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