1 /*
2 * Copyright (c) 1999, 2015, 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/arraycopynode.hpp"
34 #include "opto/callGenerator.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/countbitsnode.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/mathexactnode.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "prims/nativeLookup.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "trace/traceMacros.hpp"
52
53 class LibraryIntrinsic : public InlineCallGenerator {
54 // Extend the set of intrinsics known to the runtime:
55 public:
56 private:
57 bool _is_virtual;
58 bool _does_virtual_dispatch;
59 int8_t _predicates_count; // Intrinsic is predicated by several conditions
60 int8_t _last_predicate; // Last generated predicate
61 vmIntrinsics::ID _intrinsic_id;
62
63 public:
64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
65 : InlineCallGenerator(m),
66 _is_virtual(is_virtual),
67 _does_virtual_dispatch(does_virtual_dispatch),
68 _predicates_count((int8_t)predicates_count),
69 _last_predicate((int8_t)-1),
70 _intrinsic_id(id)
71 {
72 }
73 virtual bool is_intrinsic() const { return true; }
74 virtual bool is_virtual() const { return _is_virtual; }
75 virtual bool is_predicated() const { return _predicates_count > 0; }
76 virtual int predicates_count() const { return _predicates_count; }
77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
78 virtual JVMState* generate(JVMState* jvms);
79 virtual Node* generate_predicate(JVMState* jvms, int predicate);
80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
81 };
82
83
84 // Local helper class for LibraryIntrinsic:
85 class LibraryCallKit : public GraphKit {
86 private:
87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
88 Node* _result; // the result node, if any
89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
90
91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
92
93 public:
94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
95 : GraphKit(jvms),
96 _intrinsic(intrinsic),
97 _result(NULL)
98 {
99 // Check if this is a root compile. In that case we don't have a caller.
100 if (!jvms->has_method()) {
101 _reexecute_sp = sp();
102 } else {
103 // Find out how many arguments the interpreter needs when deoptimizing
104 // and save the stack pointer value so it can used by uncommon_trap.
105 // We find the argument count by looking at the declared signature.
106 bool ignored_will_link;
107 ciSignature* declared_signature = NULL;
108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
111 }
112 }
113
114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
115
116 ciMethod* caller() const { return jvms()->method(); }
117 int bci() const { return jvms()->bci(); }
118 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
120 ciMethod* callee() const { return _intrinsic->method(); }
121
122 bool try_to_inline(int predicate);
123 Node* try_to_predicate(int predicate);
124
125 void push_result() {
126 // Push the result onto the stack.
127 if (!stopped() && result() != NULL) {
128 BasicType bt = result()->bottom_type()->basic_type();
129 push_node(bt, result());
130 }
131 }
132
133 private:
134 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
136 }
137
138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
139 void set_result(RegionNode* region, PhiNode* value);
140 Node* result() { return _result; }
141
142 virtual int reexecute_sp() { return _reexecute_sp; }
143
144 // Helper functions to inline natives
145 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
146 Node* generate_slow_guard(Node* test, RegionNode* region);
147 Node* generate_fair_guard(Node* test, RegionNode* region);
148 Node* generate_negative_guard(Node* index, RegionNode* region,
149 // resulting CastII of index:
150 Node* *pos_index = NULL);
151 Node* generate_limit_guard(Node* offset, Node* subseq_length,
152 Node* array_length,
153 RegionNode* region);
154 Node* generate_current_thread(Node* &tls_output);
155 Node* load_mirror_from_klass(Node* klass);
156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
157 RegionNode* region, int null_path,
158 int offset);
159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::array_klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* generate_access_flags_guard(Node* kls,
174 int modifier_mask, int modifier_bits,
175 RegionNode* region);
176 Node* generate_interface_guard(Node* kls, RegionNode* region);
177 Node* generate_array_guard(Node* kls, RegionNode* region) {
178 return generate_array_guard_common(kls, region, false, false);
179 }
180 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, false, true);
182 }
183 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, true, false);
185 }
186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, true, true);
188 }
189 Node* generate_array_guard_common(Node* kls, RegionNode* region,
190 bool obj_array, bool not_array);
191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
193 bool is_virtual = false, bool is_static = false);
194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
195 return generate_method_call(method_id, false, true);
196 }
197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
198 return generate_method_call(method_id, true, false);
199 }
200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
201
202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
203 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
204 bool inline_string_compareTo();
205 bool inline_string_indexOf();
206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
207 bool inline_string_equals();
208 Node* round_double_node(Node* n);
209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
210 bool inline_math_native(vmIntrinsics::ID id);
211 bool inline_trig(vmIntrinsics::ID id);
212 bool inline_math(vmIntrinsics::ID id);
213 template <typename OverflowOp>
214 bool inline_math_overflow(Node* arg1, Node* arg2);
215 void inline_math_mathExact(Node* math, Node* test);
216 bool inline_math_addExactI(bool is_increment);
217 bool inline_math_addExactL(bool is_increment);
218 bool inline_math_multiplyExactI();
219 bool inline_math_multiplyExactL();
220 bool inline_math_negateExactI();
221 bool inline_math_negateExactL();
222 bool inline_math_subtractExactI(bool is_decrement);
223 bool inline_math_subtractExactL(bool is_decrement);
224 bool inline_exp();
225 bool inline_pow();
226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
227 bool inline_min_max(vmIntrinsics::ID id);
228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
229 // This returns Type::AnyPtr, RawPtr, or OopPtr.
230 int classify_unsafe_addr(Node* &base, Node* &offset);
231 Node* make_unsafe_address(Node* base, Node* offset);
232 // Helper for inline_unsafe_access.
233 // Generates the guards that check whether the result of
234 // Unsafe.getObject should be recorded in an SATB log buffer.
235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
237 static bool klass_needs_init_guard(Node* kls);
238 bool inline_unsafe_allocate();
239 bool inline_unsafe_copyMemory();
240 bool inline_native_currentThread();
241 #ifdef TRACE_HAVE_INTRINSICS
242 bool inline_native_classID();
243 bool inline_native_threadID();
244 #endif
245 bool inline_native_time_funcs(address method, const char* funcName);
246 bool inline_native_isInterrupted();
247 bool inline_native_Class_query(vmIntrinsics::ID id);
248 bool inline_native_subtype_check();
249
250 bool inline_native_newArray();
251 bool inline_native_getLength();
252 bool inline_array_copyOf(bool is_copyOfRange);
253 bool inline_array_equals();
254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
255 bool inline_native_clone(bool is_virtual);
256 bool inline_native_Reflection_getCallerClass();
257 // Helper function for inlining native object hash method
258 bool inline_native_hashcode(bool is_virtual, bool is_static);
259 bool inline_native_getClass();
260
261 // Helper functions for inlining arraycopy
262 bool inline_arraycopy();
263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
264 RegionNode* slow_region);
265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
267
268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
270 bool inline_unsafe_ordered_store(BasicType type);
271 bool inline_unsafe_fence(vmIntrinsics::ID id);
272 bool inline_fp_conversions(vmIntrinsics::ID id);
273 bool inline_number_methods(vmIntrinsics::ID id);
274 bool inline_reference_get();
275 bool inline_Class_cast();
276 bool inline_aescrypt_Block(vmIntrinsics::ID id);
277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
281 bool inline_ghash_processBlocks();
282 Node* get_vars_from_ghash_object(Node *ghash_object, const char *var_name);
283 bool inline_sha_implCompress(vmIntrinsics::ID id);
284 bool inline_digestBase_implCompressMB(int predicate);
285 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
286 bool long_state, address stubAddr, const char *stubName,
287 Node* src_start, Node* ofs, Node* limit);
288 Node* get_state_from_sha_object(Node *sha_object);
289 Node* get_state_from_sha5_object(Node *sha_object);
290 Node* inline_digestBase_implCompressMB_predicate(int predicate);
291 bool inline_encodeISOArray();
292 bool inline_updateCRC32();
293 bool inline_updateBytesCRC32();
294 bool inline_updateByteBufferCRC32();
295 bool inline_multiplyToLen();
296 bool inline_squareToLen();
297 bool inline_mulAdd();
298
299 bool inline_profileBoolean();
300 bool inline_isCompileConstant();
301 };
302
303
304 //---------------------------make_vm_intrinsic----------------------------
305 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
306 vmIntrinsics::ID id = m->intrinsic_id();
307 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
308
309 ccstr disable_intr = NULL;
310
311 if ((DisableIntrinsic[0] != '\0'
312 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
313 (method_has_option_value("DisableIntrinsic", disable_intr)
314 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
315 // disabled by a user request on the command line:
316 // example: -XX:DisableIntrinsic=_hashCode,_getClass
317 return NULL;
318 }
319
320 if (!m->is_loaded()) {
321 // do not attempt to inline unloaded methods
322 return NULL;
323 }
324
325 // Only a few intrinsics implement a virtual dispatch.
326 // They are expensive calls which are also frequently overridden.
327 if (is_virtual) {
328 switch (id) {
329 case vmIntrinsics::_hashCode:
330 case vmIntrinsics::_clone:
331 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
332 break;
333 default:
334 return NULL;
335 }
336 }
337
338 // -XX:-InlineNatives disables nearly all intrinsics:
339 if (!InlineNatives) {
340 switch (id) {
341 case vmIntrinsics::_indexOf:
342 case vmIntrinsics::_compareTo:
343 case vmIntrinsics::_equals:
344 case vmIntrinsics::_equalsC:
345 case vmIntrinsics::_getAndAddInt:
346 case vmIntrinsics::_getAndAddLong:
347 case vmIntrinsics::_getAndSetInt:
348 case vmIntrinsics::_getAndSetLong:
349 case vmIntrinsics::_getAndSetObject:
350 case vmIntrinsics::_loadFence:
351 case vmIntrinsics::_storeFence:
352 case vmIntrinsics::_fullFence:
353 break; // InlineNatives does not control String.compareTo
354 case vmIntrinsics::_Reference_get:
355 break; // InlineNatives does not control Reference.get
356 default:
357 return NULL;
358 }
359 }
360
361 int predicates = 0;
362 bool does_virtual_dispatch = false;
363
364 switch (id) {
365 case vmIntrinsics::_compareTo:
366 if (!SpecialStringCompareTo) return NULL;
367 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
368 break;
369 case vmIntrinsics::_indexOf:
370 if (!SpecialStringIndexOf) return NULL;
371 break;
372 case vmIntrinsics::_equals:
373 if (!SpecialStringEquals) return NULL;
374 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
375 break;
376 case vmIntrinsics::_equalsC:
377 if (!SpecialArraysEquals) return NULL;
378 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
379 break;
380 case vmIntrinsics::_arraycopy:
381 if (!InlineArrayCopy) return NULL;
382 break;
383 case vmIntrinsics::_copyMemory:
384 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
385 if (!InlineArrayCopy) return NULL;
386 break;
387 case vmIntrinsics::_hashCode:
388 if (!InlineObjectHash) return NULL;
389 does_virtual_dispatch = true;
390 break;
391 case vmIntrinsics::_clone:
392 does_virtual_dispatch = true;
393 case vmIntrinsics::_copyOf:
394 case vmIntrinsics::_copyOfRange:
395 if (!InlineObjectCopy) return NULL;
396 // These also use the arraycopy intrinsic mechanism:
397 if (!InlineArrayCopy) return NULL;
398 break;
399 case vmIntrinsics::_encodeISOArray:
400 if (!SpecialEncodeISOArray) return NULL;
401 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
402 break;
403 case vmIntrinsics::_checkIndex:
404 // We do not intrinsify this. The optimizer does fine with it.
405 return NULL;
406
407 case vmIntrinsics::_getCallerClass:
408 if (!InlineReflectionGetCallerClass) return NULL;
409 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
410 break;
411
412 case vmIntrinsics::_bitCount_i:
413 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
414 break;
415
416 case vmIntrinsics::_bitCount_l:
417 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
418 break;
419
420 case vmIntrinsics::_numberOfLeadingZeros_i:
421 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
422 break;
423
424 case vmIntrinsics::_numberOfLeadingZeros_l:
425 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
426 break;
427
428 case vmIntrinsics::_numberOfTrailingZeros_i:
429 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
430 break;
431
432 case vmIntrinsics::_numberOfTrailingZeros_l:
433 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
434 break;
435
436 case vmIntrinsics::_reverseBytes_c:
437 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
438 break;
439 case vmIntrinsics::_reverseBytes_s:
440 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
441 break;
442 case vmIntrinsics::_reverseBytes_i:
443 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
444 break;
445 case vmIntrinsics::_reverseBytes_l:
446 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
447 break;
448
449 case vmIntrinsics::_Reference_get:
450 // Use the intrinsic version of Reference.get() so that the value in
451 // the referent field can be registered by the G1 pre-barrier code.
452 // Also add memory barrier to prevent commoning reads from this field
453 // across safepoint since GC can change it value.
454 break;
455
456 case vmIntrinsics::_compareAndSwapObject:
457 #ifdef _LP64
458 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
459 #endif
460 break;
461
462 case vmIntrinsics::_compareAndSwapLong:
463 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
464 break;
465
466 case vmIntrinsics::_getAndAddInt:
467 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
468 break;
469
470 case vmIntrinsics::_getAndAddLong:
471 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
472 break;
473
474 case vmIntrinsics::_getAndSetInt:
475 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
476 break;
477
478 case vmIntrinsics::_getAndSetLong:
479 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndSetObject:
483 #ifdef _LP64
484 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
485 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
486 break;
487 #else
488 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
489 break;
490 #endif
491
492 case vmIntrinsics::_aescrypt_encryptBlock:
493 case vmIntrinsics::_aescrypt_decryptBlock:
494 if (!UseAESIntrinsics) return NULL;
495 break;
496
497 case vmIntrinsics::_multiplyToLen:
498 if (!UseMultiplyToLenIntrinsic) return NULL;
499 break;
500
501 case vmIntrinsics::_squareToLen:
502 if (!UseSquareToLenIntrinsic) return NULL;
503 break;
504
505 case vmIntrinsics::_mulAdd:
506 if (!UseMulAddIntrinsic) return NULL;
507 break;
508
509 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
510 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
511 if (!UseAESIntrinsics) return NULL;
512 // these two require the predicated logic
513 predicates = 1;
514 break;
515
516 case vmIntrinsics::_sha_implCompress:
517 if (!UseSHA1Intrinsics) return NULL;
518 break;
519
520 case vmIntrinsics::_sha2_implCompress:
521 if (!UseSHA256Intrinsics) return NULL;
522 break;
523
524 case vmIntrinsics::_sha5_implCompress:
525 if (!UseSHA512Intrinsics) return NULL;
526 break;
527
528 case vmIntrinsics::_digestBase_implCompressMB:
529 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
530 predicates = 3;
531 break;
532
533 case vmIntrinsics::_ghash_processBlocks:
534 if (!UseGHASHIntrinsics) return NULL;
535 break;
536
537 case vmIntrinsics::_updateCRC32:
538 case vmIntrinsics::_updateBytesCRC32:
539 case vmIntrinsics::_updateByteBufferCRC32:
540 if (!UseCRC32Intrinsics) return NULL;
541 break;
542
543 case vmIntrinsics::_incrementExactI:
544 case vmIntrinsics::_addExactI:
545 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
546 break;
547 case vmIntrinsics::_incrementExactL:
548 case vmIntrinsics::_addExactL:
549 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
550 break;
551 case vmIntrinsics::_decrementExactI:
552 case vmIntrinsics::_subtractExactI:
553 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
554 break;
555 case vmIntrinsics::_decrementExactL:
556 case vmIntrinsics::_subtractExactL:
557 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
558 break;
559 case vmIntrinsics::_negateExactI:
560 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
561 break;
562 case vmIntrinsics::_negateExactL:
563 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
564 break;
565 case vmIntrinsics::_multiplyExactI:
566 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
567 break;
568 case vmIntrinsics::_multiplyExactL:
569 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
570 break;
571
572 case vmIntrinsics::_getShortUnaligned:
573 case vmIntrinsics::_getCharUnaligned:
574 case vmIntrinsics::_getIntUnaligned:
575 case vmIntrinsics::_getLongUnaligned:
576 case vmIntrinsics::_putShortUnaligned:
577 case vmIntrinsics::_putCharUnaligned:
578 case vmIntrinsics::_putIntUnaligned:
579 case vmIntrinsics::_putLongUnaligned:
580 if (!UseUnalignedAccesses) return NULL;
581 break;
582
583 default:
584 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
585 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
586 break;
587 }
588
589 // -XX:-InlineClassNatives disables natives from the Class class.
590 // The flag applies to all reflective calls, notably Array.newArray
591 // (visible to Java programmers as Array.newInstance).
592 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
593 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
594 if (!InlineClassNatives) return NULL;
595 }
596
597 // -XX:-InlineThreadNatives disables natives from the Thread class.
598 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
599 if (!InlineThreadNatives) return NULL;
600 }
601
602 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
603 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
604 m->holder()->name() == ciSymbol::java_lang_Float() ||
605 m->holder()->name() == ciSymbol::java_lang_Double()) {
606 if (!InlineMathNatives) return NULL;
607 }
608
609 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
610 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
611 if (!InlineUnsafeOps) return NULL;
612 }
613
614 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
615 }
616
617 //----------------------register_library_intrinsics-----------------------
618 // Initialize this file's data structures, for each Compile instance.
619 void Compile::register_library_intrinsics() {
620 // Nothing to do here.
621 }
622
623 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
624 LibraryCallKit kit(jvms, this);
625 Compile* C = kit.C;
626 int nodes = C->unique();
627 #ifndef PRODUCT
628 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
629 char buf[1000];
630 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
631 tty->print_cr("Intrinsic %s", str);
632 }
633 #endif
634 ciMethod* callee = kit.callee();
635 const int bci = kit.bci();
636
637 // Try to inline the intrinsic.
638 if (kit.try_to_inline(_last_predicate)) {
639 if (C->print_intrinsics() || C->print_inlining()) {
640 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
641 }
642 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
643 if (C->log()) {
644 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
645 vmIntrinsics::name_at(intrinsic_id()),
646 (is_virtual() ? " virtual='1'" : ""),
647 C->unique() - nodes);
648 }
649 // Push the result from the inlined method onto the stack.
650 kit.push_result();
651 C->print_inlining_update(this);
652 return kit.transfer_exceptions_into_jvms();
653 }
654
655 // The intrinsic bailed out
656 if (C->print_intrinsics() || C->print_inlining()) {
657 if (jvms->has_method()) {
658 // Not a root compile.
659 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
660 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
661 } else {
662 // Root compile
663 tty->print("Did not generate intrinsic %s%s at bci:%d in",
664 vmIntrinsics::name_at(intrinsic_id()),
665 (is_virtual() ? " (virtual)" : ""), bci);
666 }
667 }
668 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
669 C->print_inlining_update(this);
670 return NULL;
671 }
672
673 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
674 LibraryCallKit kit(jvms, this);
675 Compile* C = kit.C;
676 int nodes = C->unique();
677 _last_predicate = predicate;
678 #ifndef PRODUCT
679 assert(is_predicated() && predicate < predicates_count(), "sanity");
680 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
681 char buf[1000];
682 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
683 tty->print_cr("Predicate for intrinsic %s", str);
684 }
685 #endif
686 ciMethod* callee = kit.callee();
687 const int bci = kit.bci();
688
689 Node* slow_ctl = kit.try_to_predicate(predicate);
690 if (!kit.failing()) {
691 if (C->print_intrinsics() || C->print_inlining()) {
692 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
693 }
694 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
695 if (C->log()) {
696 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
697 vmIntrinsics::name_at(intrinsic_id()),
698 (is_virtual() ? " virtual='1'" : ""),
699 C->unique() - nodes);
700 }
701 return slow_ctl; // Could be NULL if the check folds.
702 }
703
704 // The intrinsic bailed out
705 if (C->print_intrinsics() || C->print_inlining()) {
706 if (jvms->has_method()) {
707 // Not a root compile.
708 const char* msg = "failed to generate predicate for intrinsic";
709 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
710 } else {
711 // Root compile
712 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
713 vmIntrinsics::name_at(intrinsic_id()),
714 (is_virtual() ? " (virtual)" : ""), bci);
715 }
716 }
717 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
718 return NULL;
719 }
720
721 bool LibraryCallKit::try_to_inline(int predicate) {
722 // Handle symbolic names for otherwise undistinguished boolean switches:
723 const bool is_store = true;
724 const bool is_native_ptr = true;
725 const bool is_static = true;
726 const bool is_volatile = true;
727
728 if (!jvms()->has_method()) {
729 // Root JVMState has a null method.
730 assert(map()->memory()->Opcode() == Op_Parm, "");
731 // Insert the memory aliasing node
732 set_all_memory(reset_memory());
733 }
734 assert(merged_memory(), "");
735
736
737 switch (intrinsic_id()) {
738 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
739 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
740 case vmIntrinsics::_getClass: return inline_native_getClass();
741
742 case vmIntrinsics::_dsin:
743 case vmIntrinsics::_dcos:
744 case vmIntrinsics::_dtan:
745 case vmIntrinsics::_dabs:
746 case vmIntrinsics::_datan2:
747 case vmIntrinsics::_dsqrt:
748 case vmIntrinsics::_dexp:
749 case vmIntrinsics::_dlog:
750 case vmIntrinsics::_dlog10:
751 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
752
753 case vmIntrinsics::_min:
754 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
755
756 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
757 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
758 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
759 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
760 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
761 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
762 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
763 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
764 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
765 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
766 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
767 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
768
769 case vmIntrinsics::_arraycopy: return inline_arraycopy();
770
771 case vmIntrinsics::_compareTo: return inline_string_compareTo();
772 case vmIntrinsics::_indexOf: return inline_string_indexOf();
773 case vmIntrinsics::_equals: return inline_string_equals();
774
775 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
776 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
777 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
778 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
779 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
780 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
781 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
782 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
783 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
784
785 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
786 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
787 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
788 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
789 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
790 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
791 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
792 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
793 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
794
795 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
796 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
797 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
798 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
799 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
800 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
801 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
802 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
803
804 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
805 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
806 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
807 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
808 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
809 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
810 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
811 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
812
813 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
814 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
815 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
816 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
817 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
818 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
819 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
820 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
821 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
822
823 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
824 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
825 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
826 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
827 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
828 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
829 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
830 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
831 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
832
833 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
834 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
835 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
836 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
837
838 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
839 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
840 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
841 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
842
843 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
844 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
845 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
846
847 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
848 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
849 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
850
851 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
852 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
853 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
854 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
855 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
856
857 case vmIntrinsics::_loadFence:
858 case vmIntrinsics::_storeFence:
859 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
860
861 case vmIntrinsics::_currentThread: return inline_native_currentThread();
862 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
863
864 #ifdef TRACE_HAVE_INTRINSICS
865 case vmIntrinsics::_classID: return inline_native_classID();
866 case vmIntrinsics::_threadID: return inline_native_threadID();
867 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
868 #endif
869 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
870 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
871 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
872 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
873 case vmIntrinsics::_newArray: return inline_native_newArray();
874 case vmIntrinsics::_getLength: return inline_native_getLength();
875 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
876 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
877 case vmIntrinsics::_equalsC: return inline_array_equals();
878 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
879
880 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
881
882 case vmIntrinsics::_isInstance:
883 case vmIntrinsics::_getModifiers:
884 case vmIntrinsics::_isInterface:
885 case vmIntrinsics::_isArray:
886 case vmIntrinsics::_isPrimitive:
887 case vmIntrinsics::_getSuperclass:
888 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
889
890 case vmIntrinsics::_floatToRawIntBits:
891 case vmIntrinsics::_floatToIntBits:
892 case vmIntrinsics::_intBitsToFloat:
893 case vmIntrinsics::_doubleToRawLongBits:
894 case vmIntrinsics::_doubleToLongBits:
895 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
896
897 case vmIntrinsics::_numberOfLeadingZeros_i:
898 case vmIntrinsics::_numberOfLeadingZeros_l:
899 case vmIntrinsics::_numberOfTrailingZeros_i:
900 case vmIntrinsics::_numberOfTrailingZeros_l:
901 case vmIntrinsics::_bitCount_i:
902 case vmIntrinsics::_bitCount_l:
903 case vmIntrinsics::_reverseBytes_i:
904 case vmIntrinsics::_reverseBytes_l:
905 case vmIntrinsics::_reverseBytes_s:
906 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
907
908 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
909
910 case vmIntrinsics::_Reference_get: return inline_reference_get();
911
912 case vmIntrinsics::_Class_cast: return inline_Class_cast();
913
914 case vmIntrinsics::_aescrypt_encryptBlock:
915 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
916
917 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
918 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
919 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
920
921 case vmIntrinsics::_sha_implCompress:
922 case vmIntrinsics::_sha2_implCompress:
923 case vmIntrinsics::_sha5_implCompress:
924 return inline_sha_implCompress(intrinsic_id());
925
926 case vmIntrinsics::_digestBase_implCompressMB:
927 return inline_digestBase_implCompressMB(predicate);
928
929 case vmIntrinsics::_multiplyToLen:
930 return inline_multiplyToLen();
931
932 case vmIntrinsics::_squareToLen:
933 return inline_squareToLen();
934
935 case vmIntrinsics::_mulAdd:
936 return inline_mulAdd();
937
938 case vmIntrinsics::_ghash_processBlocks:
939 return inline_ghash_processBlocks();
940
941 case vmIntrinsics::_encodeISOArray:
942 return inline_encodeISOArray();
943
944 case vmIntrinsics::_updateCRC32:
945 return inline_updateCRC32();
946 case vmIntrinsics::_updateBytesCRC32:
947 return inline_updateBytesCRC32();
948 case vmIntrinsics::_updateByteBufferCRC32:
949 return inline_updateByteBufferCRC32();
950
951 case vmIntrinsics::_profileBoolean:
952 return inline_profileBoolean();
953 case vmIntrinsics::_isCompileConstant:
954 return inline_isCompileConstant();
955
956 default:
957 // If you get here, it may be that someone has added a new intrinsic
958 // to the list in vmSymbols.hpp without implementing it here.
959 #ifndef PRODUCT
960 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
961 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
962 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
963 }
964 #endif
965 return false;
966 }
967 }
968
969 Node* LibraryCallKit::try_to_predicate(int predicate) {
970 if (!jvms()->has_method()) {
971 // Root JVMState has a null method.
972 assert(map()->memory()->Opcode() == Op_Parm, "");
973 // Insert the memory aliasing node
974 set_all_memory(reset_memory());
975 }
976 assert(merged_memory(), "");
977
978 switch (intrinsic_id()) {
979 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
980 return inline_cipherBlockChaining_AESCrypt_predicate(false);
981 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
982 return inline_cipherBlockChaining_AESCrypt_predicate(true);
983 case vmIntrinsics::_digestBase_implCompressMB:
984 return inline_digestBase_implCompressMB_predicate(predicate);
985
986 default:
987 // If you get here, it may be that someone has added a new intrinsic
988 // to the list in vmSymbols.hpp without implementing it here.
989 #ifndef PRODUCT
990 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
991 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
992 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
993 }
994 #endif
995 Node* slow_ctl = control();
996 set_control(top()); // No fast path instrinsic
997 return slow_ctl;
998 }
999 }
1000
1001 //------------------------------set_result-------------------------------
1002 // Helper function for finishing intrinsics.
1003 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1004 record_for_igvn(region);
1005 set_control(_gvn.transform(region));
1006 set_result( _gvn.transform(value));
1007 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1008 }
1009
1010 //------------------------------generate_guard---------------------------
1011 // Helper function for generating guarded fast-slow graph structures.
1012 // The given 'test', if true, guards a slow path. If the test fails
1013 // then a fast path can be taken. (We generally hope it fails.)
1014 // In all cases, GraphKit::control() is updated to the fast path.
1015 // The returned value represents the control for the slow path.
1016 // The return value is never 'top'; it is either a valid control
1017 // or NULL if it is obvious that the slow path can never be taken.
1018 // Also, if region and the slow control are not NULL, the slow edge
1019 // is appended to the region.
1020 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1021 if (stopped()) {
1022 // Already short circuited.
1023 return NULL;
1024 }
1025
1026 // Build an if node and its projections.
1027 // If test is true we take the slow path, which we assume is uncommon.
1028 if (_gvn.type(test) == TypeInt::ZERO) {
1029 // The slow branch is never taken. No need to build this guard.
1030 return NULL;
1031 }
1032
1033 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1034
1035 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1036 if (if_slow == top()) {
1037 // The slow branch is never taken. No need to build this guard.
1038 return NULL;
1039 }
1040
1041 if (region != NULL)
1042 region->add_req(if_slow);
1043
1044 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1045 set_control(if_fast);
1046
1047 return if_slow;
1048 }
1049
1050 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1051 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1052 }
1053 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1054 return generate_guard(test, region, PROB_FAIR);
1055 }
1056
1057 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1058 Node* *pos_index) {
1059 if (stopped())
1060 return NULL; // already stopped
1061 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1062 return NULL; // index is already adequately typed
1063 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1064 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1065 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1066 if (is_neg != NULL && pos_index != NULL) {
1067 // Emulate effect of Parse::adjust_map_after_if.
1068 Node* ccast = new CastIINode(index, TypeInt::POS);
1069 ccast->set_req(0, control());
1070 (*pos_index) = _gvn.transform(ccast);
1071 }
1072 return is_neg;
1073 }
1074
1075 // Make sure that 'position' is a valid limit index, in [0..length].
1076 // There are two equivalent plans for checking this:
1077 // A. (offset + copyLength) unsigned<= arrayLength
1078 // B. offset <= (arrayLength - copyLength)
1079 // We require that all of the values above, except for the sum and
1080 // difference, are already known to be non-negative.
1081 // Plan A is robust in the face of overflow, if offset and copyLength
1082 // are both hugely positive.
1083 //
1084 // Plan B is less direct and intuitive, but it does not overflow at
1085 // all, since the difference of two non-negatives is always
1086 // representable. Whenever Java methods must perform the equivalent
1087 // check they generally use Plan B instead of Plan A.
1088 // For the moment we use Plan A.
1089 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1090 Node* subseq_length,
1091 Node* array_length,
1092 RegionNode* region) {
1093 if (stopped())
1094 return NULL; // already stopped
1095 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1096 if (zero_offset && subseq_length->eqv_uncast(array_length))
1097 return NULL; // common case of whole-array copy
1098 Node* last = subseq_length;
1099 if (!zero_offset) // last += offset
1100 last = _gvn.transform(new AddINode(last, offset));
1101 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1102 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1103 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1104 return is_over;
1105 }
1106
1107
1108 //--------------------------generate_current_thread--------------------
1109 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1110 ciKlass* thread_klass = env()->Thread_klass();
1111 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1112 Node* thread = _gvn.transform(new ThreadLocalNode());
1113 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1114 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1115 tls_output = thread;
1116 return threadObj;
1117 }
1118
1119
1120 //------------------------------make_string_method_node------------------------
1121 // Helper method for String intrinsic functions. This version is called
1122 // with str1 and str2 pointing to String object nodes.
1123 //
1124 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1125 Node* no_ctrl = NULL;
1126
1127 // Get start addr of string
1128 Node* str1_value = load_String_value(no_ctrl, str1);
1129 Node* str1_offset = load_String_offset(no_ctrl, str1);
1130 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1131
1132 // Get length of string 1
1133 Node* str1_len = load_String_length(no_ctrl, str1);
1134
1135 Node* str2_value = load_String_value(no_ctrl, str2);
1136 Node* str2_offset = load_String_offset(no_ctrl, str2);
1137 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1138
1139 Node* str2_len = NULL;
1140 Node* result = NULL;
1141
1142 switch (opcode) {
1143 case Op_StrIndexOf:
1144 // Get length of string 2
1145 str2_len = load_String_length(no_ctrl, str2);
1146
1147 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1148 str1_start, str1_len, str2_start, str2_len);
1149 break;
1150 case Op_StrComp:
1151 // Get length of string 2
1152 str2_len = load_String_length(no_ctrl, str2);
1153
1154 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1155 str1_start, str1_len, str2_start, str2_len);
1156 break;
1157 case Op_StrEquals:
1158 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1159 str1_start, str2_start, str1_len);
1160 break;
1161 default:
1162 ShouldNotReachHere();
1163 return NULL;
1164 }
1165
1166 // All these intrinsics have checks.
1167 C->set_has_split_ifs(true); // Has chance for split-if optimization
1168
1169 return _gvn.transform(result);
1170 }
1171
1172 // Helper method for String intrinsic functions. This version is called
1173 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1174 // to Int nodes containing the lenghts of str1 and str2.
1175 //
1176 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1177 Node* result = NULL;
1178 switch (opcode) {
1179 case Op_StrIndexOf:
1180 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1181 str1_start, cnt1, str2_start, cnt2);
1182 break;
1183 case Op_StrComp:
1184 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1185 str1_start, cnt1, str2_start, cnt2);
1186 break;
1187 case Op_StrEquals:
1188 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1189 str1_start, str2_start, cnt1);
1190 break;
1191 default:
1192 ShouldNotReachHere();
1193 return NULL;
1194 }
1195
1196 // All these intrinsics have checks.
1197 C->set_has_split_ifs(true); // Has chance for split-if optimization
1198
1199 return _gvn.transform(result);
1200 }
1201
1202 //------------------------------inline_string_compareTo------------------------
1203 // public int java.lang.String.compareTo(String anotherString);
1204 bool LibraryCallKit::inline_string_compareTo() {
1205 Node* receiver = null_check(argument(0));
1206 Node* arg = null_check(argument(1));
1207 if (stopped()) {
1208 return true;
1209 }
1210 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1211 return true;
1212 }
1213
1214 //------------------------------inline_string_equals------------------------
1215 bool LibraryCallKit::inline_string_equals() {
1216 Node* receiver = null_check_receiver();
1217 // NOTE: Do not null check argument for String.equals() because spec
1218 // allows to specify NULL as argument.
1219 Node* argument = this->argument(1);
1220 if (stopped()) {
1221 return true;
1222 }
1223
1224 // paths (plus control) merge
1225 RegionNode* region = new RegionNode(5);
1226 Node* phi = new PhiNode(region, TypeInt::BOOL);
1227
1228 // does source == target string?
1229 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1230 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1231
1232 Node* if_eq = generate_slow_guard(bol, NULL);
1233 if (if_eq != NULL) {
1234 // receiver == argument
1235 phi->init_req(2, intcon(1));
1236 region->init_req(2, if_eq);
1237 }
1238
1239 // get String klass for instanceOf
1240 ciInstanceKlass* klass = env()->String_klass();
1241
1242 if (!stopped()) {
1243 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1244 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1245 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1246
1247 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1248 //instanceOf == true, fallthrough
1249
1250 if (inst_false != NULL) {
1251 phi->init_req(3, intcon(0));
1252 region->init_req(3, inst_false);
1253 }
1254 }
1255
1256 if (!stopped()) {
1257 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1258
1259 // Properly cast the argument to String
1260 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1261 // This path is taken only when argument's type is String:NotNull.
1262 argument = cast_not_null(argument, false);
1263
1264 Node* no_ctrl = NULL;
1265
1266 // Get start addr of receiver
1267 Node* receiver_val = load_String_value(no_ctrl, receiver);
1268 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1269 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1270
1271 // Get length of receiver
1272 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1273
1274 // Get start addr of argument
1275 Node* argument_val = load_String_value(no_ctrl, argument);
1276 Node* argument_offset = load_String_offset(no_ctrl, argument);
1277 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1278
1279 // Get length of argument
1280 Node* argument_cnt = load_String_length(no_ctrl, argument);
1281
1282 // Check for receiver count != argument count
1283 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1284 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1285 Node* if_ne = generate_slow_guard(bol, NULL);
1286 if (if_ne != NULL) {
1287 phi->init_req(4, intcon(0));
1288 region->init_req(4, if_ne);
1289 }
1290
1291 // Check for count == 0 is done by assembler code for StrEquals.
1292
1293 if (!stopped()) {
1294 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1295 phi->init_req(1, equals);
1296 region->init_req(1, control());
1297 }
1298 }
1299
1300 // post merge
1301 set_control(_gvn.transform(region));
1302 record_for_igvn(region);
1303
1304 set_result(_gvn.transform(phi));
1305 return true;
1306 }
1307
1308 //------------------------------inline_array_equals----------------------------
1309 bool LibraryCallKit::inline_array_equals() {
1310 Node* arg1 = argument(0);
1311 Node* arg2 = argument(1);
1312 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1313 return true;
1314 }
1315
1316 // Java version of String.indexOf(constant string)
1317 // class StringDecl {
1318 // StringDecl(char[] ca) {
1319 // offset = 0;
1320 // count = ca.length;
1321 // value = ca;
1322 // }
1323 // int offset;
1324 // int count;
1325 // char[] value;
1326 // }
1327 //
1328 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1329 // int targetOffset, int cache_i, int md2) {
1330 // int cache = cache_i;
1331 // int sourceOffset = string_object.offset;
1332 // int sourceCount = string_object.count;
1333 // int targetCount = target_object.length;
1334 //
1335 // int targetCountLess1 = targetCount - 1;
1336 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1337 //
1338 // char[] source = string_object.value;
1339 // char[] target = target_object;
1340 // int lastChar = target[targetCountLess1];
1341 //
1342 // outer_loop:
1343 // for (int i = sourceOffset; i < sourceEnd; ) {
1344 // int src = source[i + targetCountLess1];
1345 // if (src == lastChar) {
1346 // // With random strings and a 4-character alphabet,
1347 // // reverse matching at this point sets up 0.8% fewer
1348 // // frames, but (paradoxically) makes 0.3% more probes.
1349 // // Since those probes are nearer the lastChar probe,
1350 // // there is may be a net D$ win with reverse matching.
1351 // // But, reversing loop inhibits unroll of inner loop
1352 // // for unknown reason. So, does running outer loop from
1353 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1354 // for (int j = 0; j < targetCountLess1; j++) {
1355 // if (target[targetOffset + j] != source[i+j]) {
1356 // if ((cache & (1 << source[i+j])) == 0) {
1357 // if (md2 < j+1) {
1358 // i += j+1;
1359 // continue outer_loop;
1360 // }
1361 // }
1362 // i += md2;
1363 // continue outer_loop;
1364 // }
1365 // }
1366 // return i - sourceOffset;
1367 // }
1368 // if ((cache & (1 << src)) == 0) {
1369 // i += targetCountLess1;
1370 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1371 // i++;
1372 // }
1373 // return -1;
1374 // }
1375
1376 //------------------------------string_indexOf------------------------
1377 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1378 jint cache_i, jint md2_i) {
1379
1380 Node* no_ctrl = NULL;
1381 float likely = PROB_LIKELY(0.9);
1382 float unlikely = PROB_UNLIKELY(0.9);
1383
1384 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1385
1386 Node* source = load_String_value(no_ctrl, string_object);
1387 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1388 Node* sourceCount = load_String_length(no_ctrl, string_object);
1389
1390 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1391 jint target_length = target_array->length();
1392 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1393 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1394
1395 // String.value field is known to be @Stable.
1396 if (UseImplicitStableValues) {
1397 target = cast_array_to_stable(target, target_type);
1398 }
1399
1400 IdealKit kit(this, false, true);
1401 #define __ kit.
1402 Node* zero = __ ConI(0);
1403 Node* one = __ ConI(1);
1404 Node* cache = __ ConI(cache_i);
1405 Node* md2 = __ ConI(md2_i);
1406 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1407 Node* targetCountLess1 = __ ConI(target_length - 1);
1408 Node* targetOffset = __ ConI(targetOffset_i);
1409 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1410
1411 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1412 Node* outer_loop = __ make_label(2 /* goto */);
1413 Node* return_ = __ make_label(1);
1414
1415 __ set(rtn,__ ConI(-1));
1416 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1417 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1418 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1419 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1420 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1421 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1422 Node* tpj = __ AddI(targetOffset, __ value(j));
1423 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1424 Node* ipj = __ AddI(__ value(i), __ value(j));
1425 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1426 __ if_then(targ, BoolTest::ne, src2); {
1427 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1428 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1429 __ increment(i, __ AddI(__ value(j), one));
1430 __ goto_(outer_loop);
1431 } __ end_if(); __ dead(j);
1432 }__ end_if(); __ dead(j);
1433 __ increment(i, md2);
1434 __ goto_(outer_loop);
1435 }__ end_if();
1436 __ increment(j, one);
1437 }__ end_loop(); __ dead(j);
1438 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1439 __ goto_(return_);
1440 }__ end_if();
1441 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1442 __ increment(i, targetCountLess1);
1443 }__ end_if();
1444 __ increment(i, one);
1445 __ bind(outer_loop);
1446 }__ end_loop(); __ dead(i);
1447 __ bind(return_);
1448
1449 // Final sync IdealKit and GraphKit.
1450 final_sync(kit);
1451 Node* result = __ value(rtn);
1452 #undef __
1453 C->set_has_loops(true);
1454 return result;
1455 }
1456
1457 //------------------------------inline_string_indexOf------------------------
1458 bool LibraryCallKit::inline_string_indexOf() {
1459 Node* receiver = argument(0);
1460 Node* arg = argument(1);
1461
1462 Node* result;
1463 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1464 UseSSE42Intrinsics) {
1465 // Generate SSE4.2 version of indexOf
1466 // We currently only have match rules that use SSE4.2
1467
1468 receiver = null_check(receiver);
1469 arg = null_check(arg);
1470 if (stopped()) {
1471 return true;
1472 }
1473
1474 // Make the merge point
1475 RegionNode* result_rgn = new RegionNode(4);
1476 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1477 Node* no_ctrl = NULL;
1478
1479 // Get start addr of source string
1480 Node* source = load_String_value(no_ctrl, receiver);
1481 Node* source_offset = load_String_offset(no_ctrl, receiver);
1482 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1483
1484 // Get length of source string
1485 Node* source_cnt = load_String_length(no_ctrl, receiver);
1486
1487 // Get start addr of substring
1488 Node* substr = load_String_value(no_ctrl, arg);
1489 Node* substr_offset = load_String_offset(no_ctrl, arg);
1490 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1491
1492 // Get length of source string
1493 Node* substr_cnt = load_String_length(no_ctrl, arg);
1494
1495 // Check for substr count > string count
1496 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1497 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1498 Node* if_gt = generate_slow_guard(bol, NULL);
1499 if (if_gt != NULL) {
1500 result_phi->init_req(2, intcon(-1));
1501 result_rgn->init_req(2, if_gt);
1502 }
1503
1504 if (!stopped()) {
1505 // Check for substr count == 0
1506 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1507 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1508 Node* if_zero = generate_slow_guard(bol, NULL);
1509 if (if_zero != NULL) {
1510 result_phi->init_req(3, intcon(0));
1511 result_rgn->init_req(3, if_zero);
1512 }
1513 }
1514
1515 if (!stopped()) {
1516 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1517 result_phi->init_req(1, result);
1518 result_rgn->init_req(1, control());
1519 }
1520 set_control(_gvn.transform(result_rgn));
1521 record_for_igvn(result_rgn);
1522 result = _gvn.transform(result_phi);
1523
1524 } else { // Use LibraryCallKit::string_indexOf
1525 // don't intrinsify if argument isn't a constant string.
1526 if (!arg->is_Con()) {
1527 return false;
1528 }
1529 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1530 if (str_type == NULL) {
1531 return false;
1532 }
1533 ciInstanceKlass* klass = env()->String_klass();
1534 ciObject* str_const = str_type->const_oop();
1535 if (str_const == NULL || str_const->klass() != klass) {
1536 return false;
1537 }
1538 ciInstance* str = str_const->as_instance();
1539 assert(str != NULL, "must be instance");
1540
1541 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1542 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1543
1544 int o;
1545 int c;
1546 if (java_lang_String::has_offset_field()) {
1547 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1548 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1549 } else {
1550 o = 0;
1551 c = pat->length();
1552 }
1553
1554 // constant strings have no offset and count == length which
1555 // simplifies the resulting code somewhat so lets optimize for that.
1556 if (o != 0 || c != pat->length()) {
1557 return false;
1558 }
1559
1560 receiver = null_check(receiver, T_OBJECT);
1561 // NOTE: No null check on the argument is needed since it's a constant String oop.
1562 if (stopped()) {
1563 return true;
1564 }
1565
1566 // The null string as a pattern always returns 0 (match at beginning of string)
1567 if (c == 0) {
1568 set_result(intcon(0));
1569 return true;
1570 }
1571
1572 // Generate default indexOf
1573 jchar lastChar = pat->char_at(o + (c - 1));
1574 int cache = 0;
1575 int i;
1576 for (i = 0; i < c - 1; i++) {
1577 assert(i < pat->length(), "out of range");
1578 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1579 }
1580
1581 int md2 = c;
1582 for (i = 0; i < c - 1; i++) {
1583 assert(i < pat->length(), "out of range");
1584 if (pat->char_at(o + i) == lastChar) {
1585 md2 = (c - 1) - i;
1586 }
1587 }
1588
1589 result = string_indexOf(receiver, pat, o, cache, md2);
1590 }
1591 set_result(result);
1592 return true;
1593 }
1594
1595 //--------------------------round_double_node--------------------------------
1596 // Round a double node if necessary.
1597 Node* LibraryCallKit::round_double_node(Node* n) {
1598 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1599 n = _gvn.transform(new RoundDoubleNode(0, n));
1600 return n;
1601 }
1602
1603 //------------------------------inline_math-----------------------------------
1604 // public static double Math.abs(double)
1605 // public static double Math.sqrt(double)
1606 // public static double Math.log(double)
1607 // public static double Math.log10(double)
1608 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1609 Node* arg = round_double_node(argument(0));
1610 Node* n;
1611 switch (id) {
1612 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1613 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1614 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1615 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1616 default: fatal_unexpected_iid(id); break;
1617 }
1618 set_result(_gvn.transform(n));
1619 return true;
1620 }
1621
1622 //------------------------------inline_trig----------------------------------
1623 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1624 // argument reduction which will turn into a fast/slow diamond.
1625 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1626 Node* arg = round_double_node(argument(0));
1627 Node* n = NULL;
1628
1629 switch (id) {
1630 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1631 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1632 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1633 default: fatal_unexpected_iid(id); break;
1634 }
1635 n = _gvn.transform(n);
1636
1637 // Rounding required? Check for argument reduction!
1638 if (Matcher::strict_fp_requires_explicit_rounding) {
1639 static const double pi_4 = 0.7853981633974483;
1640 static const double neg_pi_4 = -0.7853981633974483;
1641 // pi/2 in 80-bit extended precision
1642 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1643 // -pi/2 in 80-bit extended precision
1644 // 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};
1645 // Cutoff value for using this argument reduction technique
1646 //static const double pi_2_minus_epsilon = 1.564660403643354;
1647 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1648
1649 // Pseudocode for sin:
1650 // if (x <= Math.PI / 4.0) {
1651 // if (x >= -Math.PI / 4.0) return fsin(x);
1652 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1653 // } else {
1654 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1655 // }
1656 // return StrictMath.sin(x);
1657
1658 // Pseudocode for cos:
1659 // if (x <= Math.PI / 4.0) {
1660 // if (x >= -Math.PI / 4.0) return fcos(x);
1661 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1662 // } else {
1663 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1664 // }
1665 // return StrictMath.cos(x);
1666
1667 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1668 // requires a special machine instruction to load it. Instead we'll try
1669 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1670 // probably do the math inside the SIN encoding.
1671
1672 // Make the merge point
1673 RegionNode* r = new RegionNode(3);
1674 Node* phi = new PhiNode(r, Type::DOUBLE);
1675
1676 // Flatten arg so we need only 1 test
1677 Node *abs = _gvn.transform(new AbsDNode(arg));
1678 // Node for PI/4 constant
1679 Node *pi4 = makecon(TypeD::make(pi_4));
1680 // Check PI/4 : abs(arg)
1681 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1682 // Check: If PI/4 < abs(arg) then go slow
1683 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1684 // Branch either way
1685 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1686 set_control(opt_iff(r,iff));
1687
1688 // Set fast path result
1689 phi->init_req(2, n);
1690
1691 // Slow path - non-blocking leaf call
1692 Node* call = NULL;
1693 switch (id) {
1694 case vmIntrinsics::_dsin:
1695 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1696 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1697 "Sin", NULL, arg, top());
1698 break;
1699 case vmIntrinsics::_dcos:
1700 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1701 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1702 "Cos", NULL, arg, top());
1703 break;
1704 case vmIntrinsics::_dtan:
1705 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1706 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1707 "Tan", NULL, arg, top());
1708 break;
1709 }
1710 assert(control()->in(0) == call, "");
1711 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1712 r->init_req(1, control());
1713 phi->init_req(1, slow_result);
1714
1715 // Post-merge
1716 set_control(_gvn.transform(r));
1717 record_for_igvn(r);
1718 n = _gvn.transform(phi);
1719
1720 C->set_has_split_ifs(true); // Has chance for split-if optimization
1721 }
1722 set_result(n);
1723 return true;
1724 }
1725
1726 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1727 //-------------------
1728 //result=(result.isNaN())? funcAddr():result;
1729 // Check: If isNaN() by checking result!=result? then either trap
1730 // or go to runtime
1731 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1732 // Build the boolean node
1733 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1734
1735 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1736 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1737 // The pow or exp intrinsic returned a NaN, which requires a call
1738 // to the runtime. Recompile with the runtime call.
1739 uncommon_trap(Deoptimization::Reason_intrinsic,
1740 Deoptimization::Action_make_not_entrant);
1741 }
1742 return result;
1743 } else {
1744 // If this inlining ever returned NaN in the past, we compile a call
1745 // to the runtime to properly handle corner cases
1746
1747 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1748 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1749 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1750
1751 if (!if_slow->is_top()) {
1752 RegionNode* result_region = new RegionNode(3);
1753 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1754
1755 result_region->init_req(1, if_fast);
1756 result_val->init_req(1, result);
1757
1758 set_control(if_slow);
1759
1760 const TypePtr* no_memory_effects = NULL;
1761 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1762 no_memory_effects,
1763 x, top(), y, y ? top() : NULL);
1764 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1765 #ifdef ASSERT
1766 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1767 assert(value_top == top(), "second value must be top");
1768 #endif
1769
1770 result_region->init_req(2, control());
1771 result_val->init_req(2, value);
1772 set_control(_gvn.transform(result_region));
1773 return _gvn.transform(result_val);
1774 } else {
1775 return result;
1776 }
1777 }
1778 }
1779
1780 //------------------------------inline_exp-------------------------------------
1781 // Inline exp instructions, if possible. The Intel hardware only misses
1782 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1783 bool LibraryCallKit::inline_exp() {
1784 Node* arg = round_double_node(argument(0));
1785 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1786
1787 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1788 set_result(n);
1789
1790 C->set_has_split_ifs(true); // Has chance for split-if optimization
1791 return true;
1792 }
1793
1794 //------------------------------inline_pow-------------------------------------
1795 // Inline power instructions, if possible.
1796 bool LibraryCallKit::inline_pow() {
1797 // Pseudocode for pow
1798 // if (y == 2) {
1799 // return x * x;
1800 // } else {
1801 // if (x <= 0.0) {
1802 // long longy = (long)y;
1803 // if ((double)longy == y) { // if y is long
1804 // if (y + 1 == y) longy = 0; // huge number: even
1805 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1806 // } else {
1807 // result = NaN;
1808 // }
1809 // } else {
1810 // result = DPow(x,y);
1811 // }
1812 // if (result != result)? {
1813 // result = uncommon_trap() or runtime_call();
1814 // }
1815 // return result;
1816 // }
1817
1818 Node* x = round_double_node(argument(0));
1819 Node* y = round_double_node(argument(2));
1820
1821 Node* result = NULL;
1822
1823 Node* const_two_node = makecon(TypeD::make(2.0));
1824 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1825 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1826 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1827 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1828 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1829
1830 RegionNode* region_node = new RegionNode(3);
1831 region_node->init_req(1, if_true);
1832
1833 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1834 // special case for x^y where y == 2, we can convert it to x * x
1835 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1836
1837 // set control to if_false since we will now process the false branch
1838 set_control(if_false);
1839
1840 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1841 // Short form: skip the fancy tests and just check for NaN result.
1842 result = _gvn.transform(new PowDNode(C, control(), x, y));
1843 } else {
1844 // If this inlining ever returned NaN in the past, include all
1845 // checks + call to the runtime.
1846
1847 // Set the merge point for If node with condition of (x <= 0.0)
1848 // There are four possible paths to region node and phi node
1849 RegionNode *r = new RegionNode(4);
1850 Node *phi = new PhiNode(r, Type::DOUBLE);
1851
1852 // Build the first if node: if (x <= 0.0)
1853 // Node for 0 constant
1854 Node *zeronode = makecon(TypeD::ZERO);
1855 // Check x:0
1856 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1857 // Check: If (x<=0) then go complex path
1858 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1859 // Branch either way
1860 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1861 // Fast path taken; set region slot 3
1862 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1863 r->init_req(3,fast_taken); // Capture fast-control
1864
1865 // Fast path not-taken, i.e. slow path
1866 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1867
1868 // Set fast path result
1869 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1870 phi->init_req(3, fast_result);
1871
1872 // Complex path
1873 // Build the second if node (if y is long)
1874 // Node for (long)y
1875 Node *longy = _gvn.transform(new ConvD2LNode(y));
1876 // Node for (double)((long) y)
1877 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1878 // Check (double)((long) y) : y
1879 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1880 // Check if (y isn't long) then go to slow path
1881
1882 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1883 // Branch either way
1884 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1885 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1886
1887 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1888
1889 // Calculate DPow(abs(x), y)*(1 & (long)y)
1890 // Node for constant 1
1891 Node *conone = longcon(1);
1892 // 1& (long)y
1893 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1894
1895 // A huge number is always even. Detect a huge number by checking
1896 // if y + 1 == y and set integer to be tested for parity to 0.
1897 // Required for corner case:
1898 // (long)9.223372036854776E18 = max_jlong
1899 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1900 // max_jlong is odd but 9.223372036854776E18 is even
1901 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1902 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1903 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1904 Node* correctedsign = NULL;
1905 if (ConditionalMoveLimit != 0) {
1906 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1907 } else {
1908 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1909 RegionNode *r = new RegionNode(3);
1910 Node *phi = new PhiNode(r, TypeLong::LONG);
1911 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1912 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1913 phi->init_req(1, signnode);
1914 phi->init_req(2, longcon(0));
1915 correctedsign = _gvn.transform(phi);
1916 ylong_path = _gvn.transform(r);
1917 record_for_igvn(r);
1918 }
1919
1920 // zero node
1921 Node *conzero = longcon(0);
1922 // Check (1&(long)y)==0?
1923 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1924 // Check if (1&(long)y)!=0?, if so the result is negative
1925 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1926 // abs(x)
1927 Node *absx=_gvn.transform(new AbsDNode(x));
1928 // abs(x)^y
1929 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1930 // -abs(x)^y
1931 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1932 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1933 Node *signresult = NULL;
1934 if (ConditionalMoveLimit != 0) {
1935 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1936 } else {
1937 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1938 RegionNode *r = new RegionNode(3);
1939 Node *phi = new PhiNode(r, Type::DOUBLE);
1940 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1941 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1942 phi->init_req(1, absxpowy);
1943 phi->init_req(2, negabsxpowy);
1944 signresult = _gvn.transform(phi);
1945 ylong_path = _gvn.transform(r);
1946 record_for_igvn(r);
1947 }
1948 // Set complex path fast result
1949 r->init_req(2, ylong_path);
1950 phi->init_req(2, signresult);
1951
1952 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1953 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1954 r->init_req(1,slow_path);
1955 phi->init_req(1,slow_result);
1956
1957 // Post merge
1958 set_control(_gvn.transform(r));
1959 record_for_igvn(r);
1960 result = _gvn.transform(phi);
1961 }
1962
1963 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1964
1965 // control from finish_pow_exp is now input to the region node
1966 region_node->set_req(2, control());
1967 // the result from finish_pow_exp is now input to the phi node
1968 phi_node->init_req(2, result);
1969 set_control(_gvn.transform(region_node));
1970 record_for_igvn(region_node);
1971 set_result(_gvn.transform(phi_node));
1972
1973 C->set_has_split_ifs(true); // Has chance for split-if optimization
1974 return true;
1975 }
1976
1977 //------------------------------runtime_math-----------------------------
1978 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1979 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1980 "must be (DD)D or (D)D type");
1981
1982 // Inputs
1983 Node* a = round_double_node(argument(0));
1984 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1985
1986 const TypePtr* no_memory_effects = NULL;
1987 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1988 no_memory_effects,
1989 a, top(), b, b ? top() : NULL);
1990 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1991 #ifdef ASSERT
1992 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1993 assert(value_top == top(), "second value must be top");
1994 #endif
1995
1996 set_result(value);
1997 return true;
1998 }
1999
2000 //------------------------------inline_math_native-----------------------------
2001 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2002 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2003 switch (id) {
2004 // These intrinsics are not properly supported on all hardware
2005 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2006 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2007 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2008 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2009 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2010 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2011
2012 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2013 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2014 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2015 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2016
2017 // These intrinsics are supported on all hardware
2018 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2019 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2020
2021 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2022 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2023 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2024 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2025 #undef FN_PTR
2026
2027 // These intrinsics are not yet correctly implemented
2028 case vmIntrinsics::_datan2:
2029 return false;
2030
2031 default:
2032 fatal_unexpected_iid(id);
2033 return false;
2034 }
2035 }
2036
2037 static bool is_simple_name(Node* n) {
2038 return (n->req() == 1 // constant
2039 || (n->is_Type() && n->as_Type()->type()->singleton())
2040 || n->is_Proj() // parameter or return value
2041 || n->is_Phi() // local of some sort
2042 );
2043 }
2044
2045 //----------------------------inline_min_max-----------------------------------
2046 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2047 set_result(generate_min_max(id, argument(0), argument(1)));
2048 return true;
2049 }
2050
2051 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2052 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2053 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2054 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2055 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2056
2057 {
2058 PreserveJVMState pjvms(this);
2059 PreserveReexecuteState preexecs(this);
2060 jvms()->set_should_reexecute(true);
2061
2062 set_control(slow_path);
2063 set_i_o(i_o());
2064
2065 uncommon_trap(Deoptimization::Reason_intrinsic,
2066 Deoptimization::Action_none);
2067 }
2068
2069 set_control(fast_path);
2070 set_result(math);
2071 }
2072
2073 template <typename OverflowOp>
2074 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2075 typedef typename OverflowOp::MathOp MathOp;
2076
2077 MathOp* mathOp = new MathOp(arg1, arg2);
2078 Node* operation = _gvn.transform( mathOp );
2079 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2080 inline_math_mathExact(operation, ofcheck);
2081 return true;
2082 }
2083
2084 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2085 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2086 }
2087
2088 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2089 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2090 }
2091
2092 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2093 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2094 }
2095
2096 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2097 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2098 }
2099
2100 bool LibraryCallKit::inline_math_negateExactI() {
2101 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2102 }
2103
2104 bool LibraryCallKit::inline_math_negateExactL() {
2105 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2106 }
2107
2108 bool LibraryCallKit::inline_math_multiplyExactI() {
2109 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2110 }
2111
2112 bool LibraryCallKit::inline_math_multiplyExactL() {
2113 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2114 }
2115
2116 Node*
2117 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2118 // These are the candidate return value:
2119 Node* xvalue = x0;
2120 Node* yvalue = y0;
2121
2122 if (xvalue == yvalue) {
2123 return xvalue;
2124 }
2125
2126 bool want_max = (id == vmIntrinsics::_max);
2127
2128 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2129 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2130 if (txvalue == NULL || tyvalue == NULL) return top();
2131 // This is not really necessary, but it is consistent with a
2132 // hypothetical MaxINode::Value method:
2133 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2134
2135 // %%% This folding logic should (ideally) be in a different place.
2136 // Some should be inside IfNode, and there to be a more reliable
2137 // transformation of ?: style patterns into cmoves. We also want
2138 // more powerful optimizations around cmove and min/max.
2139
2140 // Try to find a dominating comparison of these guys.
2141 // It can simplify the index computation for Arrays.copyOf
2142 // and similar uses of System.arraycopy.
2143 // First, compute the normalized version of CmpI(x, y).
2144 int cmp_op = Op_CmpI;
2145 Node* xkey = xvalue;
2146 Node* ykey = yvalue;
2147 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2148 if (ideal_cmpxy->is_Cmp()) {
2149 // E.g., if we have CmpI(length - offset, count),
2150 // it might idealize to CmpI(length, count + offset)
2151 cmp_op = ideal_cmpxy->Opcode();
2152 xkey = ideal_cmpxy->in(1);
2153 ykey = ideal_cmpxy->in(2);
2154 }
2155
2156 // Start by locating any relevant comparisons.
2157 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2158 Node* cmpxy = NULL;
2159 Node* cmpyx = NULL;
2160 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2161 Node* cmp = start_from->fast_out(k);
2162 if (cmp->outcnt() > 0 && // must have prior uses
2163 cmp->in(0) == NULL && // must be context-independent
2164 cmp->Opcode() == cmp_op) { // right kind of compare
2165 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2166 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2167 }
2168 }
2169
2170 const int NCMPS = 2;
2171 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2172 int cmpn;
2173 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2174 if (cmps[cmpn] != NULL) break; // find a result
2175 }
2176 if (cmpn < NCMPS) {
2177 // Look for a dominating test that tells us the min and max.
2178 int depth = 0; // Limit search depth for speed
2179 Node* dom = control();
2180 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2181 if (++depth >= 100) break;
2182 Node* ifproj = dom;
2183 if (!ifproj->is_Proj()) continue;
2184 Node* iff = ifproj->in(0);
2185 if (!iff->is_If()) continue;
2186 Node* bol = iff->in(1);
2187 if (!bol->is_Bool()) continue;
2188 Node* cmp = bol->in(1);
2189 if (cmp == NULL) continue;
2190 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2191 if (cmps[cmpn] == cmp) break;
2192 if (cmpn == NCMPS) continue;
2193 BoolTest::mask btest = bol->as_Bool()->_test._test;
2194 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2195 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2196 // At this point, we know that 'x btest y' is true.
2197 switch (btest) {
2198 case BoolTest::eq:
2199 // They are proven equal, so we can collapse the min/max.
2200 // Either value is the answer. Choose the simpler.
2201 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2202 return yvalue;
2203 return xvalue;
2204 case BoolTest::lt: // x < y
2205 case BoolTest::le: // x <= y
2206 return (want_max ? yvalue : xvalue);
2207 case BoolTest::gt: // x > y
2208 case BoolTest::ge: // x >= y
2209 return (want_max ? xvalue : yvalue);
2210 }
2211 }
2212 }
2213
2214 // We failed to find a dominating test.
2215 // Let's pick a test that might GVN with prior tests.
2216 Node* best_bol = NULL;
2217 BoolTest::mask best_btest = BoolTest::illegal;
2218 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2219 Node* cmp = cmps[cmpn];
2220 if (cmp == NULL) continue;
2221 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2222 Node* bol = cmp->fast_out(j);
2223 if (!bol->is_Bool()) continue;
2224 BoolTest::mask btest = bol->as_Bool()->_test._test;
2225 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2226 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2227 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2228 best_bol = bol->as_Bool();
2229 best_btest = btest;
2230 }
2231 }
2232 }
2233
2234 Node* answer_if_true = NULL;
2235 Node* answer_if_false = NULL;
2236 switch (best_btest) {
2237 default:
2238 if (cmpxy == NULL)
2239 cmpxy = ideal_cmpxy;
2240 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2241 // and fall through:
2242 case BoolTest::lt: // x < y
2243 case BoolTest::le: // x <= y
2244 answer_if_true = (want_max ? yvalue : xvalue);
2245 answer_if_false = (want_max ? xvalue : yvalue);
2246 break;
2247 case BoolTest::gt: // x > y
2248 case BoolTest::ge: // x >= y
2249 answer_if_true = (want_max ? xvalue : yvalue);
2250 answer_if_false = (want_max ? yvalue : xvalue);
2251 break;
2252 }
2253
2254 jint hi, lo;
2255 if (want_max) {
2256 // We can sharpen the minimum.
2257 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2258 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2259 } else {
2260 // We can sharpen the maximum.
2261 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2262 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2263 }
2264
2265 // Use a flow-free graph structure, to avoid creating excess control edges
2266 // which could hinder other optimizations.
2267 // Since Math.min/max is often used with arraycopy, we want
2268 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2269 Node* cmov = CMoveNode::make(NULL, best_bol,
2270 answer_if_false, answer_if_true,
2271 TypeInt::make(lo, hi, widen));
2272
2273 return _gvn.transform(cmov);
2274
2275 /*
2276 // This is not as desirable as it may seem, since Min and Max
2277 // nodes do not have a full set of optimizations.
2278 // And they would interfere, anyway, with 'if' optimizations
2279 // and with CMoveI canonical forms.
2280 switch (id) {
2281 case vmIntrinsics::_min:
2282 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2283 case vmIntrinsics::_max:
2284 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2285 default:
2286 ShouldNotReachHere();
2287 }
2288 */
2289 }
2290
2291 inline int
2292 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2293 const TypePtr* base_type = TypePtr::NULL_PTR;
2294 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2295 if (base_type == NULL) {
2296 // Unknown type.
2297 return Type::AnyPtr;
2298 } else if (base_type == TypePtr::NULL_PTR) {
2299 // Since this is a NULL+long form, we have to switch to a rawptr.
2300 base = _gvn.transform(new CastX2PNode(offset));
2301 offset = MakeConX(0);
2302 return Type::RawPtr;
2303 } else if (base_type->base() == Type::RawPtr) {
2304 return Type::RawPtr;
2305 } else if (base_type->isa_oopptr()) {
2306 // Base is never null => always a heap address.
2307 if (base_type->ptr() == TypePtr::NotNull) {
2308 return Type::OopPtr;
2309 }
2310 // Offset is small => always a heap address.
2311 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2312 if (offset_type != NULL &&
2313 base_type->offset() == 0 && // (should always be?)
2314 offset_type->_lo >= 0 &&
2315 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2316 return Type::OopPtr;
2317 }
2318 // Otherwise, it might either be oop+off or NULL+addr.
2319 return Type::AnyPtr;
2320 } else {
2321 // No information:
2322 return Type::AnyPtr;
2323 }
2324 }
2325
2326 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2327 int kind = classify_unsafe_addr(base, offset);
2328 if (kind == Type::RawPtr) {
2329 return basic_plus_adr(top(), base, offset);
2330 } else {
2331 return basic_plus_adr(base, offset);
2332 }
2333 }
2334
2335 //--------------------------inline_number_methods-----------------------------
2336 // inline int Integer.numberOfLeadingZeros(int)
2337 // inline int Long.numberOfLeadingZeros(long)
2338 //
2339 // inline int Integer.numberOfTrailingZeros(int)
2340 // inline int Long.numberOfTrailingZeros(long)
2341 //
2342 // inline int Integer.bitCount(int)
2343 // inline int Long.bitCount(long)
2344 //
2345 // inline char Character.reverseBytes(char)
2346 // inline short Short.reverseBytes(short)
2347 // inline int Integer.reverseBytes(int)
2348 // inline long Long.reverseBytes(long)
2349 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2350 Node* arg = argument(0);
2351 Node* n;
2352 switch (id) {
2353 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2354 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2355 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2356 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2357 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2358 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2359 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2360 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2361 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2362 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2363 default: fatal_unexpected_iid(id); break;
2364 }
2365 set_result(_gvn.transform(n));
2366 return true;
2367 }
2368
2369 //----------------------------inline_unsafe_access----------------------------
2370
2371 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2372
2373 // Helper that guards and inserts a pre-barrier.
2374 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2375 Node* pre_val, bool need_mem_bar) {
2376 // We could be accessing the referent field of a reference object. If so, when G1
2377 // is enabled, we need to log the value in the referent field in an SATB buffer.
2378 // This routine performs some compile time filters and generates suitable
2379 // runtime filters that guard the pre-barrier code.
2380 // Also add memory barrier for non volatile load from the referent field
2381 // to prevent commoning of loads across safepoint.
2382 if (!UseG1GC && !need_mem_bar)
2383 return;
2384
2385 // Some compile time checks.
2386
2387 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2388 const TypeX* otype = offset->find_intptr_t_type();
2389 if (otype != NULL && otype->is_con() &&
2390 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2391 // Constant offset but not the reference_offset so just return
2392 return;
2393 }
2394
2395 // We only need to generate the runtime guards for instances.
2396 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2397 if (btype != NULL) {
2398 if (btype->isa_aryptr()) {
2399 // Array type so nothing to do
2400 return;
2401 }
2402
2403 const TypeInstPtr* itype = btype->isa_instptr();
2404 if (itype != NULL) {
2405 // Can the klass of base_oop be statically determined to be
2406 // _not_ a sub-class of Reference and _not_ Object?
2407 ciKlass* klass = itype->klass();
2408 if ( klass->is_loaded() &&
2409 !klass->is_subtype_of(env()->Reference_klass()) &&
2410 !env()->Object_klass()->is_subtype_of(klass)) {
2411 return;
2412 }
2413 }
2414 }
2415
2416 // The compile time filters did not reject base_oop/offset so
2417 // we need to generate the following runtime filters
2418 //
2419 // if (offset == java_lang_ref_Reference::_reference_offset) {
2420 // if (instance_of(base, java.lang.ref.Reference)) {
2421 // pre_barrier(_, pre_val, ...);
2422 // }
2423 // }
2424
2425 float likely = PROB_LIKELY( 0.999);
2426 float unlikely = PROB_UNLIKELY(0.999);
2427
2428 IdealKit ideal(this);
2429 #define __ ideal.
2430
2431 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2432
2433 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2434 // Update graphKit memory and control from IdealKit.
2435 sync_kit(ideal);
2436
2437 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2438 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2439
2440 // Update IdealKit memory and control from graphKit.
2441 __ sync_kit(this);
2442
2443 Node* one = __ ConI(1);
2444 // is_instof == 0 if base_oop == NULL
2445 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2446
2447 // Update graphKit from IdeakKit.
2448 sync_kit(ideal);
2449
2450 // Use the pre-barrier to record the value in the referent field
2451 pre_barrier(false /* do_load */,
2452 __ ctrl(),
2453 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2454 pre_val /* pre_val */,
2455 T_OBJECT);
2456 if (need_mem_bar) {
2457 // Add memory barrier to prevent commoning reads from this field
2458 // across safepoint since GC can change its value.
2459 insert_mem_bar(Op_MemBarCPUOrder);
2460 }
2461 // Update IdealKit from graphKit.
2462 __ sync_kit(this);
2463
2464 } __ end_if(); // _ref_type != ref_none
2465 } __ end_if(); // offset == referent_offset
2466
2467 // Final sync IdealKit and GraphKit.
2468 final_sync(ideal);
2469 #undef __
2470 }
2471
2472
2473 // Interpret Unsafe.fieldOffset cookies correctly:
2474 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2475
2476 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2477 // Attempt to infer a sharper value type from the offset and base type.
2478 ciKlass* sharpened_klass = NULL;
2479
2480 // See if it is an instance field, with an object type.
2481 if (alias_type->field() != NULL) {
2482 assert(!is_native_ptr, "native pointer op cannot use a java address");
2483 if (alias_type->field()->type()->is_klass()) {
2484 sharpened_klass = alias_type->field()->type()->as_klass();
2485 }
2486 }
2487
2488 // See if it is a narrow oop array.
2489 if (adr_type->isa_aryptr()) {
2490 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2491 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2492 if (elem_type != NULL) {
2493 sharpened_klass = elem_type->klass();
2494 }
2495 }
2496 }
2497
2498 // The sharpened class might be unloaded if there is no class loader
2499 // contraint in place.
2500 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2501 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2502
2503 #ifndef PRODUCT
2504 if (C->print_intrinsics() || C->print_inlining()) {
2505 tty->print(" from base type: "); adr_type->dump();
2506 tty->print(" sharpened value: "); tjp->dump();
2507 }
2508 #endif
2509 // Sharpen the value type.
2510 return tjp;
2511 }
2512 return NULL;
2513 }
2514
2515 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2516 if (callee()->is_static()) return false; // caller must have the capability!
2517
2518 #ifndef PRODUCT
2519 {
2520 ResourceMark rm;
2521 // Check the signatures.
2522 ciSignature* sig = callee()->signature();
2523 #ifdef ASSERT
2524 if (!is_store) {
2525 // Object getObject(Object base, int/long offset), etc.
2526 BasicType rtype = sig->return_type()->basic_type();
2527 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2528 rtype = T_ADDRESS; // it is really a C void*
2529 assert(rtype == type, "getter must return the expected value");
2530 if (!is_native_ptr) {
2531 assert(sig->count() == 2, "oop getter has 2 arguments");
2532 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2533 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2534 } else {
2535 assert(sig->count() == 1, "native getter has 1 argument");
2536 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2537 }
2538 } else {
2539 // void putObject(Object base, int/long offset, Object x), etc.
2540 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2541 if (!is_native_ptr) {
2542 assert(sig->count() == 3, "oop putter has 3 arguments");
2543 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2544 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2545 } else {
2546 assert(sig->count() == 2, "native putter has 2 arguments");
2547 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2548 }
2549 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2550 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2551 vtype = T_ADDRESS; // it is really a C void*
2552 assert(vtype == type, "putter must accept the expected value");
2553 }
2554 #endif // ASSERT
2555 }
2556 #endif //PRODUCT
2557
2558 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2559
2560 Node* receiver = argument(0); // type: oop
2561
2562 // Build address expression.
2563 Node* adr;
2564 Node* heap_base_oop = top();
2565 Node* offset = top();
2566 Node* val;
2567
2568 if (!is_native_ptr) {
2569 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2570 Node* base = argument(1); // type: oop
2571 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2572 offset = argument(2); // type: long
2573 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2574 // to be plain byte offsets, which are also the same as those accepted
2575 // by oopDesc::field_base.
2576 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2577 "fieldOffset must be byte-scaled");
2578 // 32-bit machines ignore the high half!
2579 offset = ConvL2X(offset);
2580 adr = make_unsafe_address(base, offset);
2581 heap_base_oop = base;
2582 val = is_store ? argument(4) : NULL;
2583 } else {
2584 Node* ptr = argument(1); // type: long
2585 ptr = ConvL2X(ptr); // adjust Java long to machine word
2586 adr = make_unsafe_address(NULL, ptr);
2587 val = is_store ? argument(3) : NULL;
2588 }
2589
2590 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2591
2592 // First guess at the value type.
2593 const Type *value_type = Type::get_const_basic_type(type);
2594
2595 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2596 // there was not enough information to nail it down.
2597 Compile::AliasType* alias_type = C->alias_type(adr_type);
2598 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2599
2600 // We will need memory barriers unless we can determine a unique
2601 // alias category for this reference. (Note: If for some reason
2602 // the barriers get omitted and the unsafe reference begins to "pollute"
2603 // the alias analysis of the rest of the graph, either Compile::can_alias
2604 // or Compile::must_alias will throw a diagnostic assert.)
2605 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2606
2607 // If we are reading the value of the referent field of a Reference
2608 // object (either by using Unsafe directly or through reflection)
2609 // then, if G1 is enabled, we need to record the referent in an
2610 // SATB log buffer using the pre-barrier mechanism.
2611 // Also we need to add memory barrier to prevent commoning reads
2612 // from this field across safepoint since GC can change its value.
2613 bool need_read_barrier = !is_native_ptr && !is_store &&
2614 offset != top() && heap_base_oop != top();
2615
2616 if (!is_store && type == T_OBJECT) {
2617 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2618 if (tjp != NULL) {
2619 value_type = tjp;
2620 }
2621 }
2622
2623 receiver = null_check(receiver);
2624 if (stopped()) {
2625 return true;
2626 }
2627 // Heap pointers get a null-check from the interpreter,
2628 // as a courtesy. However, this is not guaranteed by Unsafe,
2629 // and it is not possible to fully distinguish unintended nulls
2630 // from intended ones in this API.
2631
2632 if (is_volatile) {
2633 // We need to emit leading and trailing CPU membars (see below) in
2634 // addition to memory membars when is_volatile. This is a little
2635 // too strong, but avoids the need to insert per-alias-type
2636 // volatile membars (for stores; compare Parse::do_put_xxx), which
2637 // we cannot do effectively here because we probably only have a
2638 // rough approximation of type.
2639 need_mem_bar = true;
2640 // For Stores, place a memory ordering barrier now.
2641 if (is_store) {
2642 insert_mem_bar(Op_MemBarRelease);
2643 } else {
2644 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2645 insert_mem_bar(Op_MemBarVolatile);
2646 }
2647 }
2648 }
2649
2650 // Memory barrier to prevent normal and 'unsafe' accesses from
2651 // bypassing each other. Happens after null checks, so the
2652 // exception paths do not take memory state from the memory barrier,
2653 // so there's no problems making a strong assert about mixing users
2654 // of safe & unsafe memory.
2655 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2656
2657 if (!is_store) {
2658 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2659 // To be valid, unsafe loads may depend on other conditions than
2660 // the one that guards them: pin the Load node
2661 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2662 // load value
2663 switch (type) {
2664 case T_BOOLEAN:
2665 case T_CHAR:
2666 case T_BYTE:
2667 case T_SHORT:
2668 case T_INT:
2669 case T_LONG:
2670 case T_FLOAT:
2671 case T_DOUBLE:
2672 break;
2673 case T_OBJECT:
2674 if (need_read_barrier) {
2675 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2676 }
2677 break;
2678 case T_ADDRESS:
2679 // Cast to an int type.
2680 p = _gvn.transform(new CastP2XNode(NULL, p));
2681 p = ConvX2UL(p);
2682 break;
2683 default:
2684 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2685 break;
2686 }
2687 // The load node has the control of the preceding MemBarCPUOrder. All
2688 // following nodes will have the control of the MemBarCPUOrder inserted at
2689 // the end of this method. So, pushing the load onto the stack at a later
2690 // point is fine.
2691 set_result(p);
2692 } else {
2693 // place effect of store into memory
2694 switch (type) {
2695 case T_DOUBLE:
2696 val = dstore_rounding(val);
2697 break;
2698 case T_ADDRESS:
2699 // Repackage the long as a pointer.
2700 val = ConvL2X(val);
2701 val = _gvn.transform(new CastX2PNode(val));
2702 break;
2703 }
2704
2705 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2706 if (type != T_OBJECT ) {
2707 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2708 } else {
2709 // Possibly an oop being stored to Java heap or native memory
2710 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2711 // oop to Java heap.
2712 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2713 } else {
2714 // We can't tell at compile time if we are storing in the Java heap or outside
2715 // of it. So we need to emit code to conditionally do the proper type of
2716 // store.
2717
2718 IdealKit ideal(this);
2719 #define __ ideal.
2720 // QQQ who knows what probability is here??
2721 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2722 // Sync IdealKit and graphKit.
2723 sync_kit(ideal);
2724 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2725 // Update IdealKit memory.
2726 __ sync_kit(this);
2727 } __ else_(); {
2728 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2729 } __ end_if();
2730 // Final sync IdealKit and GraphKit.
2731 final_sync(ideal);
2732 #undef __
2733 }
2734 }
2735 }
2736
2737 if (is_volatile) {
2738 if (!is_store) {
2739 insert_mem_bar(Op_MemBarAcquire);
2740 } else {
2741 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2742 insert_mem_bar(Op_MemBarVolatile);
2743 }
2744 }
2745 }
2746
2747 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2748
2749 return true;
2750 }
2751
2752 //----------------------------inline_unsafe_load_store----------------------------
2753 // This method serves a couple of different customers (depending on LoadStoreKind):
2754 //
2755 // LS_cmpxchg:
2756 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2757 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2758 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2759 //
2760 // LS_xadd:
2761 // public int getAndAddInt( Object o, long offset, int delta)
2762 // public long getAndAddLong(Object o, long offset, long delta)
2763 //
2764 // LS_xchg:
2765 // int getAndSet(Object o, long offset, int newValue)
2766 // long getAndSet(Object o, long offset, long newValue)
2767 // Object getAndSet(Object o, long offset, Object newValue)
2768 //
2769 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2770 // This basic scheme here is the same as inline_unsafe_access, but
2771 // differs in enough details that combining them would make the code
2772 // overly confusing. (This is a true fact! I originally combined
2773 // them, but even I was confused by it!) As much code/comments as
2774 // possible are retained from inline_unsafe_access though to make
2775 // the correspondences clearer. - dl
2776
2777 if (callee()->is_static()) return false; // caller must have the capability!
2778
2779 #ifndef PRODUCT
2780 BasicType rtype;
2781 {
2782 ResourceMark rm;
2783 // Check the signatures.
2784 ciSignature* sig = callee()->signature();
2785 rtype = sig->return_type()->basic_type();
2786 if (kind == LS_xadd || kind == LS_xchg) {
2787 // Check the signatures.
2788 #ifdef ASSERT
2789 assert(rtype == type, "get and set must return the expected type");
2790 assert(sig->count() == 3, "get and set has 3 arguments");
2791 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2792 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2793 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2794 #endif // ASSERT
2795 } else if (kind == LS_cmpxchg) {
2796 // Check the signatures.
2797 #ifdef ASSERT
2798 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2799 assert(sig->count() == 4, "CAS has 4 arguments");
2800 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2801 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2802 #endif // ASSERT
2803 } else {
2804 ShouldNotReachHere();
2805 }
2806 }
2807 #endif //PRODUCT
2808
2809 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2810
2811 // Get arguments:
2812 Node* receiver = NULL;
2813 Node* base = NULL;
2814 Node* offset = NULL;
2815 Node* oldval = NULL;
2816 Node* newval = NULL;
2817 if (kind == LS_cmpxchg) {
2818 const bool two_slot_type = type2size[type] == 2;
2819 receiver = argument(0); // type: oop
2820 base = argument(1); // type: oop
2821 offset = argument(2); // type: long
2822 oldval = argument(4); // type: oop, int, or long
2823 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2824 } else if (kind == LS_xadd || kind == LS_xchg){
2825 receiver = argument(0); // type: oop
2826 base = argument(1); // type: oop
2827 offset = argument(2); // type: long
2828 oldval = NULL;
2829 newval = argument(4); // type: oop, int, or long
2830 }
2831
2832 // Null check receiver.
2833 receiver = null_check(receiver);
2834 if (stopped()) {
2835 return true;
2836 }
2837
2838 // Build field offset expression.
2839 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2840 // to be plain byte offsets, which are also the same as those accepted
2841 // by oopDesc::field_base.
2842 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2843 // 32-bit machines ignore the high half of long offsets
2844 offset = ConvL2X(offset);
2845 Node* adr = make_unsafe_address(base, offset);
2846 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2847
2848 // For CAS, unlike inline_unsafe_access, there seems no point in
2849 // trying to refine types. Just use the coarse types here.
2850 const Type *value_type = Type::get_const_basic_type(type);
2851 Compile::AliasType* alias_type = C->alias_type(adr_type);
2852 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2853
2854 if (kind == LS_xchg && type == T_OBJECT) {
2855 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2856 if (tjp != NULL) {
2857 value_type = tjp;
2858 }
2859 }
2860
2861 int alias_idx = C->get_alias_index(adr_type);
2862
2863 // Memory-model-wise, a LoadStore acts like a little synchronized
2864 // block, so needs barriers on each side. These don't translate
2865 // into actual barriers on most machines, but we still need rest of
2866 // compiler to respect ordering.
2867
2868 insert_mem_bar(Op_MemBarRelease);
2869 insert_mem_bar(Op_MemBarCPUOrder);
2870
2871 // 4984716: MemBars must be inserted before this
2872 // memory node in order to avoid a false
2873 // dependency which will confuse the scheduler.
2874 Node *mem = memory(alias_idx);
2875
2876 // For now, we handle only those cases that actually exist: ints,
2877 // longs, and Object. Adding others should be straightforward.
2878 Node* load_store;
2879 switch(type) {
2880 case T_INT:
2881 if (kind == LS_xadd) {
2882 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2883 } else if (kind == LS_xchg) {
2884 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2885 } else if (kind == LS_cmpxchg) {
2886 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2887 } else {
2888 ShouldNotReachHere();
2889 }
2890 break;
2891 case T_LONG:
2892 if (kind == LS_xadd) {
2893 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2894 } else if (kind == LS_xchg) {
2895 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2896 } else if (kind == LS_cmpxchg) {
2897 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2898 } else {
2899 ShouldNotReachHere();
2900 }
2901 break;
2902 case T_OBJECT:
2903 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2904 // could be delayed during Parse (for example, in adjust_map_after_if()).
2905 // Execute transformation here to avoid barrier generation in such case.
2906 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2907 newval = _gvn.makecon(TypePtr::NULL_PTR);
2908
2909 // Reference stores need a store barrier.
2910 if (kind == LS_xchg) {
2911 // If pre-barrier must execute before the oop store, old value will require do_load here.
2912 if (!can_move_pre_barrier()) {
2913 pre_barrier(true /* do_load*/,
2914 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2915 NULL /* pre_val*/,
2916 T_OBJECT);
2917 } // Else move pre_barrier to use load_store value, see below.
2918 } else if (kind == LS_cmpxchg) {
2919 // Same as for newval above:
2920 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2921 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2922 }
2923 // The only known value which might get overwritten is oldval.
2924 pre_barrier(false /* do_load */,
2925 control(), NULL, NULL, max_juint, NULL, NULL,
2926 oldval /* pre_val */,
2927 T_OBJECT);
2928 } else {
2929 ShouldNotReachHere();
2930 }
2931
2932 #ifdef _LP64
2933 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2934 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2935 if (kind == LS_xchg) {
2936 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2937 newval_enc, adr_type, value_type->make_narrowoop()));
2938 } else {
2939 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2940 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2941 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2942 newval_enc, oldval_enc));
2943 }
2944 } else
2945 #endif
2946 {
2947 if (kind == LS_xchg) {
2948 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2949 } else {
2950 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2951 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2952 }
2953 }
2954 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2955 break;
2956 default:
2957 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2958 break;
2959 }
2960
2961 // SCMemProjNodes represent the memory state of a LoadStore. Their
2962 // main role is to prevent LoadStore nodes from being optimized away
2963 // when their results aren't used.
2964 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2965 set_memory(proj, alias_idx);
2966
2967 if (type == T_OBJECT && kind == LS_xchg) {
2968 #ifdef _LP64
2969 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2970 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2971 }
2972 #endif
2973 if (can_move_pre_barrier()) {
2974 // Don't need to load pre_val. The old value is returned by load_store.
2975 // The pre_barrier can execute after the xchg as long as no safepoint
2976 // gets inserted between them.
2977 pre_barrier(false /* do_load */,
2978 control(), NULL, NULL, max_juint, NULL, NULL,
2979 load_store /* pre_val */,
2980 T_OBJECT);
2981 }
2982 }
2983
2984 // Add the trailing membar surrounding the access
2985 insert_mem_bar(Op_MemBarCPUOrder);
2986 insert_mem_bar(Op_MemBarAcquire);
2987
2988 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2989 set_result(load_store);
2990 return true;
2991 }
2992
2993 //----------------------------inline_unsafe_ordered_store----------------------
2994 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2995 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2996 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2997 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2998 // This is another variant of inline_unsafe_access, differing in
2999 // that it always issues store-store ("release") barrier and ensures
3000 // store-atomicity (which only matters for "long").
3001
3002 if (callee()->is_static()) return false; // caller must have the capability!
3003
3004 #ifndef PRODUCT
3005 {
3006 ResourceMark rm;
3007 // Check the signatures.
3008 ciSignature* sig = callee()->signature();
3009 #ifdef ASSERT
3010 BasicType rtype = sig->return_type()->basic_type();
3011 assert(rtype == T_VOID, "must return void");
3012 assert(sig->count() == 3, "has 3 arguments");
3013 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3014 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3015 #endif // ASSERT
3016 }
3017 #endif //PRODUCT
3018
3019 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3020
3021 // Get arguments:
3022 Node* receiver = argument(0); // type: oop
3023 Node* base = argument(1); // type: oop
3024 Node* offset = argument(2); // type: long
3025 Node* val = argument(4); // type: oop, int, or long
3026
3027 // Null check receiver.
3028 receiver = null_check(receiver);
3029 if (stopped()) {
3030 return true;
3031 }
3032
3033 // Build field offset expression.
3034 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3035 // 32-bit machines ignore the high half of long offsets
3036 offset = ConvL2X(offset);
3037 Node* adr = make_unsafe_address(base, offset);
3038 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3039 const Type *value_type = Type::get_const_basic_type(type);
3040 Compile::AliasType* alias_type = C->alias_type(adr_type);
3041
3042 insert_mem_bar(Op_MemBarRelease);
3043 insert_mem_bar(Op_MemBarCPUOrder);
3044 // Ensure that the store is atomic for longs:
3045 const bool require_atomic_access = true;
3046 Node* store;
3047 if (type == T_OBJECT) // reference stores need a store barrier.
3048 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3049 else {
3050 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3051 }
3052 insert_mem_bar(Op_MemBarCPUOrder);
3053 return true;
3054 }
3055
3056 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3057 // Regardless of form, don't allow previous ld/st to move down,
3058 // then issue acquire, release, or volatile mem_bar.
3059 insert_mem_bar(Op_MemBarCPUOrder);
3060 switch(id) {
3061 case vmIntrinsics::_loadFence:
3062 insert_mem_bar(Op_LoadFence);
3063 return true;
3064 case vmIntrinsics::_storeFence:
3065 insert_mem_bar(Op_StoreFence);
3066 return true;
3067 case vmIntrinsics::_fullFence:
3068 insert_mem_bar(Op_MemBarVolatile);
3069 return true;
3070 default:
3071 fatal_unexpected_iid(id);
3072 return false;
3073 }
3074 }
3075
3076 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3077 if (!kls->is_Con()) {
3078 return true;
3079 }
3080 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3081 if (klsptr == NULL) {
3082 return true;
3083 }
3084 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3085 // don't need a guard for a klass that is already initialized
3086 return !ik->is_initialized();
3087 }
3088
3089 //----------------------------inline_unsafe_allocate---------------------------
3090 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3091 bool LibraryCallKit::inline_unsafe_allocate() {
3092 if (callee()->is_static()) return false; // caller must have the capability!
3093
3094 null_check_receiver(); // null-check, then ignore
3095 Node* cls = null_check(argument(1));
3096 if (stopped()) return true;
3097
3098 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3099 kls = null_check(kls);
3100 if (stopped()) return true; // argument was like int.class
3101
3102 Node* test = NULL;
3103 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3104 // Note: The argument might still be an illegal value like
3105 // Serializable.class or Object[].class. The runtime will handle it.
3106 // But we must make an explicit check for initialization.
3107 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3108 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3109 // can generate code to load it as unsigned byte.
3110 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3111 Node* bits = intcon(InstanceKlass::fully_initialized);
3112 test = _gvn.transform(new SubINode(inst, bits));
3113 // The 'test' is non-zero if we need to take a slow path.
3114 }
3115
3116 Node* obj = new_instance(kls, test);
3117 set_result(obj);
3118 return true;
3119 }
3120
3121 #ifdef TRACE_HAVE_INTRINSICS
3122 /*
3123 * oop -> myklass
3124 * myklass->trace_id |= USED
3125 * return myklass->trace_id & ~0x3
3126 */
3127 bool LibraryCallKit::inline_native_classID() {
3128 null_check_receiver(); // null-check, then ignore
3129 Node* cls = null_check(argument(1), T_OBJECT);
3130 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3131 kls = null_check(kls, T_OBJECT);
3132 ByteSize offset = TRACE_ID_OFFSET;
3133 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3134 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3135 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3136 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3137 Node* clsused = longcon(0x01l); // set the class bit
3138 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3139
3140 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3141 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3142 set_result(andl);
3143 return true;
3144 }
3145
3146 bool LibraryCallKit::inline_native_threadID() {
3147 Node* tls_ptr = NULL;
3148 Node* cur_thr = generate_current_thread(tls_ptr);
3149 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3150 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3151 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3152
3153 Node* threadid = NULL;
3154 size_t thread_id_size = OSThread::thread_id_size();
3155 if (thread_id_size == (size_t) BytesPerLong) {
3156 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3157 } else if (thread_id_size == (size_t) BytesPerInt) {
3158 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3159 } else {
3160 ShouldNotReachHere();
3161 }
3162 set_result(threadid);
3163 return true;
3164 }
3165 #endif
3166
3167 //------------------------inline_native_time_funcs--------------
3168 // inline code for System.currentTimeMillis() and System.nanoTime()
3169 // these have the same type and signature
3170 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3171 const TypeFunc* tf = OptoRuntime::void_long_Type();
3172 const TypePtr* no_memory_effects = NULL;
3173 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3174 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3175 #ifdef ASSERT
3176 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3177 assert(value_top == top(), "second value must be top");
3178 #endif
3179 set_result(value);
3180 return true;
3181 }
3182
3183 //------------------------inline_native_currentThread------------------
3184 bool LibraryCallKit::inline_native_currentThread() {
3185 Node* junk = NULL;
3186 set_result(generate_current_thread(junk));
3187 return true;
3188 }
3189
3190 //------------------------inline_native_isInterrupted------------------
3191 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3192 bool LibraryCallKit::inline_native_isInterrupted() {
3193 // Add a fast path to t.isInterrupted(clear_int):
3194 // (t == Thread.current() &&
3195 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3196 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3197 // So, in the common case that the interrupt bit is false,
3198 // we avoid making a call into the VM. Even if the interrupt bit
3199 // is true, if the clear_int argument is false, we avoid the VM call.
3200 // However, if the receiver is not currentThread, we must call the VM,
3201 // because there must be some locking done around the operation.
3202
3203 // We only go to the fast case code if we pass two guards.
3204 // Paths which do not pass are accumulated in the slow_region.
3205
3206 enum {
3207 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3208 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3209 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3210 PATH_LIMIT
3211 };
3212
3213 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3214 // out of the function.
3215 insert_mem_bar(Op_MemBarCPUOrder);
3216
3217 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3218 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3219
3220 RegionNode* slow_region = new RegionNode(1);
3221 record_for_igvn(slow_region);
3222
3223 // (a) Receiving thread must be the current thread.
3224 Node* rec_thr = argument(0);
3225 Node* tls_ptr = NULL;
3226 Node* cur_thr = generate_current_thread(tls_ptr);
3227 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3228 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3229
3230 generate_slow_guard(bol_thr, slow_region);
3231
3232 // (b) Interrupt bit on TLS must be false.
3233 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3234 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3235 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3236
3237 // Set the control input on the field _interrupted read to prevent it floating up.
3238 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3239 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3240 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3241
3242 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3243
3244 // First fast path: if (!TLS._interrupted) return false;
3245 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3246 result_rgn->init_req(no_int_result_path, false_bit);
3247 result_val->init_req(no_int_result_path, intcon(0));
3248
3249 // drop through to next case
3250 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3251
3252 #ifndef TARGET_OS_FAMILY_windows
3253 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3254 Node* clr_arg = argument(1);
3255 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3256 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3257 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3258
3259 // Second fast path: ... else if (!clear_int) return true;
3260 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3261 result_rgn->init_req(no_clear_result_path, false_arg);
3262 result_val->init_req(no_clear_result_path, intcon(1));
3263
3264 // drop through to next case
3265 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3266 #else
3267 // To return true on Windows you must read the _interrupted field
3268 // and check the the event state i.e. take the slow path.
3269 #endif // TARGET_OS_FAMILY_windows
3270
3271 // (d) Otherwise, go to the slow path.
3272 slow_region->add_req(control());
3273 set_control( _gvn.transform(slow_region));
3274
3275 if (stopped()) {
3276 // There is no slow path.
3277 result_rgn->init_req(slow_result_path, top());
3278 result_val->init_req(slow_result_path, top());
3279 } else {
3280 // non-virtual because it is a private non-static
3281 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3282
3283 Node* slow_val = set_results_for_java_call(slow_call);
3284 // this->control() comes from set_results_for_java_call
3285
3286 Node* fast_io = slow_call->in(TypeFunc::I_O);
3287 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3288
3289 // These two phis are pre-filled with copies of of the fast IO and Memory
3290 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3291 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3292
3293 result_rgn->init_req(slow_result_path, control());
3294 result_io ->init_req(slow_result_path, i_o());
3295 result_mem->init_req(slow_result_path, reset_memory());
3296 result_val->init_req(slow_result_path, slow_val);
3297
3298 set_all_memory(_gvn.transform(result_mem));
3299 set_i_o( _gvn.transform(result_io));
3300 }
3301
3302 C->set_has_split_ifs(true); // Has chance for split-if optimization
3303 set_result(result_rgn, result_val);
3304 return true;
3305 }
3306
3307 //---------------------------load_mirror_from_klass----------------------------
3308 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3309 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3310 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3311 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3312 }
3313
3314 //-----------------------load_klass_from_mirror_common-------------------------
3315 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3316 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3317 // and branch to the given path on the region.
3318 // If never_see_null, take an uncommon trap on null, so we can optimistically
3319 // compile for the non-null case.
3320 // If the region is NULL, force never_see_null = true.
3321 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3322 bool never_see_null,
3323 RegionNode* region,
3324 int null_path,
3325 int offset) {
3326 if (region == NULL) never_see_null = true;
3327 Node* p = basic_plus_adr(mirror, offset);
3328 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3329 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3330 Node* null_ctl = top();
3331 kls = null_check_oop(kls, &null_ctl, never_see_null);
3332 if (region != NULL) {
3333 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3334 region->init_req(null_path, null_ctl);
3335 } else {
3336 assert(null_ctl == top(), "no loose ends");
3337 }
3338 return kls;
3339 }
3340
3341 //--------------------(inline_native_Class_query helpers)---------------------
3342 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3343 // Fall through if (mods & mask) == bits, take the guard otherwise.
3344 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3345 // Branch around if the given klass has the given modifier bit set.
3346 // Like generate_guard, adds a new path onto the region.
3347 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3348 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3349 Node* mask = intcon(modifier_mask);
3350 Node* bits = intcon(modifier_bits);
3351 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3352 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3353 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3354 return generate_fair_guard(bol, region);
3355 }
3356 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3357 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3358 }
3359
3360 //-------------------------inline_native_Class_query-------------------
3361 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3362 const Type* return_type = TypeInt::BOOL;
3363 Node* prim_return_value = top(); // what happens if it's a primitive class?
3364 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3365 bool expect_prim = false; // most of these guys expect to work on refs
3366
3367 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3368
3369 Node* mirror = argument(0);
3370 Node* obj = top();
3371
3372 switch (id) {
3373 case vmIntrinsics::_isInstance:
3374 // nothing is an instance of a primitive type
3375 prim_return_value = intcon(0);
3376 obj = argument(1);
3377 break;
3378 case vmIntrinsics::_getModifiers:
3379 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3380 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3381 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3382 break;
3383 case vmIntrinsics::_isInterface:
3384 prim_return_value = intcon(0);
3385 break;
3386 case vmIntrinsics::_isArray:
3387 prim_return_value = intcon(0);
3388 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3389 break;
3390 case vmIntrinsics::_isPrimitive:
3391 prim_return_value = intcon(1);
3392 expect_prim = true; // obviously
3393 break;
3394 case vmIntrinsics::_getSuperclass:
3395 prim_return_value = null();
3396 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3397 break;
3398 case vmIntrinsics::_getClassAccessFlags:
3399 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3400 return_type = TypeInt::INT; // not bool! 6297094
3401 break;
3402 default:
3403 fatal_unexpected_iid(id);
3404 break;
3405 }
3406
3407 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3408 if (mirror_con == NULL) return false; // cannot happen?
3409
3410 #ifndef PRODUCT
3411 if (C->print_intrinsics() || C->print_inlining()) {
3412 ciType* k = mirror_con->java_mirror_type();
3413 if (k) {
3414 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3415 k->print_name();
3416 tty->cr();
3417 }
3418 }
3419 #endif
3420
3421 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3422 RegionNode* region = new RegionNode(PATH_LIMIT);
3423 record_for_igvn(region);
3424 PhiNode* phi = new PhiNode(region, return_type);
3425
3426 // The mirror will never be null of Reflection.getClassAccessFlags, however
3427 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3428 // if it is. See bug 4774291.
3429
3430 // For Reflection.getClassAccessFlags(), the null check occurs in
3431 // the wrong place; see inline_unsafe_access(), above, for a similar
3432 // situation.
3433 mirror = null_check(mirror);
3434 // If mirror or obj is dead, only null-path is taken.
3435 if (stopped()) return true;
3436
3437 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3438
3439 // Now load the mirror's klass metaobject, and null-check it.
3440 // Side-effects region with the control path if the klass is null.
3441 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3442 // If kls is null, we have a primitive mirror.
3443 phi->init_req(_prim_path, prim_return_value);
3444 if (stopped()) { set_result(region, phi); return true; }
3445 bool safe_for_replace = (region->in(_prim_path) == top());
3446
3447 Node* p; // handy temp
3448 Node* null_ctl;
3449
3450 // Now that we have the non-null klass, we can perform the real query.
3451 // For constant classes, the query will constant-fold in LoadNode::Value.
3452 Node* query_value = top();
3453 switch (id) {
3454 case vmIntrinsics::_isInstance:
3455 // nothing is an instance of a primitive type
3456 query_value = gen_instanceof(obj, kls, safe_for_replace);
3457 break;
3458
3459 case vmIntrinsics::_getModifiers:
3460 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3461 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3462 break;
3463
3464 case vmIntrinsics::_isInterface:
3465 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3466 if (generate_interface_guard(kls, region) != NULL)
3467 // A guard was added. If the guard is taken, it was an interface.
3468 phi->add_req(intcon(1));
3469 // If we fall through, it's a plain class.
3470 query_value = intcon(0);
3471 break;
3472
3473 case vmIntrinsics::_isArray:
3474 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3475 if (generate_array_guard(kls, region) != NULL)
3476 // A guard was added. If the guard is taken, it was an array.
3477 phi->add_req(intcon(1));
3478 // If we fall through, it's a plain class.
3479 query_value = intcon(0);
3480 break;
3481
3482 case vmIntrinsics::_isPrimitive:
3483 query_value = intcon(0); // "normal" path produces false
3484 break;
3485
3486 case vmIntrinsics::_getSuperclass:
3487 // The rules here are somewhat unfortunate, but we can still do better
3488 // with random logic than with a JNI call.
3489 // Interfaces store null or Object as _super, but must report null.
3490 // Arrays store an intermediate super as _super, but must report Object.
3491 // Other types can report the actual _super.
3492 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3493 if (generate_interface_guard(kls, region) != NULL)
3494 // A guard was added. If the guard is taken, it was an interface.
3495 phi->add_req(null());
3496 if (generate_array_guard(kls, region) != NULL)
3497 // A guard was added. If the guard is taken, it was an array.
3498 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3499 // If we fall through, it's a plain class. Get its _super.
3500 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3501 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3502 null_ctl = top();
3503 kls = null_check_oop(kls, &null_ctl);
3504 if (null_ctl != top()) {
3505 // If the guard is taken, Object.superClass is null (both klass and mirror).
3506 region->add_req(null_ctl);
3507 phi ->add_req(null());
3508 }
3509 if (!stopped()) {
3510 query_value = load_mirror_from_klass(kls);
3511 }
3512 break;
3513
3514 case vmIntrinsics::_getClassAccessFlags:
3515 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3516 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3517 break;
3518
3519 default:
3520 fatal_unexpected_iid(id);
3521 break;
3522 }
3523
3524 // Fall-through is the normal case of a query to a real class.
3525 phi->init_req(1, query_value);
3526 region->init_req(1, control());
3527
3528 C->set_has_split_ifs(true); // Has chance for split-if optimization
3529 set_result(region, phi);
3530 return true;
3531 }
3532
3533 //-------------------------inline_Class_cast-------------------
3534 bool LibraryCallKit::inline_Class_cast() {
3535 Node* mirror = argument(0); // Class
3536 Node* obj = argument(1);
3537 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3538 if (mirror_con == NULL) {
3539 return false; // dead path (mirror->is_top()).
3540 }
3541 if (obj == NULL || obj->is_top()) {
3542 return false; // dead path
3543 }
3544 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3545
3546 // First, see if Class.cast() can be folded statically.
3547 // java_mirror_type() returns non-null for compile-time Class constants.
3548 ciType* tm = mirror_con->java_mirror_type();
3549 if (tm != NULL && tm->is_klass() &&
3550 tp != NULL && tp->klass() != NULL) {
3551 if (!tp->klass()->is_loaded()) {
3552 // Don't use intrinsic when class is not loaded.
3553 return false;
3554 } else {
3555 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3556 if (static_res == Compile::SSC_always_true) {
3557 // isInstance() is true - fold the code.
3558 set_result(obj);
3559 return true;
3560 } else if (static_res == Compile::SSC_always_false) {
3561 // Don't use intrinsic, have to throw ClassCastException.
3562 // If the reference is null, the non-intrinsic bytecode will
3563 // be optimized appropriately.
3564 return false;
3565 }
3566 }
3567 }
3568
3569 // Bailout intrinsic and do normal inlining if exception path is frequent.
3570 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3571 return false;
3572 }
3573
3574 // Generate dynamic checks.
3575 // Class.cast() is java implementation of _checkcast bytecode.
3576 // Do checkcast (Parse::do_checkcast()) optimizations here.
3577
3578 mirror = null_check(mirror);
3579 // If mirror is dead, only null-path is taken.
3580 if (stopped()) {
3581 return true;
3582 }
3583
3584 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3585 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3586 RegionNode* region = new RegionNode(PATH_LIMIT);
3587 record_for_igvn(region);
3588
3589 // Now load the mirror's klass metaobject, and null-check it.
3590 // If kls is null, we have a primitive mirror and
3591 // nothing is an instance of a primitive type.
3592 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3593
3594 Node* res = top();
3595 if (!stopped()) {
3596 Node* bad_type_ctrl = top();
3597 // Do checkcast optimizations.
3598 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3599 region->init_req(_bad_type_path, bad_type_ctrl);
3600 }
3601 if (region->in(_prim_path) != top() ||
3602 region->in(_bad_type_path) != top()) {
3603 // Let Interpreter throw ClassCastException.
3604 PreserveJVMState pjvms(this);
3605 set_control(_gvn.transform(region));
3606 uncommon_trap(Deoptimization::Reason_intrinsic,
3607 Deoptimization::Action_maybe_recompile);
3608 }
3609 if (!stopped()) {
3610 set_result(res);
3611 }
3612 return true;
3613 }
3614
3615
3616 //--------------------------inline_native_subtype_check------------------------
3617 // This intrinsic takes the JNI calls out of the heart of
3618 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3619 bool LibraryCallKit::inline_native_subtype_check() {
3620 // Pull both arguments off the stack.
3621 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3622 args[0] = argument(0);
3623 args[1] = argument(1);
3624 Node* klasses[2]; // corresponding Klasses: superk, subk
3625 klasses[0] = klasses[1] = top();
3626
3627 enum {
3628 // A full decision tree on {superc is prim, subc is prim}:
3629 _prim_0_path = 1, // {P,N} => false
3630 // {P,P} & superc!=subc => false
3631 _prim_same_path, // {P,P} & superc==subc => true
3632 _prim_1_path, // {N,P} => false
3633 _ref_subtype_path, // {N,N} & subtype check wins => true
3634 _both_ref_path, // {N,N} & subtype check loses => false
3635 PATH_LIMIT
3636 };
3637
3638 RegionNode* region = new RegionNode(PATH_LIMIT);
3639 Node* phi = new PhiNode(region, TypeInt::BOOL);
3640 record_for_igvn(region);
3641
3642 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3643 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3644 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3645
3646 // First null-check both mirrors and load each mirror's klass metaobject.
3647 int which_arg;
3648 for (which_arg = 0; which_arg <= 1; which_arg++) {
3649 Node* arg = args[which_arg];
3650 arg = null_check(arg);
3651 if (stopped()) break;
3652 args[which_arg] = arg;
3653
3654 Node* p = basic_plus_adr(arg, class_klass_offset);
3655 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3656 klasses[which_arg] = _gvn.transform(kls);
3657 }
3658
3659 // Having loaded both klasses, test each for null.
3660 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3661 for (which_arg = 0; which_arg <= 1; which_arg++) {
3662 Node* kls = klasses[which_arg];
3663 Node* null_ctl = top();
3664 kls = null_check_oop(kls, &null_ctl, never_see_null);
3665 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3666 region->init_req(prim_path, null_ctl);
3667 if (stopped()) break;
3668 klasses[which_arg] = kls;
3669 }
3670
3671 if (!stopped()) {
3672 // now we have two reference types, in klasses[0..1]
3673 Node* subk = klasses[1]; // the argument to isAssignableFrom
3674 Node* superk = klasses[0]; // the receiver
3675 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3676 // now we have a successful reference subtype check
3677 region->set_req(_ref_subtype_path, control());
3678 }
3679
3680 // If both operands are primitive (both klasses null), then
3681 // we must return true when they are identical primitives.
3682 // It is convenient to test this after the first null klass check.
3683 set_control(region->in(_prim_0_path)); // go back to first null check
3684 if (!stopped()) {
3685 // Since superc is primitive, make a guard for the superc==subc case.
3686 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3687 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3688 generate_guard(bol_eq, region, PROB_FAIR);
3689 if (region->req() == PATH_LIMIT+1) {
3690 // A guard was added. If the added guard is taken, superc==subc.
3691 region->swap_edges(PATH_LIMIT, _prim_same_path);
3692 region->del_req(PATH_LIMIT);
3693 }
3694 region->set_req(_prim_0_path, control()); // Not equal after all.
3695 }
3696
3697 // these are the only paths that produce 'true':
3698 phi->set_req(_prim_same_path, intcon(1));
3699 phi->set_req(_ref_subtype_path, intcon(1));
3700
3701 // pull together the cases:
3702 assert(region->req() == PATH_LIMIT, "sane region");
3703 for (uint i = 1; i < region->req(); i++) {
3704 Node* ctl = region->in(i);
3705 if (ctl == NULL || ctl == top()) {
3706 region->set_req(i, top());
3707 phi ->set_req(i, top());
3708 } else if (phi->in(i) == NULL) {
3709 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3710 }
3711 }
3712
3713 set_control(_gvn.transform(region));
3714 set_result(_gvn.transform(phi));
3715 return true;
3716 }
3717
3718 //---------------------generate_array_guard_common------------------------
3719 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3720 bool obj_array, bool not_array) {
3721
3722 if (stopped()) {
3723 return NULL;
3724 }
3725
3726 // If obj_array/non_array==false/false:
3727 // Branch around if the given klass is in fact an array (either obj or prim).
3728 // If obj_array/non_array==false/true:
3729 // Branch around if the given klass is not an array klass of any kind.
3730 // If obj_array/non_array==true/true:
3731 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3732 // If obj_array/non_array==true/false:
3733 // Branch around if the kls is an oop array (Object[] or subtype)
3734 //
3735 // Like generate_guard, adds a new path onto the region.
3736 jint layout_con = 0;
3737 Node* layout_val = get_layout_helper(kls, layout_con);
3738 if (layout_val == NULL) {
3739 bool query = (obj_array
3740 ? Klass::layout_helper_is_objArray(layout_con)
3741 : Klass::layout_helper_is_array(layout_con));
3742 if (query == not_array) {
3743 return NULL; // never a branch
3744 } else { // always a branch
3745 Node* always_branch = control();
3746 if (region != NULL)
3747 region->add_req(always_branch);
3748 set_control(top());
3749 return always_branch;
3750 }
3751 }
3752 // Now test the correct condition.
3753 jint nval = (obj_array
3754 ? ((jint)Klass::_lh_array_tag_type_value
3755 << Klass::_lh_array_tag_shift)
3756 : Klass::_lh_neutral_value);
3757 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3758 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3759 // invert the test if we are looking for a non-array
3760 if (not_array) btest = BoolTest(btest).negate();
3761 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3762 return generate_fair_guard(bol, region);
3763 }
3764
3765
3766 //-----------------------inline_native_newArray--------------------------
3767 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3768 bool LibraryCallKit::inline_native_newArray() {
3769 Node* mirror = argument(0);
3770 Node* count_val = argument(1);
3771
3772 mirror = null_check(mirror);
3773 // If mirror or obj is dead, only null-path is taken.
3774 if (stopped()) return true;
3775
3776 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3777 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3778 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3779 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3780 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3781
3782 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3783 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3784 result_reg, _slow_path);
3785 Node* normal_ctl = control();
3786 Node* no_array_ctl = result_reg->in(_slow_path);
3787
3788 // Generate code for the slow case. We make a call to newArray().
3789 set_control(no_array_ctl);
3790 if (!stopped()) {
3791 // Either the input type is void.class, or else the
3792 // array klass has not yet been cached. Either the
3793 // ensuing call will throw an exception, or else it
3794 // will cache the array klass for next time.
3795 PreserveJVMState pjvms(this);
3796 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3797 Node* slow_result = set_results_for_java_call(slow_call);
3798 // this->control() comes from set_results_for_java_call
3799 result_reg->set_req(_slow_path, control());
3800 result_val->set_req(_slow_path, slow_result);
3801 result_io ->set_req(_slow_path, i_o());
3802 result_mem->set_req(_slow_path, reset_memory());
3803 }
3804
3805 set_control(normal_ctl);
3806 if (!stopped()) {
3807 // Normal case: The array type has been cached in the java.lang.Class.
3808 // The following call works fine even if the array type is polymorphic.
3809 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3810 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3811 result_reg->init_req(_normal_path, control());
3812 result_val->init_req(_normal_path, obj);
3813 result_io ->init_req(_normal_path, i_o());
3814 result_mem->init_req(_normal_path, reset_memory());
3815 }
3816
3817 // Return the combined state.
3818 set_i_o( _gvn.transform(result_io) );
3819 set_all_memory( _gvn.transform(result_mem));
3820
3821 C->set_has_split_ifs(true); // Has chance for split-if optimization
3822 set_result(result_reg, result_val);
3823 return true;
3824 }
3825
3826 //----------------------inline_native_getLength--------------------------
3827 // public static native int java.lang.reflect.Array.getLength(Object array);
3828 bool LibraryCallKit::inline_native_getLength() {
3829 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3830
3831 Node* array = null_check(argument(0));
3832 // If array is dead, only null-path is taken.
3833 if (stopped()) return true;
3834
3835 // Deoptimize if it is a non-array.
3836 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3837
3838 if (non_array != NULL) {
3839 PreserveJVMState pjvms(this);
3840 set_control(non_array);
3841 uncommon_trap(Deoptimization::Reason_intrinsic,
3842 Deoptimization::Action_maybe_recompile);
3843 }
3844
3845 // If control is dead, only non-array-path is taken.
3846 if (stopped()) return true;
3847
3848 // The works fine even if the array type is polymorphic.
3849 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3850 Node* result = load_array_length(array);
3851
3852 C->set_has_split_ifs(true); // Has chance for split-if optimization
3853 set_result(result);
3854 return true;
3855 }
3856
3857 //------------------------inline_array_copyOf----------------------------
3858 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3859 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3860 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3861 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3862
3863 // Get the arguments.
3864 Node* original = argument(0);
3865 Node* start = is_copyOfRange? argument(1): intcon(0);
3866 Node* end = is_copyOfRange? argument(2): argument(1);
3867 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3868
3869 Node* newcopy;
3870
3871 // Set the original stack and the reexecute bit for the interpreter to reexecute
3872 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3873 { PreserveReexecuteState preexecs(this);
3874 jvms()->set_should_reexecute(true);
3875
3876 array_type_mirror = null_check(array_type_mirror);
3877 original = null_check(original);
3878
3879 // Check if a null path was taken unconditionally.
3880 if (stopped()) return true;
3881
3882 Node* orig_length = load_array_length(original);
3883
3884 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3885 klass_node = null_check(klass_node);
3886
3887 RegionNode* bailout = new RegionNode(1);
3888 record_for_igvn(bailout);
3889
3890 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3891 // Bail out if that is so.
3892 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3893 if (not_objArray != NULL) {
3894 // Improve the klass node's type from the new optimistic assumption:
3895 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3896 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3897 Node* cast = new CastPPNode(klass_node, akls);
3898 cast->init_req(0, control());
3899 klass_node = _gvn.transform(cast);
3900 }
3901
3902 // Bail out if either start or end is negative.
3903 generate_negative_guard(start, bailout, &start);
3904 generate_negative_guard(end, bailout, &end);
3905
3906 Node* length = end;
3907 if (_gvn.type(start) != TypeInt::ZERO) {
3908 length = _gvn.transform(new SubINode(end, start));
3909 }
3910
3911 // Bail out if length is negative.
3912 // Without this the new_array would throw
3913 // NegativeArraySizeException but IllegalArgumentException is what
3914 // should be thrown
3915 generate_negative_guard(length, bailout, &length);
3916
3917 if (bailout->req() > 1) {
3918 PreserveJVMState pjvms(this);
3919 set_control(_gvn.transform(bailout));
3920 uncommon_trap(Deoptimization::Reason_intrinsic,
3921 Deoptimization::Action_maybe_recompile);
3922 }
3923
3924 if (!stopped()) {
3925 // How many elements will we copy from the original?
3926 // The answer is MinI(orig_length - start, length).
3927 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3928 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3929
3930 // Generate a direct call to the right arraycopy function(s).
3931 // We know the copy is disjoint but we might not know if the
3932 // oop stores need checking.
3933 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3934 // This will fail a store-check if x contains any non-nulls.
3935
3936 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3937 // loads/stores but it is legal only if we're sure the
3938 // Arrays.copyOf would succeed. So we need all input arguments
3939 // to the copyOf to be validated, including that the copy to the
3940 // new array won't trigger an ArrayStoreException. That subtype
3941 // check can be optimized if we know something on the type of
3942 // the input array from type speculation.
3943 if (_gvn.type(klass_node)->singleton()) {
3944 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3945 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3946
3947 int test = C->static_subtype_check(superk, subk);
3948 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3949 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3950 if (t_original->speculative_type() != NULL) {
3951 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3952 }
3953 }
3954 }
3955
3956 bool validated = false;
3957 // Reason_class_check rather than Reason_intrinsic because we
3958 // want to intrinsify even if this traps.
3959 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3960 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3961 klass_node);
3962
3963 if (not_subtype_ctrl != top()) {
3964 PreserveJVMState pjvms(this);
3965 set_control(not_subtype_ctrl);
3966 uncommon_trap(Deoptimization::Reason_class_check,
3967 Deoptimization::Action_make_not_entrant);
3968 assert(stopped(), "Should be stopped");
3969 }
3970 validated = true;
3971 }
3972
3973 if (!stopped()) {
3974 newcopy = new_array(klass_node, length, 0); // no arguments to push
3975
3976 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3977 load_object_klass(original), klass_node);
3978 if (!is_copyOfRange) {
3979 ac->set_copyof(validated);
3980 } else {
3981 ac->set_copyofrange(validated);
3982 }
3983 Node* n = _gvn.transform(ac);
3984 if (n == ac) {
3985 ac->connect_outputs(this);
3986 } else {
3987 assert(validated, "shouldn't transform if all arguments not validated");
3988 set_all_memory(n);
3989 }
3990 }
3991 }
3992 } // original reexecute is set back here
3993
3994 C->set_has_split_ifs(true); // Has chance for split-if optimization
3995 if (!stopped()) {
3996 set_result(newcopy);
3997 }
3998 return true;
3999 }
4000
4001
4002 //----------------------generate_virtual_guard---------------------------
4003 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4004 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4005 RegionNode* slow_region) {
4006 ciMethod* method = callee();
4007 int vtable_index = method->vtable_index();
4008 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4009 err_msg_res("bad index %d", vtable_index));
4010 // Get the Method* out of the appropriate vtable entry.
4011 int entry_offset = (InstanceKlass::vtable_start_offset() +
4012 vtable_index*vtableEntry::size()) * wordSize +
4013 vtableEntry::method_offset_in_bytes();
4014 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4015 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4016
4017 // Compare the target method with the expected method (e.g., Object.hashCode).
4018 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4019
4020 Node* native_call = makecon(native_call_addr);
4021 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4022 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4023
4024 return generate_slow_guard(test_native, slow_region);
4025 }
4026
4027 //-----------------------generate_method_call----------------------------
4028 // Use generate_method_call to make a slow-call to the real
4029 // method if the fast path fails. An alternative would be to
4030 // use a stub like OptoRuntime::slow_arraycopy_Java.
4031 // This only works for expanding the current library call,
4032 // not another intrinsic. (E.g., don't use this for making an
4033 // arraycopy call inside of the copyOf intrinsic.)
4034 CallJavaNode*
4035 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4036 // When compiling the intrinsic method itself, do not use this technique.
4037 guarantee(callee() != C->method(), "cannot make slow-call to self");
4038
4039 ciMethod* method = callee();
4040 // ensure the JVMS we have will be correct for this call
4041 guarantee(method_id == method->intrinsic_id(), "must match");
4042
4043 const TypeFunc* tf = TypeFunc::make(method);
4044 CallJavaNode* slow_call;
4045 if (is_static) {
4046 assert(!is_virtual, "");
4047 slow_call = new CallStaticJavaNode(C, tf,
4048 SharedRuntime::get_resolve_static_call_stub(),
4049 method, bci());
4050 } else if (is_virtual) {
4051 null_check_receiver();
4052 int vtable_index = Method::invalid_vtable_index;
4053 if (UseInlineCaches) {
4054 // Suppress the vtable call
4055 } else {
4056 // hashCode and clone are not a miranda methods,
4057 // so the vtable index is fixed.
4058 // No need to use the linkResolver to get it.
4059 vtable_index = method->vtable_index();
4060 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4061 err_msg_res("bad index %d", vtable_index));
4062 }
4063 slow_call = new CallDynamicJavaNode(tf,
4064 SharedRuntime::get_resolve_virtual_call_stub(),
4065 method, vtable_index, bci());
4066 } else { // neither virtual nor static: opt_virtual
4067 null_check_receiver();
4068 slow_call = new CallStaticJavaNode(C, tf,
4069 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4070 method, bci());
4071 slow_call->set_optimized_virtual(true);
4072 }
4073 set_arguments_for_java_call(slow_call);
4074 set_edges_for_java_call(slow_call);
4075 return slow_call;
4076 }
4077
4078
4079 /**
4080 * Build special case code for calls to hashCode on an object. This call may
4081 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4082 * slightly different code.
4083 */
4084 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4085 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4086 assert(!(is_virtual && is_static), "either virtual, special, or static");
4087
4088 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4089
4090 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4091 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4092 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4093 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4094 Node* obj = NULL;
4095 if (!is_static) {
4096 // Check for hashing null object
4097 obj = null_check_receiver();
4098 if (stopped()) return true; // unconditionally null
4099 result_reg->init_req(_null_path, top());
4100 result_val->init_req(_null_path, top());
4101 } else {
4102 // Do a null check, and return zero if null.
4103 // System.identityHashCode(null) == 0
4104 obj = argument(0);
4105 Node* null_ctl = top();
4106 obj = null_check_oop(obj, &null_ctl);
4107 result_reg->init_req(_null_path, null_ctl);
4108 result_val->init_req(_null_path, _gvn.intcon(0));
4109 }
4110
4111 // Unconditionally null? Then return right away.
4112 if (stopped()) {
4113 set_control( result_reg->in(_null_path));
4114 if (!stopped())
4115 set_result(result_val->in(_null_path));
4116 return true;
4117 }
4118
4119 // We only go to the fast case code if we pass a number of guards. The
4120 // paths which do not pass are accumulated in the slow_region.
4121 RegionNode* slow_region = new RegionNode(1);
4122 record_for_igvn(slow_region);
4123
4124 // If this is a virtual call, we generate a funny guard. We pull out
4125 // the vtable entry corresponding to hashCode() from the target object.
4126 // If the target method which we are calling happens to be the native
4127 // Object hashCode() method, we pass the guard. We do not need this
4128 // guard for non-virtual calls -- the caller is known to be the native
4129 // Object hashCode().
4130 if (is_virtual) {
4131 // After null check, get the object's klass.
4132 Node* obj_klass = load_object_klass(obj);
4133 generate_virtual_guard(obj_klass, slow_region);
4134 }
4135
4136 // Get the header out of the object, use LoadMarkNode when available
4137 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4138 // The control of the load must be NULL. Otherwise, the load can move before
4139 // the null check after castPP removal.
4140 Node* no_ctrl = NULL;
4141 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4142
4143 // Test the header to see if it is unlocked.
4144 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4145 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4146 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4147 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4148 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4149
4150 generate_slow_guard(test_unlocked, slow_region);
4151
4152 // Get the hash value and check to see that it has been properly assigned.
4153 // We depend on hash_mask being at most 32 bits and avoid the use of
4154 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4155 // vm: see markOop.hpp.
4156 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4157 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4158 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4159 // This hack lets the hash bits live anywhere in the mark object now, as long
4160 // as the shift drops the relevant bits into the low 32 bits. Note that
4161 // Java spec says that HashCode is an int so there's no point in capturing
4162 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4163 hshifted_header = ConvX2I(hshifted_header);
4164 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4165
4166 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4167 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4168 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4169
4170 generate_slow_guard(test_assigned, slow_region);
4171
4172 Node* init_mem = reset_memory();
4173 // fill in the rest of the null path:
4174 result_io ->init_req(_null_path, i_o());
4175 result_mem->init_req(_null_path, init_mem);
4176
4177 result_val->init_req(_fast_path, hash_val);
4178 result_reg->init_req(_fast_path, control());
4179 result_io ->init_req(_fast_path, i_o());
4180 result_mem->init_req(_fast_path, init_mem);
4181
4182 // Generate code for the slow case. We make a call to hashCode().
4183 set_control(_gvn.transform(slow_region));
4184 if (!stopped()) {
4185 // No need for PreserveJVMState, because we're using up the present state.
4186 set_all_memory(init_mem);
4187 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4188 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4189 Node* slow_result = set_results_for_java_call(slow_call);
4190 // this->control() comes from set_results_for_java_call
4191 result_reg->init_req(_slow_path, control());
4192 result_val->init_req(_slow_path, slow_result);
4193 result_io ->set_req(_slow_path, i_o());
4194 result_mem ->set_req(_slow_path, reset_memory());
4195 }
4196
4197 // Return the combined state.
4198 set_i_o( _gvn.transform(result_io) );
4199 set_all_memory( _gvn.transform(result_mem));
4200
4201 set_result(result_reg, result_val);
4202 return true;
4203 }
4204
4205 //---------------------------inline_native_getClass----------------------------
4206 // public final native Class<?> java.lang.Object.getClass();
4207 //
4208 // Build special case code for calls to getClass on an object.
4209 bool LibraryCallKit::inline_native_getClass() {
4210 Node* obj = null_check_receiver();
4211 if (stopped()) return true;
4212 set_result(load_mirror_from_klass(load_object_klass(obj)));
4213 return true;
4214 }
4215
4216 //-----------------inline_native_Reflection_getCallerClass---------------------
4217 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4218 //
4219 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4220 //
4221 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4222 // in that it must skip particular security frames and checks for
4223 // caller sensitive methods.
4224 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4225 #ifndef PRODUCT
4226 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4227 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4228 }
4229 #endif
4230
4231 if (!jvms()->has_method()) {
4232 #ifndef PRODUCT
4233 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4234 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4235 }
4236 #endif
4237 return false;
4238 }
4239
4240 // Walk back up the JVM state to find the caller at the required
4241 // depth.
4242 JVMState* caller_jvms = jvms();
4243
4244 // Cf. JVM_GetCallerClass
4245 // NOTE: Start the loop at depth 1 because the current JVM state does
4246 // not include the Reflection.getCallerClass() frame.
4247 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4248 ciMethod* m = caller_jvms->method();
4249 switch (n) {
4250 case 0:
4251 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4252 break;
4253 case 1:
4254 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4255 if (!m->caller_sensitive()) {
4256 #ifndef PRODUCT
4257 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4258 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4259 }
4260 #endif
4261 return false; // bail-out; let JVM_GetCallerClass do the work
4262 }
4263 break;
4264 default:
4265 if (!m->is_ignored_by_security_stack_walk()) {
4266 // We have reached the desired frame; return the holder class.
4267 // Acquire method holder as java.lang.Class and push as constant.
4268 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4269 ciInstance* caller_mirror = caller_klass->java_mirror();
4270 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4271
4272 #ifndef PRODUCT
4273 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4274 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());
4275 tty->print_cr(" JVM state at this point:");
4276 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4277 ciMethod* m = jvms()->of_depth(i)->method();
4278 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4279 }
4280 }
4281 #endif
4282 return true;
4283 }
4284 break;
4285 }
4286 }
4287
4288 #ifndef PRODUCT
4289 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4290 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4291 tty->print_cr(" JVM state at this point:");
4292 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4293 ciMethod* m = jvms()->of_depth(i)->method();
4294 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4295 }
4296 }
4297 #endif
4298
4299 return false; // bail-out; let JVM_GetCallerClass do the work
4300 }
4301
4302 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4303 Node* arg = argument(0);
4304 Node* result;
4305
4306 switch (id) {
4307 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4308 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4309 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4310 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4311
4312 case vmIntrinsics::_doubleToLongBits: {
4313 // two paths (plus control) merge in a wood
4314 RegionNode *r = new RegionNode(3);
4315 Node *phi = new PhiNode(r, TypeLong::LONG);
4316
4317 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4318 // Build the boolean node
4319 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4320
4321 // Branch either way.
4322 // NaN case is less traveled, which makes all the difference.
4323 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4324 Node *opt_isnan = _gvn.transform(ifisnan);
4325 assert( opt_isnan->is_If(), "Expect an IfNode");
4326 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4327 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4328
4329 set_control(iftrue);
4330
4331 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4332 Node *slow_result = longcon(nan_bits); // return NaN
4333 phi->init_req(1, _gvn.transform( slow_result ));
4334 r->init_req(1, iftrue);
4335
4336 // Else fall through
4337 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4338 set_control(iffalse);
4339
4340 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4341 r->init_req(2, iffalse);
4342
4343 // Post merge
4344 set_control(_gvn.transform(r));
4345 record_for_igvn(r);
4346
4347 C->set_has_split_ifs(true); // Has chance for split-if optimization
4348 result = phi;
4349 assert(result->bottom_type()->isa_long(), "must be");
4350 break;
4351 }
4352
4353 case vmIntrinsics::_floatToIntBits: {
4354 // two paths (plus control) merge in a wood
4355 RegionNode *r = new RegionNode(3);
4356 Node *phi = new PhiNode(r, TypeInt::INT);
4357
4358 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4359 // Build the boolean node
4360 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4361
4362 // Branch either way.
4363 // NaN case is less traveled, which makes all the difference.
4364 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4365 Node *opt_isnan = _gvn.transform(ifisnan);
4366 assert( opt_isnan->is_If(), "Expect an IfNode");
4367 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4368 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4369
4370 set_control(iftrue);
4371
4372 static const jint nan_bits = 0x7fc00000;
4373 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4374 phi->init_req(1, _gvn.transform( slow_result ));
4375 r->init_req(1, iftrue);
4376
4377 // Else fall through
4378 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4379 set_control(iffalse);
4380
4381 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4382 r->init_req(2, iffalse);
4383
4384 // Post merge
4385 set_control(_gvn.transform(r));
4386 record_for_igvn(r);
4387
4388 C->set_has_split_ifs(true); // Has chance for split-if optimization
4389 result = phi;
4390 assert(result->bottom_type()->isa_int(), "must be");
4391 break;
4392 }
4393
4394 default:
4395 fatal_unexpected_iid(id);
4396 break;
4397 }
4398 set_result(_gvn.transform(result));
4399 return true;
4400 }
4401
4402 #ifdef _LP64
4403 #define XTOP ,top() /*additional argument*/
4404 #else //_LP64
4405 #define XTOP /*no additional argument*/
4406 #endif //_LP64
4407
4408 //----------------------inline_unsafe_copyMemory-------------------------
4409 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4410 bool LibraryCallKit::inline_unsafe_copyMemory() {
4411 if (callee()->is_static()) return false; // caller must have the capability!
4412 null_check_receiver(); // null-check receiver
4413 if (stopped()) return true;
4414
4415 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4416
4417 Node* src_ptr = argument(1); // type: oop
4418 Node* src_off = ConvL2X(argument(2)); // type: long
4419 Node* dst_ptr = argument(4); // type: oop
4420 Node* dst_off = ConvL2X(argument(5)); // type: long
4421 Node* size = ConvL2X(argument(7)); // type: long
4422
4423 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4424 "fieldOffset must be byte-scaled");
4425
4426 Node* src = make_unsafe_address(src_ptr, src_off);
4427 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4428
4429 // Conservatively insert a memory barrier on all memory slices.
4430 // Do not let writes of the copy source or destination float below the copy.
4431 insert_mem_bar(Op_MemBarCPUOrder);
4432
4433 // Call it. Note that the length argument is not scaled.
4434 make_runtime_call(RC_LEAF|RC_NO_FP,
4435 OptoRuntime::fast_arraycopy_Type(),
4436 StubRoutines::unsafe_arraycopy(),
4437 "unsafe_arraycopy",
4438 TypeRawPtr::BOTTOM,
4439 src, dst, size XTOP);
4440
4441 // Do not let reads of the copy destination float above the copy.
4442 insert_mem_bar(Op_MemBarCPUOrder);
4443
4444 return true;
4445 }
4446
4447 //------------------------clone_coping-----------------------------------
4448 // Helper function for inline_native_clone.
4449 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4450 assert(obj_size != NULL, "");
4451 Node* raw_obj = alloc_obj->in(1);
4452 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4453
4454 AllocateNode* alloc = NULL;
4455 if (ReduceBulkZeroing) {
4456 // We will be completely responsible for initializing this object -
4457 // mark Initialize node as complete.
4458 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4459 // The object was just allocated - there should be no any stores!
4460 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4461 // Mark as complete_with_arraycopy so that on AllocateNode
4462 // expansion, we know this AllocateNode is initialized by an array
4463 // copy and a StoreStore barrier exists after the array copy.
4464 alloc->initialization()->set_complete_with_arraycopy();
4465 }
4466
4467 // Copy the fastest available way.
4468 // TODO: generate fields copies for small objects instead.
4469 Node* src = obj;
4470 Node* dest = alloc_obj;
4471 Node* size = _gvn.transform(obj_size);
4472
4473 // Exclude the header but include array length to copy by 8 bytes words.
4474 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4475 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4476 instanceOopDesc::base_offset_in_bytes();
4477 // base_off:
4478 // 8 - 32-bit VM
4479 // 12 - 64-bit VM, compressed klass
4480 // 16 - 64-bit VM, normal klass
4481 if (base_off % BytesPerLong != 0) {
4482 assert(UseCompressedClassPointers, "");
4483 if (is_array) {
4484 // Exclude length to copy by 8 bytes words.
4485 base_off += sizeof(int);
4486 } else {
4487 // Include klass to copy by 8 bytes words.
4488 base_off = instanceOopDesc::klass_offset_in_bytes();
4489 }
4490 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4491 }
4492 src = basic_plus_adr(src, base_off);
4493 dest = basic_plus_adr(dest, base_off);
4494
4495 // Compute the length also, if needed:
4496 Node* countx = size;
4497 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4498 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4499
4500 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4501
4502 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4503 ac->set_clonebasic();
4504 Node* n = _gvn.transform(ac);
4505 if (n == ac) {
4506 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4507 } else {
4508 set_all_memory(n);
4509 }
4510
4511 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4512 if (card_mark) {
4513 assert(!is_array, "");
4514 // Put in store barrier for any and all oops we are sticking
4515 // into this object. (We could avoid this if we could prove
4516 // that the object type contains no oop fields at all.)
4517 Node* no_particular_value = NULL;
4518 Node* no_particular_field = NULL;
4519 int raw_adr_idx = Compile::AliasIdxRaw;
4520 post_barrier(control(),
4521 memory(raw_adr_type),
4522 alloc_obj,
4523 no_particular_field,
4524 raw_adr_idx,
4525 no_particular_value,
4526 T_OBJECT,
4527 false);
4528 }
4529
4530 // Do not let reads from the cloned object float above the arraycopy.
4531 if (alloc != NULL) {
4532 // Do not let stores that initialize this object be reordered with
4533 // a subsequent store that would make this object accessible by
4534 // other threads.
4535 // Record what AllocateNode this StoreStore protects so that
4536 // escape analysis can go from the MemBarStoreStoreNode to the
4537 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4538 // based on the escape status of the AllocateNode.
4539 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4540 } else {
4541 insert_mem_bar(Op_MemBarCPUOrder);
4542 }
4543 }
4544
4545 //------------------------inline_native_clone----------------------------
4546 // protected native Object java.lang.Object.clone();
4547 //
4548 // Here are the simple edge cases:
4549 // null receiver => normal trap
4550 // virtual and clone was overridden => slow path to out-of-line clone
4551 // not cloneable or finalizer => slow path to out-of-line Object.clone
4552 //
4553 // The general case has two steps, allocation and copying.
4554 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4555 //
4556 // Copying also has two cases, oop arrays and everything else.
4557 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4558 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4559 //
4560 // These steps fold up nicely if and when the cloned object's klass
4561 // can be sharply typed as an object array, a type array, or an instance.
4562 //
4563 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4564 PhiNode* result_val;
4565
4566 // Set the reexecute bit for the interpreter to reexecute
4567 // the bytecode that invokes Object.clone if deoptimization happens.
4568 { PreserveReexecuteState preexecs(this);
4569 jvms()->set_should_reexecute(true);
4570
4571 Node* obj = null_check_receiver();
4572 if (stopped()) return true;
4573
4574 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4575
4576 // If we are going to clone an instance, we need its exact type to
4577 // know the number and types of fields to convert the clone to
4578 // loads/stores. Maybe a speculative type can help us.
4579 if (!obj_type->klass_is_exact() &&
4580 obj_type->speculative_type() != NULL &&
4581 obj_type->speculative_type()->is_instance_klass()) {
4582 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4583 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4584 !spec_ik->has_injected_fields()) {
4585 ciKlass* k = obj_type->klass();
4586 if (!k->is_instance_klass() ||
4587 k->as_instance_klass()->is_interface() ||
4588 k->as_instance_klass()->has_subklass()) {
4589 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4590 }
4591 }
4592 }
4593
4594 Node* obj_klass = load_object_klass(obj);
4595 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4596 const TypeOopPtr* toop = ((tklass != NULL)
4597 ? tklass->as_instance_type()
4598 : TypeInstPtr::NOTNULL);
4599
4600 // Conservatively insert a memory barrier on all memory slices.
4601 // Do not let writes into the original float below the clone.
4602 insert_mem_bar(Op_MemBarCPUOrder);
4603
4604 // paths into result_reg:
4605 enum {
4606 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4607 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4608 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4609 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4610 PATH_LIMIT
4611 };
4612 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4613 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4614 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4615 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4616 record_for_igvn(result_reg);
4617
4618 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4619 int raw_adr_idx = Compile::AliasIdxRaw;
4620
4621 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4622 if (array_ctl != NULL) {
4623 // It's an array.
4624 PreserveJVMState pjvms(this);
4625 set_control(array_ctl);
4626 Node* obj_length = load_array_length(obj);
4627 Node* obj_size = NULL;
4628 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4629
4630 if (!use_ReduceInitialCardMarks()) {
4631 // If it is an oop array, it requires very special treatment,
4632 // because card marking is required on each card of the array.
4633 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4634 if (is_obja != NULL) {
4635 PreserveJVMState pjvms2(this);
4636 set_control(is_obja);
4637 // Generate a direct call to the right arraycopy function(s).
4638 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4639 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4640 ac->set_cloneoop();
4641 Node* n = _gvn.transform(ac);
4642 assert(n == ac, "cannot disappear");
4643 ac->connect_outputs(this);
4644
4645 result_reg->init_req(_objArray_path, control());
4646 result_val->init_req(_objArray_path, alloc_obj);
4647 result_i_o ->set_req(_objArray_path, i_o());
4648 result_mem ->set_req(_objArray_path, reset_memory());
4649 }
4650 }
4651 // Otherwise, there are no card marks to worry about.
4652 // (We can dispense with card marks if we know the allocation
4653 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4654 // causes the non-eden paths to take compensating steps to
4655 // simulate a fresh allocation, so that no further
4656 // card marks are required in compiled code to initialize
4657 // the object.)
4658
4659 if (!stopped()) {
4660 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4661
4662 // Present the results of the copy.
4663 result_reg->init_req(_array_path, control());
4664 result_val->init_req(_array_path, alloc_obj);
4665 result_i_o ->set_req(_array_path, i_o());
4666 result_mem ->set_req(_array_path, reset_memory());
4667 }
4668 }
4669
4670 // We only go to the instance fast case code if we pass a number of guards.
4671 // The paths which do not pass are accumulated in the slow_region.
4672 RegionNode* slow_region = new RegionNode(1);
4673 record_for_igvn(slow_region);
4674 if (!stopped()) {
4675 // It's an instance (we did array above). Make the slow-path tests.
4676 // If this is a virtual call, we generate a funny guard. We grab
4677 // the vtable entry corresponding to clone() from the target object.
4678 // If the target method which we are calling happens to be the
4679 // Object clone() method, we pass the guard. We do not need this
4680 // guard for non-virtual calls; the caller is known to be the native
4681 // Object clone().
4682 if (is_virtual) {
4683 generate_virtual_guard(obj_klass, slow_region);
4684 }
4685
4686 // The object must be cloneable and must not have a finalizer.
4687 // Both of these conditions may be checked in a single test.
4688 // We could optimize the cloneable test further, but we don't care.
4689 generate_access_flags_guard(obj_klass,
4690 // Test both conditions:
4691 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4692 // Must be cloneable but not finalizer:
4693 JVM_ACC_IS_CLONEABLE,
4694 slow_region);
4695 }
4696
4697 if (!stopped()) {
4698 // It's an instance, and it passed the slow-path tests.
4699 PreserveJVMState pjvms(this);
4700 Node* obj_size = NULL;
4701 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4702 // is reexecuted if deoptimization occurs and there could be problems when merging
4703 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4704 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4705
4706 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4707
4708 // Present the results of the slow call.
4709 result_reg->init_req(_instance_path, control());
4710 result_val->init_req(_instance_path, alloc_obj);
4711 result_i_o ->set_req(_instance_path, i_o());
4712 result_mem ->set_req(_instance_path, reset_memory());
4713 }
4714
4715 // Generate code for the slow case. We make a call to clone().
4716 set_control(_gvn.transform(slow_region));
4717 if (!stopped()) {
4718 PreserveJVMState pjvms(this);
4719 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4720 Node* slow_result = set_results_for_java_call(slow_call);
4721 // this->control() comes from set_results_for_java_call
4722 result_reg->init_req(_slow_path, control());
4723 result_val->init_req(_slow_path, slow_result);
4724 result_i_o ->set_req(_slow_path, i_o());
4725 result_mem ->set_req(_slow_path, reset_memory());
4726 }
4727
4728 // Return the combined state.
4729 set_control( _gvn.transform(result_reg));
4730 set_i_o( _gvn.transform(result_i_o));
4731 set_all_memory( _gvn.transform(result_mem));
4732 } // original reexecute is set back here
4733
4734 set_result(_gvn.transform(result_val));
4735 return true;
4736 }
4737
4738 // If we have a tighly coupled allocation, the arraycopy may take care
4739 // of the array initialization. If one of the guards we insert between
4740 // the allocation and the arraycopy causes a deoptimization, an
4741 // unitialized array will escape the compiled method. To prevent that
4742 // we set the JVM state for uncommon traps between the allocation and
4743 // the arraycopy to the state before the allocation so, in case of
4744 // deoptimization, we'll reexecute the allocation and the
4745 // initialization.
4746 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4747 if (alloc != NULL) {
4748 ciMethod* trap_method = alloc->jvms()->method();
4749 int trap_bci = alloc->jvms()->bci();
4750
4751 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4752 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4753 // Make sure there's no store between the allocation and the
4754 // arraycopy otherwise visible side effects could be rexecuted
4755 // in case of deoptimization and cause incorrect execution.
4756 bool no_interfering_store = true;
4757 Node* mem = alloc->in(TypeFunc::Memory);
4758 if (mem->is_MergeMem()) {
4759 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4760 Node* n = mms.memory();
4761 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4762 assert(n->is_Store(), "what else?");
4763 no_interfering_store = false;
4764 break;
4765 }
4766 }
4767 } else {
4768 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4769 Node* n = mms.memory();
4770 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4771 assert(n->is_Store(), "what else?");
4772 no_interfering_store = false;
4773 break;
4774 }
4775 }
4776 }
4777
4778 if (no_interfering_store) {
4779 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4780 uint size = alloc->req();
4781 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4782 old_jvms->set_map(sfpt);
4783 for (uint i = 0; i < size; i++) {
4784 sfpt->init_req(i, alloc->in(i));
4785 }
4786 // re-push array length for deoptimization
4787 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4788 old_jvms->set_sp(old_jvms->sp()+1);
4789 old_jvms->set_monoff(old_jvms->monoff()+1);
4790 old_jvms->set_scloff(old_jvms->scloff()+1);
4791 old_jvms->set_endoff(old_jvms->endoff()+1);
4792 old_jvms->set_should_reexecute(true);
4793
4794 sfpt->set_i_o(map()->i_o());
4795 sfpt->set_memory(map()->memory());
4796 sfpt->set_control(map()->control());
4797
4798 JVMState* saved_jvms = jvms();
4799 saved_reexecute_sp = _reexecute_sp;
4800
4801 set_jvms(sfpt->jvms());
4802 _reexecute_sp = jvms()->sp();
4803
4804 return saved_jvms;
4805 }
4806 }
4807 }
4808 return NULL;
4809 }
4810
4811 // In case of a deoptimization, we restart execution at the
4812 // allocation, allocating a new array. We would leave an uninitialized
4813 // array in the heap that GCs wouldn't expect. Move the allocation
4814 // after the traps so we don't allocate the array if we
4815 // deoptimize. This is possible because tightly_coupled_allocation()
4816 // guarantees there's no observer of the allocated array at this point
4817 // and the control flow is simple enough.
4818 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4819 if (saved_jvms != NULL && !stopped()) {
4820 assert(alloc != NULL, "only with a tightly coupled allocation");
4821 // restore JVM state to the state at the arraycopy
4822 saved_jvms->map()->set_control(map()->control());
4823 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4824 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4825 // If we've improved the types of some nodes (null check) while
4826 // emitting the guards, propagate them to the current state
4827 map()->replaced_nodes().apply(saved_jvms->map());
4828 set_jvms(saved_jvms);
4829 _reexecute_sp = saved_reexecute_sp;
4830
4831 // Remove the allocation from above the guards
4832 CallProjections callprojs;
4833 alloc->extract_projections(&callprojs, true);
4834 InitializeNode* init = alloc->initialization();
4835 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4836 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4837 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4838 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4839
4840 // move the allocation here (after the guards)
4841 _gvn.hash_delete(alloc);
4842 alloc->set_req(TypeFunc::Control, control());
4843 alloc->set_req(TypeFunc::I_O, i_o());
4844 Node *mem = reset_memory();
4845 set_all_memory(mem);
4846 alloc->set_req(TypeFunc::Memory, mem);
4847 set_control(init->proj_out(TypeFunc::Control));
4848 set_i_o(callprojs.fallthrough_ioproj);
4849
4850 // Update memory as done in GraphKit::set_output_for_allocation()
4851 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4852 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4853 if (ary_type->isa_aryptr() && length_type != NULL) {
4854 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4855 }
4856 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4857 int elemidx = C->get_alias_index(telemref);
4858 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4859 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4860
4861 Node* allocx = _gvn.transform(alloc);
4862 assert(allocx == alloc, "where has the allocation gone?");
4863 assert(dest->is_CheckCastPP(), "not an allocation result?");
4864
4865 _gvn.hash_delete(dest);
4866 dest->set_req(0, control());
4867 Node* destx = _gvn.transform(dest);
4868 assert(destx == dest, "where has the allocation result gone?");
4869 }
4870 }
4871
4872
4873 //------------------------------inline_arraycopy-----------------------
4874 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4875 // Object dest, int destPos,
4876 // int length);
4877 bool LibraryCallKit::inline_arraycopy() {
4878 // Get the arguments.
4879 Node* src = argument(0); // type: oop
4880 Node* src_offset = argument(1); // type: int
4881 Node* dest = argument(2); // type: oop
4882 Node* dest_offset = argument(3); // type: int
4883 Node* length = argument(4); // type: int
4884
4885
4886 // Check for allocation before we add nodes that would confuse
4887 // tightly_coupled_allocation()
4888 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4889
4890 int saved_reexecute_sp = -1;
4891 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4892 // See arraycopy_restore_alloc_state() comment
4893 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4894 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4895 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4896 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4897
4898 // The following tests must be performed
4899 // (1) src and dest are arrays.
4900 // (2) src and dest arrays must have elements of the same BasicType
4901 // (3) src and dest must not be null.
4902 // (4) src_offset must not be negative.
4903 // (5) dest_offset must not be negative.
4904 // (6) length must not be negative.
4905 // (7) src_offset + length must not exceed length of src.
4906 // (8) dest_offset + length must not exceed length of dest.
4907 // (9) each element of an oop array must be assignable
4908
4909 // (3) src and dest must not be null.
4910 // always do this here because we need the JVM state for uncommon traps
4911 Node* null_ctl = top();
4912 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4913 assert(null_ctl->is_top(), "no null control here");
4914 dest = null_check(dest, T_ARRAY);
4915
4916 if (!can_emit_guards) {
4917 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4918 // guards but the arraycopy node could still take advantage of a
4919 // tightly allocated allocation. tightly_coupled_allocation() is
4920 // called again to make sure it takes the null check above into
4921 // account: the null check is mandatory and if it caused an
4922 // uncommon trap to be emitted then the allocation can't be
4923 // considered tightly coupled in this context.
4924 alloc = tightly_coupled_allocation(dest, NULL);
4925 }
4926
4927 bool validated = false;
4928
4929 const Type* src_type = _gvn.type(src);
4930 const Type* dest_type = _gvn.type(dest);
4931 const TypeAryPtr* top_src = src_type->isa_aryptr();
4932 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4933
4934 // Do we have the type of src?
4935 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4936 // Do we have the type of dest?
4937 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4938 // Is the type for src from speculation?
4939 bool src_spec = false;
4940 // Is the type for dest from speculation?
4941 bool dest_spec = false;
4942
4943 if ((!has_src || !has_dest) && can_emit_guards) {
4944 // We don't have sufficient type information, let's see if
4945 // speculative types can help. We need to have types for both src
4946 // and dest so that it pays off.
4947
4948 // Do we already have or could we have type information for src
4949 bool could_have_src = has_src;
4950 // Do we already have or could we have type information for dest
4951 bool could_have_dest = has_dest;
4952
4953 ciKlass* src_k = NULL;
4954 if (!has_src) {
4955 src_k = src_type->speculative_type_not_null();
4956 if (src_k != NULL && src_k->is_array_klass()) {
4957 could_have_src = true;
4958 }
4959 }
4960
4961 ciKlass* dest_k = NULL;
4962 if (!has_dest) {
4963 dest_k = dest_type->speculative_type_not_null();
4964 if (dest_k != NULL && dest_k->is_array_klass()) {
4965 could_have_dest = true;
4966 }
4967 }
4968
4969 if (could_have_src && could_have_dest) {
4970 // This is going to pay off so emit the required guards
4971 if (!has_src) {
4972 src = maybe_cast_profiled_obj(src, src_k, true);
4973 src_type = _gvn.type(src);
4974 top_src = src_type->isa_aryptr();
4975 has_src = (top_src != NULL && top_src->klass() != NULL);
4976 src_spec = true;
4977 }
4978 if (!has_dest) {
4979 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4980 dest_type = _gvn.type(dest);
4981 top_dest = dest_type->isa_aryptr();
4982 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4983 dest_spec = true;
4984 }
4985 }
4986 }
4987
4988 if (has_src && has_dest && can_emit_guards) {
4989 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4990 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4991 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4992 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4993
4994 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4995 // If both arrays are object arrays then having the exact types
4996 // for both will remove the need for a subtype check at runtime
4997 // before the call and may make it possible to pick a faster copy
4998 // routine (without a subtype check on every element)
4999 // Do we have the exact type of src?
5000 bool could_have_src = src_spec;
5001 // Do we have the exact type of dest?
5002 bool could_have_dest = dest_spec;
5003 ciKlass* src_k = top_src->klass();
5004 ciKlass* dest_k = top_dest->klass();
5005 if (!src_spec) {
5006 src_k = src_type->speculative_type_not_null();
5007 if (src_k != NULL && src_k->is_array_klass()) {
5008 could_have_src = true;
5009 }
5010 }
5011 if (!dest_spec) {
5012 dest_k = dest_type->speculative_type_not_null();
5013 if (dest_k != NULL && dest_k->is_array_klass()) {
5014 could_have_dest = true;
5015 }
5016 }
5017 if (could_have_src && could_have_dest) {
5018 // If we can have both exact types, emit the missing guards
5019 if (could_have_src && !src_spec) {
5020 src = maybe_cast_profiled_obj(src, src_k, true);
5021 }
5022 if (could_have_dest && !dest_spec) {
5023 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5024 }
5025 }
5026 }
5027 }
5028
5029 ciMethod* trap_method = method();
5030 int trap_bci = bci();
5031 if (saved_jvms != NULL) {
5032 trap_method = alloc->jvms()->method();
5033 trap_bci = alloc->jvms()->bci();
5034 }
5035
5036 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5037 can_emit_guards &&
5038 !src->is_top() && !dest->is_top()) {
5039 // validate arguments: enables transformation the ArrayCopyNode
5040 validated = true;
5041
5042 RegionNode* slow_region = new RegionNode(1);
5043 record_for_igvn(slow_region);
5044
5045 // (1) src and dest are arrays.
5046 generate_non_array_guard(load_object_klass(src), slow_region);
5047 generate_non_array_guard(load_object_klass(dest), slow_region);
5048
5049 // (2) src and dest arrays must have elements of the same BasicType
5050 // done at macro expansion or at Ideal transformation time
5051
5052 // (4) src_offset must not be negative.
5053 generate_negative_guard(src_offset, slow_region);
5054
5055 // (5) dest_offset must not be negative.
5056 generate_negative_guard(dest_offset, slow_region);
5057
5058 // (7) src_offset + length must not exceed length of src.
5059 generate_limit_guard(src_offset, length,
5060 load_array_length(src),
5061 slow_region);
5062
5063 // (8) dest_offset + length must not exceed length of dest.
5064 generate_limit_guard(dest_offset, length,
5065 load_array_length(dest),
5066 slow_region);
5067
5068 // (9) each element of an oop array must be assignable
5069 Node* src_klass = load_object_klass(src);
5070 Node* dest_klass = load_object_klass(dest);
5071 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5072
5073 if (not_subtype_ctrl != top()) {
5074 PreserveJVMState pjvms(this);
5075 set_control(not_subtype_ctrl);
5076 uncommon_trap(Deoptimization::Reason_intrinsic,
5077 Deoptimization::Action_make_not_entrant);
5078 assert(stopped(), "Should be stopped");
5079 }
5080 {
5081 PreserveJVMState pjvms(this);
5082 set_control(_gvn.transform(slow_region));
5083 uncommon_trap(Deoptimization::Reason_intrinsic,
5084 Deoptimization::Action_make_not_entrant);
5085 assert(stopped(), "Should be stopped");
5086 }
5087 }
5088
5089 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5090
5091 if (stopped()) {
5092 return true;
5093 }
5094
5095 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5096 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5097 // so the compiler has a chance to eliminate them: during macro expansion,
5098 // we have to set their control (CastPP nodes are eliminated).
5099 load_object_klass(src), load_object_klass(dest),
5100 load_array_length(src), load_array_length(dest));
5101
5102 ac->set_arraycopy(validated);
5103
5104 Node* n = _gvn.transform(ac);
5105 if (n == ac) {
5106 ac->connect_outputs(this);
5107 } else {
5108 assert(validated, "shouldn't transform if all arguments not validated");
5109 set_all_memory(n);
5110 }
5111
5112 return true;
5113 }
5114
5115
5116 // Helper function which determines if an arraycopy immediately follows
5117 // an allocation, with no intervening tests or other escapes for the object.
5118 AllocateArrayNode*
5119 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5120 RegionNode* slow_region) {
5121 if (stopped()) return NULL; // no fast path
5122 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5123
5124 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5125 if (alloc == NULL) return NULL;
5126
5127 Node* rawmem = memory(Compile::AliasIdxRaw);
5128 // Is the allocation's memory state untouched?
5129 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5130 // Bail out if there have been raw-memory effects since the allocation.
5131 // (Example: There might have been a call or safepoint.)
5132 return NULL;
5133 }
5134 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5135 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5136 return NULL;
5137 }
5138
5139 // There must be no unexpected observers of this allocation.
5140 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5141 Node* obs = ptr->fast_out(i);
5142 if (obs != this->map()) {
5143 return NULL;
5144 }
5145 }
5146
5147 // This arraycopy must unconditionally follow the allocation of the ptr.
5148 Node* alloc_ctl = ptr->in(0);
5149 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5150
5151 Node* ctl = control();
5152 while (ctl != alloc_ctl) {
5153 // There may be guards which feed into the slow_region.
5154 // Any other control flow means that we might not get a chance
5155 // to finish initializing the allocated object.
5156 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5157 IfNode* iff = ctl->in(0)->as_If();
5158 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5159 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5160 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5161 ctl = iff->in(0); // This test feeds the known slow_region.
5162 continue;
5163 }
5164 // One more try: Various low-level checks bottom out in
5165 // uncommon traps. If the debug-info of the trap omits
5166 // any reference to the allocation, as we've already
5167 // observed, then there can be no objection to the trap.
5168 bool found_trap = false;
5169 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5170 Node* obs = not_ctl->fast_out(j);
5171 if (obs->in(0) == not_ctl && obs->is_Call() &&
5172 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5173 found_trap = true; break;
5174 }
5175 }
5176 if (found_trap) {
5177 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5178 continue;
5179 }
5180 }
5181 return NULL;
5182 }
5183
5184 // If we get this far, we have an allocation which immediately
5185 // precedes the arraycopy, and we can take over zeroing the new object.
5186 // The arraycopy will finish the initialization, and provide
5187 // a new control state to which we will anchor the destination pointer.
5188
5189 return alloc;
5190 }
5191
5192 //-------------inline_encodeISOArray-----------------------------------
5193 // encode char[] to byte[] in ISO_8859_1
5194 bool LibraryCallKit::inline_encodeISOArray() {
5195 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5196 // no receiver since it is static method
5197 Node *src = argument(0);
5198 Node *src_offset = argument(1);
5199 Node *dst = argument(2);
5200 Node *dst_offset = argument(3);
5201 Node *length = argument(4);
5202
5203 const Type* src_type = src->Value(&_gvn);
5204 const Type* dst_type = dst->Value(&_gvn);
5205 const TypeAryPtr* top_src = src_type->isa_aryptr();
5206 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5207 if (top_src == NULL || top_src->klass() == NULL ||
5208 top_dest == NULL || top_dest->klass() == NULL) {
5209 // failed array check
5210 return false;
5211 }
5212
5213 // Figure out the size and type of the elements we will be copying.
5214 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5215 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5216 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5217 return false;
5218 }
5219 Node* src_start = array_element_address(src, src_offset, src_elem);
5220 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5221 // 'src_start' points to src array + scaled offset
5222 // 'dst_start' points to dst array + scaled offset
5223
5224 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5225 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5226 enc = _gvn.transform(enc);
5227 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5228 set_memory(res_mem, mtype);
5229 set_result(enc);
5230 return true;
5231 }
5232
5233 //-------------inline_multiplyToLen-----------------------------------
5234 bool LibraryCallKit::inline_multiplyToLen() {
5235 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5236
5237 address stubAddr = StubRoutines::multiplyToLen();
5238 if (stubAddr == NULL) {
5239 return false; // Intrinsic's stub is not implemented on this platform
5240 }
5241 const char* stubName = "multiplyToLen";
5242
5243 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5244
5245 Node* x = argument(1);
5246 Node* xlen = argument(2);
5247 Node* y = argument(3);
5248 Node* ylen = argument(4);
5249 Node* z = argument(5);
5250
5251 const Type* x_type = x->Value(&_gvn);
5252 const Type* y_type = y->Value(&_gvn);
5253 const TypeAryPtr* top_x = x_type->isa_aryptr();
5254 const TypeAryPtr* top_y = y_type->isa_aryptr();
5255 if (top_x == NULL || top_x->klass() == NULL ||
5256 top_y == NULL || top_y->klass() == NULL) {
5257 // failed array check
5258 return false;
5259 }
5260
5261 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5262 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5263 if (x_elem != T_INT || y_elem != T_INT) {
5264 return false;
5265 }
5266
5267 // Set the original stack and the reexecute bit for the interpreter to reexecute
5268 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5269 // on the return from z array allocation in runtime.
5270 { PreserveReexecuteState preexecs(this);
5271 jvms()->set_should_reexecute(true);
5272
5273 Node* x_start = array_element_address(x, intcon(0), x_elem);
5274 Node* y_start = array_element_address(y, intcon(0), y_elem);
5275 // 'x_start' points to x array + scaled xlen
5276 // 'y_start' points to y array + scaled ylen
5277
5278 // Allocate the result array
5279 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5280 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5281 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5282
5283 IdealKit ideal(this);
5284
5285 #define __ ideal.
5286 Node* one = __ ConI(1);
5287 Node* zero = __ ConI(0);
5288 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5289 __ set(need_alloc, zero);
5290 __ set(z_alloc, z);
5291 __ if_then(z, BoolTest::eq, null()); {
5292 __ increment (need_alloc, one);
5293 } __ else_(); {
5294 // Update graphKit memory and control from IdealKit.
5295 sync_kit(ideal);
5296 Node* zlen_arg = load_array_length(z);
5297 // Update IdealKit memory and control from graphKit.
5298 __ sync_kit(this);
5299 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5300 __ increment (need_alloc, one);
5301 } __ end_if();
5302 } __ end_if();
5303
5304 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5305 // Update graphKit memory and control from IdealKit.
5306 sync_kit(ideal);
5307 Node * narr = new_array(klass_node, zlen, 1);
5308 // Update IdealKit memory and control from graphKit.
5309 __ sync_kit(this);
5310 __ set(z_alloc, narr);
5311 } __ end_if();
5312
5313 sync_kit(ideal);
5314 z = __ value(z_alloc);
5315 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5316 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5317 // Final sync IdealKit and GraphKit.
5318 final_sync(ideal);
5319 #undef __
5320
5321 Node* z_start = array_element_address(z, intcon(0), T_INT);
5322
5323 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5324 OptoRuntime::multiplyToLen_Type(),
5325 stubAddr, stubName, TypePtr::BOTTOM,
5326 x_start, xlen, y_start, ylen, z_start, zlen);
5327 } // original reexecute is set back here
5328
5329 C->set_has_split_ifs(true); // Has chance for split-if optimization
5330 set_result(z);
5331 return true;
5332 }
5333
5334 //-------------inline_squareToLen------------------------------------
5335 bool LibraryCallKit::inline_squareToLen() {
5336 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5337
5338 address stubAddr = StubRoutines::squareToLen();
5339 if (stubAddr == NULL) {
5340 return false; // Intrinsic's stub is not implemented on this platform
5341 }
5342 const char* stubName = "squareToLen";
5343
5344 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5345
5346 Node* x = argument(0);
5347 Node* len = argument(1);
5348 Node* z = argument(2);
5349 Node* zlen = argument(3);
5350
5351 const Type* x_type = x->Value(&_gvn);
5352 const Type* z_type = z->Value(&_gvn);
5353 const TypeAryPtr* top_x = x_type->isa_aryptr();
5354 const TypeAryPtr* top_z = z_type->isa_aryptr();
5355 if (top_x == NULL || top_x->klass() == NULL ||
5356 top_z == NULL || top_z->klass() == NULL) {
5357 // failed array check
5358 return false;
5359 }
5360
5361 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5362 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5363 if (x_elem != T_INT || z_elem != T_INT) {
5364 return false;
5365 }
5366
5367
5368 Node* x_start = array_element_address(x, intcon(0), x_elem);
5369 Node* z_start = array_element_address(z, intcon(0), z_elem);
5370
5371 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5372 OptoRuntime::squareToLen_Type(),
5373 stubAddr, stubName, TypePtr::BOTTOM,
5374 x_start, len, z_start, zlen);
5375
5376 set_result(z);
5377 return true;
5378 }
5379
5380 //-------------inline_mulAdd------------------------------------------
5381 bool LibraryCallKit::inline_mulAdd() {
5382 assert(UseMulAddIntrinsic, "not implementated on this platform");
5383
5384 address stubAddr = StubRoutines::mulAdd();
5385 if (stubAddr == NULL) {
5386 return false; // Intrinsic's stub is not implemented on this platform
5387 }
5388 const char* stubName = "mulAdd";
5389
5390 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5391
5392 Node* out = argument(0);
5393 Node* in = argument(1);
5394 Node* offset = argument(2);
5395 Node* len = argument(3);
5396 Node* k = argument(4);
5397
5398 const Type* out_type = out->Value(&_gvn);
5399 const Type* in_type = in->Value(&_gvn);
5400 const TypeAryPtr* top_out = out_type->isa_aryptr();
5401 const TypeAryPtr* top_in = in_type->isa_aryptr();
5402 if (top_out == NULL || top_out->klass() == NULL ||
5403 top_in == NULL || top_in->klass() == NULL) {
5404 // failed array check
5405 return false;
5406 }
5407
5408 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5409 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5410 if (out_elem != T_INT || in_elem != T_INT) {
5411 return false;
5412 }
5413
5414 Node* outlen = load_array_length(out);
5415 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5416 Node* out_start = array_element_address(out, intcon(0), out_elem);
5417 Node* in_start = array_element_address(in, intcon(0), in_elem);
5418
5419 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5420 OptoRuntime::mulAdd_Type(),
5421 stubAddr, stubName, TypePtr::BOTTOM,
5422 out_start,in_start, new_offset, len, k);
5423 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5424 set_result(result);
5425 return true;
5426 }
5427
5428
5429 /**
5430 * Calculate CRC32 for byte.
5431 * int java.util.zip.CRC32.update(int crc, int b)
5432 */
5433 bool LibraryCallKit::inline_updateCRC32() {
5434 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5435 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5436 // no receiver since it is static method
5437 Node* crc = argument(0); // type: int
5438 Node* b = argument(1); // type: int
5439
5440 /*
5441 * int c = ~ crc;
5442 * b = timesXtoThe32[(b ^ c) & 0xFF];
5443 * b = b ^ (c >>> 8);
5444 * crc = ~b;
5445 */
5446
5447 Node* M1 = intcon(-1);
5448 crc = _gvn.transform(new XorINode(crc, M1));
5449 Node* result = _gvn.transform(new XorINode(crc, b));
5450 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5451
5452 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5453 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5454 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5455 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5456
5457 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5458 result = _gvn.transform(new XorINode(crc, result));
5459 result = _gvn.transform(new XorINode(result, M1));
5460 set_result(result);
5461 return true;
5462 }
5463
5464 /**
5465 * Calculate CRC32 for byte[] array.
5466 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5467 */
5468 bool LibraryCallKit::inline_updateBytesCRC32() {
5469 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5470 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5471 // no receiver since it is static method
5472 Node* crc = argument(0); // type: int
5473 Node* src = argument(1); // type: oop
5474 Node* offset = argument(2); // type: int
5475 Node* length = argument(3); // type: int
5476
5477 const Type* src_type = src->Value(&_gvn);
5478 const TypeAryPtr* top_src = src_type->isa_aryptr();
5479 if (top_src == NULL || top_src->klass() == NULL) {
5480 // failed array check
5481 return false;
5482 }
5483
5484 // Figure out the size and type of the elements we will be copying.
5485 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5486 if (src_elem != T_BYTE) {
5487 return false;
5488 }
5489
5490 // 'src_start' points to src array + scaled offset
5491 Node* src_start = array_element_address(src, offset, src_elem);
5492
5493 // We assume that range check is done by caller.
5494 // TODO: generate range check (offset+length < src.length) in debug VM.
5495
5496 // Call the stub.
5497 address stubAddr = StubRoutines::updateBytesCRC32();
5498 const char *stubName = "updateBytesCRC32";
5499
5500 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5501 stubAddr, stubName, TypePtr::BOTTOM,
5502 crc, src_start, length);
5503 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5504 set_result(result);
5505 return true;
5506 }
5507
5508 /**
5509 * Calculate CRC32 for ByteBuffer.
5510 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5511 */
5512 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5513 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5514 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5515 // no receiver since it is static method
5516 Node* crc = argument(0); // type: int
5517 Node* src = argument(1); // type: long
5518 Node* offset = argument(3); // type: int
5519 Node* length = argument(4); // type: int
5520
5521 src = ConvL2X(src); // adjust Java long to machine word
5522 Node* base = _gvn.transform(new CastX2PNode(src));
5523 offset = ConvI2X(offset);
5524
5525 // 'src_start' points to src array + scaled offset
5526 Node* src_start = basic_plus_adr(top(), base, offset);
5527
5528 // Call the stub.
5529 address stubAddr = StubRoutines::updateBytesCRC32();
5530 const char *stubName = "updateBytesCRC32";
5531
5532 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5533 stubAddr, stubName, TypePtr::BOTTOM,
5534 crc, src_start, length);
5535 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5536 set_result(result);
5537 return true;
5538 }
5539
5540 //----------------------------inline_reference_get----------------------------
5541 // public T java.lang.ref.Reference.get();
5542 bool LibraryCallKit::inline_reference_get() {
5543 const int referent_offset = java_lang_ref_Reference::referent_offset;
5544 guarantee(referent_offset > 0, "should have already been set");
5545
5546 // Get the argument:
5547 Node* reference_obj = null_check_receiver();
5548 if (stopped()) return true;
5549
5550 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5551
5552 ciInstanceKlass* klass = env()->Object_klass();
5553 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5554
5555 Node* no_ctrl = NULL;
5556 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5557
5558 // Use the pre-barrier to record the value in the referent field
5559 pre_barrier(false /* do_load */,
5560 control(),
5561 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5562 result /* pre_val */,
5563 T_OBJECT);
5564
5565 // Add memory barrier to prevent commoning reads from this field
5566 // across safepoint since GC can change its value.
5567 insert_mem_bar(Op_MemBarCPUOrder);
5568
5569 set_result(result);
5570 return true;
5571 }
5572
5573
5574 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5575 bool is_exact=true, bool is_static=false) {
5576
5577 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5578 assert(tinst != NULL, "obj is null");
5579 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5580 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5581
5582 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5583 ciSymbol::make(fieldTypeString),
5584 is_static);
5585 if (field == NULL) return (Node *) NULL;
5586 assert (field != NULL, "undefined field");
5587
5588 // Next code copied from Parse::do_get_xxx():
5589
5590 // Compute address and memory type.
5591 int offset = field->offset_in_bytes();
5592 bool is_vol = field->is_volatile();
5593 ciType* field_klass = field->type();
5594 assert(field_klass->is_loaded(), "should be loaded");
5595 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5596 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5597 BasicType bt = field->layout_type();
5598
5599 // Build the resultant type of the load
5600 const Type *type;
5601 if (bt == T_OBJECT) {
5602 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5603 } else {
5604 type = Type::get_const_basic_type(bt);
5605 }
5606
5607 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5608 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5609 }
5610 // Build the load.
5611 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5612 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5613 // If reference is volatile, prevent following memory ops from
5614 // floating up past the volatile read. Also prevents commoning
5615 // another volatile read.
5616 if (is_vol) {
5617 // Memory barrier includes bogus read of value to force load BEFORE membar
5618 insert_mem_bar(Op_MemBarAcquire, loadedField);
5619 }
5620 return loadedField;
5621 }
5622
5623
5624 //------------------------------inline_aescrypt_Block-----------------------
5625 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5626 address stubAddr;
5627 const char *stubName;
5628 assert(UseAES, "need AES instruction support");
5629
5630 switch(id) {
5631 case vmIntrinsics::_aescrypt_encryptBlock:
5632 stubAddr = StubRoutines::aescrypt_encryptBlock();
5633 stubName = "aescrypt_encryptBlock";
5634 break;
5635 case vmIntrinsics::_aescrypt_decryptBlock:
5636 stubAddr = StubRoutines::aescrypt_decryptBlock();
5637 stubName = "aescrypt_decryptBlock";
5638 break;
5639 }
5640 if (stubAddr == NULL) return false;
5641
5642 Node* aescrypt_object = argument(0);
5643 Node* src = argument(1);
5644 Node* src_offset = argument(2);
5645 Node* dest = argument(3);
5646 Node* dest_offset = argument(4);
5647
5648 // (1) src and dest are arrays.
5649 const Type* src_type = src->Value(&_gvn);
5650 const Type* dest_type = dest->Value(&_gvn);
5651 const TypeAryPtr* top_src = src_type->isa_aryptr();
5652 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5653 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5654
5655 // for the quick and dirty code we will skip all the checks.
5656 // we are just trying to get the call to be generated.
5657 Node* src_start = src;
5658 Node* dest_start = dest;
5659 if (src_offset != NULL || dest_offset != NULL) {
5660 assert(src_offset != NULL && dest_offset != NULL, "");
5661 src_start = array_element_address(src, src_offset, T_BYTE);
5662 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5663 }
5664
5665 // now need to get the start of its expanded key array
5666 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5667 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5668 if (k_start == NULL) return false;
5669
5670 if (Matcher::pass_original_key_for_aes()) {
5671 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5672 // compatibility issues between Java key expansion and SPARC crypto instructions
5673 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5674 if (original_k_start == NULL) return false;
5675
5676 // Call the stub.
5677 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5678 stubAddr, stubName, TypePtr::BOTTOM,
5679 src_start, dest_start, k_start, original_k_start);
5680 } else {
5681 // Call the stub.
5682 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5683 stubAddr, stubName, TypePtr::BOTTOM,
5684 src_start, dest_start, k_start);
5685 }
5686
5687 return true;
5688 }
5689
5690 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5691 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5692 address stubAddr;
5693 const char *stubName;
5694
5695 assert(UseAES, "need AES instruction support");
5696
5697 switch(id) {
5698 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5699 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5700 stubName = "cipherBlockChaining_encryptAESCrypt";
5701 break;
5702 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5703 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5704 stubName = "cipherBlockChaining_decryptAESCrypt";
5705 break;
5706 }
5707 if (stubAddr == NULL) return false;
5708
5709 Node* cipherBlockChaining_object = argument(0);
5710 Node* src = argument(1);
5711 Node* src_offset = argument(2);
5712 Node* len = argument(3);
5713 Node* dest = argument(4);
5714 Node* dest_offset = argument(5);
5715
5716 // (1) src and dest are arrays.
5717 const Type* src_type = src->Value(&_gvn);
5718 const Type* dest_type = dest->Value(&_gvn);
5719 const TypeAryPtr* top_src = src_type->isa_aryptr();
5720 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5721 assert (top_src != NULL && top_src->klass() != NULL
5722 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5723
5724 // checks are the responsibility of the caller
5725 Node* src_start = src;
5726 Node* dest_start = dest;
5727 if (src_offset != NULL || dest_offset != NULL) {
5728 assert(src_offset != NULL && dest_offset != NULL, "");
5729 src_start = array_element_address(src, src_offset, T_BYTE);
5730 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5731 }
5732
5733 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5734 // (because of the predicated logic executed earlier).
5735 // so we cast it here safely.
5736 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5737
5738 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5739 if (embeddedCipherObj == NULL) return false;
5740
5741 // cast it to what we know it will be at runtime
5742 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5743 assert(tinst != NULL, "CBC obj is null");
5744 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5745 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5746 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5747
5748 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5749 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5750 const TypeOopPtr* xtype = aklass->as_instance_type();
5751 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5752 aescrypt_object = _gvn.transform(aescrypt_object);
5753
5754 // we need to get the start of the aescrypt_object's expanded key array
5755 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5756 if (k_start == NULL) return false;
5757
5758 // similarly, get the start address of the r vector
5759 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5760 if (objRvec == NULL) return false;
5761 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5762
5763 Node* cbcCrypt;
5764 if (Matcher::pass_original_key_for_aes()) {
5765 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5766 // compatibility issues between Java key expansion and SPARC crypto instructions
5767 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5768 if (original_k_start == NULL) return false;
5769
5770 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5771 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5772 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5773 stubAddr, stubName, TypePtr::BOTTOM,
5774 src_start, dest_start, k_start, r_start, len, original_k_start);
5775 } else {
5776 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5777 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5778 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5779 stubAddr, stubName, TypePtr::BOTTOM,
5780 src_start, dest_start, k_start, r_start, len);
5781 }
5782
5783 // return cipher length (int)
5784 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5785 set_result(retvalue);
5786 return true;
5787 }
5788
5789 //------------------------------get_key_start_from_aescrypt_object-----------------------
5790 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5791 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5792 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5793 if (objAESCryptKey == NULL) return (Node *) NULL;
5794
5795 // now have the array, need to get the start address of the K array
5796 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5797 return k_start;
5798 }
5799
5800 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5801 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5802 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5803 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5804 if (objAESCryptKey == NULL) return (Node *) NULL;
5805
5806 // now have the array, need to get the start address of the lastKey array
5807 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5808 return original_k_start;
5809 }
5810
5811 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5812 // Return node representing slow path of predicate check.
5813 // the pseudo code we want to emulate with this predicate is:
5814 // for encryption:
5815 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5816 // for decryption:
5817 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5818 // note cipher==plain is more conservative than the original java code but that's OK
5819 //
5820 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5821 // The receiver was checked for NULL already.
5822 Node* objCBC = argument(0);
5823
5824 // Load embeddedCipher field of CipherBlockChaining object.
5825 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5826
5827 // get AESCrypt klass for instanceOf check
5828 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5829 // will have same classloader as CipherBlockChaining object
5830 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5831 assert(tinst != NULL, "CBCobj is null");
5832 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5833
5834 // we want to do an instanceof comparison against the AESCrypt class
5835 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5836 if (!klass_AESCrypt->is_loaded()) {
5837 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5838 Node* ctrl = control();
5839 set_control(top()); // no regular fast path
5840 return ctrl;
5841 }
5842 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5843
5844 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5845 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5846 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5847
5848 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5849
5850 // for encryption, we are done
5851 if (!decrypting)
5852 return instof_false; // even if it is NULL
5853
5854 // for decryption, we need to add a further check to avoid
5855 // taking the intrinsic path when cipher and plain are the same
5856 // see the original java code for why.
5857 RegionNode* region = new RegionNode(3);
5858 region->init_req(1, instof_false);
5859 Node* src = argument(1);
5860 Node* dest = argument(4);
5861 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5862 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5863 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5864 region->init_req(2, src_dest_conjoint);
5865
5866 record_for_igvn(region);
5867 return _gvn.transform(region);
5868 }
5869
5870 //------------------------------get_vars_from_ghash_object-----------------------
5871 Node * LibraryCallKit::get_vars_from_ghash_object(Node *ghash_object, const char *var_name) {
5872 Node* ghash_var = load_field_from_object(ghash_object, var_name, "[J", /*is_exact*/ false);
5873 assert (ghash_var != NULL, "wrong version of sun.security.provider.GHASH");
5874 if (ghash_var == NULL) return (Node *) NULL;
5875
5876 // now have the array, need to get the start address of the array
5877 Node* var = array_element_address(ghash_var, intcon(0), T_LONG);
5878 return var;
5879 }
5880
5881 //------------------------------inline_ghash_processBlocks
5882 bool LibraryCallKit::inline_ghash_processBlocks() {
5883 address stubAddr;
5884 const char *stubName;
5885 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5886
5887 stubAddr = StubRoutines::ghash_processBlocks();
5888 stubName = "ghash_processBlocks";
5889
5890 Node* ghash_object = argument(0);
5891 Node* data = argument(1);
5892 Node* offset = argument(2);
5893 Node* len = argument(3);
5894
5895 Node* state_start = get_vars_from_ghash_object(ghash_object, "state");
5896 assert(state_start, "Unable to load GHASH state");
5897 Node* subkeyH_start = get_vars_from_ghash_object(ghash_object, "subkeyH");
5898 assert(subkeyH_start, "Unable to load GHASH subkeyH");
5899 Node* data_start = array_element_address(data, offset, T_BYTE);
5900 assert(data_start, "data is NULL");
5901
5902 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
5903 OptoRuntime::ghash_processBlocks_Type(),
5904 stubAddr, stubName, TypePtr::BOTTOM,
5905 state_start, subkeyH_start, data_start, len);
5906 return true;
5907 }
5908
5909 //------------------------------inline_sha_implCompress-----------------------
5910 //
5911 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5912 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5913 //
5914 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
5915 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
5916 //
5917 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
5918 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
5919 //
5920 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
5921 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
5922
5923 Node* sha_obj = argument(0);
5924 Node* src = argument(1); // type oop
5925 Node* ofs = argument(2); // type int
5926
5927 const Type* src_type = src->Value(&_gvn);
5928 const TypeAryPtr* top_src = src_type->isa_aryptr();
5929 if (top_src == NULL || top_src->klass() == NULL) {
5930 // failed array check
5931 return false;
5932 }
5933 // Figure out the size and type of the elements we will be copying.
5934 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5935 if (src_elem != T_BYTE) {
5936 return false;
5937 }
5938 // 'src_start' points to src array + offset
5939 Node* src_start = array_element_address(src, ofs, src_elem);
5940 Node* state = NULL;
5941 address stubAddr;
5942 const char *stubName;
5943
5944 switch(id) {
5945 case vmIntrinsics::_sha_implCompress:
5946 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
5947 state = get_state_from_sha_object(sha_obj);
5948 stubAddr = StubRoutines::sha1_implCompress();
5949 stubName = "sha1_implCompress";
5950 break;
5951 case vmIntrinsics::_sha2_implCompress:
5952 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
5953 state = get_state_from_sha_object(sha_obj);
5954 stubAddr = StubRoutines::sha256_implCompress();
5955 stubName = "sha256_implCompress";
5956 break;
5957 case vmIntrinsics::_sha5_implCompress:
5958 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
5959 state = get_state_from_sha5_object(sha_obj);
5960 stubAddr = StubRoutines::sha512_implCompress();
5961 stubName = "sha512_implCompress";
5962 break;
5963 default:
5964 fatal_unexpected_iid(id);
5965 return false;
5966 }
5967 if (state == NULL) return false;
5968
5969 // Call the stub.
5970 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
5971 stubAddr, stubName, TypePtr::BOTTOM,
5972 src_start, state);
5973
5974 return true;
5975 }
5976
5977 //------------------------------inline_digestBase_implCompressMB-----------------------
5978 //
5979 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
5980 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
5981 //
5982 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
5983 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5984 "need SHA1/SHA256/SHA512 instruction support");
5985 assert((uint)predicate < 3, "sanity");
5986 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
5987
5988 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
5989 Node* src = argument(1); // byte[] array
5990 Node* ofs = argument(2); // type int
5991 Node* limit = argument(3); // type int
5992
5993 const Type* src_type = src->Value(&_gvn);
5994 const TypeAryPtr* top_src = src_type->isa_aryptr();
5995 if (top_src == NULL || top_src->klass() == NULL) {
5996 // failed array check
5997 return false;
5998 }
5999 // Figure out the size and type of the elements we will be copying.
6000 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6001 if (src_elem != T_BYTE) {
6002 return false;
6003 }
6004 // 'src_start' points to src array + offset
6005 Node* src_start = array_element_address(src, ofs, src_elem);
6006
6007 const char* klass_SHA_name = NULL;
6008 const char* stub_name = NULL;
6009 address stub_addr = NULL;
6010 bool long_state = false;
6011
6012 switch (predicate) {
6013 case 0:
6014 if (UseSHA1Intrinsics) {
6015 klass_SHA_name = "sun/security/provider/SHA";
6016 stub_name = "sha1_implCompressMB";
6017 stub_addr = StubRoutines::sha1_implCompressMB();
6018 }
6019 break;
6020 case 1:
6021 if (UseSHA256Intrinsics) {
6022 klass_SHA_name = "sun/security/provider/SHA2";
6023 stub_name = "sha256_implCompressMB";
6024 stub_addr = StubRoutines::sha256_implCompressMB();
6025 }
6026 break;
6027 case 2:
6028 if (UseSHA512Intrinsics) {
6029 klass_SHA_name = "sun/security/provider/SHA5";
6030 stub_name = "sha512_implCompressMB";
6031 stub_addr = StubRoutines::sha512_implCompressMB();
6032 long_state = true;
6033 }
6034 break;
6035 default:
6036 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6037 }
6038 if (klass_SHA_name != NULL) {
6039 // get DigestBase klass to lookup for SHA klass
6040 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6041 assert(tinst != NULL, "digestBase_obj is not instance???");
6042 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6043
6044 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6045 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6046 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6047 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6048 }
6049 return false;
6050 }
6051 //------------------------------inline_sha_implCompressMB-----------------------
6052 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6053 bool long_state, address stubAddr, const char *stubName,
6054 Node* src_start, Node* ofs, Node* limit) {
6055 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6056 const TypeOopPtr* xtype = aklass->as_instance_type();
6057 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6058 sha_obj = _gvn.transform(sha_obj);
6059
6060 Node* state;
6061 if (long_state) {
6062 state = get_state_from_sha5_object(sha_obj);
6063 } else {
6064 state = get_state_from_sha_object(sha_obj);
6065 }
6066 if (state == NULL) return false;
6067
6068 // Call the stub.
6069 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6070 OptoRuntime::digestBase_implCompressMB_Type(),
6071 stubAddr, stubName, TypePtr::BOTTOM,
6072 src_start, state, ofs, limit);
6073 // return ofs (int)
6074 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6075 set_result(result);
6076
6077 return true;
6078 }
6079
6080 //------------------------------get_state_from_sha_object-----------------------
6081 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6082 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6083 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6084 if (sha_state == NULL) return (Node *) NULL;
6085
6086 // now have the array, need to get the start address of the state array
6087 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6088 return state;
6089 }
6090
6091 //------------------------------get_state_from_sha5_object-----------------------
6092 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6093 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6094 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6095 if (sha_state == NULL) return (Node *) NULL;
6096
6097 // now have the array, need to get the start address of the state array
6098 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6099 return state;
6100 }
6101
6102 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6103 // Return node representing slow path of predicate check.
6104 // the pseudo code we want to emulate with this predicate is:
6105 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6106 //
6107 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6108 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6109 "need SHA1/SHA256/SHA512 instruction support");
6110 assert((uint)predicate < 3, "sanity");
6111
6112 // The receiver was checked for NULL already.
6113 Node* digestBaseObj = argument(0);
6114
6115 // get DigestBase klass for instanceOf check
6116 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6117 assert(tinst != NULL, "digestBaseObj is null");
6118 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6119
6120 const char* klass_SHA_name = NULL;
6121 switch (predicate) {
6122 case 0:
6123 if (UseSHA1Intrinsics) {
6124 // we want to do an instanceof comparison against the SHA class
6125 klass_SHA_name = "sun/security/provider/SHA";
6126 }
6127 break;
6128 case 1:
6129 if (UseSHA256Intrinsics) {
6130 // we want to do an instanceof comparison against the SHA2 class
6131 klass_SHA_name = "sun/security/provider/SHA2";
6132 }
6133 break;
6134 case 2:
6135 if (UseSHA512Intrinsics) {
6136 // we want to do an instanceof comparison against the SHA5 class
6137 klass_SHA_name = "sun/security/provider/SHA5";
6138 }
6139 break;
6140 default:
6141 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6142 }
6143
6144 ciKlass* klass_SHA = NULL;
6145 if (klass_SHA_name != NULL) {
6146 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6147 }
6148 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6149 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6150 Node* ctrl = control();
6151 set_control(top()); // no intrinsic path
6152 return ctrl;
6153 }
6154 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6155
6156 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6157 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6158 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6159 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6160
6161 return instof_false; // even if it is NULL
6162 }
6163
6164 bool LibraryCallKit::inline_profileBoolean() {
6165 Node* counts = argument(1);
6166 const TypeAryPtr* ary = NULL;
6167 ciArray* aobj = NULL;
6168 if (counts->is_Con()
6169 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6170 && (aobj = ary->const_oop()->as_array()) != NULL
6171 && (aobj->length() == 2)) {
6172 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6173 jint false_cnt = aobj->element_value(0).as_int();
6174 jint true_cnt = aobj->element_value(1).as_int();
6175
6176 method()->set_injected_profile(true);
6177
6178 if (C->log() != NULL) {
6179 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6180 false_cnt, true_cnt);
6181 }
6182
6183 if (false_cnt + true_cnt == 0) {
6184 // According to profile, never executed.
6185 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6186 Deoptimization::Action_reinterpret);
6187 return true;
6188 }
6189
6190 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6191 // is a number of each value occurrences.
6192 Node* result = argument(0);
6193 if (false_cnt == 0 || true_cnt == 0) {
6194 // According to profile, one value has been never seen.
6195 int expected_val = (false_cnt == 0) ? 1 : 0;
6196
6197 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6198 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6199
6200 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6201 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6202 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6203
6204 { // Slow path: uncommon trap for never seen value and then reexecute
6205 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6206 // the value has been seen at least once.
6207 PreserveJVMState pjvms(this);
6208 PreserveReexecuteState preexecs(this);
6209 jvms()->set_should_reexecute(true);
6210
6211 set_control(slow_path);
6212 set_i_o(i_o());
6213
6214 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6215 Deoptimization::Action_reinterpret);
6216 }
6217 // The guard for never seen value enables sharpening of the result and
6218 // returning a constant. It allows to eliminate branches on the same value
6219 // later on.
6220 set_control(fast_path);
6221 result = intcon(expected_val);
6222 }
6223 // Stop profiling.
6224 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6225 // By replacing method body with profile data (represented as ProfileBooleanNode
6226 // on IR level) we effectively disable profiling.
6227 // It enables full speed execution once optimized code is generated.
6228 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6229 C->record_for_igvn(profile);
6230 set_result(profile);
6231 return true;
6232 } else {
6233 // Continue profiling.
6234 // Profile data isn't available at the moment. So, execute method's bytecode version.
6235 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6236 // is compiled and counters aren't available since corresponding MethodHandle
6237 // isn't a compile-time constant.
6238 return false;
6239 }
6240 }
6241
6242 bool LibraryCallKit::inline_isCompileConstant() {
6243 Node* n = argument(0);
6244 set_result(n->is_Con() ? intcon(1) : intcon(0));
6245 return true;
6246 }