics::_getAndSetInt:
463 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
464 break;
465
466 case vmIntrinsics::_getAndSetLong:
467 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
468 break;
469
470 case vmIntrinsics::_getAndSetObject:
471 #ifdef _LP64
472 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
473 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
474 break;
475 #else
476 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
477 break;
478 #endif
479
480 case vmIntrinsics::_aescrypt_encryptBlock:
481 case vmIntrinsics::_aescrypt_decryptBlock:
482 if (!UseAESIntrinsics) return NULL;
483 break;
484
485 case vmIntrinsics::_multiplyToLen:
486 if (!UseMultiplyToLenIntrinsic) return NULL;
487 break;
488
489 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
490 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
491 if (!UseAESIntrinsics) return NULL;
492 // these two require the predicated logic
493 predicates = 1;
494 break;
495
496 case vmIntrinsics::_sha_implCompress:
497 if (!UseSHA1Intrinsics) return NULL;
498 break;
499
500 case vmIntrinsics::_sha2_implCompress:
501 if (!UseSHA256Intrinsics) return NULL;
502 break;
503
504 case vmIntrinsics::_sha5_implCompress:
505 if (!UseSHA512Intrinsics) return NULL;
506 break;
507
508 case vmIntrinsics::_digestBase_implCompressMB:
509 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
510 predicates = 3;
511 break;
512
513 case vmIntrinsics::_updateCRC32:
514 case vmIntrinsics::_updateBytesCRC32:
515 case vmIntrinsics::_updateByteBufferCRC32:
516 if (!UseCRC32Intrinsics) return NULL;
517 break;
518
519 case vmIntrinsics::_incrementExactI:
520 case vmIntrinsics::_addExactI:
521 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
522 break;
523 case vmIntrinsics::_incrementExactL:
524 case vmIntrinsics::_addExactL:
525 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
526 break;
527 case vmIntrinsics::_decrementExactI:
528 case vmIntrinsics::_subtractExactI:
529 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
530 break;
531 case vmIntrinsics::_decrementExactL:
532 case vmIntrinsics::_subtractExactL:
533 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
534 break;
535 case vmIntrinsics::_negateExactI:
536 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
537 break;
538 case vmIntrinsics::_negateExactL:
539 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
540 break;
541 case vmIntrinsics::_multiplyExactI:
542 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
543 break;
544 case vmIntrinsics::_multiplyExactL:
545 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
546 break;
547
548 default:
549 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
550 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
551 break;
552 }
553
554 // -XX:-InlineClassNatives disables natives from the Class class.
555 // The flag applies to all reflective calls, notably Array.newArray
556 // (visible to Java programmers as Array.newInstance).
557 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
558 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
559 if (!InlineClassNatives) return NULL;
560 }
561
562 // -XX:-InlineThreadNatives disables natives from the Thread class.
563 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
564 if (!InlineThreadNatives) return NULL;
565 }
566
567 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
568 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
569 m->holder()->name() == ciSymbol::java_lang_Float() ||
570 m->holder()->name() == ciSymbol::java_lang_Double()) {
571 if (!InlineMathNatives) return NULL;
572 }
573
574 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
575 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
576 if (!InlineUnsafeOps) return NULL;
577 }
578
579 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
580 }
581
582 //----------------------register_library_intrinsics-----------------------
583 // Initialize this file's data structures, for each Compile instance.
584 void Compile::register_library_intrinsics() {
585 // Nothing to do here.
586 }
587
588 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
589 LibraryCallKit kit(jvms, this);
590 Compile* C = kit.C;
591 int nodes = C->unique();
592 #ifndef PRODUCT
593 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
594 char buf[1000];
595 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
596 tty->print_cr("Intrinsic %s", str);
597 }
598 #endif
599 ciMethod* callee = kit.callee();
600 const int bci = kit.bci();
601
602 // Try to inline the intrinsic.
603 if (kit.try_to_inline(_last_predicate)) {
604 if (C->print_intrinsics() || C->print_inlining()) {
605 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
606 }
607 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
608 if (C->log()) {
609 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
610 vmIntrinsics::name_at(intrinsic_id()),
611 (is_virtual() ? " virtual='1'" : ""),
612 C->unique() - nodes);
613 }
614 // Push the result from the inlined method onto the stack.
615 kit.push_result();
616 C->print_inlining_update(this);
617 return kit.transfer_exceptions_into_jvms();
618 }
619
620 // The intrinsic bailed out
621 if (C->print_intrinsics() || C->print_inlining()) {
622 if (jvms->has_method()) {
623 // Not a root compile.
624 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
625 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
626 } else {
627 // Root compile
628 tty->print("Did not generate intrinsic %s%s at bci:%d in",
629 vmIntrinsics::name_at(intrinsic_id()),
630 (is_virtual() ? " (virtual)" : ""), bci);
631 }
632 }
633 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
634 C->print_inlining_update(this);
635 return NULL;
636 }
637
638 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
639 LibraryCallKit kit(jvms, this);
640 Compile* C = kit.C;
641 int nodes = C->unique();
642 _last_predicate = predicate;
643 #ifndef PRODUCT
644 assert(is_predicated() && predicate < predicates_count(), "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(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, predicate)" : "(intrinsic, predicate)");
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(int predicate) {
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::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
849
850 case vmIntrinsics::_floatToRawIntBits:
851 case vmIntrinsics::_floatToIntBits:
852 case vmIntrinsics::_intBitsToFloat:
853 case vmIntrinsics::_doubleToRawLongBits:
854 case vmIntrinsics::_doubleToLongBits:
855 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
856
857 case vmIntrinsics::_numberOfLeadingZeros_i:
858 case vmIntrinsics::_numberOfLeadingZeros_l:
859 case vmIntrinsics::_numberOfTrailingZeros_i:
860 case vmIntrinsics::_numberOfTrailingZeros_l:
861 case vmIntrinsics::_bitCount_i:
862 case vmIntrinsics::_bitCount_l:
863 case vmIntrinsics::_reverseBytes_i:
864 case vmIntrinsics::_reverseBytes_l:
865 case vmIntrinsics::_reverseBytes_s:
866 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
867
868 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
869
870 case vmIntrinsics::_Reference_get: return inline_reference_get();
871
872 case vmIntrinsics::_aescrypt_encryptBlock:
873 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
874
875 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
876 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
877 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
878
879 case vmIntrinsics::_sha_implCompress:
880 case vmIntrinsics::_sha2_implCompress:
881 case vmIntrinsics::_sha5_implCompress:
882 return inline_sha_implCompress(intrinsic_id());
883
884 case vmIntrinsics::_digestBase_implCompressMB:
885 return inline_digestBase_implCompressMB(predicate);
886
887 case vmIntrinsics::_multiplyToLen:
888 return inline_multiplyToLen();
889
890 case vmIntrinsics::_encodeISOArray:
891 return inline_encodeISOArray();
892
893 case vmIntrinsics::_updateCRC32:
894 return inline_updateCRC32();
895 case vmIntrinsics::_updateBytesCRC32:
896 return inline_updateBytesCRC32();
897 case vmIntrinsics::_updateByteBufferCRC32:
898 return inline_updateByteBufferCRC32();
899
900 default:
901 // If you get here, it may be that someone has added a new intrinsic
902 // to the list in vmSymbols.hpp without implementing it here.
903 #ifndef PRODUCT
904 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
905 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
906 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
907 }
908 #endif
909 return false;
910 }
911 }
912
913 Node* LibraryCallKit::try_to_predicate(int predicate) {
914 if (!jvms()->has_method()) {
915 // Root JVMState has a null method.
916 assert(map()->memory()->Opcode() == Op_Parm, "");
917 // Insert the memory aliasing node
918 set_all_memory(reset_memory());
919 }
920 assert(merged_memory(), "");
921
922 switch (intrinsic_id()) {
923 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
924 return inline_cipherBlockChaining_AESCrypt_predicate(false);
925 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
926 return inline_cipherBlockChaining_AESCrypt_predicate(true);
927 case vmIntrinsics::_digestBase_implCompressMB:
928 return inline_digestBase_implCompressMB_predicate(predicate);
929
930 default:
931 // If you get here, it may be that someone has added a new intrinsic
932 // to the list in vmSymbols.hpp without implementing it here.
933 #ifndef PRODUCT
934 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
935 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
936 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
937 }
938 #endif
939 Node* slow_ctl = control();
940 set_control(top()); // No fast path instrinsic
941 return slow_ctl;
942 }
943 }
944
945 //------------------------------set_result-------------------------------
946 // Helper function for finishing intrinsics.
947 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
948 record_for_igvn(region);
949 set_control(_gvn.transform(region));
950 set_result( _gvn.transform(value));
951 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
952 }
953
954 //------------------------------generate_guard---------------------------
955 // Helper function for generating guarded fast-slow graph structures.
956 // The given 'test', if true, guards a slow path. If the test fails
957 // then a fast path can be taken. (We generally hope it fails.)
958 // In all cases, GraphKit::control() is updated to the fast path.
959 // The returned value represents the control for the slow path.
960 // The return value is never 'top'; it is either a valid control
961 // or NULL if it is obvious that the slow path can never be taken.
962 // Also, if region and the slow control are not NULL, the slow edge
963 // is appended to the region.
964 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
965 if (stopped()) {
966 // Already short circuited.
967 return NULL;
968 }
969
970 // Build an if node and its projections.
971 // If test is true we take the slow path, which we assume is uncommon.
972 if (_gvn.type(test) == TypeInt::ZERO) {
973 // The slow branch is never taken. No need to build this guard.
974 return NULL;
975 }
976
977 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
978
979 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
980 if (if_slow == top()) {
981 // The slow branch is never taken. No need to build this guard.
982 return NULL;
983 }
984
985 if (region != NULL)
986 region->add_req(if_slow);
987
988 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
989 set_control(if_fast);
990
991 return if_slow;
992 }
993
994 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
995 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
996 }
997 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
998 return generate_guard(test, region, PROB_FAIR);
999 }
1000
1001 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1002 Node* *pos_index) {
1003 if (stopped())
1004 return NULL; // already stopped
1005 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1006 return NULL; // index is already adequately typed
1007 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1008 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1009 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1010 if (is_neg != NULL && pos_index != NULL) {
1011 // Emulate effect of Parse::adjust_map_after_if.
1012 Node* ccast = new CastIINode(index, TypeInt::POS);
1013 ccast->set_req(0, control());
1014 (*pos_index) = _gvn.transform(ccast);
1015 }
1016 return is_neg;
1017 }
1018
1019 // Make sure that 'position' is a valid limit index, in [0..length].
1020 // There are two equivalent plans for checking this:
1021 // A. (offset + copyLength) unsigned<= arrayLength
1022 // B. offset <= (arrayLength - copyLength)
1023 // We require that all of the values above, except for the sum and
1024 // difference, are already known to be non-negative.
1025 // Plan A is robust in the face of overflow, if offset and copyLength
1026 // are both hugely positive.
1027 //
1028 // Plan B is less direct and intuitive, but it does not overflow at
1029 // all, since the difference of two non-negatives is always
1030 // representable. Whenever Java methods must perform the equivalent
1031 // check they generally use Plan B instead of Plan A.
1032 // For the moment we use Plan A.
1033 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1034 Node* subseq_length,
1035 Node* array_length,
1036 RegionNode* region) {
1037 if (stopped())
1038 return NULL; // already stopped
1039 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1040 if (zero_offset && subseq_length->eqv_uncast(array_length))
1041 return NULL; // common case of whole-array copy
1042 Node* last = subseq_length;
1043 if (!zero_offset) // last += offset
1044 last = _gvn.transform(new AddINode(last, offset));
1045 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1046 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1047 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1048 return is_over;
1049 }
1050
1051
1052 //--------------------------generate_current_thread--------------------
1053 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1054 ciKlass* thread_klass = env()->Thread_klass();
1055 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1056 Node* thread = _gvn.transform(new ThreadLocalNode());
1057 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1058 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1059 tls_output = thread;
1060 return threadObj;
1061 }
1062
1063
1064 //------------------------------make_string_method_node------------------------
1065 // Helper method for String intrinsic functions. This version is called
1066 // with str1 and str2 pointing to String object nodes.
1067 //
1068 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1069 Node* no_ctrl = NULL;
1070
1071 // Get start addr of string
1072 Node* str1_value = load_String_value(no_ctrl, str1);
1073 Node* str1_offset = load_String_offset(no_ctrl, str1);
1074 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1075
1076 // Get length of string 1
1077 Node* str1_len = load_String_length(no_ctrl, str1);
1078
1079 Node* str2_value = load_String_value(no_ctrl, str2);
1080 Node* str2_offset = load_String_offset(no_ctrl, str2);
1081 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1082
1083 Node* str2_len = NULL;
1084 Node* result = NULL;
1085
1086 switch (opcode) {
1087 case Op_StrIndexOf:
1088 // Get length of string 2
1089 str2_len = load_String_length(no_ctrl, str2);
1090
1091 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1092 str1_start, str1_len, str2_start, str2_len);
1093 break;
1094 case Op_StrComp:
1095 // Get length of string 2
1096 str2_len = load_String_length(no_ctrl, str2);
1097
1098 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1099 str1_start, str1_len, str2_start, str2_len);
1100 break;
1101 case Op_StrEquals:
1102 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1103 str1_start, str2_start, str1_len);
1104 break;
1105 default:
1106 ShouldNotReachHere();
1107 return NULL;
1108 }
1109
1110 // All these intrinsics have checks.
1111 C->set_has_split_ifs(true); // Has chance for split-if optimization
1112
1113 return _gvn.transform(result);
1114 }
1115
1116 // Helper method for String intrinsic functions. This version is called
1117 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1118 // to Int nodes containing the lenghts of str1 and str2.
1119 //
1120 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1121 Node* result = NULL;
1122 switch (opcode) {
1123 case Op_StrIndexOf:
1124 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1125 str1_start, cnt1, str2_start, cnt2);
1126 break;
1127 case Op_StrComp:
1128 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1129 str1_start, cnt1, str2_start, cnt2);
1130 break;
1131 case Op_StrEquals:
1132 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1133 str1_start, str2_start, cnt1);
1134 break;
1135 default:
1136 ShouldNotReachHere();
1137 return NULL;
1138 }
1139
1140 // All these intrinsics have checks.
1141 C->set_has_split_ifs(true); // Has chance for split-if optimization
1142
1143 return _gvn.transform(result);
1144 }
1145
1146 //------------------------------inline_string_compareTo------------------------
1147 // public int java.lang.String.compareTo(String anotherString);
1148 bool LibraryCallKit::inline_string_compareTo() {
1149 Node* receiver = null_check(argument(0));
1150 Node* arg = null_check(argument(1));
1151 if (stopped()) {
1152 return true;
1153 }
1154 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1155 return true;
1156 }
1157
1158 //------------------------------inline_string_equals------------------------
1159 bool LibraryCallKit::inline_string_equals() {
1160 Node* receiver = null_check_receiver();
1161 // NOTE: Do not null check argument for String.equals() because spec
1162 // allows to specify NULL as argument.
1163 Node* argument = this->argument(1);
1164 if (stopped()) {
1165 return true;
1166 }
1167
1168 // paths (plus control) merge
1169 RegionNode* region = new RegionNode(5);
1170 Node* phi = new PhiNode(region, TypeInt::BOOL);
1171
1172 // does source == target string?
1173 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1174 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1175
1176 Node* if_eq = generate_slow_guard(bol, NULL);
1177 if (if_eq != NULL) {
1178 // receiver == argument
1179 phi->init_req(2, intcon(1));
1180 region->init_req(2, if_eq);
1181 }
1182
1183 // get String klass for instanceOf
1184 ciInstanceKlass* klass = env()->String_klass();
1185
1186 if (!stopped()) {
1187 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1188 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1189 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1190
1191 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1192 //instanceOf == true, fallthrough
1193
1194 if (inst_false != NULL) {
1195 phi->init_req(3, intcon(0));
1196 region->init_req(3, inst_false);
1197 }
1198 }
1199
1200 if (!stopped()) {
1201 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1202
1203 // Properly cast the argument to String
1204 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1205 // This path is taken only when argument's type is String:NotNull.
1206 argument = cast_not_null(argument, false);
1207
1208 Node* no_ctrl = NULL;
1209
1210 // Get start addr of receiver
1211 Node* receiver_val = load_String_value(no_ctrl, receiver);
1212 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1213 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1214
1215 // Get length of receiver
1216 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1217
1218 // Get start addr of argument
1219 Node* argument_val = load_String_value(no_ctrl, argument);
1220 Node* argument_offset = load_String_offset(no_ctrl, argument);
1221 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1222
1223 // Get length of argument
1224 Node* argument_cnt = load_String_length(no_ctrl, argument);
1225
1226 // Check for receiver count != argument count
1227 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1228 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1229 Node* if_ne = generate_slow_guard(bol, NULL);
1230 if (if_ne != NULL) {
1231 phi->init_req(4, intcon(0));
1232 region->init_req(4, if_ne);
1233 }
1234
1235 // Check for count == 0 is done by assembler code for StrEquals.
1236
1237 if (!stopped()) {
1238 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1239 phi->init_req(1, equals);
1240 region->init_req(1, control());
1241 }
1242 }
1243
1244 // post merge
1245 set_control(_gvn.transform(region));
1246 record_for_igvn(region);
1247
1248 set_result(_gvn.transform(phi));
1249 return true;
1250 }
1251
1252 //------------------------------inline_array_equals----------------------------
1253 bool LibraryCallKit::inline_array_equals() {
1254 Node* arg1 = argument(0);
1255 Node* arg2 = argument(1);
1256 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1257 return true;
1258 }
1259
1260 // Java version of String.indexOf(constant string)
1261 // class StringDecl {
1262 // StringDecl(char[] ca) {
1263 // offset = 0;
1264 // count = ca.length;
1265 // value = ca;
1266 // }
1267 // int offset;
1268 // int count;
1269 // char[] value;
1270 // }
1271 //
1272 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1273 // int targetOffset, int cache_i, int md2) {
1274 // int cache = cache_i;
1275 // int sourceOffset = string_object.offset;
1276 // int sourceCount = string_object.count;
1277 // int targetCount = target_object.length;
1278 //
1279 // int targetCountLess1 = targetCount - 1;
1280 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1281 //
1282 // char[] source = string_object.value;
1283 // char[] target = target_object;
1284 // int lastChar = target[targetCountLess1];
1285 //
1286 // outer_loop:
1287 // for (int i = sourceOffset; i < sourceEnd; ) {
1288 // int src = source[i + targetCountLess1];
1289 // if (src == lastChar) {
1290 // // With random strings and a 4-character alphabet,
1291 // // reverse matching at this point sets up 0.8% fewer
1292 // // frames, but (paradoxically) makes 0.3% more probes.
1293 // // Since those probes are nearer the lastChar probe,
1294 // // there is may be a net D$ win with reverse matching.
1295 // // But, reversing loop inhibits unroll of inner loop
1296 // // for unknown reason. So, does running outer loop from
1297 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1298 // for (int j = 0; j < targetCountLess1; j++) {
1299 // if (target[targetOffset + j] != source[i+j]) {
1300 // if ((cache & (1 << source[i+j])) == 0) {
1301 // if (md2 < j+1) {
1302 // i += j+1;
1303 // continue outer_loop;
1304 // }
1305 // }
1306 // i += md2;
1307 // continue outer_loop;
1308 // }
1309 // }
1310 // return i - sourceOffset;
1311 // }
1312 // if ((cache & (1 << src)) == 0) {
1313 // i += targetCountLess1;
1314 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1315 // i++;
1316 // }
1317 // return -1;
1318 // }
1319
1320 //------------------------------string_indexOf------------------------
1321 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1322 jint cache_i, jint md2_i) {
1323
1324 Node* no_ctrl = NULL;
1325 float likely = PROB_LIKELY(0.9);
1326 float unlikely = PROB_UNLIKELY(0.9);
1327
1328 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1329
1330 Node* source = load_String_value(no_ctrl, string_object);
1331 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1332 Node* sourceCount = load_String_length(no_ctrl, string_object);
1333
1334 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1335 jint target_length = target_array->length();
1336 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1337 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1338
1339 // String.value field is known to be @Stable.
1340 if (UseImplicitStableValues) {
1341 target = cast_array_to_stable(target, target_type);
1342 }
1343
1344 IdealKit kit(this, false, true);
1345 #define __ kit.
1346 Node* zero = __ ConI(0);
1347 Node* one = __ ConI(1);
1348 Node* cache = __ ConI(cache_i);
1349 Node* md2 = __ ConI(md2_i);
1350 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1351 Node* targetCount = __ ConI(target_length);
1352 Node* targetCountLess1 = __ ConI(target_length - 1);
1353 Node* targetOffset = __ ConI(targetOffset_i);
1354 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1355
1356 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1357 Node* outer_loop = __ make_label(2 /* goto */);
1358 Node* return_ = __ make_label(1);
1359
1360 __ set(rtn,__ ConI(-1));
1361 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1362 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1363 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1364 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1365 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1366 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1367 Node* tpj = __ AddI(targetOffset, __ value(j));
1368 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1369 Node* ipj = __ AddI(__ value(i), __ value(j));
1370 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1371 __ if_then(targ, BoolTest::ne, src2); {
1372 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1373 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1374 __ increment(i, __ AddI(__ value(j), one));
1375 __ goto_(outer_loop);
1376 } __ end_if(); __ dead(j);
1377 }__ end_if(); __ dead(j);
1378 __ increment(i, md2);
1379 __ goto_(outer_loop);
1380 }__ end_if();
1381 __ increment(j, one);
1382 }__ end_loop(); __ dead(j);
1383 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1384 __ goto_(return_);
1385 }__ end_if();
1386 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1387 __ increment(i, targetCountLess1);
1388 }__ end_if();
1389 __ increment(i, one);
1390 __ bind(outer_loop);
1391 }__ end_loop(); __ dead(i);
1392 __ bind(return_);
1393
1394 // Final sync IdealKit and GraphKit.
1395 final_sync(kit);
1396 Node* result = __ value(rtn);
1397 #undef __
1398 C->set_has_loops(true);
1399 return result;
1400 }
1401
1402 //------------------------------inline_string_indexOf------------------------
1403 bool LibraryCallKit::inline_string_indexOf() {
1404 Node* receiver = argument(0);
1405 Node* arg = argument(1);
1406
1407 Node* result;
1408 // Disable the use of pcmpestri until it can be guaranteed that
1409 // the load doesn't cross into the uncommited space.
1410 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1411 UseSSE42Intrinsics) {
1412 // Generate SSE4.2 version of indexOf
1413 // We currently only have match rules that use SSE4.2
1414
1415 receiver = null_check(receiver);
1416 arg = null_check(arg);
1417 if (stopped()) {
1418 return true;
1419 }
1420
1421 ciInstanceKlass* str_klass = env()->String_klass();
1422 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1423
1424 // Make the merge point
1425 RegionNode* result_rgn = new RegionNode(4);
1426 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1427 Node* no_ctrl = NULL;
1428
1429 // Get start addr of source string
1430 Node* source = load_String_value(no_ctrl, receiver);
1431 Node* source_offset = load_String_offset(no_ctrl, receiver);
1432 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1433
1434 // Get length of source string
1435 Node* source_cnt = load_String_length(no_ctrl, receiver);
1436
1437 // Get start addr of substring
1438 Node* substr = load_String_value(no_ctrl, arg);
1439 Node* substr_offset = load_String_offset(no_ctrl, arg);
1440 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1441
1442 // Get length of source string
1443 Node* substr_cnt = load_String_length(no_ctrl, arg);
1444
1445 // Check for substr count > string count
1446 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1447 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1448 Node* if_gt = generate_slow_guard(bol, NULL);
1449 if (if_gt != NULL) {
1450 result_phi->init_req(2, intcon(-1));
1451 result_rgn->init_req(2, if_gt);
1452 }
1453
1454 if (!stopped()) {
1455 // Check for substr count == 0
1456 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1457 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1458 Node* if_zero = generate_slow_guard(bol, NULL);
1459 if (if_zero != NULL) {
1460 result_phi->init_req(3, intcon(0));
1461 result_rgn->init_req(3, if_zero);
1462 }
1463 }
1464
1465 if (!stopped()) {
1466 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1467 result_phi->init_req(1, result);
1468 result_rgn->init_req(1, control());
1469 }
1470 set_control(_gvn.transform(result_rgn));
1471 record_for_igvn(result_rgn);
1472 result = _gvn.transform(result_phi);
1473
1474 } else { // Use LibraryCallKit::string_indexOf
1475 // don't intrinsify if argument isn't a constant string.
1476 if (!arg->is_Con()) {
1477 return false;
1478 }
1479 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1480 if (str_type == NULL) {
1481 return false;
1482 }
1483 ciInstanceKlass* klass = env()->String_klass();
1484 ciObject* str_const = str_type->const_oop();
1485 if (str_const == NULL || str_const->klass() != klass) {
1486 return false;
1487 }
1488 ciInstance* str = str_const->as_instance();
1489 assert(str != NULL, "must be instance");
1490
1491 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1492 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1493
1494 int o;
1495 int c;
1496 if (java_lang_String::has_offset_field()) {
1497 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1498 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1499 } else {
1500 o = 0;
1501 c = pat->length();
1502 }
1503
1504 // constant strings have no offset and count == length which
1505 // simplifies the resulting code somewhat so lets optimize for that.
1506 if (o != 0 || c != pat->length()) {
1507 return false;
1508 }
1509
1510 receiver = null_check(receiver, T_OBJECT);
1511 // NOTE: No null check on the argument is needed since it's a constant String oop.
1512 if (stopped()) {
1513 return true;
1514 }
1515
1516 // The null string as a pattern always returns 0 (match at beginning of string)
1517 if (c == 0) {
1518 set_result(intcon(0));
1519 return true;
1520 }
1521
1522 // Generate default indexOf
1523 jchar lastChar = pat->char_at(o + (c - 1));
1524 int cache = 0;
1525 int i;
1526 for (i = 0; i < c - 1; i++) {
1527 assert(i < pat->length(), "out of range");
1528 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1529 }
1530
1531 int md2 = c;
1532 for (i = 0; i < c - 1; i++) {
1533 assert(i < pat->length(), "out of range");
1534 if (pat->char_at(o + i) == lastChar) {
1535 md2 = (c - 1) - i;
1536 }
1537 }
1538
1539 result = string_indexOf(receiver, pat, o, cache, md2);
1540 }
1541 set_result(result);
1542 return true;
1543 }
1544
1545 //--------------------------round_double_node--------------------------------
1546 // Round a double node if necessary.
1547 Node* LibraryCallKit::round_double_node(Node* n) {
1548 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1549 n = _gvn.transform(new RoundDoubleNode(0, n));
1550 return n;
1551 }
1552
1553 //------------------------------inline_math-----------------------------------
1554 // public static double Math.abs(double)
1555 // public static double Math.sqrt(double)
1556 // public static double Math.log(double)
1557 // public static double Math.log10(double)
1558 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1559 Node* arg = round_double_node(argument(0));
1560 Node* n;
1561 switch (id) {
1562 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1563 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1564 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1565 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1566 default: fatal_unexpected_iid(id); break;
1567 }
1568 set_result(_gvn.transform(n));
1569 return true;
1570 }
1571
1572 //------------------------------inline_trig----------------------------------
1573 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1574 // argument reduction which will turn into a fast/slow diamond.
1575 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1576 Node* arg = round_double_node(argument(0));
1577 Node* n = NULL;
1578
1579 switch (id) {
1580 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1581 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1582 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1583 default: fatal_unexpected_iid(id); break;
1584 }
1585 n = _gvn.transform(n);
1586
1587 // Rounding required? Check for argument reduction!
1588 if (Matcher::strict_fp_requires_explicit_rounding) {
1589 static const double pi_4 = 0.7853981633974483;
1590 static const double neg_pi_4 = -0.7853981633974483;
1591 // pi/2 in 80-bit extended precision
1592 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1593 // -pi/2 in 80-bit extended precision
1594 // 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};
1595 // Cutoff value for using this argument reduction technique
1596 //static const double pi_2_minus_epsilon = 1.564660403643354;
1597 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1598
1599 // Pseudocode for sin:
1600 // if (x <= Math.PI / 4.0) {
1601 // if (x >= -Math.PI / 4.0) return fsin(x);
1602 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1603 // } else {
1604 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1605 // }
1606 // return StrictMath.sin(x);
1607
1608 // Pseudocode for cos:
1609 // if (x <= Math.PI / 4.0) {
1610 // if (x >= -Math.PI / 4.0) return fcos(x);
1611 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1612 // } else {
1613 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1614 // }
1615 // return StrictMath.cos(x);
1616
1617 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1618 // requires a special machine instruction to load it. Instead we'll try
1619 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1620 // probably do the math inside the SIN encoding.
1621
1622 // Make the merge point
1623 RegionNode* r = new RegionNode(3);
1624 Node* phi = new PhiNode(r, Type::DOUBLE);
1625
1626 // Flatten arg so we need only 1 test
1627 Node *abs = _gvn.transform(new AbsDNode(arg));
1628 // Node for PI/4 constant
1629 Node *pi4 = makecon(TypeD::make(pi_4));
1630 // Check PI/4 : abs(arg)
1631 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1632 // Check: If PI/4 < abs(arg) then go slow
1633 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1634 // Branch either way
1635 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1636 set_control(opt_iff(r,iff));
1637
1638 // Set fast path result
1639 phi->init_req(2, n);
1640
1641 // Slow path - non-blocking leaf call
1642 Node* call = NULL;
1643 switch (id) {
1644 case vmIntrinsics::_dsin:
1645 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1646 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1647 "Sin", NULL, arg, top());
1648 break;
1649 case vmIntrinsics::_dcos:
1650 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1651 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1652 "Cos", NULL, arg, top());
1653 break;
1654 case vmIntrinsics::_dtan:
1655 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1656 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1657 "Tan", NULL, arg, top());
1658 break;
1659 }
1660 assert(control()->in(0) == call, "");
1661 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1662 r->init_req(1, control());
1663 phi->init_req(1, slow_result);
1664
1665 // Post-merge
1666 set_control(_gvn.transform(r));
1667 record_for_igvn(r);
1668 n = _gvn.transform(phi);
1669
1670 C->set_has_split_ifs(true); // Has chance for split-if optimization
1671 }
1672 set_result(n);
1673 return true;
1674 }
1675
1676 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1677 //-------------------
1678 //result=(result.isNaN())? funcAddr():result;
1679 // Check: If isNaN() by checking result!=result? then either trap
1680 // or go to runtime
1681 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1682 // Build the boolean node
1683 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1684
1685 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1686 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1687 // The pow or exp intrinsic returned a NaN, which requires a call
1688 // to the runtime. Recompile with the runtime call.
1689 uncommon_trap(Deoptimization::Reason_intrinsic,
1690 Deoptimization::Action_make_not_entrant);
1691 }
1692 return result;
1693 } else {
1694 // If this inlining ever returned NaN in the past, we compile a call
1695 // to the runtime to properly handle corner cases
1696
1697 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1698 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1699 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1700
1701 if (!if_slow->is_top()) {
1702 RegionNode* result_region = new RegionNode(3);
1703 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1704
1705 result_region->init_req(1, if_fast);
1706 result_val->init_req(1, result);
1707
1708 set_control(if_slow);
1709
1710 const TypePtr* no_memory_effects = NULL;
1711 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1712 no_memory_effects,
1713 x, top(), y, y ? top() : NULL);
1714 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1715 #ifdef ASSERT
1716 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1717 assert(value_top == top(), "second value must be top");
1718 #endif
1719
1720 result_region->init_req(2, control());
1721 result_val->init_req(2, value);
1722 set_control(_gvn.transform(result_region));
1723 return _gvn.transform(result_val);
1724 } else {
1725 return result;
1726 }
1727 }
1728 }
1729
1730 //------------------------------inline_exp-------------------------------------
1731 // Inline exp instructions, if possible. The Intel hardware only misses
1732 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1733 bool LibraryCallKit::inline_exp() {
1734 Node* arg = round_double_node(argument(0));
1735 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1736
1737 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1738 set_result(n);
1739
1740 C->set_has_split_ifs(true); // Has chance for split-if optimization
1741 return true;
1742 }
1743
1744 //------------------------------inline_pow-------------------------------------
1745 // Inline power instructions, if possible.
1746 bool LibraryCallKit::inline_pow() {
1747 // Pseudocode for pow
1748 // if (y == 2) {
1749 // return x * x;
1750 // } else {
1751 // if (x <= 0.0) {
1752 // long longy = (long)y;
1753 // if ((double)longy == y) { // if y is long
1754 // if (y + 1 == y) longy = 0; // huge number: even
1755 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1756 // } else {
1757 // result = NaN;
1758 // }
1759 // } else {
1760 // result = DPow(x,y);
1761 // }
1762 // if (result != result)? {
1763 // result = uncommon_trap() or runtime_call();
1764 // }
1765 // return result;
1766 // }
1767
1768 Node* x = round_double_node(argument(0));
1769 Node* y = round_double_node(argument(2));
1770
1771 Node* result = NULL;
1772
1773 Node* const_two_node = makecon(TypeD::make(2.0));
1774 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1775 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1776 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1777 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1778 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1779
1780 RegionNode* region_node = new RegionNode(3);
1781 region_node->init_req(1, if_true);
1782
1783 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1784 // special case for x^y where y == 2, we can convert it to x * x
1785 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1786
1787 // set control to if_false since we will now process the false branch
1788 set_control(if_false);
1789
1790 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1791 // Short form: skip the fancy tests and just check for NaN result.
1792 result = _gvn.transform(new PowDNode(C, control(), x, y));
1793 } else {
1794 // If this inlining ever returned NaN in the past, include all
1795 // checks + call to the runtime.
1796
1797 // Set the merge point for If node with condition of (x <= 0.0)
1798 // There are four possible paths to region node and phi node
1799 RegionNode *r = new RegionNode(4);
1800 Node *phi = new PhiNode(r, Type::DOUBLE);
1801
1802 // Build the first if node: if (x <= 0.0)
1803 // Node for 0 constant
1804 Node *zeronode = makecon(TypeD::ZERO);
1805 // Check x:0
1806 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1807 // Check: If (x<=0) then go complex path
1808 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1809 // Branch either way
1810 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1811 // Fast path taken; set region slot 3
1812 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1813 r->init_req(3,fast_taken); // Capture fast-control
1814
1815 // Fast path not-taken, i.e. slow path
1816 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1817
1818 // Set fast path result
1819 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1820 phi->init_req(3, fast_result);
1821
1822 // Complex path
1823 // Build the second if node (if y is long)
1824 // Node for (long)y
1825 Node *longy = _gvn.transform(new ConvD2LNode(y));
1826 // Node for (double)((long) y)
1827 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1828 // Check (double)((long) y) : y
1829 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1830 // Check if (y isn't long) then go to slow path
1831
1832 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1833 // Branch either way
1834 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1835 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1836
1837 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1838
1839 // Calculate DPow(abs(x), y)*(1 & (long)y)
1840 // Node for constant 1
1841 Node *conone = longcon(1);
1842 // 1& (long)y
1843 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1844
1845 // A huge number is always even. Detect a huge number by checking
1846 // if y + 1 == y and set integer to be tested for parity to 0.
1847 // Required for corner case:
1848 // (long)9.223372036854776E18 = max_jlong
1849 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1850 // max_jlong is odd but 9.223372036854776E18 is even
1851 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1852 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1853 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1854 Node* correctedsign = NULL;
1855 if (ConditionalMoveLimit != 0) {
1856 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1857 } else {
1858 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1859 RegionNode *r = new RegionNode(3);
1860 Node *phi = new PhiNode(r, TypeLong::LONG);
1861 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1862 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1863 phi->init_req(1, signnode);
1864 phi->init_req(2, longcon(0));
1865 correctedsign = _gvn.transform(phi);
1866 ylong_path = _gvn.transform(r);
1867 record_for_igvn(r);
1868 }
1869
1870 // zero node
1871 Node *conzero = longcon(0);
1872 // Check (1&(long)y)==0?
1873 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1874 // Check if (1&(long)y)!=0?, if so the result is negative
1875 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1876 // abs(x)
1877 Node *absx=_gvn.transform(new AbsDNode(x));
1878 // abs(x)^y
1879 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1880 // -abs(x)^y
1881 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1882 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1883 Node *signresult = NULL;
1884 if (ConditionalMoveLimit != 0) {
1885 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1886 } else {
1887 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1888 RegionNode *r = new RegionNode(3);
1889 Node *phi = new PhiNode(r, Type::DOUBLE);
1890 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1891 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1892 phi->init_req(1, absxpowy);
1893 phi->init_req(2, negabsxpowy);
1894 signresult = _gvn.transform(phi);
1895 ylong_path = _gvn.transform(r);
1896 record_for_igvn(r);
1897 }
1898 // Set complex path fast result
1899 r->init_req(2, ylong_path);
1900 phi->init_req(2, signresult);
1901
1902 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1903 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1904 r->init_req(1,slow_path);
1905 phi->init_req(1,slow_result);
1906
1907 // Post merge
1908 set_control(_gvn.transform(r));
1909 record_for_igvn(r);
1910 result = _gvn.transform(phi);
1911 }
1912
1913 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1914
1915 // control from finish_pow_exp is now input to the region node
1916 region_node->set_req(2, control());
1917 // the result from finish_pow_exp is now input to the phi node
1918 phi_node->init_req(2, result);
1919 set_control(_gvn.transform(region_node));
1920 record_for_igvn(region_node);
1921 set_result(_gvn.transform(phi_node));
1922
1923 C->set_has_split_ifs(true); // Has chance for split-if optimization
1924 return true;
1925 }
1926
1927 //------------------------------runtime_math-----------------------------
1928 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1929 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1930 "must be (DD)D or (D)D type");
1931
1932 // Inputs
1933 Node* a = round_double_node(argument(0));
1934 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1935
1936 const TypePtr* no_memory_effects = NULL;
1937 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1938 no_memory_effects,
1939 a, top(), b, b ? top() : NULL);
1940 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1941 #ifdef ASSERT
1942 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1943 assert(value_top == top(), "second value must be top");
1944 #endif
1945
1946 set_result(value);
1947 return true;
1948 }
1949
1950 //------------------------------inline_math_native-----------------------------
1951 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1952 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1953 switch (id) {
1954 // These intrinsics are not properly supported on all hardware
1955 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1956 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1957 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1958 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1959 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1960 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1961
1962 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1963 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1964 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1965 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1966
1967 // These intrinsics are supported on all hardware
1968 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1969 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1970
1971 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1972 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1973 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1974 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1975 #undef FN_PTR
1976
1977 // These intrinsics are not yet correctly implemented
1978 case vmIntrinsics::_datan2:
1979 return false;
1980
1981 default:
1982 fatal_unexpected_iid(id);
1983 return false;
1984 }
1985 }
1986
1987 static bool is_simple_name(Node* n) {
1988 return (n->req() == 1 // constant
1989 || (n->is_Type() && n->as_Type()->type()->singleton())
1990 || n->is_Proj() // parameter or return value
1991 || n->is_Phi() // local of some sort
1992 );
1993 }
1994
1995 //----------------------------inline_min_max-----------------------------------
1996 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1997 set_result(generate_min_max(id, argument(0), argument(1)));
1998 return true;
1999 }
2000
2001 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2002 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2003 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2004 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2005 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2006
2007 {
2008 PreserveJVMState pjvms(this);
2009 PreserveReexecuteState preexecs(this);
2010 jvms()->set_should_reexecute(true);
2011
2012 set_control(slow_path);
2013 set_i_o(i_o());
2014
2015 uncommon_trap(Deoptimization::Reason_intrinsic,
2016 Deoptimization::Action_none);
2017 }
2018
2019 set_control(fast_path);
2020 set_result(math);
2021 }
2022
2023 template <typename OverflowOp>
2024 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2025 typedef typename OverflowOp::MathOp MathOp;
2026
2027 MathOp* mathOp = new MathOp(arg1, arg2);
2028 Node* operation = _gvn.transform( mathOp );
2029 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2030 inline_math_mathExact(operation, ofcheck);
2031 return true;
2032 }
2033
2034 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2035 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2036 }
2037
2038 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2039 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2040 }
2041
2042 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2043 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2044 }
2045
2046 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2047 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2048 }
2049
2050 bool LibraryCallKit::inline_math_negateExactI() {
2051 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2052 }
2053
2054 bool LibraryCallKit::inline_math_negateExactL() {
2055 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2056 }
2057
2058 bool LibraryCallKit::inline_math_multiplyExactI() {
2059 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2060 }
2061
2062 bool LibraryCallKit::inline_math_multiplyExactL() {
2063 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2064 }
2065
2066 Node*
2067 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2068 // These are the candidate return value:
2069 Node* xvalue = x0;
2070 Node* yvalue = y0;
2071
2072 if (xvalue == yvalue) {
2073 return xvalue;
2074 }
2075
2076 bool want_max = (id == vmIntrinsics::_max);
2077
2078 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2079 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2080 if (txvalue == NULL || tyvalue == NULL) return top();
2081 // This is not really necessary, but it is consistent with a
2082 // hypothetical MaxINode::Value method:
2083 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2084
2085 // %%% This folding logic should (ideally) be in a different place.
2086 // Some should be inside IfNode, and there to be a more reliable
2087 // transformation of ?: style patterns into cmoves. We also want
2088 // more powerful optimizations around cmove and min/max.
2089
2090 // Try to find a dominating comparison of these guys.
2091 // It can simplify the index computation for Arrays.copyOf
2092 // and similar uses of System.arraycopy.
2093 // First, compute the normalized version of CmpI(x, y).
2094 int cmp_op = Op_CmpI;
2095 Node* xkey = xvalue;
2096 Node* ykey = yvalue;
2097 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2098 if (ideal_cmpxy->is_Cmp()) {
2099 // E.g., if we have CmpI(length - offset, count),
2100 // it might idealize to CmpI(length, count + offset)
2101 cmp_op = ideal_cmpxy->Opcode();
2102 xkey = ideal_cmpxy->in(1);
2103 ykey = ideal_cmpxy->in(2);
2104 }
2105
2106 // Start by locating any relevant comparisons.
2107 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2108 Node* cmpxy = NULL;
2109 Node* cmpyx = NULL;
2110 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2111 Node* cmp = start_from->fast_out(k);
2112 if (cmp->outcnt() > 0 && // must have prior uses
2113 cmp->in(0) == NULL && // must be context-independent
2114 cmp->Opcode() == cmp_op) { // right kind of compare
2115 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2116 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2117 }
2118 }
2119
2120 const int NCMPS = 2;
2121 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2122 int cmpn;
2123 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2124 if (cmps[cmpn] != NULL) break; // find a result
2125 }
2126 if (cmpn < NCMPS) {
2127 // Look for a dominating test that tells us the min and max.
2128 int depth = 0; // Limit search depth for speed
2129 Node* dom = control();
2130 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2131 if (++depth >= 100) break;
2132 Node* ifproj = dom;
2133 if (!ifproj->is_Proj()) continue;
2134 Node* iff = ifproj->in(0);
2135 if (!iff->is_If()) continue;
2136 Node* bol = iff->in(1);
2137 if (!bol->is_Bool()) continue;
2138 Node* cmp = bol->in(1);
2139 if (cmp == NULL) continue;
2140 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2141 if (cmps[cmpn] == cmp) break;
2142 if (cmpn == NCMPS) continue;
2143 BoolTest::mask btest = bol->as_Bool()->_test._test;
2144 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2145 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2146 // At this point, we know that 'x btest y' is true.
2147 switch (btest) {
2148 case BoolTest::eq:
2149 // They are proven equal, so we can collapse the min/max.
2150 // Either value is the answer. Choose the simpler.
2151 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2152 return yvalue;
2153 return xvalue;
2154 case BoolTest::lt: // x < y
2155 case BoolTest::le: // x <= y
2156 return (want_max ? yvalue : xvalue);
2157 case BoolTest::gt: // x > y
2158 case BoolTest::ge: // x >= y
2159 return (want_max ? xvalue : yvalue);
2160 }
2161 }
2162 }
2163
2164 // We failed to find a dominating test.
2165 // Let's pick a test that might GVN with prior tests.
2166 Node* best_bol = NULL;
2167 BoolTest::mask best_btest = BoolTest::illegal;
2168 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2169 Node* cmp = cmps[cmpn];
2170 if (cmp == NULL) continue;
2171 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2172 Node* bol = cmp->fast_out(j);
2173 if (!bol->is_Bool()) continue;
2174 BoolTest::mask btest = bol->as_Bool()->_test._test;
2175 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2176 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2177 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2178 best_bol = bol->as_Bool();
2179 best_btest = btest;
2180 }
2181 }
2182 }
2183
2184 Node* answer_if_true = NULL;
2185 Node* answer_if_false = NULL;
2186 switch (best_btest) {
2187 default:
2188 if (cmpxy == NULL)
2189 cmpxy = ideal_cmpxy;
2190 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2191 // and fall through:
2192 case BoolTest::lt: // x < y
2193 case BoolTest::le: // x <= y
2194 answer_if_true = (want_max ? yvalue : xvalue);
2195 answer_if_false = (want_max ? xvalue : yvalue);
2196 break;
2197 case BoolTest::gt: // x > y
2198 case BoolTest::ge: // x >= y
2199 answer_if_true = (want_max ? xvalue : yvalue);
2200 answer_if_false = (want_max ? yvalue : xvalue);
2201 break;
2202 }
2203
2204 jint hi, lo;
2205 if (want_max) {
2206 // We can sharpen the minimum.
2207 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2208 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2209 } else {
2210 // We can sharpen the maximum.
2211 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2212 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2213 }
2214
2215 // Use a flow-free graph structure, to avoid creating excess control edges
2216 // which could hinder other optimizations.
2217 // Since Math.min/max is often used with arraycopy, we want
2218 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2219 Node* cmov = CMoveNode::make(NULL, best_bol,
2220 answer_if_false, answer_if_true,
2221 TypeInt::make(lo, hi, widen));
2222
2223 return _gvn.transform(cmov);
2224
2225 /*
2226 // This is not as desirable as it may seem, since Min and Max
2227 // nodes do not have a full set of optimizations.
2228 // And they would interfere, anyway, with 'if' optimizations
2229 // and with CMoveI canonical forms.
2230 switch (id) {
2231 case vmIntrinsics::_min:
2232 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2233 case vmIntrinsics::_max:
2234 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2235 default:
2236 ShouldNotReachHere();
2237 }
2238 */
2239 }
2240
2241 inline int
2242 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2243 const TypePtr* base_type = TypePtr::NULL_PTR;
2244 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2245 if (base_type == NULL) {
2246 // Unknown type.
2247 return Type::AnyPtr;
2248 } else if (base_type == TypePtr::NULL_PTR) {
2249 // Since this is a NULL+long form, we have to switch to a rawptr.
2250 base = _gvn.transform(new CastX2PNode(offset));
2251 offset = MakeConX(0);
2252 return Type::RawPtr;
2253 } else if (base_type->base() == Type::RawPtr) {
2254 return Type::RawPtr;
2255 } else if (base_type->isa_oopptr()) {
2256 // Base is never null => always a heap address.
2257 if (base_type->ptr() == TypePtr::NotNull) {
2258 return Type::OopPtr;
2259 }
2260 // Offset is small => always a heap address.
2261 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2262 if (offset_type != NULL &&
2263 base_type->offset() == 0 && // (should always be?)
2264 offset_type->_lo >= 0 &&
2265 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2266 return Type::OopPtr;
2267 }
2268 // Otherwise, it might either be oop+off or NULL+addr.
2269 return Type::AnyPtr;
2270 } else {
2271 // No information:
2272 return Type::AnyPtr;
2273 }
2274 }
2275
2276 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2277 int kind = classify_unsafe_addr(base, offset);
2278 if (kind == Type::RawPtr) {
2279 return basic_plus_adr(top(), base, offset);
2280 } else {
2281 return basic_plus_adr(base, offset);
2282 }
2283 }
2284
2285 //--------------------------inline_number_methods-----------------------------
2286 // inline int Integer.numberOfLeadingZeros(int)
2287 // inline int Long.numberOfLeadingZeros(long)
2288 //
2289 // inline int Integer.numberOfTrailingZeros(int)
2290 // inline int Long.numberOfTrailingZeros(long)
2291 //
2292 // inline int Integer.bitCount(int)
2293 // inline int Long.bitCount(long)
2294 //
2295 // inline char Character.reverseBytes(char)
2296 // inline short Short.reverseBytes(short)
2297 // inline int Integer.reverseBytes(int)
2298 // inline long Long.reverseBytes(long)
2299 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2300 Node* arg = argument(0);
2301 Node* n;
2302 switch (id) {
2303 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2304 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2305 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2306 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2307 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2308 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2309 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2310 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2311 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2312 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2313 default: fatal_unexpected_iid(id); break;
2314 }
2315 set_result(_gvn.transform(n));
2316 return true;
2317 }
2318
2319 //----------------------------inline_unsafe_access----------------------------
2320
2321 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2322
2323 // Helper that guards and inserts a pre-barrier.
2324 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2325 Node* pre_val, bool need_mem_bar) {
2326 // We could be accessing the referent field of a reference object. If so, when G1
2327 // is enabled, we need to log the value in the referent field in an SATB buffer.
2328 // This routine performs some compile time filters and generates suitable
2329 // runtime filters that guard the pre-barrier code.
2330 // Also add memory barrier for non volatile load from the referent field
2331 // to prevent commoning of loads across safepoint.
2332 if (!UseG1GC && !need_mem_bar)
2333 return;
2334
2335 // Some compile time checks.
2336
2337 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2338 const TypeX* otype = offset->find_intptr_t_type();
2339 if (otype != NULL && otype->is_con() &&
2340 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2341 // Constant offset but not the reference_offset so just return
2342 return;
2343 }
2344
2345 // We only need to generate the runtime guards for instances.
2346 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2347 if (btype != NULL) {
2348 if (btype->isa_aryptr()) {
2349 // Array type so nothing to do
2350 return;
2351 }
2352
2353 const TypeInstPtr* itype = btype->isa_instptr();
2354 if (itype != NULL) {
2355 // Can the klass of base_oop be statically determined to be
2356 // _not_ a sub-class of Reference and _not_ Object?
2357 ciKlass* klass = itype->klass();
2358 if ( klass->is_loaded() &&
2359 !klass->is_subtype_of(env()->Reference_klass()) &&
2360 !env()->Object_klass()->is_subtype_of(klass)) {
2361 return;
2362 }
2363 }
2364 }
2365
2366 // The compile time filters did not reject base_oop/offset so
2367 // we need to generate the following runtime filters
2368 //
2369 // if (offset == java_lang_ref_Reference::_reference_offset) {
2370 // if (instance_of(base, java.lang.ref.Reference)) {
2371 // pre_barrier(_, pre_val, ...);
2372 // }
2373 // }
2374
2375 float likely = PROB_LIKELY( 0.999);
2376 float unlikely = PROB_UNLIKELY(0.999);
2377
2378 IdealKit ideal(this);
2379 #define __ ideal.
2380
2381 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2382
2383 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2384 // Update graphKit memory and control from IdealKit.
2385 sync_kit(ideal);
2386
2387 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2388 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2389
2390 // Update IdealKit memory and control from graphKit.
2391 __ sync_kit(this);
2392
2393 Node* one = __ ConI(1);
2394 // is_instof == 0 if base_oop == NULL
2395 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2396
2397 // Update graphKit from IdeakKit.
2398 sync_kit(ideal);
2399
2400 // Use the pre-barrier to record the value in the referent field
2401 pre_barrier(false /* do_load */,
2402 __ ctrl(),
2403 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2404 pre_val /* pre_val */,
2405 T_OBJECT);
2406 if (need_mem_bar) {
2407 // Add memory barrier to prevent commoning reads from this field
2408 // across safepoint since GC can change its value.
2409 insert_mem_bar(Op_MemBarCPUOrder);
2410 }
2411 // Update IdealKit from graphKit.
2412 __ sync_kit(this);
2413
2414 } __ end_if(); // _ref_type != ref_none
2415 } __ end_if(); // offset == referent_offset
2416
2417 // Final sync IdealKit and GraphKit.
2418 final_sync(ideal);
2419 #undef __
2420 }
2421
2422
2423 // Interpret Unsafe.fieldOffset cookies correctly:
2424 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2425
2426 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2427 // Attempt to infer a sharper value type from the offset and base type.
2428 ciKlass* sharpened_klass = NULL;
2429
2430 // See if it is an instance field, with an object type.
2431 if (alias_type->field() != NULL) {
2432 assert(!is_native_ptr, "native pointer op cannot use a java address");
2433 if (alias_type->field()->type()->is_klass()) {
2434 sharpened_klass = alias_type->field()->type()->as_klass();
2435 }
2436 }
2437
2438 // See if it is a narrow oop array.
2439 if (adr_type->isa_aryptr()) {
2440 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2441 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2442 if (elem_type != NULL) {
2443 sharpened_klass = elem_type->klass();
2444 }
2445 }
2446 }
2447
2448 // The sharpened class might be unloaded if there is no class loader
2449 // contraint in place.
2450 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2451 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2452
2453 #ifndef PRODUCT
2454 if (C->print_intrinsics() || C->print_inlining()) {
2455 tty->print(" from base type: "); adr_type->dump();
2456 tty->print(" sharpened value: "); tjp->dump();
2457 }
2458 #endif
2459 // Sharpen the value type.
2460 return tjp;
2461 }
2462 return NULL;
2463 }
2464
2465 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2466 if (callee()->is_static()) return false; // caller must have the capability!
2467
2468 #ifndef PRODUCT
2469 {
2470 ResourceMark rm;
2471 // Check the signatures.
2472 ciSignature* sig = callee()->signature();
2473 #ifdef ASSERT
2474 if (!is_store) {
2475 // Object getObject(Object base, int/long offset), etc.
2476 BasicType rtype = sig->return_type()->basic_type();
2477 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2478 rtype = T_ADDRESS; // it is really a C void*
2479 assert(rtype == type, "getter must return the expected value");
2480 if (!is_native_ptr) {
2481 assert(sig->count() == 2, "oop getter has 2 arguments");
2482 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2483 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2484 } else {
2485 assert(sig->count() == 1, "native getter has 1 argument");
2486 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2487 }
2488 } else {
2489 // void putObject(Object base, int/long offset, Object x), etc.
2490 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2491 if (!is_native_ptr) {
2492 assert(sig->count() == 3, "oop putter has 3 arguments");
2493 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2494 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2495 } else {
2496 assert(sig->count() == 2, "native putter has 2 arguments");
2497 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2498 }
2499 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2500 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2501 vtype = T_ADDRESS; // it is really a C void*
2502 assert(vtype == type, "putter must accept the expected value");
2503 }
2504 #endif // ASSERT
2505 }
2506 #endif //PRODUCT
2507
2508 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2509
2510 Node* receiver = argument(0); // type: oop
2511
2512 // Build address expression. See the code in inline_unsafe_prefetch.
2513 Node* adr;
2514 Node* heap_base_oop = top();
2515 Node* offset = top();
2516 Node* val;
2517
2518 if (!is_native_ptr) {
2519 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2520 Node* base = argument(1); // type: oop
2521 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2522 offset = argument(2); // type: long
2523 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2524 // to be plain byte offsets, which are also the same as those accepted
2525 // by oopDesc::field_base.
2526 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2527 "fieldOffset must be byte-scaled");
2528 // 32-bit machines ignore the high half!
2529 offset = ConvL2X(offset);
2530 adr = make_unsafe_address(base, offset);
2531 heap_base_oop = base;
2532 val = is_store ? argument(4) : NULL;
2533 } else {
2534 Node* ptr = argument(1); // type: long
2535 ptr = ConvL2X(ptr); // adjust Java long to machine word
2536 adr = make_unsafe_address(NULL, ptr);
2537 val = is_store ? argument(3) : NULL;
2538 }
2539
2540 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2541
2542 // First guess at the value type.
2543 const Type *value_type = Type::get_const_basic_type(type);
2544
2545 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2546 // there was not enough information to nail it down.
2547 Compile::AliasType* alias_type = C->alias_type(adr_type);
2548 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2549
2550 // We will need memory barriers unless we can determine a unique
2551 // alias category for this reference. (Note: If for some reason
2552 // the barriers get omitted and the unsafe reference begins to "pollute"
2553 // the alias analysis of the rest of the graph, either Compile::can_alias
2554 // or Compile::must_alias will throw a diagnostic assert.)
2555 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2556
2557 // If we are reading the value of the referent field of a Reference
2558 // object (either by using Unsafe directly or through reflection)
2559 // then, if G1 is enabled, we need to record the referent in an
2560 // SATB log buffer using the pre-barrier mechanism.
2561 // Also we need to add memory barrier to prevent commoning reads
2562 // from this field across safepoint since GC can change its value.
2563 bool need_read_barrier = !is_native_ptr && !is_store &&
2564 offset != top() && heap_base_oop != top();
2565
2566 if (!is_store && type == T_OBJECT) {
2567 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2568 if (tjp != NULL) {
2569 value_type = tjp;
2570 }
2571 }
2572
2573 receiver = null_check(receiver);
2574 if (stopped()) {
2575 return true;
2576 }
2577 // Heap pointers get a null-check from the interpreter,
2578 // as a courtesy. However, this is not guaranteed by Unsafe,
2579 // and it is not possible to fully distinguish unintended nulls
2580 // from intended ones in this API.
2581
2582 if (is_volatile) {
2583 // We need to emit leading and trailing CPU membars (see below) in
2584 // addition to memory membars when is_volatile. This is a little
2585 // too strong, but avoids the need to insert per-alias-type
2586 // volatile membars (for stores; compare Parse::do_put_xxx), which
2587 // we cannot do effectively here because we probably only have a
2588 // rough approximation of type.
2589 need_mem_bar = true;
2590 // For Stores, place a memory ordering barrier now.
2591 if (UseBarriersForVolatile) {
2592 if (is_store) {
2593 insert_mem_bar(Op_MemBarRelease);
2594 } else {
2595 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2596 insert_mem_bar(Op_MemBarVolatile);
2597 }
2598 }
2599 }
2600 }
2601
2602 // Memory barrier to prevent normal and 'unsafe' accesses from
2603 // bypassing each other. Happens after null checks, so the
2604 // exception paths do not take memory state from the memory barrier,
2605 // so there's no problems making a strong assert about mixing users
2606 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2607 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2608 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2609
2610 if (!is_store) {
2611 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2612 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2613 // load value
2614 switch (type) {
2615 case T_BOOLEAN:
2616 case T_CHAR:
2617 case T_BYTE:
2618 case T_SHORT:
2619 case T_INT:
2620 case T_LONG:
2621 case T_FLOAT:
2622 case T_DOUBLE:
2623 break;
2624 case T_OBJECT:
2625 if (need_read_barrier) {
2626 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2627 }
2628 break;
2629 case T_ADDRESS:
2630 // Cast to an int type.
2631 p = _gvn.transform(new CastP2XNode(NULL, p));
2632 p = ConvX2UL(p);
2633 break;
2634 default:
2635 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2636 break;
2637 }
2638 // The load node has the control of the preceding MemBarCPUOrder. All
2639 // following nodes will have the control of the MemBarCPUOrder inserted at
2640 // the end of this method. So, pushing the load onto the stack at a later
2641 // point is fine.
2642 set_result(p);
2643 } else {
2644 // place effect of store into memory
2645 switch (type) {
2646 case T_DOUBLE:
2647 val = dstore_rounding(val);
2648 break;
2649 case T_ADDRESS:
2650 // Repackage the long as a pointer.
2651 val = ConvL2X(val);
2652 val = _gvn.transform(new CastX2PNode(val));
2653 break;
2654 }
2655
2656 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2657 if (type != T_OBJECT ) {
2658 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2659 } else {
2660 // Possibly an oop being stored to Java heap or native memory
2661 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2662 // oop to Java heap.
2663 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2664 } else {
2665 // We can't tell at compile time if we are storing in the Java heap or outside
2666 // of it. So we need to emit code to conditionally do the proper type of
2667 // store.
2668
2669 IdealKit ideal(this);
2670 #define __ ideal.
2671 // QQQ who knows what probability is here??
2672 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2673 // Sync IdealKit and graphKit.
2674 sync_kit(ideal);
2675 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2676 // Update IdealKit memory.
2677 __ sync_kit(this);
2678 } __ else_(); {
2679 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2680 } __ end_if();
2681 // Final sync IdealKit and GraphKit.
2682 final_sync(ideal);
2683 #undef __
2684 }
2685 }
2686 }
2687
2688 if (is_volatile && UseBarriersForVolatile) {
2689 if (!is_store) {
2690 insert_mem_bar(Op_MemBarAcquire);
2691 } else {
2692 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2693 insert_mem_bar(Op_MemBarVolatile);
2694 }
2695 }
2696 }
2697 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2698
2699 return true;
2700 }
2701
2702 //----------------------------inline_unsafe_prefetch----------------------------
2703
2704 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2705 #ifndef PRODUCT
2706 {
2707 ResourceMark rm;
2708 // Check the signatures.
2709 ciSignature* sig = callee()->signature();
2710 #ifdef ASSERT
2711 // Object getObject(Object base, int/long offset), etc.
2712 BasicType rtype = sig->return_type()->basic_type();
2713 if (!is_native_ptr) {
2714 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2715 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2716 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2717 } else {
2718 assert(sig->count() == 1, "native prefetch has 1 argument");
2719 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2720 }
2721 #endif // ASSERT
2722 }
2723 #endif // !PRODUCT
2724
2725 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2726
2727 const int idx = is_static ? 0 : 1;
2728 if (!is_static) {
2729 null_check_receiver();
2730 if (stopped()) {
2731 return true;
2732 }
2733 }
2734
2735 // Build address expression. See the code in inline_unsafe_access.
2736 Node *adr;
2737 if (!is_native_ptr) {
2738 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2739 Node* base = argument(idx + 0); // type: oop
2740 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2741 Node* offset = argument(idx + 1); // type: long
2742 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2743 // to be plain byte offsets, which are also the same as those accepted
2744 // by oopDesc::field_base.
2745 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2746 "fieldOffset must be byte-scaled");
2747 // 32-bit machines ignore the high half!
2748 offset = ConvL2X(offset);
2749 adr = make_unsafe_address(base, offset);
2750 } else {
2751 Node* ptr = argument(idx + 0); // type: long
2752 ptr = ConvL2X(ptr); // adjust Java long to machine word
2753 adr = make_unsafe_address(NULL, ptr);
2754 }
2755
2756 // Generate the read or write prefetch
2757 Node *prefetch;
2758 if (is_store) {
2759 prefetch = new PrefetchWriteNode(i_o(), adr);
2760 } else {
2761 prefetch = new PrefetchReadNode(i_o(), adr);
2762 }
2763 prefetch->init_req(0, control());
2764 set_i_o(_gvn.transform(prefetch));
2765
2766 return true;
2767 }
2768
2769 //----------------------------inline_unsafe_load_store----------------------------
2770 // This method serves a couple of different customers (depending on LoadStoreKind):
2771 //
2772 // LS_cmpxchg:
2773 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2774 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2775 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2776 //
2777 // LS_xadd:
2778 // public int getAndAddInt( Object o, long offset, int delta)
2779 // public long getAndAddLong(Object o, long offset, long delta)
2780 //
2781 // LS_xchg:
2782 // int getAndSet(Object o, long offset, int newValue)
2783 // long getAndSet(Object o, long offset, long newValue)
2784 // Object getAndSet(Object o, long offset, Object newValue)
2785 //
2786 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2787 // This basic scheme here is the same as inline_unsafe_access, but
2788 // differs in enough details that combining them would make the code
2789 // overly confusing. (This is a true fact! I originally combined
2790 // them, but even I was confused by it!) As much code/comments as
2791 // possible are retained from inline_unsafe_access though to make
2792 // the correspondences clearer. - dl
2793
2794 if (callee()->is_static()) return false; // caller must have the capability!
2795
2796 #ifndef PRODUCT
2797 BasicType rtype;
2798 {
2799 ResourceMark rm;
2800 // Check the signatures.
2801 ciSignature* sig = callee()->signature();
2802 rtype = sig->return_type()->basic_type();
2803 if (kind == LS_xadd || kind == LS_xchg) {
2804 // Check the signatures.
2805 #ifdef ASSERT
2806 assert(rtype == type, "get and set must return the expected type");
2807 assert(sig->count() == 3, "get and set has 3 arguments");
2808 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2809 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2810 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2811 #endif // ASSERT
2812 } else if (kind == LS_cmpxchg) {
2813 // Check the signatures.
2814 #ifdef ASSERT
2815 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2816 assert(sig->count() == 4, "CAS has 4 arguments");
2817 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2818 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2819 #endif // ASSERT
2820 } else {
2821 ShouldNotReachHere();
2822 }
2823 }
2824 #endif //PRODUCT
2825
2826 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2827
2828 // Get arguments:
2829 Node* receiver = NULL;
2830 Node* base = NULL;
2831 Node* offset = NULL;
2832 Node* oldval = NULL;
2833 Node* newval = NULL;
2834 if (kind == LS_cmpxchg) {
2835 const bool two_slot_type = type2size[type] == 2;
2836 receiver = argument(0); // type: oop
2837 base = argument(1); // type: oop
2838 offset = argument(2); // type: long
2839 oldval = argument(4); // type: oop, int, or long
2840 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2841 } else if (kind == LS_xadd || kind == LS_xchg){
2842 receiver = argument(0); // type: oop
2843 base = argument(1); // type: oop
2844 offset = argument(2); // type: long
2845 oldval = NULL;
2846 newval = argument(4); // type: oop, int, or long
2847 }
2848
2849 // Null check receiver.
2850 receiver = null_check(receiver);
2851 if (stopped()) {
2852 return true;
2853 }
2854
2855 // Build field offset expression.
2856 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2857 // to be plain byte offsets, which are also the same as those accepted
2858 // by oopDesc::field_base.
2859 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2860 // 32-bit machines ignore the high half of long offsets
2861 offset = ConvL2X(offset);
2862 Node* adr = make_unsafe_address(base, offset);
2863 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2864
2865 // For CAS, unlike inline_unsafe_access, there seems no point in
2866 // trying to refine types. Just use the coarse types here.
2867 const Type *value_type = Type::get_const_basic_type(type);
2868 Compile::AliasType* alias_type = C->alias_type(adr_type);
2869 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2870
2871 if (kind == LS_xchg && type == T_OBJECT) {
2872 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2873 if (tjp != NULL) {
2874 value_type = tjp;
2875 }
2876 }
2877
2878 int alias_idx = C->get_alias_index(adr_type);
2879
2880 // Memory-model-wise, a LoadStore acts like a little synchronized
2881 // block, so needs barriers on each side. These don't translate
2882 // into actual barriers on most machines, but we still need rest of
2883 // compiler to respect ordering.
2884
2885 insert_mem_bar(Op_MemBarRelease);
2886 insert_mem_bar(Op_MemBarCPUOrder);
2887
2888 // 4984716: MemBars must be inserted before this
2889 // memory node in order to avoid a false
2890 // dependency which will confuse the scheduler.
2891 Node *mem = memory(alias_idx);
2892
2893 // For now, we handle only those cases that actually exist: ints,
2894 // longs, and Object. Adding others should be straightforward.
2895 Node* load_store;
2896 switch(type) {
2897 case T_INT:
2898 if (kind == LS_xadd) {
2899 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2900 } else if (kind == LS_xchg) {
2901 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2902 } else if (kind == LS_cmpxchg) {
2903 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2904 } else {
2905 ShouldNotReachHere();
2906 }
2907 break;
2908 case T_LONG:
2909 if (kind == LS_xadd) {
2910 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2911 } else if (kind == LS_xchg) {
2912 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2913 } else if (kind == LS_cmpxchg) {
2914 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2915 } else {
2916 ShouldNotReachHere();
2917 }
2918 break;
2919 case T_OBJECT:
2920 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2921 // could be delayed during Parse (for example, in adjust_map_after_if()).
2922 // Execute transformation here to avoid barrier generation in such case.
2923 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2924 newval = _gvn.makecon(TypePtr::NULL_PTR);
2925
2926 // Reference stores need a store barrier.
2927 if (kind == LS_xchg) {
2928 // If pre-barrier must execute before the oop store, old value will require do_load here.
2929 if (!can_move_pre_barrier()) {
2930 pre_barrier(true /* do_load*/,
2931 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2932 NULL /* pre_val*/,
2933 T_OBJECT);
2934 } // Else move pre_barrier to use load_store value, see below.
2935 } else if (kind == LS_cmpxchg) {
2936 // Same as for newval above:
2937 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2938 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2939 }
2940 // The only known value which might get overwritten is oldval.
2941 pre_barrier(false /* do_load */,
2942 control(), NULL, NULL, max_juint, NULL, NULL,
2943 oldval /* pre_val */,
2944 T_OBJECT);
2945 } else {
2946 ShouldNotReachHere();
2947 }
2948
2949 #ifdef _LP64
2950 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2951 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2952 if (kind == LS_xchg) {
2953 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2954 newval_enc, adr_type, value_type->make_narrowoop()));
2955 } else {
2956 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2957 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2958 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2959 newval_enc, oldval_enc));
2960 }
2961 } else
2962 #endif
2963 {
2964 if (kind == LS_xchg) {
2965 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2966 } else {
2967 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2968 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2969 }
2970 }
2971 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2972 break;
2973 default:
2974 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2975 break;
2976 }
2977
2978 // SCMemProjNodes represent the memory state of a LoadStore. Their
2979 // main role is to prevent LoadStore nodes from being optimized away
2980 // when their results aren't used.
2981 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2982 set_memory(proj, alias_idx);
2983
2984 if (type == T_OBJECT && kind == LS_xchg) {
2985 #ifdef _LP64
2986 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2987 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2988 }
2989 #endif
2990 if (can_move_pre_barrier()) {
2991 // Don't need to load pre_val. The old value is returned by load_store.
2992 // The pre_barrier can execute after the xchg as long as no safepoint
2993 // gets inserted between them.
2994 pre_barrier(false /* do_load */,
2995 control(), NULL, NULL, max_juint, NULL, NULL,
2996 load_store /* pre_val */,
2997 T_OBJECT);
2998 }
2999 }
3000
3001 // Add the trailing membar surrounding the access
3002 insert_mem_bar(Op_MemBarCPUOrder);
3003 insert_mem_bar(Op_MemBarAcquire);
3004
3005 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3006 set_result(load_store);
3007 return true;
3008 }
3009
3010 //----------------------------inline_unsafe_ordered_store----------------------
3011 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3012 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3013 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3014 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3015 // This is another variant of inline_unsafe_access, differing in
3016 // that it always issues store-store ("release") barrier and ensures
3017 // store-atomicity (which only matters for "long").
3018
3019 if (callee()->is_static()) return false; // caller must have the capability!
3020
3021 #ifndef PRODUCT
3022 {
3023 ResourceMark rm;
3024 // Check the signatures.
3025 ciSignature* sig = callee()->signature();
3026 #ifdef ASSERT
3027 BasicType rtype = sig->return_type()->basic_type();
3028 assert(rtype == T_VOID, "must return void");
3029 assert(sig->count() == 3, "has 3 arguments");
3030 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3031 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3032 #endif // ASSERT
3033 }
3034 #endif //PRODUCT
3035
3036 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3037
3038 // Get arguments:
3039 Node* receiver = argument(0); // type: oop
3040 Node* base = argument(1); // type: oop
3041 Node* offset = argument(2); // type: long
3042 Node* val = argument(4); // type: oop, int, or long
3043
3044 // Null check receiver.
3045 receiver = null_check(receiver);
3046 if (stopped()) {
3047 return true;
3048 }
3049
3050 // Build field offset expression.
3051 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3052 // 32-bit machines ignore the high half of long offsets
3053 offset = ConvL2X(offset);
3054 Node* adr = make_unsafe_address(base, offset);
3055 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3056 const Type *value_type = Type::get_const_basic_type(type);
3057 Compile::AliasType* alias_type = C->alias_type(adr_type);
3058
3059 insert_mem_bar(Op_MemBarRelease);
3060 insert_mem_bar(Op_MemBarCPUOrder);
3061 // Ensure that the store is atomic for longs:
3062 const bool require_atomic_access = true;
3063 Node* store;
3064 if (type == T_OBJECT) // reference stores need a store barrier.
3065 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3066 else {
3067 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3068 }
3069 insert_mem_bar(Op_MemBarCPUOrder);
3070 return true;
3071 }
3072
3073 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3074 // Regardless of form, don't allow previous ld/st to move down,
3075 // then issue acquire, release, or volatile mem_bar.
3076 insert_mem_bar(Op_MemBarCPUOrder);
3077 switch(id) {
3078 case vmIntrinsics::_loadFence:
3079 insert_mem_bar(Op_LoadFence);
3080 return true;
3081 case vmIntrinsics::_storeFence:
3082 insert_mem_bar(Op_StoreFence);
3083 return true;
3084 case vmIntrinsics::_fullFence:
3085 insert_mem_bar(Op_MemBarVolatile);
3086 return true;
3087 default:
3088 fatal_unexpected_iid(id);
3089 return false;
3090 }
3091 }
3092
3093 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3094 if (!kls->is_Con()) {
3095 return true;
3096 }
3097 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3098 if (klsptr == NULL) {
3099 return true;
3100 }
3101 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3102 // don't need a guard for a klass that is already initialized
3103 return !ik->is_initialized();
3104 }
3105
3106 //----------------------------inline_unsafe_allocate---------------------------
3107 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3108 bool LibraryCallKit::inline_unsafe_allocate() {
3109 if (callee()->is_static()) return false; // caller must have the capability!
3110
3111 null_check_receiver(); // null-check, then ignore
3112 Node* cls = null_check(argument(1));
3113 if (stopped()) return true;
3114
3115 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3116 kls = null_check(kls);
3117 if (stopped()) return true; // argument was like int.class
3118
3119 Node* test = NULL;
3120 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3121 // Note: The argument might still be an illegal value like
3122 // Serializable.class or Object[].class. The runtime will handle it.
3123 // But we must make an explicit check for initialization.
3124 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3125 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3126 // can generate code to load it as unsigned byte.
3127 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3128 Node* bits = intcon(InstanceKlass::fully_initialized);
3129 test = _gvn.transform(new SubINode(inst, bits));
3130 // The 'test' is non-zero if we need to take a slow path.
3131 }
3132
3133 Node* obj = new_instance(kls, test);
3134 set_result(obj);
3135 return true;
3136 }
3137
3138 #ifdef TRACE_HAVE_INTRINSICS
3139 /*
3140 * oop -> myklass
3141 * myklass->trace_id |= USED
3142 * return myklass->trace_id & ~0x3
3143 */
3144 bool LibraryCallKit::inline_native_classID() {
3145 null_check_receiver(); // null-check, then ignore
3146 Node* cls = null_check(argument(1), T_OBJECT);
3147 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3148 kls = null_check(kls, T_OBJECT);
3149 ByteSize offset = TRACE_ID_OFFSET;
3150 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3151 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3152 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3153 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3154 Node* clsused = longcon(0x01l); // set the class bit
3155 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3156
3157 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3158 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3159 set_result(andl);
3160 return true;
3161 }
3162
3163 bool LibraryCallKit::inline_native_threadID() {
3164 Node* tls_ptr = NULL;
3165 Node* cur_thr = generate_current_thread(tls_ptr);
3166 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3167 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3168 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3169
3170 Node* threadid = NULL;
3171 size_t thread_id_size = OSThread::thread_id_size();
3172 if (thread_id_size == (size_t) BytesPerLong) {
3173 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3174 } else if (thread_id_size == (size_t) BytesPerInt) {
3175 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3176 } else {
3177 ShouldNotReachHere();
3178 }
3179 set_result(threadid);
3180 return true;
3181 }
3182 #endif
3183
3184 //------------------------inline_native_time_funcs--------------
3185 // inline code for System.currentTimeMillis() and System.nanoTime()
3186 // these have the same type and signature
3187 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3188 const TypeFunc* tf = OptoRuntime::void_long_Type();
3189 const TypePtr* no_memory_effects = NULL;
3190 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3191 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3192 #ifdef ASSERT
3193 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3194 assert(value_top == top(), "second value must be top");
3195 #endif
3196 set_result(value);
3197 return true;
3198 }
3199
3200 //------------------------inline_native_currentThread------------------
3201 bool LibraryCallKit::inline_native_currentThread() {
3202 Node* junk = NULL;
3203 set_result(generate_current_thread(junk));
3204 return true;
3205 }
3206
3207 //------------------------inline_native_isInterrupted------------------
3208 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3209 bool LibraryCallKit::inline_native_isInterrupted() {
3210 // Add a fast path to t.isInterrupted(clear_int):
3211 // (t == Thread.current() &&
3212 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3213 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3214 // So, in the common case that the interrupt bit is false,
3215 // we avoid making a call into the VM. Even if the interrupt bit
3216 // is true, if the clear_int argument is false, we avoid the VM call.
3217 // However, if the receiver is not currentThread, we must call the VM,
3218 // because there must be some locking done around the operation.
3219
3220 // We only go to the fast case code if we pass two guards.
3221 // Paths which do not pass are accumulated in the slow_region.
3222
3223 enum {
3224 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3225 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3226 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3227 PATH_LIMIT
3228 };
3229
3230 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3231 // out of the function.
3232 insert_mem_bar(Op_MemBarCPUOrder);
3233
3234 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3235 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3236
3237 RegionNode* slow_region = new RegionNode(1);
3238 record_for_igvn(slow_region);
3239
3240 // (a) Receiving thread must be the current thread.
3241 Node* rec_thr = argument(0);
3242 Node* tls_ptr = NULL;
3243 Node* cur_thr = generate_current_thread(tls_ptr);
3244 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3245 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3246
3247 generate_slow_guard(bol_thr, slow_region);
3248
3249 // (b) Interrupt bit on TLS must be false.
3250 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3251 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3252 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3253
3254 // Set the control input on the field _interrupted read to prevent it floating up.
3255 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3256 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3257 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3258
3259 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3260
3261 // First fast path: if (!TLS._interrupted) return false;
3262 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3263 result_rgn->init_req(no_int_result_path, false_bit);
3264 result_val->init_req(no_int_result_path, intcon(0));
3265
3266 // drop through to next case
3267 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3268
3269 #ifndef TARGET_OS_FAMILY_windows
3270 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3271 Node* clr_arg = argument(1);
3272 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3273 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3274 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3275
3276 // Second fast path: ... else if (!clear_int) return true;
3277 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3278 result_rgn->init_req(no_clear_result_path, false_arg);
3279 result_val->init_req(no_clear_result_path, intcon(1));
3280
3281 // drop through to next case
3282 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3283 #else
3284 // To return true on Windows you must read the _interrupted field
3285 // and check the the event state i.e. take the slow path.
3286 #endif // TARGET_OS_FAMILY_windows
3287
3288 // (d) Otherwise, go to the slow path.
3289 slow_region->add_req(control());
3290 set_control( _gvn.transform(slow_region));
3291
3292 if (stopped()) {
3293 // There is no slow path.
3294 result_rgn->init_req(slow_result_path, top());
3295 result_val->init_req(slow_result_path, top());
3296 } else {
3297 // non-virtual because it is a private non-static
3298 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3299
3300 Node* slow_val = set_results_for_java_call(slow_call);
3301 // this->control() comes from set_results_for_java_call
3302
3303 Node* fast_io = slow_call->in(TypeFunc::I_O);
3304 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3305
3306 // These two phis are pre-filled with copies of of the fast IO and Memory
3307 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3308 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3309
3310 result_rgn->init_req(slow_result_path, control());
3311 result_io ->init_req(slow_result_path, i_o());
3312 result_mem->init_req(slow_result_path, reset_memory());
3313 result_val->init_req(slow_result_path, slow_val);
3314
3315 set_all_memory(_gvn.transform(result_mem));
3316 set_i_o( _gvn.transform(result_io));
3317 }
3318
3319 C->set_has_split_ifs(true); // Has chance for split-if optimization
3320 set_result(result_rgn, result_val);
3321 return true;
3322 }
3323
3324 //---------------------------load_mirror_from_klass----------------------------
3325 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3326 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3327 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3328 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3329 }
3330
3331 //-----------------------load_klass_from_mirror_common-------------------------
3332 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3333 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3334 // and branch to the given path on the region.
3335 // If never_see_null, take an uncommon trap on null, so we can optimistically
3336 // compile for the non-null case.
3337 // If the region is NULL, force never_see_null = true.
3338 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3339 bool never_see_null,
3340 RegionNode* region,
3341 int null_path,
3342 int offset) {
3343 if (region == NULL) never_see_null = true;
3344 Node* p = basic_plus_adr(mirror, offset);
3345 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3346 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3347 Node* null_ctl = top();
3348 kls = null_check_oop(kls, &null_ctl, never_see_null);
3349 if (region != NULL) {
3350 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3351 region->init_req(null_path, null_ctl);
3352 } else {
3353 assert(null_ctl == top(), "no loose ends");
3354 }
3355 return kls;
3356 }
3357
3358 //--------------------(inline_native_Class_query helpers)---------------------
3359 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3360 // Fall through if (mods & mask) == bits, take the guard otherwise.
3361 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3362 // Branch around if the given klass has the given modifier bit set.
3363 // Like generate_guard, adds a new path onto the region.
3364 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3365 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3366 Node* mask = intcon(modifier_mask);
3367 Node* bits = intcon(modifier_bits);
3368 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3369 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3370 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3371 return generate_fair_guard(bol, region);
3372 }
3373 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3374 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3375 }
3376
3377 //-------------------------inline_native_Class_query-------------------
3378 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3379 const Type* return_type = TypeInt::BOOL;
3380 Node* prim_return_value = top(); // what happens if it's a primitive class?
3381 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3382 bool expect_prim = false; // most of these guys expect to work on refs
3383
3384 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3385
3386 Node* mirror = argument(0);
3387 Node* obj = top();
3388
3389 switch (id) {
3390 case vmIntrinsics::_isInstance:
3391 // nothing is an instance of a primitive type
3392 prim_return_value = intcon(0);
3393 obj = argument(1);
3394 break;
3395 case vmIntrinsics::_getModifiers:
3396 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3397 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3398 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3399 break;
3400 case vmIntrinsics::_isInterface:
3401 prim_return_value = intcon(0);
3402 break;
3403 case vmIntrinsics::_isArray:
3404 prim_return_value = intcon(0);
3405 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3406 break;
3407 case vmIntrinsics::_isPrimitive:
3408 prim_return_value = intcon(1);
3409 expect_prim = true; // obviously
3410 break;
3411 case vmIntrinsics::_getSuperclass:
3412 prim_return_value = null();
3413 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3414 break;
3415 case vmIntrinsics::_getClassAccessFlags:
3416 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3417 return_type = TypeInt::INT; // not bool! 6297094
3418 break;
3419 default:
3420 fatal_unexpected_iid(id);
3421 break;
3422 }
3423
3424 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3425 if (mirror_con == NULL) return false; // cannot happen?
3426
3427 #ifndef PRODUCT
3428 if (C->print_intrinsics() || C->print_inlining()) {
3429 ciType* k = mirror_con->java_mirror_type();
3430 if (k) {
3431 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3432 k->print_name();
3433 tty->cr();
3434 }
3435 }
3436 #endif
3437
3438 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3439 RegionNode* region = new RegionNode(PATH_LIMIT);
3440 record_for_igvn(region);
3441 PhiNode* phi = new PhiNode(region, return_type);
3442
3443 // The mirror will never be null of Reflection.getClassAccessFlags, however
3444 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3445 // if it is. See bug 4774291.
3446
3447 // For Reflection.getClassAccessFlags(), the null check occurs in
3448 // the wrong place; see inline_unsafe_access(), above, for a similar
3449 // situation.
3450 mirror = null_check(mirror);
3451 // If mirror or obj is dead, only null-path is taken.
3452 if (stopped()) return true;
3453
3454 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3455
3456 // Now load the mirror's klass metaobject, and null-check it.
3457 // Side-effects region with the control path if the klass is null.
3458 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3459 // If kls is null, we have a primitive mirror.
3460 phi->init_req(_prim_path, prim_return_value);
3461 if (stopped()) { set_result(region, phi); return true; }
3462 bool safe_for_replace = (region->in(_prim_path) == top());
3463
3464 Node* p; // handy temp
3465 Node* null_ctl;
3466
3467 // Now that we have the non-null klass, we can perform the real query.
3468 // For constant classes, the query will constant-fold in LoadNode::Value.
3469 Node* query_value = top();
3470 switch (id) {
3471 case vmIntrinsics::_isInstance:
3472 // nothing is an instance of a primitive type
3473 query_value = gen_instanceof(obj, kls, safe_for_replace);
3474 break;
3475
3476 case vmIntrinsics::_getModifiers:
3477 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3478 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3479 break;
3480
3481 case vmIntrinsics::_isInterface:
3482 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3483 if (generate_interface_guard(kls, region) != NULL)
3484 // A guard was added. If the guard is taken, it was an interface.
3485 phi->add_req(intcon(1));
3486 // If we fall through, it's a plain class.
3487 query_value = intcon(0);
3488 break;
3489
3490 case vmIntrinsics::_isArray:
3491 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3492 if (generate_array_guard(kls, region) != NULL)
3493 // A guard was added. If the guard is taken, it was an array.
3494 phi->add_req(intcon(1));
3495 // If we fall through, it's a plain class.
3496 query_value = intcon(0);
3497 break;
3498
3499 case vmIntrinsics::_isPrimitive:
3500 query_value = intcon(0); // "normal" path produces false
3501 break;
3502
3503 case vmIntrinsics::_getSuperclass:
3504 // The rules here are somewhat unfortunate, but we can still do better
3505 // with random logic than with a JNI call.
3506 // Interfaces store null or Object as _super, but must report null.
3507 // Arrays store an intermediate super as _super, but must report Object.
3508 // Other types can report the actual _super.
3509 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3510 if (generate_interface_guard(kls, region) != NULL)
3511 // A guard was added. If the guard is taken, it was an interface.
3512 phi->add_req(null());
3513 if (generate_array_guard(kls, region) != NULL)
3514 // A guard was added. If the guard is taken, it was an array.
3515 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3516 // If we fall through, it's a plain class. Get its _super.
3517 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3518 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3519 null_ctl = top();
3520 kls = null_check_oop(kls, &null_ctl);
3521 if (null_ctl != top()) {
3522 // If the guard is taken, Object.superClass is null (both klass and mirror).
3523 region->add_req(null_ctl);
3524 phi ->add_req(null());
3525 }
3526 if (!stopped()) {
3527 query_value = load_mirror_from_klass(kls);
3528 }
3529 break;
3530
3531 case vmIntrinsics::_getClassAccessFlags:
3532 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3533 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3534 break;
3535
3536 default:
3537 fatal_unexpected_iid(id);
3538 break;
3539 }
3540
3541 // Fall-through is the normal case of a query to a real class.
3542 phi->init_req(1, query_value);
3543 region->init_req(1, control());
3544
3545 C->set_has_split_ifs(true); // Has chance for split-if optimization
3546 set_result(region, phi);
3547 return true;
3548 }
3549
3550 //--------------------------inline_native_subtype_check------------------------
3551 // This intrinsic takes the JNI calls out of the heart of
3552 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3553 bool LibraryCallKit::inline_native_subtype_check() {
3554 // Pull both arguments off the stack.
3555 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3556 args[0] = argument(0);
3557 args[1] = argument(1);
3558 Node* klasses[2]; // corresponding Klasses: superk, subk
3559 klasses[0] = klasses[1] = top();
3560
3561 enum {
3562 // A full decision tree on {superc is prim, subc is prim}:
3563 _prim_0_path = 1, // {P,N} => false
3564 // {P,P} & superc!=subc => false
3565 _prim_same_path, // {P,P} & superc==subc => true
3566 _prim_1_path, // {N,P} => false
3567 _ref_subtype_path, // {N,N} & subtype check wins => true
3568 _both_ref_path, // {N,N} & subtype check loses => false
3569 PATH_LIMIT
3570 };
3571
3572 RegionNode* region = new RegionNode(PATH_LIMIT);
3573 Node* phi = new PhiNode(region, TypeInt::BOOL);
3574 record_for_igvn(region);
3575
3576 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3577 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3578 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3579
3580 // First null-check both mirrors and load each mirror's klass metaobject.
3581 int which_arg;
3582 for (which_arg = 0; which_arg <= 1; which_arg++) {
3583 Node* arg = args[which_arg];
3584 arg = null_check(arg);
3585 if (stopped()) break;
3586 args[which_arg] = arg;
3587
3588 Node* p = basic_plus_adr(arg, class_klass_offset);
3589 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3590 klasses[which_arg] = _gvn.transform(kls);
3591 }
3592
3593 // Having loaded both klasses, test each for null.
3594 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3595 for (which_arg = 0; which_arg <= 1; which_arg++) {
3596 Node* kls = klasses[which_arg];
3597 Node* null_ctl = top();
3598 kls = null_check_oop(kls, &null_ctl, never_see_null);
3599 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3600 region->init_req(prim_path, null_ctl);
3601 if (stopped()) break;
3602 klasses[which_arg] = kls;
3603 }
3604
3605 if (!stopped()) {
3606 // now we have two reference types, in klasses[0..1]
3607 Node* subk = klasses[1]; // the argument to isAssignableFrom
3608 Node* superk = klasses[0]; // the receiver
3609 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3610 // now we have a successful reference subtype check
3611 region->set_req(_ref_subtype_path, control());
3612 }
3613
3614 // If both operands are primitive (both klasses null), then
3615 // we must return true when they are identical primitives.
3616 // It is convenient to test this after the first null klass check.
3617 set_control(region->in(_prim_0_path)); // go back to first null check
3618 if (!stopped()) {
3619 // Since superc is primitive, make a guard for the superc==subc case.
3620 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3621 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3622 generate_guard(bol_eq, region, PROB_FAIR);
3623 if (region->req() == PATH_LIMIT+1) {
3624 // A guard was added. If the added guard is taken, superc==subc.
3625 region->swap_edges(PATH_LIMIT, _prim_same_path);
3626 region->del_req(PATH_LIMIT);
3627 }
3628 region->set_req(_prim_0_path, control()); // Not equal after all.
3629 }
3630
3631 // these are the only paths that produce 'true':
3632 phi->set_req(_prim_same_path, intcon(1));
3633 phi->set_req(_ref_subtype_path, intcon(1));
3634
3635 // pull together the cases:
3636 assert(region->req() == PATH_LIMIT, "sane region");
3637 for (uint i = 1; i < region->req(); i++) {
3638 Node* ctl = region->in(i);
3639 if (ctl == NULL || ctl == top()) {
3640 region->set_req(i, top());
3641 phi ->set_req(i, top());
3642 } else if (phi->in(i) == NULL) {
3643 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3644 }
3645 }
3646
3647 set_control(_gvn.transform(region));
3648 set_result(_gvn.transform(phi));
3649 return true;
3650 }
3651
3652 //---------------------generate_array_guard_common------------------------
3653 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3654 bool obj_array, bool not_array) {
3655 // If obj_array/non_array==false/false:
3656 // Branch around if the given klass is in fact an array (either obj or prim).
3657 // If obj_array/non_array==false/true:
3658 // Branch around if the given klass is not an array klass of any kind.
3659 // If obj_array/non_array==true/true:
3660 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3661 // If obj_array/non_array==true/false:
3662 // Branch around if the kls is an oop array (Object[] or subtype)
3663 //
3664 // Like generate_guard, adds a new path onto the region.
3665 jint layout_con = 0;
3666 Node* layout_val = get_layout_helper(kls, layout_con);
3667 if (layout_val == NULL) {
3668 bool query = (obj_array
3669 ? Klass::layout_helper_is_objArray(layout_con)
3670 : Klass::layout_helper_is_array(layout_con));
3671 if (query == not_array) {
3672 return NULL; // never a branch
3673 } else { // always a branch
3674 Node* always_branch = control();
3675 if (region != NULL)
3676 region->add_req(always_branch);
3677 set_control(top());
3678 return always_branch;
3679 }
3680 }
3681 // Now test the correct condition.
3682 jint nval = (obj_array
3683 ? ((jint)Klass::_lh_array_tag_type_value
3684 << Klass::_lh_array_tag_shift)
3685 : Klass::_lh_neutral_value);
3686 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3687 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3688 // invert the test if we are looking for a non-array
3689 if (not_array) btest = BoolTest(btest).negate();
3690 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3691 return generate_fair_guard(bol, region);
3692 }
3693
3694
3695 //-----------------------inline_native_newArray--------------------------
3696 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3697 bool LibraryCallKit::inline_native_newArray() {
3698 Node* mirror = argument(0);
3699 Node* count_val = argument(1);
3700
3701 mirror = null_check(mirror);
3702 // If mirror or obj is dead, only null-path is taken.
3703 if (stopped()) return true;
3704
3705 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3706 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3707 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3708 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3709 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3710
3711 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3712 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3713 result_reg, _slow_path);
3714 Node* normal_ctl = control();
3715 Node* no_array_ctl = result_reg->in(_slow_path);
3716
3717 // Generate code for the slow case. We make a call to newArray().
3718 set_control(no_array_ctl);
3719 if (!stopped()) {
3720 // Either the input type is void.class, or else the
3721 // array klass has not yet been cached. Either the
3722 // ensuing call will throw an exception, or else it
3723 // will cache the array klass for next time.
3724 PreserveJVMState pjvms(this);
3725 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3726 Node* slow_result = set_results_for_java_call(slow_call);
3727 // this->control() comes from set_results_for_java_call
3728 result_reg->set_req(_slow_path, control());
3729 result_val->set_req(_slow_path, slow_result);
3730 result_io ->set_req(_slow_path, i_o());
3731 result_mem->set_req(_slow_path, reset_memory());
3732 }
3733
3734 set_control(normal_ctl);
3735 if (!stopped()) {
3736 // Normal case: The array type has been cached in the java.lang.Class.
3737 // The following call works fine even if the array type is polymorphic.
3738 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3739 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3740 result_reg->init_req(_normal_path, control());
3741 result_val->init_req(_normal_path, obj);
3742 result_io ->init_req(_normal_path, i_o());
3743 result_mem->init_req(_normal_path, reset_memory());
3744 }
3745
3746 // Return the combined state.
3747 set_i_o( _gvn.transform(result_io) );
3748 set_all_memory( _gvn.transform(result_mem));
3749
3750 C->set_has_split_ifs(true); // Has chance for split-if optimization
3751 set_result(result_reg, result_val);
3752 return true;
3753 }
3754
3755 //----------------------inline_native_getLength--------------------------
3756 // public static native int java.lang.reflect.Array.getLength(Object array);
3757 bool LibraryCallKit::inline_native_getLength() {
3758 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3759
3760 Node* array = null_check(argument(0));
3761 // If array is dead, only null-path is taken.
3762 if (stopped()) return true;
3763
3764 // Deoptimize if it is a non-array.
3765 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3766
3767 if (non_array != NULL) {
3768 PreserveJVMState pjvms(this);
3769 set_control(non_array);
3770 uncommon_trap(Deoptimization::Reason_intrinsic,
3771 Deoptimization::Action_maybe_recompile);
3772 }
3773
3774 // If control is dead, only non-array-path is taken.
3775 if (stopped()) return true;
3776
3777 // The works fine even if the array type is polymorphic.
3778 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3779 Node* result = load_array_length(array);
3780
3781 C->set_has_split_ifs(true); // Has chance for split-if optimization
3782 set_result(result);
3783 return true;
3784 }
3785
3786 //------------------------inline_array_copyOf----------------------------
3787 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3788 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3789 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3790 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3791
3792 // Get the arguments.
3793 Node* original = argument(0);
3794 Node* start = is_copyOfRange? argument(1): intcon(0);
3795 Node* end = is_copyOfRange? argument(2): argument(1);
3796 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3797
3798 Node* newcopy;
3799
3800 // Set the original stack and the reexecute bit for the interpreter to reexecute
3801 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3802 { PreserveReexecuteState preexecs(this);
3803 jvms()->set_should_reexecute(true);
3804
3805 array_type_mirror = null_check(array_type_mirror);
3806 original = null_check(original);
3807
3808 // Check if a null path was taken unconditionally.
3809 if (stopped()) return true;
3810
3811 Node* orig_length = load_array_length(original);
3812
3813 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3814 klass_node = null_check(klass_node);
3815
3816 RegionNode* bailout = new RegionNode(1);
3817 record_for_igvn(bailout);
3818
3819 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3820 // Bail out if that is so.
3821 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3822 if (not_objArray != NULL) {
3823 // Improve the klass node's type from the new optimistic assumption:
3824 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3825 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3826 Node* cast = new CastPPNode(klass_node, akls);
3827 cast->init_req(0, control());
3828 klass_node = _gvn.transform(cast);
3829 }
3830
3831 // Bail out if either start or end is negative.
3832 generate_negative_guard(start, bailout, &start);
3833 generate_negative_guard(end, bailout, &end);
3834
3835 Node* length = end;
3836 if (_gvn.type(start) != TypeInt::ZERO) {
3837 length = _gvn.transform(new SubINode(end, start));
3838 }
3839
3840 // Bail out if length is negative.
3841 // Without this the new_array would throw
3842 // NegativeArraySizeException but IllegalArgumentException is what
3843 // should be thrown
3844 generate_negative_guard(length, bailout, &length);
3845
3846 if (bailout->req() > 1) {
3847 PreserveJVMState pjvms(this);
3848 set_control(_gvn.transform(bailout));
3849 uncommon_trap(Deoptimization::Reason_intrinsic,
3850 Deoptimization::Action_maybe_recompile);
3851 }
3852
3853 if (!stopped()) {
3854 // How many elements will we copy from the original?
3855 // The answer is MinI(orig_length - start, length).
3856 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3857 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3858
3859 newcopy = new_array(klass_node, length, 0); // no arguments to push
3860
3861 // Generate a direct call to the right arraycopy function(s).
3862 // We know the copy is disjoint but we might not know if the
3863 // oop stores need checking.
3864 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3865 // This will fail a store-check if x contains any non-nulls.
3866
3867 Node* alloc = tightly_coupled_allocation(newcopy, NULL);
3868
3869 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL,
3870 load_object_klass(original), klass_node);
3871 if (!is_copyOfRange) {
3872 ac->set_copyof();
3873 } else {
3874 ac->set_copyofrange();
3875 }
3876 Node* n = _gvn.transform(ac);
3877 assert(n == ac, "cannot disappear");
3878 ac->connect_outputs(this);
3879 }
3880 } // original reexecute is set back here
3881
3882 C->set_has_split_ifs(true); // Has chance for split-if optimization
3883 if (!stopped()) {
3884 set_result(newcopy);
3885 }
3886 return true;
3887 }
3888
3889
3890 //----------------------generate_virtual_guard---------------------------
3891 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3892 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3893 RegionNode* slow_region) {
3894 ciMethod* method = callee();
3895 int vtable_index = method->vtable_index();
3896 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3897 err_msg_res("bad index %d", vtable_index));
3898 // Get the Method* out of the appropriate vtable entry.
3899 int entry_offset = (InstanceKlass::vtable_start_offset() +
3900 vtable_index*vtableEntry::size()) * wordSize +
3901 vtableEntry::method_offset_in_bytes();
3902 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3903 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3904
3905 // Compare the target method with the expected method (e.g., Object.hashCode).
3906 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3907
3908 Node* native_call = makecon(native_call_addr);
3909 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3910 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3911
3912 return generate_slow_guard(test_native, slow_region);
3913 }
3914
3915 //-----------------------generate_method_call----------------------------
3916 // Use generate_method_call to make a slow-call to the real
3917 // method if the fast path fails. An alternative would be to
3918 // use a stub like OptoRuntime::slow_arraycopy_Java.
3919 // This only works for expanding the current library call,
3920 // not another intrinsic. (E.g., don't use this for making an
3921 // arraycopy call inside of the copyOf intrinsic.)
3922 CallJavaNode*
3923 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3924 // When compiling the intrinsic method itself, do not use this technique.
3925 guarantee(callee() != C->method(), "cannot make slow-call to self");
3926
3927 ciMethod* method = callee();
3928 // ensure the JVMS we have will be correct for this call
3929 guarantee(method_id == method->intrinsic_id(), "must match");
3930
3931 const TypeFunc* tf = TypeFunc::make(method);
3932 CallJavaNode* slow_call;
3933 if (is_static) {
3934 assert(!is_virtual, "");
3935 slow_call = new CallStaticJavaNode(C, tf,
3936 SharedRuntime::get_resolve_static_call_stub(),
3937 method, bci());
3938 } else if (is_virtual) {
3939 null_check_receiver();
3940 int vtable_index = Method::invalid_vtable_index;
3941 if (UseInlineCaches) {
3942 // Suppress the vtable call
3943 } else {
3944 // hashCode and clone are not a miranda methods,
3945 // so the vtable index is fixed.
3946 // No need to use the linkResolver to get it.
3947 vtable_index = method->vtable_index();
3948 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3949 err_msg_res("bad index %d", vtable_index));
3950 }
3951 slow_call = new CallDynamicJavaNode(tf,
3952 SharedRuntime::get_resolve_virtual_call_stub(),
3953 method, vtable_index, bci());
3954 } else { // neither virtual nor static: opt_virtual
3955 null_check_receiver();
3956 slow_call = new CallStaticJavaNode(C, tf,
3957 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3958 method, bci());
3959 slow_call->set_optimized_virtual(true);
3960 }
3961 set_arguments_for_java_call(slow_call);
3962 set_edges_for_java_call(slow_call);
3963 return slow_call;
3964 }
3965
3966
3967 /**
3968 * Build special case code for calls to hashCode on an object. This call may
3969 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3970 * slightly different code.
3971 */
3972 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3973 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3974 assert(!(is_virtual && is_static), "either virtual, special, or static");
3975
3976 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3977
3978 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3979 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3980 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3981 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3982 Node* obj = NULL;
3983 if (!is_static) {
3984 // Check for hashing null object
3985 obj = null_check_receiver();
3986 if (stopped()) return true; // unconditionally null
3987 result_reg->init_req(_null_path, top());
3988 result_val->init_req(_null_path, top());
3989 } else {
3990 // Do a null check, and return zero if null.
3991 // System.identityHashCode(null) == 0
3992 obj = argument(0);
3993 Node* null_ctl = top();
3994 obj = null_check_oop(obj, &null_ctl);
3995 result_reg->init_req(_null_path, null_ctl);
3996 result_val->init_req(_null_path, _gvn.intcon(0));
3997 }
3998
3999 // Unconditionally null? Then return right away.
4000 if (stopped()) {
4001 set_control( result_reg->in(_null_path));
4002 if (!stopped())
4003 set_result(result_val->in(_null_path));
4004 return true;
4005 }
4006
4007 // We only go to the fast case code if we pass a number of guards. The
4008 // paths which do not pass are accumulated in the slow_region.
4009 RegionNode* slow_region = new RegionNode(1);
4010 record_for_igvn(slow_region);
4011
4012 // If this is a virtual call, we generate a funny guard. We pull out
4013 // the vtable entry corresponding to hashCode() from the target object.
4014 // If the target method which we are calling happens to be the native
4015 // Object hashCode() method, we pass the guard. We do not need this
4016 // guard for non-virtual calls -- the caller is known to be the native
4017 // Object hashCode().
4018 if (is_virtual) {
4019 // After null check, get the object's klass.
4020 Node* obj_klass = load_object_klass(obj);
4021 generate_virtual_guard(obj_klass, slow_region);
4022 }
4023
4024 // Get the header out of the object, use LoadMarkNode when available
4025 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4026 // The control of the load must be NULL. Otherwise, the load can move before
4027 // the null check after castPP removal.
4028 Node* no_ctrl = NULL;
4029 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4030
4031 // Test the header to see if it is unlocked.
4032 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4033 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4034 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4035 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4036 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4037
4038 generate_slow_guard(test_unlocked, slow_region);
4039
4040 // Get the hash value and check to see that it has been properly assigned.
4041 // We depend on hash_mask being at most 32 bits and avoid the use of
4042 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4043 // vm: see markOop.hpp.
4044 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4045 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4046 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4047 // This hack lets the hash bits live anywhere in the mark object now, as long
4048 // as the shift drops the relevant bits into the low 32 bits. Note that
4049 // Java spec says that HashCode is an int so there's no point in capturing
4050 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4051 hshifted_header = ConvX2I(hshifted_header);
4052 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4053
4054 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4055 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4056 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4057
4058 generate_slow_guard(test_assigned, slow_region);
4059
4060 Node* init_mem = reset_memory();
4061 // fill in the rest of the null path:
4062 result_io ->init_req(_null_path, i_o());
4063 result_mem->init_req(_null_path, init_mem);
4064
4065 result_val->init_req(_fast_path, hash_val);
4066 result_reg->init_req(_fast_path, control());
4067 result_io ->init_req(_fast_path, i_o());
4068 result_mem->init_req(_fast_path, init_mem);
4069
4070 // Generate code for the slow case. We make a call to hashCode().
4071 set_control(_gvn.transform(slow_region));
4072 if (!stopped()) {
4073 // No need for PreserveJVMState, because we're using up the present state.
4074 set_all_memory(init_mem);
4075 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4076 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4077 Node* slow_result = set_results_for_java_call(slow_call);
4078 // this->control() comes from set_results_for_java_call
4079 result_reg->init_req(_slow_path, control());
4080 result_val->init_req(_slow_path, slow_result);
4081 result_io ->set_req(_slow_path, i_o());
4082 result_mem ->set_req(_slow_path, reset_memory());
4083 }
4084
4085 // Return the combined state.
4086 set_i_o( _gvn.transform(result_io) );
4087 set_all_memory( _gvn.transform(result_mem));
4088
4089 set_result(result_reg, result_val);
4090 return true;
4091 }
4092
4093 //---------------------------inline_native_getClass----------------------------
4094 // public final native Class<?> java.lang.Object.getClass();
4095 //
4096 // Build special case code for calls to getClass on an object.
4097 bool LibraryCallKit::inline_native_getClass() {
4098 Node* obj = null_check_receiver();
4099 if (stopped()) return true;
4100 set_result(load_mirror_from_klass(load_object_klass(obj)));
4101 return true;
4102 }
4103
4104 //-----------------inline_native_Reflection_getCallerClass---------------------
4105 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4106 //
4107 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4108 //
4109 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4110 // in that it must skip particular security frames and checks for
4111 // caller sensitive methods.
4112 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4113 #ifndef PRODUCT
4114 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4115 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4116 }
4117 #endif
4118
4119 if (!jvms()->has_method()) {
4120 #ifndef PRODUCT
4121 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4122 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4123 }
4124 #endif
4125 return false;
4126 }
4127
4128 // Walk back up the JVM state to find the caller at the required
4129 // depth.
4130 JVMState* caller_jvms = jvms();
4131
4132 // Cf. JVM_GetCallerClass
4133 // NOTE: Start the loop at depth 1 because the current JVM state does
4134 // not include the Reflection.getCallerClass() frame.
4135 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4136 ciMethod* m = caller_jvms->method();
4137 switch (n) {
4138 case 0:
4139 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4140 break;
4141 case 1:
4142 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4143 if (!m->caller_sensitive()) {
4144 #ifndef PRODUCT
4145 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4146 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4147 }
4148 #endif
4149 return false; // bail-out; let JVM_GetCallerClass do the work
4150 }
4151 break;
4152 default:
4153 if (!m->is_ignored_by_security_stack_walk()) {
4154 // We have reached the desired frame; return the holder class.
4155 // Acquire method holder as java.lang.Class and push as constant.
4156 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4157 ciInstance* caller_mirror = caller_klass->java_mirror();
4158 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4159
4160 #ifndef PRODUCT
4161 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4162 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());
4163 tty->print_cr(" JVM state at this point:");
4164 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4165 ciMethod* m = jvms()->of_depth(i)->method();
4166 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4167 }
4168 }
4169 #endif
4170 return true;
4171 }
4172 break;
4173 }
4174 }
4175
4176 #ifndef PRODUCT
4177 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4178 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4179 tty->print_cr(" JVM state at this point:");
4180 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4181 ciMethod* m = jvms()->of_depth(i)->method();
4182 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4183 }
4184 }
4185 #endif
4186
4187 return false; // bail-out; let JVM_GetCallerClass do the work
4188 }
4189
4190 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4191 Node* arg = argument(0);
4192 Node* result;
4193
4194 switch (id) {
4195 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4196 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4197 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4198 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4199
4200 case vmIntrinsics::_doubleToLongBits: {
4201 // two paths (plus control) merge in a wood
4202 RegionNode *r = new RegionNode(3);
4203 Node *phi = new PhiNode(r, TypeLong::LONG);
4204
4205 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4206 // Build the boolean node
4207 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4208
4209 // Branch either way.
4210 // NaN case is less traveled, which makes all the difference.
4211 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4212 Node *opt_isnan = _gvn.transform(ifisnan);
4213 assert( opt_isnan->is_If(), "Expect an IfNode");
4214 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4215 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4216
4217 set_control(iftrue);
4218
4219 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4220 Node *slow_result = longcon(nan_bits); // return NaN
4221 phi->init_req(1, _gvn.transform( slow_result ));
4222 r->init_req(1, iftrue);
4223
4224 // Else fall through
4225 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4226 set_control(iffalse);
4227
4228 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4229 r->init_req(2, iffalse);
4230
4231 // Post merge
4232 set_control(_gvn.transform(r));
4233 record_for_igvn(r);
4234
4235 C->set_has_split_ifs(true); // Has chance for split-if optimization
4236 result = phi;
4237 assert(result->bottom_type()->isa_long(), "must be");
4238 break;
4239 }
4240
4241 case vmIntrinsics::_floatToIntBits: {
4242 // two paths (plus control) merge in a wood
4243 RegionNode *r = new RegionNode(3);
4244 Node *phi = new PhiNode(r, TypeInt::INT);
4245
4246 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4247 // Build the boolean node
4248 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4249
4250 // Branch either way.
4251 // NaN case is less traveled, which makes all the difference.
4252 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4253 Node *opt_isnan = _gvn.transform(ifisnan);
4254 assert( opt_isnan->is_If(), "Expect an IfNode");
4255 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4256 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4257
4258 set_control(iftrue);
4259
4260 static const jint nan_bits = 0x7fc00000;
4261 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4262 phi->init_req(1, _gvn.transform( slow_result ));
4263 r->init_req(1, iftrue);
4264
4265 // Else fall through
4266 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4267 set_control(iffalse);
4268
4269 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4270 r->init_req(2, iffalse);
4271
4272 // Post merge
4273 set_control(_gvn.transform(r));
4274 record_for_igvn(r);
4275
4276 C->set_has_split_ifs(true); // Has chance for split-if optimization
4277 result = phi;
4278 assert(result->bottom_type()->isa_int(), "must be");
4279 break;
4280 }
4281
4282 default:
4283 fatal_unexpected_iid(id);
4284 break;
4285 }
4286 set_result(_gvn.transform(result));
4287 return true;
4288 }
4289
4290 #ifdef _LP64
4291 #define XTOP ,top() /*additional argument*/
4292 #else //_LP64
4293 #define XTOP /*no additional argument*/
4294 #endif //_LP64
4295
4296 //----------------------inline_unsafe_copyMemory-------------------------
4297 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4298 bool LibraryCallKit::inline_unsafe_copyMemory() {
4299 if (callee()->is_static()) return false; // caller must have the capability!
4300 null_check_receiver(); // null-check receiver
4301 if (stopped()) return true;
4302
4303 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4304
4305 Node* src_ptr = argument(1); // type: oop
4306 Node* src_off = ConvL2X(argument(2)); // type: long
4307 Node* dst_ptr = argument(4); // type: oop
4308 Node* dst_off = ConvL2X(argument(5)); // type: long
4309 Node* size = ConvL2X(argument(7)); // type: long
4310
4311 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4312 "fieldOffset must be byte-scaled");
4313
4314 Node* src = make_unsafe_address(src_ptr, src_off);
4315 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4316
4317 // Conservatively insert a memory barrier on all memory slices.
4318 // Do not let writes of the copy source or destination float below the copy.
4319 insert_mem_bar(Op_MemBarCPUOrder);
4320
4321 // Call it. Note that the length argument is not scaled.
4322 make_runtime_call(RC_LEAF|RC_NO_FP,
4323 OptoRuntime::fast_arraycopy_Type(),
4324 StubRoutines::unsafe_arraycopy(),
4325 "unsafe_arraycopy",
4326 TypeRawPtr::BOTTOM,
4327 src, dst, size XTOP);
4328
4329 // Do not let reads of the copy destination float above the copy.
4330 insert_mem_bar(Op_MemBarCPUOrder);
4331
4332 return true;
4333 }
4334
4335 //------------------------clone_coping-----------------------------------
4336 // Helper function for inline_native_clone.
4337 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4338 assert(obj_size != NULL, "");
4339 Node* raw_obj = alloc_obj->in(1);
4340 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4341
4342 AllocateNode* alloc = NULL;
4343 if (ReduceBulkZeroing) {
4344 // We will be completely responsible for initializing this object -
4345 // mark Initialize node as complete.
4346 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4347 // The object was just allocated - there should be no any stores!
4348 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4349 // Mark as complete_with_arraycopy so that on AllocateNode
4350 // expansion, we know this AllocateNode is initialized by an array
4351 // copy and a StoreStore barrier exists after the array copy.
4352 alloc->initialization()->set_complete_with_arraycopy();
4353 }
4354
4355 // Copy the fastest available way.
4356 // TODO: generate fields copies for small objects instead.
4357 Node* src = obj;
4358 Node* dest = alloc_obj;
4359 Node* size = _gvn.transform(obj_size);
4360
4361 // Exclude the header but include array length to copy by 8 bytes words.
4362 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4363 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4364 instanceOopDesc::base_offset_in_bytes();
4365 // base_off:
4366 // 8 - 32-bit VM
4367 // 12 - 64-bit VM, compressed klass
4368 // 16 - 64-bit VM, normal klass
4369 if (base_off % BytesPerLong != 0) {
4370 assert(UseCompressedClassPointers, "");
4371 if (is_array) {
4372 // Exclude length to copy by 8 bytes words.
4373 base_off += sizeof(int);
4374 } else {
4375 // Include klass to copy by 8 bytes words.
4376 base_off = instanceOopDesc::klass_offset_in_bytes();
4377 }
4378 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4379 }
4380 src = basic_plus_adr(src, base_off);
4381 dest = basic_plus_adr(dest, base_off);
4382
4383 // Compute the length also, if needed:
4384 Node* countx = size;
4385 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4386 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4387
4388 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4389
4390 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4391 ac->set_clonebasic();
4392 Node* n = _gvn.transform(ac);
4393 assert(n == ac, "cannot disappear");
4394 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4395
4396 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4397 if (card_mark) {
4398 assert(!is_array, "");
4399 // Put in store barrier for any and all oops we are sticking
4400 // into this object. (We could avoid this if we could prove
4401 // that the object type contains no oop fields at all.)
4402 Node* no_particular_value = NULL;
4403 Node* no_particular_field = NULL;
4404 int raw_adr_idx = Compile::AliasIdxRaw;
4405 post_barrier(control(),
4406 memory(raw_adr_type),
4407 alloc_obj,
4408 no_particular_field,
4409 raw_adr_idx,
4410 no_particular_value,
4411 T_OBJECT,
4412 false);
4413 }
4414
4415 // Do not let reads from the cloned object float above the arraycopy.
4416 if (alloc != NULL) {
4417 // Do not let stores that initialize this object be reordered with
4418 // a subsequent store that would make this object accessible by
4419 // other threads.
4420 // Record what AllocateNode this StoreStore protects so that
4421 // escape analysis can go from the MemBarStoreStoreNode to the
4422 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4423 // based on the escape status of the AllocateNode.
4424 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4425 } else {
4426 insert_mem_bar(Op_MemBarCPUOrder);
4427 }
4428 }
4429
4430 //------------------------inline_native_clone----------------------------
4431 // protected native Object java.lang.Object.clone();
4432 //
4433 // Here are the simple edge cases:
4434 // null receiver => normal trap
4435 // virtual and clone was overridden => slow path to out-of-line clone
4436 // not cloneable or finalizer => slow path to out-of-line Object.clone
4437 //
4438 // The general case has two steps, allocation and copying.
4439 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4440 //
4441 // Copying also has two cases, oop arrays and everything else.
4442 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4443 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4444 //
4445 // These steps fold up nicely if and when the cloned object's klass
4446 // can be sharply typed as an object array, a type array, or an instance.
4447 //
4448 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4449 PhiNode* result_val;
4450
4451 // Set the reexecute bit for the interpreter to reexecute
4452 // the bytecode that invokes Object.clone if deoptimization happens.
4453 { PreserveReexecuteState preexecs(this);
4454 jvms()->set_should_reexecute(true);
4455
4456 Node* obj = null_check_receiver();
4457 if (stopped()) return true;
4458
4459 Node* obj_klass = load_object_klass(obj);
4460 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4461 const TypeOopPtr* toop = ((tklass != NULL)
4462 ? tklass->as_instance_type()
4463 : TypeInstPtr::NOTNULL);
4464
4465 // Conservatively insert a memory barrier on all memory slices.
4466 // Do not let writes into the original float below the clone.
4467 insert_mem_bar(Op_MemBarCPUOrder);
4468
4469 // paths into result_reg:
4470 enum {
4471 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4472 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4473 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4474 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4475 PATH_LIMIT
4476 };
4477 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4478 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4479 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4480 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4481 record_for_igvn(result_reg);
4482
4483 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4484 int raw_adr_idx = Compile::AliasIdxRaw;
4485
4486 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4487 if (array_ctl != NULL) {
4488 // It's an array.
4489 PreserveJVMState pjvms(this);
4490 set_control(array_ctl);
4491 Node* obj_length = load_array_length(obj);
4492 Node* obj_size = NULL;
4493 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4494
4495 if (!use_ReduceInitialCardMarks()) {
4496 // If it is an oop array, it requires very special treatment,
4497 // because card marking is required on each card of the array.
4498 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4499 if (is_obja != NULL) {
4500 PreserveJVMState pjvms2(this);
4501 set_control(is_obja);
4502 // Generate a direct call to the right arraycopy function(s).
4503 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4504 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4505 ac->set_cloneoop();
4506 Node* n = _gvn.transform(ac);
4507 assert(n == ac, "cannot disappear");
4508 ac->connect_outputs(this);
4509
4510 result_reg->init_req(_objArray_path, control());
4511 result_val->init_req(_objArray_path, alloc_obj);
4512 result_i_o ->set_req(_objArray_path, i_o());
4513 result_mem ->set_req(_objArray_path, reset_memory());
4514 }
4515 }
4516 // Otherwise, there are no card marks to worry about.
4517 // (We can dispense with card marks if we know the allocation
4518 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4519 // causes the non-eden paths to take compensating steps to
4520 // simulate a fresh allocation, so that no further
4521 // card marks are required in compiled code to initialize
4522 // the object.)
4523
4524 if (!stopped()) {
4525 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4526
4527 // Present the results of the copy.
4528 result_reg->init_req(_array_path, control());
4529 result_val->init_req(_array_path, alloc_obj);
4530 result_i_o ->set_req(_array_path, i_o());
4531 result_mem ->set_req(_array_path, reset_memory());
4532 }
4533 }
4534
4535 // We only go to the instance fast case code if we pass a number of guards.
4536 // The paths which do not pass are accumulated in the slow_region.
4537 RegionNode* slow_region = new RegionNode(1);
4538 record_for_igvn(slow_region);
4539 if (!stopped()) {
4540 // It's an instance (we did array above). Make the slow-path tests.
4541 // If this is a virtual call, we generate a funny guard. We grab
4542 // the vtable entry corresponding to clone() from the target object.
4543 // If the target method which we are calling happens to be the
4544 // Object clone() method, we pass the guard. We do not need this
4545 // guard for non-virtual calls; the caller is known to be the native
4546 // Object clone().
4547 if (is_virtual) {
4548 generate_virtual_guard(obj_klass, slow_region);
4549 }
4550
4551 // The object must be cloneable and must not have a finalizer.
4552 // Both of these conditions may be checked in a single test.
4553 // We could optimize the cloneable test further, but we don't care.
4554 generate_access_flags_guard(obj_klass,
4555 // Test both conditions:
4556 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4557 // Must be cloneable but not finalizer:
4558 JVM_ACC_IS_CLONEABLE,
4559 slow_region);
4560 }
4561
4562 if (!stopped()) {
4563 // It's an instance, and it passed the slow-path tests.
4564 PreserveJVMState pjvms(this);
4565 Node* obj_size = NULL;
4566 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4567 // is reexecuted if deoptimization occurs and there could be problems when merging
4568 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4569 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4570
4571 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4572
4573 // Present the results of the slow call.
4574 result_reg->init_req(_instance_path, control());
4575 result_val->init_req(_instance_path, alloc_obj);
4576 result_i_o ->set_req(_instance_path, i_o());
4577 result_mem ->set_req(_instance_path, reset_memory());
4578 }
4579
4580 // Generate code for the slow case. We make a call to clone().
4581 set_control(_gvn.transform(slow_region));
4582 if (!stopped()) {
4583 PreserveJVMState pjvms(this);
4584 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4585 Node* slow_result = set_results_for_java_call(slow_call);
4586 // this->control() comes from set_results_for_java_call
4587 result_reg->init_req(_slow_path, control());
4588 result_val->init_req(_slow_path, slow_result);
4589 result_i_o ->set_req(_slow_path, i_o());
4590 result_mem ->set_req(_slow_path, reset_memory());
4591 }
4592
4593 // Return the combined state.
4594 set_control( _gvn.transform(result_reg));
4595 set_i_o( _gvn.transform(result_i_o));
4596 set_all_memory( _gvn.transform(result_mem));
4597 } // original reexecute is set back here
4598
4599 set_result(_gvn.transform(result_val));
4600 return true;
4601 }
4602
4603 //------------------------------inline_arraycopy-----------------------
4604 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4605 // Object dest, int destPos,
4606 // int length);
4607 bool LibraryCallKit::inline_arraycopy() {
4608 // Get the arguments.
4609 Node* src = argument(0); // type: oop
4610 Node* src_offset = argument(1); // type: int
4611 Node* dest = argument(2); // type: oop
4612 Node* dest_offset = argument(3); // type: int
4613 Node* length = argument(4); // type: int
4614
4615 // The following tests must be performed
4616 // (1) src and dest are arrays.
4617 // (2) src and dest arrays must have elements of the same BasicType
4618 // (3) src and dest must not be null.
4619 // (4) src_offset must not be negative.
4620 // (5) dest_offset must not be negative.
4621 // (6) length must not be negative.
4622 // (7) src_offset + length must not exceed length of src.
4623 // (8) dest_offset + length must not exceed length of dest.
4624 // (9) each element of an oop array must be assignable
4625
4626 // (3) src and dest must not be null.
4627 // always do this here because we need the JVM state for uncommon traps
4628 src = null_check(src, T_ARRAY);
4629 dest = null_check(dest, T_ARRAY);
4630
4631 bool notest = false;
4632
4633 const Type* src_type = _gvn.type(src);
4634 const Type* dest_type = _gvn.type(dest);
4635 const TypeAryPtr* top_src = src_type->isa_aryptr();
4636 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4637
4638 // Do we have the type of src?
4639 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4640 // Do we have the type of dest?
4641 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4642 // Is the type for src from speculation?
4643 bool src_spec = false;
4644 // Is the type for dest from speculation?
4645 bool dest_spec = false;
4646
4647 if (!has_src || !has_dest) {
4648 // We don't have sufficient type information, let's see if
4649 // speculative types can help. We need to have types for both src
4650 // and dest so that it pays off.
4651
4652 // Do we already have or could we have type information for src
4653 bool could_have_src = has_src;
4654 // Do we already have or could we have type information for dest
4655 bool could_have_dest = has_dest;
4656
4657 ciKlass* src_k = NULL;
4658 if (!has_src) {
4659 src_k = src_type->speculative_type_not_null();
4660 if (src_k != NULL && src_k->is_array_klass()) {
4661 could_have_src = true;
4662 }
4663 }
4664
4665 ciKlass* dest_k = NULL;
4666 if (!has_dest) {
4667 dest_k = dest_type->speculative_type_not_null();
4668 if (dest_k != NULL && dest_k->is_array_klass()) {
4669 could_have_dest = true;
4670 }
4671 }
4672
4673 if (could_have_src && could_have_dest) {
4674 // This is going to pay off so emit the required guards
4675 if (!has_src) {
4676 src = maybe_cast_profiled_obj(src, src_k);
4677 src_type = _gvn.type(src);
4678 top_src = src_type->isa_aryptr();
4679 has_src = (top_src != NULL && top_src->klass() != NULL);
4680 src_spec = true;
4681 }
4682 if (!has_dest) {
4683 dest = maybe_cast_profiled_obj(dest, dest_k);
4684 dest_type = _gvn.type(dest);
4685 top_dest = dest_type->isa_aryptr();
4686 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4687 dest_spec = true;
4688 }
4689 }
4690 }
4691
4692 if (has_src && has_dest) {
4693 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4694 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4695 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4696 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4697
4698 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4699 // If both arrays are object arrays then having the exact types
4700 // for both will remove the need for a subtype check at runtime
4701 // before the call and may make it possible to pick a faster copy
4702 // routine (without a subtype check on every element)
4703 // Do we have the exact type of src?
4704 bool could_have_src = src_spec;
4705 // Do we have the exact type of dest?
4706 bool could_have_dest = dest_spec;
4707 ciKlass* src_k = top_src->klass();
4708 ciKlass* dest_k = top_dest->klass();
4709 if (!src_spec) {
4710 src_k = src_type->speculative_type_not_null();
4711 if (src_k != NULL && src_k->is_array_klass()) {
4712 could_have_src = true;
4713 }
4714 }
4715 if (!dest_spec) {
4716 dest_k = dest_type->speculative_type_not_null();
4717 if (dest_k != NULL && dest_k->is_array_klass()) {
4718 could_have_dest = true;
4719 }
4720 }
4721 if (could_have_src && could_have_dest) {
4722 // If we can have both exact types, emit the missing guards
4723 if (could_have_src && !src_spec) {
4724 src = maybe_cast_profiled_obj(src, src_k);
4725 }
4726 if (could_have_dest && !dest_spec) {
4727 dest = maybe_cast_profiled_obj(dest, dest_k);
4728 }
4729 }
4730 }
4731 }
4732
4733 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) {
4734 // validate arguments: enables transformation the ArrayCopyNode
4735 notest = true;
4736
4737 RegionNode* slow_region = new RegionNode(1);
4738 record_for_igvn(slow_region);
4739
4740 // (1) src and dest are arrays.
4741 generate_non_array_guard(load_object_klass(src), slow_region);
4742 generate_non_array_guard(load_object_klass(dest), slow_region);
4743
4744 // (2) src and dest arrays must have elements of the same BasicType
4745 // done at macro expansion or at Ideal transformation time
4746
4747 // (4) src_offset must not be negative.
4748 generate_negative_guard(src_offset, slow_region);
4749
4750 // (5) dest_offset must not be negative.
4751 generate_negative_guard(dest_offset, slow_region);
4752
4753 // (7) src_offset + length must not exceed length of src.
4754 generate_limit_guard(src_offset, length,
4755 load_array_length(src),
4756 slow_region);
4757
4758 // (8) dest_offset + length must not exceed length of dest.
4759 generate_limit_guard(dest_offset, length,
4760 load_array_length(dest),
4761 slow_region);
4762
4763 // (9) each element of an oop array must be assignable
4764 Node* src_klass = load_object_klass(src);
4765 Node* dest_klass = load_object_klass(dest);
4766 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4767
4768 if (not_subtype_ctrl != top()) {
4769 PreserveJVMState pjvms(this);
4770 set_control(not_subtype_ctrl);
4771 uncommon_trap(Deoptimization::Reason_intrinsic,
4772 Deoptimization::Action_make_not_entrant);
4773 assert(stopped(), "Should be stopped");
4774 }
4775 {
4776 PreserveJVMState pjvms(this);
4777 set_control(_gvn.transform(slow_region));
4778 uncommon_trap(Deoptimization::Reason_intrinsic,
4779 Deoptimization::Action_make_not_entrant);
4780 assert(stopped(), "Should be stopped");
4781 }
4782 }
4783
4784 if (stopped()) {
4785 return true;
4786 }
4787
4788 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4789 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4790 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4791 // so the compiler has a chance to eliminate them: during macro expansion,
4792 // we have to set their control (CastPP nodes are eliminated).
4793 load_object_klass(src), load_object_klass(dest),
4794 load_array_length(src), load_array_length(dest));
4795
4796 if (notest) {
4797 ac->set_arraycopy_notest();
4798 }
4799
4800 Node* n = _gvn.transform(ac);
4801 assert(n == ac, "cannot disappear");
4802 ac->connect_outputs(this);
4803
4804 return true;
4805 }
4806
4807
4808 // Helper function which determines if an arraycopy immediately follows
4809 // an allocation, with no intervening tests or other escapes for the object.
4810 AllocateArrayNode*
4811 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4812 RegionNode* slow_region) {
4813 if (stopped()) return NULL; // no fast path
4814 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4815
4816 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4817 if (alloc == NULL) return NULL;
4818
4819 Node* rawmem = memory(Compile::AliasIdxRaw);
4820 // Is the allocation's memory state untouched?
4821 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4822 // Bail out if there have been raw-memory effects since the allocation.
4823 // (Example: There might have been a call or safepoint.)
4824 return NULL;
4825 }
4826 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4827 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4828 return NULL;
4829 }
4830
4831 // There must be no unexpected observers of this allocation.
4832 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4833 Node* obs = ptr->fast_out(i);
4834 if (obs != this->map()) {
4835 return NULL;
4836 }
4837 }
4838
4839 // This arraycopy must unconditionally follow the allocation of the ptr.
4840 Node* alloc_ctl = ptr->in(0);
4841 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4842
4843 Node* ctl = control();
4844 while (ctl != alloc_ctl) {
4845 // There may be guards which feed into the slow_region.
4846 // Any other control flow means that we might not get a chance
4847 // to finish initializing the allocated object.
4848 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4849 IfNode* iff = ctl->in(0)->as_If();
4850 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4851 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4852 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4853 ctl = iff->in(0); // This test feeds the known slow_region.
4854 continue;
4855 }
4856 // One more try: Various low-level checks bottom out in
4857 // uncommon traps. If the debug-info of the trap omits
4858 // any reference to the allocation, as we've already
4859 // observed, then there can be no objection to the trap.
4860 bool found_trap = false;
4861 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4862 Node* obs = not_ctl->fast_out(j);
4863 if (obs->in(0) == not_ctl && obs->is_Call() &&
4864 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4865 found_trap = true; break;
4866 }
4867 }
4868 if (found_trap) {
4869 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4870 continue;
4871 }
4872 }
4873 return NULL;
4874 }
4875
4876 // If we get this far, we have an allocation which immediately
4877 // precedes the arraycopy, and we can take over zeroing the new object.
4878 // The arraycopy will finish the initialization, and provide
4879 // a new control state to which we will anchor the destination pointer.
4880
4881 return alloc;
4882 }
4883
4884 //-------------inline_encodeISOArray-----------------------------------
4885 // encode char[] to byte[] in ISO_8859_1
4886 bool LibraryCallKit::inline_encodeISOArray() {
4887 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4888 // no receiver since it is static method
4889 Node *src = argument(0);
4890 Node *src_offset = argument(1);
4891 Node *dst = argument(2);
4892 Node *dst_offset = argument(3);
4893 Node *length = argument(4);
4894
4895 const Type* src_type = src->Value(&_gvn);
4896 const Type* dst_type = dst->Value(&_gvn);
4897 const TypeAryPtr* top_src = src_type->isa_aryptr();
4898 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4899 if (top_src == NULL || top_src->klass() == NULL ||
4900 top_dest == NULL || top_dest->klass() == NULL) {
4901 // failed array check
4902 return false;
4903 }
4904
4905 // Figure out the size and type of the elements we will be copying.
4906 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4907 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4908 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
4909 return false;
4910 }
4911 Node* src_start = array_element_address(src, src_offset, src_elem);
4912 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
4913 // 'src_start' points to src array + scaled offset
4914 // 'dst_start' points to dst array + scaled offset
4915
4916 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
4917 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
4918 enc = _gvn.transform(enc);
4919 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
4920 set_memory(res_mem, mtype);
4921 set_result(enc);
4922 return true;
4923 }
4924
4925 //-------------inline_multiplyToLen-----------------------------------
4926 bool LibraryCallKit::inline_multiplyToLen() {
4927 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
4928
4929 address stubAddr = StubRoutines::multiplyToLen();
4930 if (stubAddr == NULL) {
4931 return false; // Intrinsic's stub is not implemented on this platform
4932 }
4933 const char* stubName = "multiplyToLen";
4934
4935 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
4936
4937 Node* x = argument(1);
4938 Node* xlen = argument(2);
4939 Node* y = argument(3);
4940 Node* ylen = argument(4);
4941 Node* z = argument(5);
4942
4943 const Type* x_type = x->Value(&_gvn);
4944 const Type* y_type = y->Value(&_gvn);
4945 const TypeAryPtr* top_x = x_type->isa_aryptr();
4946 const TypeAryPtr* top_y = y_type->isa_aryptr();
4947 if (top_x == NULL || top_x->klass() == NULL ||
4948 top_y == NULL || top_y->klass() == NULL) {
4949 // failed array check
4950 return false;
4951 }
4952
4953 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4954 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4955 if (x_elem != T_INT || y_elem != T_INT) {
4956 return false;
4957 }
4958
4959 // Set the original stack and the reexecute bit for the interpreter to reexecute
4960 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
4961 // on the return from z array allocation in runtime.
4962 { PreserveReexecuteState preexecs(this);
4963 jvms()->set_should_reexecute(true);
4964
4965 Node* x_start = array_element_address(x, intcon(0), x_elem);
4966 Node* y_start = array_element_address(y, intcon(0), y_elem);
4967 // 'x_start' points to x array + scaled xlen
4968 // 'y_start' points to y array + scaled ylen
4969
4970 // Allocate the result array
4971 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
4972 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
4973 Node* klass_node = makecon(TypeKlassPtr::make(klass));
4974
4975 IdealKit ideal(this);
4976
4977 #define __ ideal.
4978 Node* one = __ ConI(1);
4979 Node* zero = __ ConI(0);
4980 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
4981 __ set(need_alloc, zero);
4982 __ set(z_alloc, z);
4983 __ if_then(z, BoolTest::eq, null()); {
4984 __ increment (need_alloc, one);
4985 } __ else_(); {
4986 // Update graphKit memory and control from IdealKit.
4987 sync_kit(ideal);
4988 Node* zlen_arg = load_array_length(z);
4989 // Update IdealKit memory and control from graphKit.
4990 __ sync_kit(this);
4991 __ if_then(zlen_arg, BoolTest::lt, zlen); {
4992 __ increment (need_alloc, one);
4993 } __ end_if();
4994 } __ end_if();
4995
4996 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
4997 // Update graphKit memory and control from IdealKit.
4998 sync_kit(ideal);
4999 Node * narr = new_array(klass_node, zlen, 1);
5000 // Update IdealKit memory and control from graphKit.
5001 __ sync_kit(this);
5002 __ set(z_alloc, narr);
5003 } __ end_if();
5004
5005 sync_kit(ideal);
5006 z = __ value(z_alloc);
5007 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5008 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5009 // Final sync IdealKit and GraphKit.
5010 final_sync(ideal);
5011 #undef __
5012
5013 Node* z_start = array_element_address(z, intcon(0), T_INT);
5014
5015 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5016 OptoRuntime::multiplyToLen_Type(),
5017 stubAddr, stubName, TypePtr::BOTTOM,
5018 x_start, xlen, y_start, ylen, z_start, zlen);
5019 } // original reexecute is set back here
5020
5021 C->set_has_split_ifs(true); // Has chance for split-if optimization
5022 set_result(z);
5023 return true;
5024 }
5025
5026
5027 /**
5028 * Calculate CRC32 for byte.
5029 * int java.util.zip.CRC32.update(int crc, int b)
5030 */
5031 bool LibraryCallKit::inline_updateCRC32() {
5032 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5033 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5034 // no receiver since it is static method
5035 Node* crc = argument(0); // type: int
5036 Node* b = argument(1); // type: int
5037
5038 /*
5039 * int c = ~ crc;
5040 * b = timesXtoThe32[(b ^ c) & 0xFF];
5041 * b = b ^ (c >>> 8);
5042 * crc = ~b;
5043 */
5044
5045 Node* M1 = intcon(-1);
5046 crc = _gvn.transform(new XorINode(crc, M1));
5047 Node* result = _gvn.transform(new XorINode(crc, b));
5048 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5049
5050 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5051 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5052 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5053 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5054
5055 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5056 result = _gvn.transform(new XorINode(crc, result));
5057 result = _gvn.transform(new XorINode(result, M1));
5058 set_result(result);
5059 return true;
5060 }
5061
5062 /**
5063 * Calculate CRC32 for byte[] array.
5064 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5065 */
5066 bool LibraryCallKit::inline_updateBytesCRC32() {
5067 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5068 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5069 // no receiver since it is static method
5070 Node* crc = argument(0); // type: int
5071 Node* src = argument(1); // type: oop
5072 Node* offset = argument(2); // type: int
5073 Node* length = argument(3); // type: int
5074
5075 const Type* src_type = src->Value(&_gvn);
5076 const TypeAryPtr* top_src = src_type->isa_aryptr();
5077 if (top_src == NULL || top_src->klass() == NULL) {
5078 // failed array check
5079 return false;
5080 }
5081
5082 // Figure out the size and type of the elements we will be copying.
5083 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5084 if (src_elem != T_BYTE) {
5085 return false;
5086 }
5087
5088 // 'src_start' points to src array + scaled offset
5089 Node* src_start = array_element_address(src, offset, src_elem);
5090
5091 // We assume that range check is done by caller.
5092 // TODO: generate range check (offset+length < src.length) in debug VM.
5093
5094 // Call the stub.
5095 address stubAddr = StubRoutines::updateBytesCRC32();
5096 const char *stubName = "updateBytesCRC32";
5097
5098 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5099 stubAddr, stubName, TypePtr::BOTTOM,
5100 crc, src_start, length);
5101 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5102 set_result(result);
5103 return true;
5104 }
5105
5106 /**
5107 * Calculate CRC32 for ByteBuffer.
5108 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5109 */
5110 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5111 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5112 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5113 // no receiver since it is static method
5114 Node* crc = argument(0); // type: int
5115 Node* src = argument(1); // type: long
5116 Node* offset = argument(3); // type: int
5117 Node* length = argument(4); // type: int
5118
5119 src = ConvL2X(src); // adjust Java long to machine word
5120 Node* base = _gvn.transform(new CastX2PNode(src));
5121 offset = ConvI2X(offset);
5122
5123 // 'src_start' points to src array + scaled offset
5124 Node* src_start = basic_plus_adr(top(), base, offset);
5125
5126 // Call the stub.
5127 address stubAddr = StubRoutines::updateBytesCRC32();
5128 const char *stubName = "updateBytesCRC32";
5129
5130 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5131 stubAddr, stubName, TypePtr::BOTTOM,
5132 crc, src_start, length);
5133 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5134 set_result(result);
5135 return true;
5136 }
5137
5138 //----------------------------inline_reference_get----------------------------
5139 // public T java.lang.ref.Reference.get();
5140 bool LibraryCallKit::inline_reference_get() {
5141 const int referent_offset = java_lang_ref_Reference::referent_offset;
5142 guarantee(referent_offset > 0, "should have already been set");
5143
5144 // Get the argument:
5145 Node* reference_obj = null_check_receiver();
5146 if (stopped()) return true;
5147
5148 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5149
5150 ciInstanceKlass* klass = env()->Object_klass();
5151 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5152
5153 Node* no_ctrl = NULL;
5154 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5155
5156 // Use the pre-barrier to record the value in the referent field
5157 pre_barrier(false /* do_load */,
5158 control(),
5159 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5160 result /* pre_val */,
5161 T_OBJECT);
5162
5163 // Add memory barrier to prevent commoning reads from this field
5164 // across safepoint since GC can change its value.
5165 insert_mem_bar(Op_MemBarCPUOrder);
5166
5167 set_result(result);
5168 return true;
5169 }
5170
5171
5172 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5173 bool is_exact=true, bool is_static=false) {
5174
5175 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5176 assert(tinst != NULL, "obj is null");
5177 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5178 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5179
5180 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5181 ciSymbol::make(fieldTypeString),
5182 is_static);
5183 if (field == NULL) return (Node *) NULL;
5184 assert (field != NULL, "undefined field");
5185
5186 // Next code copied from Parse::do_get_xxx():
5187
5188 // Compute address and memory type.
5189 int offset = field->offset_in_bytes();
5190 bool is_vol = field->is_volatile();
5191 ciType* field_klass = field->type();
5192 assert(field_klass->is_loaded(), "should be loaded");
5193 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5194 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5195 BasicType bt = field->layout_type();
5196
5197 // Build the resultant type of the load
5198 const Type *type;
5199 if (bt == T_OBJECT) {
5200 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5201 } else {
5202 type = Type::get_const_basic_type(bt);
5203 }
5204
5205 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5206 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5207 }
5208 // Build the load.
5209 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5210 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
5211 // If reference is volatile, prevent following memory ops from
5212 // floating up past the volatile read. Also prevents commoning
5213 // another volatile read.
5214 if (is_vol) {
5215 // Memory barrier includes bogus read of value to force load BEFORE membar
5216 insert_mem_bar(Op_MemBarAcquire, loadedField);
5217 }
5218 return loadedField;
5219 }
5220
5221
5222 //------------------------------inline_aescrypt_Block-----------------------
5223 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5224 address stubAddr;
5225 const char *stubName;
5226 assert(UseAES, "need AES instruction support");
5227
5228 switch(id) {
5229 case vmIntrinsics::_aescrypt_encryptBlock:
5230 stubAddr = StubRoutines::aescrypt_encryptBlock();
5231 stubName = "aescrypt_encryptBlock";
5232 break;
5233 case vmIntrinsics::_aescrypt_decryptBlock:
5234 stubAddr = StubRoutines::aescrypt_decryptBlock();
5235 stubName = "aescrypt_decryptBlock";
5236 break;
5237 }
5238 if (stubAddr == NULL) return false;
5239
5240 Node* aescrypt_object = argument(0);
5241 Node* src = argument(1);
5242 Node* src_offset = argument(2);
5243 Node* dest = argument(3);
5244 Node* dest_offset = argument(4);
5245
5246 // (1) src and dest are arrays.
5247 const Type* src_type = src->Value(&_gvn);
5248 const Type* dest_type = dest->Value(&_gvn);
5249 const TypeAryPtr* top_src = src_type->isa_aryptr();
5250 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5251 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5252
5253 // for the quick and dirty code we will skip all the checks.
5254 // we are just trying to get the call to be generated.
5255 Node* src_start = src;
5256 Node* dest_start = dest;
5257 if (src_offset != NULL || dest_offset != NULL) {
5258 assert(src_offset != NULL && dest_offset != NULL, "");
5259 src_start = array_element_address(src, src_offset, T_BYTE);
5260 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5261 }
5262
5263 // now need to get the start of its expanded key array
5264 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5265 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5266 if (k_start == NULL) return false;
5267
5268 if (Matcher::pass_original_key_for_aes()) {
5269 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5270 // compatibility issues between Java key expansion and SPARC crypto instructions
5271 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5272 if (original_k_start == NULL) return false;
5273
5274 // Call the stub.
5275 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5276 stubAddr, stubName, TypePtr::BOTTOM,
5277 src_start, dest_start, k_start, original_k_start);
5278 } else {
5279 // Call the stub.
5280 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5281 stubAddr, stubName, TypePtr::BOTTOM,
5282 src_start, dest_start, k_start);
5283 }
5284
5285 return true;
5286 }
5287
5288 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5289 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5290 address stubAddr;
5291 const char *stubName;
5292
5293 assert(UseAES, "need AES instruction support");
5294
5295 switch(id) {
5296 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5297 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5298 stubName = "cipherBlockChaining_encryptAESCrypt";
5299 break;
5300 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5301 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5302 stubName = "cipherBlockChaining_decryptAESCrypt";
5303 break;
5304 }
5305 if (stubAddr == NULL) return false;
5306
5307 Node* cipherBlockChaining_object = argument(0);
5308 Node* src = argument(1);
5309 Node* src_offset = argument(2);
5310 Node* len = argument(3);
5311 Node* dest = argument(4);
5312 Node* dest_offset = argument(5);
5313
5314 // (1) src and dest are arrays.
5315 const Type* src_type = src->Value(&_gvn);
5316 const Type* dest_type = dest->Value(&_gvn);
5317 const TypeAryPtr* top_src = src_type->isa_aryptr();
5318 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5319 assert (top_src != NULL && top_src->klass() != NULL
5320 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5321
5322 // checks are the responsibility of the caller
5323 Node* src_start = src;
5324 Node* dest_start = dest;
5325 if (src_offset != NULL || dest_offset != NULL) {
5326 assert(src_offset != NULL && dest_offset != NULL, "");
5327 src_start = array_element_address(src, src_offset, T_BYTE);
5328 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5329 }
5330
5331 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5332 // (because of the predicated logic executed earlier).
5333 // so we cast it here safely.
5334 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5335
5336 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5337 if (embeddedCipherObj == NULL) return false;
5338
5339 // cast it to what we know it will be at runtime
5340 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5341 assert(tinst != NULL, "CBC obj is null");
5342 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5343 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5344 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5345
5346 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5347 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5348 const TypeOopPtr* xtype = aklass->as_instance_type();
5349 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5350 aescrypt_object = _gvn.transform(aescrypt_object);
5351
5352 // we need to get the start of the aescrypt_object's expanded key array
5353 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5354 if (k_start == NULL) return false;
5355
5356 // similarly, get the start address of the r vector
5357 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5358 if (objRvec == NULL) return false;
5359 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5360
5361 Node* cbcCrypt;
5362 if (Matcher::pass_original_key_for_aes()) {
5363 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5364 // compatibility issues between Java key expansion and SPARC crypto instructions
5365 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5366 if (original_k_start == NULL) return false;
5367
5368 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5369 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5370 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5371 stubAddr, stubName, TypePtr::BOTTOM,
5372 src_start, dest_start, k_start, r_start, len, original_k_start);
5373 } else {
5374 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5375 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5376 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5377 stubAddr, stubName, TypePtr::BOTTOM,
5378 src_start, dest_start, k_start, r_start, len);
5379 }
5380
5381 // return cipher length (int)
5382 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5383 set_result(retvalue);
5384 return true;
5385 }
5386
5387 //------------------------------get_key_start_from_aescrypt_object-----------------------
5388 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5389 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5390 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5391 if (objAESCryptKey == NULL) return (Node *) NULL;
5392
5393 // now have the array, need to get the start address of the K array
5394 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5395 return k_start;
5396 }
5397
5398 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5399 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5400 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5401 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5402 if (objAESCryptKey == NULL) return (Node *) NULL;
5403
5404 // now have the array, need to get the start address of the lastKey array
5405 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5406 return original_k_start;
5407 }
5408
5409 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5410 // Return node representing slow path of predicate check.
5411 // the pseudo code we want to emulate with this predicate is:
5412 // for encryption:
5413 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5414 // for decryption:
5415 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5416 // note cipher==plain is more conservative than the original java code but that's OK
5417 //
5418 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5419 // The receiver was checked for NULL already.
5420 Node* objCBC = argument(0);
5421
5422 // Load embeddedCipher field of CipherBlockChaining object.
5423 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5424
5425 // get AESCrypt klass for instanceOf check
5426 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5427 // will have same classloader as CipherBlockChaining object
5428 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5429 assert(tinst != NULL, "CBCobj is null");
5430 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5431
5432 // we want to do an instanceof comparison against the AESCrypt class
5433 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5434 if (!klass_AESCrypt->is_loaded()) {
5435 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5436 Node* ctrl = control();
5437 set_control(top()); // no regular fast path
5438 return ctrl;
5439 }
5440 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5441
5442 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5443 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5444 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5445
5446 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5447
5448 // for encryption, we are done
5449 if (!decrypting)
5450 return instof_false; // even if it is NULL
5451
5452 // for decryption, we need to add a further check to avoid
5453 // taking the intrinsic path when cipher and plain are the same
5454 // see the original java code for why.
5455 RegionNode* region = new RegionNode(3);
5456 region->init_req(1, instof_false);
5457 Node* src = argument(1);
5458 Node* dest = argument(4);
5459 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5460 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5461 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5462 region->init_req(2, src_dest_conjoint);
5463
5464 record_for_igvn(region);
5465 return _gvn.transform(region);
5466 }
5467
5468 //------------------------------inline_sha_implCompress-----------------------
5469 //
5470 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5471 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5472 //
5473 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
5474 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
5475 //
5476 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
5477 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
5478 //
5479 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
5480 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
5481
5482 Node* sha_obj = argument(0);
5483 Node* src = argument(1); // type oop
5484 Node* ofs = argument(2); // type int
5485
5486 const Type* src_type = src->Value(&_gvn);
5487 const TypeAryPtr* top_src = src_type->isa_aryptr();
5488 if (top_src == NULL || top_src->klass() == NULL) {
5489 // failed array check
5490 return false;
5491 }
5492 // Figure out the size and type of the elements we will be copying.
5493 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5494 if (src_elem != T_BYTE) {
5495 return false;
5496 }
5497 // 'src_start' points to src array + offset
5498 Node* src_start = array_element_address(src, ofs, src_elem);
5499 Node* state = NULL;
5500 address stubAddr;
5501 const char *stubName;
5502
5503 switch(id) {
5504 case vmIntrinsics::_sha_implCompress:
5505 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
5506 state = get_state_from_sha_object(sha_obj);
5507 stubAddr = StubRoutines::sha1_implCompress();
5508 stubName = "sha1_implCompress";
5509 break;
5510 case vmIntrinsics::_sha2_implCompress:
5511 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
5512 state = get_state_from_sha_object(sha_obj);
5513 stubAddr = StubRoutines::sha256_implCompress();
5514 stubName = "sha256_implCompress";
5515 break;
5516 case vmIntrinsics::_sha5_implCompress:
5517 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
5518 state = get_state_from_sha5_object(sha_obj);
5519 stubAddr = StubRoutines::sha512_implCompress();
5520 stubName = "sha512_implCompress";
5521 break;
5522 default:
5523 fatal_unexpected_iid(id);
5524 return false;
5525 }
5526 if (state == NULL) return false;
5527
5528 // Call the stub.
5529 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
5530 stubAddr, stubName, TypePtr::BOTTOM,
5531 src_start, state);
5532
5533 return true;
5534 }
5535
5536 //------------------------------inline_digestBase_implCompressMB-----------------------
5537 //
5538 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
5539 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
5540 //
5541 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
5542 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5543 "need SHA1/SHA256/SHA512 instruction support");
5544 assert((uint)predicate < 3, "sanity");
5545 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
5546
5547 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
5548 Node* src = argument(1); // byte[] array
5549 Node* ofs = argument(2); // type int
5550 Node* limit = argument(3); // type int
5551
5552 const Type* src_type = src->Value(&_gvn);
5553 const TypeAryPtr* top_src = src_type->isa_aryptr();
5554 if (top_src == NULL || top_src->klass() == NULL) {
5555 // failed array check
5556 return false;
5557 }
5558 // Figure out the size and type of the elements we will be copying.
5559 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5560 if (src_elem != T_BYTE) {
5561 return false;
5562 }
5563 // 'src_start' points to src array + offset
5564 Node* src_start = array_element_address(src, ofs, src_elem);
5565
5566 const char* klass_SHA_name = NULL;
5567 const char* stub_name = NULL;
5568 address stub_addr = NULL;
5569 bool long_state = false;
5570
5571 switch (predicate) {
5572 case 0:
5573 if (UseSHA1Intrinsics) {
5574 klass_SHA_name = "sun/security/provider/SHA";
5575 stub_name = "sha1_implCompressMB";
5576 stub_addr = StubRoutines::sha1_implCompressMB();
5577 }
5578 break;
5579 case 1:
5580 if (UseSHA256Intrinsics) {
5581 klass_SHA_name = "sun/security/provider/SHA2";
5582 stub_name = "sha256_implCompressMB";
5583 stub_addr = StubRoutines::sha256_implCompressMB();
5584 }
5585 break;
5586 case 2:
5587 if (UseSHA512Intrinsics) {
5588 klass_SHA_name = "sun/security/provider/SHA5";
5589 stub_name = "sha512_implCompressMB";
5590 stub_addr = StubRoutines::sha512_implCompressMB();
5591 long_state = true;
5592 }
5593 break;
5594 default:
5595 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5596 }
5597 if (klass_SHA_name != NULL) {
5598 // get DigestBase klass to lookup for SHA klass
5599 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
5600 assert(tinst != NULL, "digestBase_obj is not instance???");
5601 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5602
5603 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5604 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
5605 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5606 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
5607 }
5608 return false;
5609 }
5610 //------------------------------inline_sha_implCompressMB-----------------------
5611 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
5612 bool long_state, address stubAddr, const char *stubName,
5613 Node* src_start, Node* ofs, Node* limit) {
5614 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
5615 const TypeOopPtr* xtype = aklass->as_instance_type();
5616 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
5617 sha_obj = _gvn.transform(sha_obj);
5618
5619 Node* state;
5620 if (long_state) {
5621 state = get_state_from_sha5_object(sha_obj);
5622 } else {
5623 state = get_state_from_sha_object(sha_obj);
5624 }
5625 if (state == NULL) return false;
5626
5627 // Call the stub.
5628 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5629 OptoRuntime::digestBase_implCompressMB_Type(),
5630 stubAddr, stubName, TypePtr::BOTTOM,
5631 src_start, state, ofs, limit);
5632 // return ofs (int)
5633 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5634 set_result(result);
5635
5636 return true;
5637 }
5638
5639 //------------------------------get_state_from_sha_object-----------------------
5640 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
5641 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
5642 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
5643 if (sha_state == NULL) return (Node *) NULL;
5644
5645 // now have the array, need to get the start address of the state array
5646 Node* state = array_element_address(sha_state, intcon(0), T_INT);
5647 return state;
5648 }
5649
5650 //------------------------------get_state_from_sha5_object-----------------------
5651 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
5652 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
5653 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
5654 if (sha_state == NULL) return (Node *) NULL;
5655
5656 // now have the array, need to get the start address of the state array
5657 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
5658 return state;
5659 }
5660
5661 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
5662 // Return node representing slow path of predicate check.
5663 // the pseudo code we want to emulate with this predicate is:
5664 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
5665 //
5666 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
5667 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5668 "need SHA1/SHA256/SHA512 instruction support");
5669 assert((uint)predicate < 3, "sanity");
5670
5671 // The receiver was checked for NULL already.
5672 Node* digestBaseObj = argument(0);
5673
5674 // get DigestBase klass for instanceOf check
5675 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
5676 assert(tinst != NULL, "digestBaseObj is null");
5677 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5678
5679 const char* klass_SHA_name = NULL;
5680 switch (predicate) {
5681 case 0:
5682 if (UseSHA1Intrinsics) {
5683 // we want to do an instanceof comparison against the SHA class
5684 klass_SHA_name = "sun/security/provider/SHA";
5685 }
5686 break;
5687 case 1:
5688 if (UseSHA256Intrinsics) {
5689 // we want to do an instanceof comparison against the SHA2 class
5690 klass_SHA_name = "sun/security/provider/SHA2";
5691 }
5692 break;
5693 case 2:
5694 if (UseSHA512Intrinsics) {
5695 // we want to do an instanceof comparison against the SHA5 class
5696 klass_SHA_name = "sun/security/provider/SHA5";
5697 }
5698 break;
5699 default:
5700 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5701 }
5702
5703 ciKlass* klass_SHA = NULL;
5704 if (klass_SHA_name != NULL) {
5705 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5706 }
5707 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
5708 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
5709 Node* ctrl = control();
5710 set_control(top()); // no intrinsic path
5711 return ctrl;
5712 }
5713 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5714
5715 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
5716 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
5717 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5718 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5719
5720 return instof_false; // even if it is NULL
5721 }