1 /*
2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callGenerator.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/countbitsnode.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/mathexactnode.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "prims/nativeLookup.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "trace/traceMacros.hpp"
52
53 class LibraryIntrinsic : public InlineCallGenerator {
54 // Extend the set of intrinsics known to the runtime:
55 public:
56 private:
57 bool _is_virtual;
58 bool _does_virtual_dispatch;
59 int8_t _predicates_count; // Intrinsic is predicated by several conditions
60 int8_t _last_predicate; // Last generated predicate
61 vmIntrinsics::ID _intrinsic_id;
62
63 public:
64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
65 : InlineCallGenerator(m),
66 _is_virtual(is_virtual),
67 _does_virtual_dispatch(does_virtual_dispatch),
68 _predicates_count((int8_t)predicates_count),
69 _last_predicate((int8_t)-1),
70 _intrinsic_id(id)
71 {
72 }
73 virtual bool is_intrinsic() const { return true; }
74 virtual bool is_virtual() const { return _is_virtual; }
75 virtual bool is_predicated() const { return _predicates_count > 0; }
76 virtual int predicates_count() const { return _predicates_count; }
77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
78 virtual JVMState* generate(JVMState* jvms);
79 virtual Node* generate_predicate(JVMState* jvms, int predicate);
80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
81 };
82
83
84 // Local helper class for LibraryIntrinsic:
85 class LibraryCallKit : public GraphKit {
86 private:
87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
88 Node* _result; // the result node, if any
89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
90
91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
92
93 public:
94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
95 : GraphKit(jvms),
96 _intrinsic(intrinsic),
97 _result(NULL)
98 {
99 // Check if this is a root compile. In that case we don't have a caller.
100 if (!jvms->has_method()) {
101 _reexecute_sp = sp();
102 } else {
103 // Find out how many arguments the interpreter needs when deoptimizing
104 // and save the stack pointer value so it can used by uncommon_trap.
105 // We find the argument count by looking at the declared signature.
106 bool ignored_will_link;
107 ciSignature* declared_signature = NULL;
108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
111 }
112 }
113
114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
115
116 ciMethod* caller() const { return jvms()->method(); }
117 int bci() const { return jvms()->bci(); }
118 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
120 ciMethod* callee() const { return _intrinsic->method(); }
121
122 bool try_to_inline(int predicate);
123 Node* try_to_predicate(int predicate);
124
125 void push_result() {
126 // Push the result onto the stack.
127 if (!stopped() && result() != NULL) {
128 BasicType bt = result()->bottom_type()->basic_type();
129 push_node(bt, result());
130 }
131 }
132
133 private:
134 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
136 }
137
138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
139 void set_result(RegionNode* region, PhiNode* value);
140 Node* result() { return _result; }
141
142 virtual int reexecute_sp() { return _reexecute_sp; }
143
144 // Helper functions to inline natives
145 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
146 Node* generate_slow_guard(Node* test, RegionNode* region);
147 Node* generate_fair_guard(Node* test, RegionNode* region);
148 Node* generate_negative_guard(Node* index, RegionNode* region,
149 // resulting CastII of index:
150 Node* *pos_index = NULL);
151 Node* generate_limit_guard(Node* offset, Node* subseq_length,
152 Node* array_length,
153 RegionNode* region);
154 Node* generate_current_thread(Node* &tls_output);
155 Node* load_mirror_from_klass(Node* klass);
156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
157 RegionNode* region, int null_path,
158 int offset);
159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::array_klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* generate_access_flags_guard(Node* kls,
174 int modifier_mask, int modifier_bits,
175 RegionNode* region);
176 Node* generate_interface_guard(Node* kls, RegionNode* region);
177 Node* generate_array_guard(Node* kls, RegionNode* region) {
178 return generate_array_guard_common(kls, region, false, false);
179 }
180 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, false, true);
182 }
183 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, true, false);
185 }
186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, true, true);
188 }
189 Node* generate_array_guard_common(Node* kls, RegionNode* region,
190 bool obj_array, bool not_array);
191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
193 bool is_virtual = false, bool is_static = false);
194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
195 return generate_method_call(method_id, false, true);
196 }
197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
198 return generate_method_call(method_id, true, false);
199 }
200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
201
202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
203 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
204 bool inline_string_compareTo();
205 bool inline_string_indexOf();
206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
207 bool inline_string_equals();
208 Node* round_double_node(Node* n);
209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
210 bool inline_math_native(vmIntrinsics::ID id);
211 bool inline_trig(vmIntrinsics::ID id);
212 bool inline_math(vmIntrinsics::ID id);
213 template <typename OverflowOp>
214 bool inline_math_overflow(Node* arg1, Node* arg2);
215 void inline_math_mathExact(Node* math, Node* test);
216 bool inline_math_addExactI(bool is_increment);
217 bool inline_math_addExactL(bool is_increment);
218 bool inline_math_multiplyExactI();
219 bool inline_math_multiplyExactL();
220 bool inline_math_negateExactI();
221 bool inline_math_negateExactL();
222 bool inline_math_subtractExactI(bool is_decrement);
223 bool inline_math_subtractExactL(bool is_decrement);
224 bool inline_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 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
265 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
266
267 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
268 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
269 bool inline_unsafe_ordered_store(BasicType type);
270 bool inline_unsafe_fence(vmIntrinsics::ID id);
271 bool inline_fp_conversions(vmIntrinsics::ID id);
272 bool inline_number_methods(vmIntrinsics::ID id);
273 bool inline_reference_get();
274 bool inline_Class_cast();
275 bool inline_aescrypt_Block(vmIntrinsics::ID id);
276 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
277 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
278 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
279 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
280 bool inline_ghash_processBlocks();
281 bool inline_sha_implCompress(vmIntrinsics::ID id);
282 bool inline_digestBase_implCompressMB(int predicate);
283 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
284 bool long_state, address stubAddr, const char *stubName,
285 Node* src_start, Node* ofs, Node* limit);
286 Node* get_state_from_sha_object(Node *sha_object);
287 Node* get_state_from_sha5_object(Node *sha_object);
288 Node* inline_digestBase_implCompressMB_predicate(int predicate);
289 bool inline_encodeISOArray();
290 bool inline_updateCRC32();
291 bool inline_updateBytesCRC32();
292 bool inline_updateByteBufferCRC32();
293 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
294 bool inline_updateBytesCRC32C();
295 bool inline_updateDirectByteBufferCRC32C();
296 bool inline_multiplyToLen();
297 bool inline_squareToLen();
298 bool inline_mulAdd();
299 bool inline_montgomeryMultiply();
300 bool inline_montgomerySquare();
301
302 bool inline_profileBoolean();
303 bool inline_isCompileConstant();
304 };
305
306
307 //---------------------------make_vm_intrinsic----------------------------
308 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
309 vmIntrinsics::ID id = m->intrinsic_id();
310 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
311
312 ccstr disable_intr = NULL;
313
314 if ((DisableIntrinsic[0] != '\0'
315 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
316 (method_has_option_value("DisableIntrinsic", disable_intr)
317 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
318 // disabled by a user request on the command line:
319 // example: -XX:DisableIntrinsic=_hashCode,_getClass
320 return NULL;
321 }
322
323 if (!m->is_loaded()) {
324 // do not attempt to inline unloaded methods
325 return NULL;
326 }
327
328 // Only a few intrinsics implement a virtual dispatch.
329 // They are expensive calls which are also frequently overridden.
330 if (is_virtual) {
331 switch (id) {
332 case vmIntrinsics::_hashCode:
333 case vmIntrinsics::_clone:
334 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
335 break;
336 default:
337 return NULL;
338 }
339 }
340
341 // -XX:-InlineNatives disables nearly all intrinsics:
342 if (!InlineNatives) {
343 switch (id) {
344 case vmIntrinsics::_indexOf:
345 case vmIntrinsics::_compareTo:
346 case vmIntrinsics::_equals:
347 case vmIntrinsics::_equalsC:
348 case vmIntrinsics::_getAndAddInt:
349 case vmIntrinsics::_getAndAddLong:
350 case vmIntrinsics::_getAndSetInt:
351 case vmIntrinsics::_getAndSetLong:
352 case vmIntrinsics::_getAndSetObject:
353 case vmIntrinsics::_loadFence:
354 case vmIntrinsics::_storeFence:
355 case vmIntrinsics::_fullFence:
356 break; // InlineNatives does not control String.compareTo
357 case vmIntrinsics::_Reference_get:
358 break; // InlineNatives does not control Reference.get
359 default:
360 return NULL;
361 }
362 }
363
364 int predicates = 0;
365 bool does_virtual_dispatch = false;
366
367 switch (id) {
368 case vmIntrinsics::_compareTo:
369 if (!SpecialStringCompareTo) return NULL;
370 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
371 break;
372 case vmIntrinsics::_indexOf:
373 if (!SpecialStringIndexOf) return NULL;
374 break;
375 case vmIntrinsics::_equals:
376 if (!SpecialStringEquals) return NULL;
377 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
378 break;
379 case vmIntrinsics::_equalsC:
380 if (!SpecialArraysEquals) return NULL;
381 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
382 break;
383 case vmIntrinsics::_arraycopy:
384 if (!InlineArrayCopy) return NULL;
385 break;
386 case vmIntrinsics::_copyMemory:
387 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
388 if (!InlineArrayCopy) return NULL;
389 break;
390 case vmIntrinsics::_hashCode:
391 if (!InlineObjectHash) return NULL;
392 does_virtual_dispatch = true;
393 break;
394 case vmIntrinsics::_clone:
395 does_virtual_dispatch = true;
396 case vmIntrinsics::_copyOf:
397 case vmIntrinsics::_copyOfRange:
398 if (!InlineObjectCopy) return NULL;
399 // These also use the arraycopy intrinsic mechanism:
400 if (!InlineArrayCopy) return NULL;
401 break;
402 case vmIntrinsics::_encodeISOArray:
403 if (!SpecialEncodeISOArray) return NULL;
404 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
405 break;
406 case vmIntrinsics::_checkIndex:
407 // We do not intrinsify this. The optimizer does fine with it.
408 return NULL;
409
410 case vmIntrinsics::_getCallerClass:
411 if (!InlineReflectionGetCallerClass) return NULL;
412 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
413 break;
414
415 case vmIntrinsics::_bitCount_i:
416 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
417 break;
418
419 case vmIntrinsics::_bitCount_l:
420 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
421 break;
422
423 case vmIntrinsics::_numberOfLeadingZeros_i:
424 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
425 break;
426
427 case vmIntrinsics::_numberOfLeadingZeros_l:
428 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
429 break;
430
431 case vmIntrinsics::_numberOfTrailingZeros_i:
432 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
433 break;
434
435 case vmIntrinsics::_numberOfTrailingZeros_l:
436 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
437 break;
438
439 case vmIntrinsics::_reverseBytes_c:
440 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
441 break;
442 case vmIntrinsics::_reverseBytes_s:
443 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
444 break;
445 case vmIntrinsics::_reverseBytes_i:
446 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
447 break;
448 case vmIntrinsics::_reverseBytes_l:
449 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
450 break;
451
452 case vmIntrinsics::_Reference_get:
453 // Use the intrinsic version of Reference.get() so that the value in
454 // the referent field can be registered by the G1 pre-barrier code.
455 // Also add memory barrier to prevent commoning reads from this field
456 // across safepoint since GC can change it value.
457 break;
458
459 case vmIntrinsics::_compareAndSwapObject:
460 #ifdef _LP64
461 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
462 #endif
463 break;
464
465 case vmIntrinsics::_compareAndSwapLong:
466 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
467 break;
468
469 case vmIntrinsics::_getAndAddInt:
470 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
471 break;
472
473 case vmIntrinsics::_getAndAddLong:
474 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
475 break;
476
477 case vmIntrinsics::_getAndSetInt:
478 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
479 break;
480
481 case vmIntrinsics::_getAndSetLong:
482 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
483 break;
484
485 case vmIntrinsics::_getAndSetObject:
486 #ifdef _LP64
487 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
488 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
489 break;
490 #else
491 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
492 break;
493 #endif
494
495 case vmIntrinsics::_aescrypt_encryptBlock:
496 case vmIntrinsics::_aescrypt_decryptBlock:
497 if (!UseAESIntrinsics) return NULL;
498 break;
499
500 case vmIntrinsics::_multiplyToLen:
501 if (!UseMultiplyToLenIntrinsic) return NULL;
502 break;
503
504 case vmIntrinsics::_squareToLen:
505 if (!UseSquareToLenIntrinsic) return NULL;
506 break;
507
508 case vmIntrinsics::_mulAdd:
509 if (!UseMulAddIntrinsic) return NULL;
510 break;
511
512 case vmIntrinsics::_montgomeryMultiply:
513 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
514 break;
515 case vmIntrinsics::_montgomerySquare:
516 if (!UseMontgomerySquareIntrinsic) return NULL;
517 break;
518
519 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
520 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
521 if (!UseAESIntrinsics) return NULL;
522 // these two require the predicated logic
523 predicates = 1;
524 break;
525
526 case vmIntrinsics::_sha_implCompress:
527 if (!UseSHA1Intrinsics) return NULL;
528 break;
529
530 case vmIntrinsics::_sha2_implCompress:
531 if (!UseSHA256Intrinsics) return NULL;
532 break;
533
534 case vmIntrinsics::_sha5_implCompress:
535 if (!UseSHA512Intrinsics) return NULL;
536 break;
537
538 case vmIntrinsics::_digestBase_implCompressMB:
539 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
540 predicates = 3;
541 break;
542
543 case vmIntrinsics::_ghash_processBlocks:
544 if (!UseGHASHIntrinsics) return NULL;
545 break;
546
547 case vmIntrinsics::_updateCRC32:
548 case vmIntrinsics::_updateBytesCRC32:
549 case vmIntrinsics::_updateByteBufferCRC32:
550 if (!UseCRC32Intrinsics) return NULL;
551 break;
552
553 case vmIntrinsics::_updateBytesCRC32C:
554 case vmIntrinsics::_updateDirectByteBufferCRC32C:
555 if (!UseCRC32CIntrinsics) return NULL;
556 break;
557
558 case vmIntrinsics::_incrementExactI:
559 case vmIntrinsics::_addExactI:
560 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
561 break;
562 case vmIntrinsics::_incrementExactL:
563 case vmIntrinsics::_addExactL:
564 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
565 break;
566 case vmIntrinsics::_decrementExactI:
567 case vmIntrinsics::_subtractExactI:
568 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
569 break;
570 case vmIntrinsics::_decrementExactL:
571 case vmIntrinsics::_subtractExactL:
572 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
573 break;
574 case vmIntrinsics::_negateExactI:
575 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
576 break;
577 case vmIntrinsics::_negateExactL:
578 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
579 break;
580 case vmIntrinsics::_multiplyExactI:
581 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
582 break;
583 case vmIntrinsics::_multiplyExactL:
584 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
585 break;
586
587 case vmIntrinsics::_getShortUnaligned:
588 case vmIntrinsics::_getCharUnaligned:
589 case vmIntrinsics::_getIntUnaligned:
590 case vmIntrinsics::_getLongUnaligned:
591 case vmIntrinsics::_putShortUnaligned:
592 case vmIntrinsics::_putCharUnaligned:
593 case vmIntrinsics::_putIntUnaligned:
594 case vmIntrinsics::_putLongUnaligned:
595 if (!UseUnalignedAccesses) return NULL;
596 break;
597
598 default:
599 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
600 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
601 break;
602 }
603
604 // -XX:-InlineClassNatives disables natives from the Class class.
605 // The flag applies to all reflective calls, notably Array.newArray
606 // (visible to Java programmers as Array.newInstance).
607 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
608 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
609 if (!InlineClassNatives) return NULL;
610 }
611
612 // -XX:-InlineThreadNatives disables natives from the Thread class.
613 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
614 if (!InlineThreadNatives) return NULL;
615 }
616
617 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
618 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
619 m->holder()->name() == ciSymbol::java_lang_Float() ||
620 m->holder()->name() == ciSymbol::java_lang_Double()) {
621 if (!InlineMathNatives) return NULL;
622 }
623
624 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
625 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
626 if (!InlineUnsafeOps) return NULL;
627 }
628
629 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
630 }
631
632 //----------------------register_library_intrinsics-----------------------
633 // Initialize this file's data structures, for each Compile instance.
634 void Compile::register_library_intrinsics() {
635 // Nothing to do here.
636 }
637
638 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
639 LibraryCallKit kit(jvms, this);
640 Compile* C = kit.C;
641 int nodes = C->unique();
642 #ifndef PRODUCT
643 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
644 char buf[1000];
645 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
646 tty->print_cr("Intrinsic %s", str);
647 }
648 #endif
649 ciMethod* callee = kit.callee();
650 const int bci = kit.bci();
651
652 // Try to inline the intrinsic.
653 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
654 kit.try_to_inline(_last_predicate)) {
655 if (C->print_intrinsics() || C->print_inlining()) {
656 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
657 }
658 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
659 if (C->log()) {
660 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
661 vmIntrinsics::name_at(intrinsic_id()),
662 (is_virtual() ? " virtual='1'" : ""),
663 C->unique() - nodes);
664 }
665 // Push the result from the inlined method onto the stack.
666 kit.push_result();
667 C->print_inlining_update(this);
668 return kit.transfer_exceptions_into_jvms();
669 }
670
671 // The intrinsic bailed out
672 if (C->print_intrinsics() || C->print_inlining()) {
673 if (jvms->has_method()) {
674 // Not a root compile.
675 const char* msg;
676 if (callee->intrinsic_candidate()) {
677 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
678 } else {
679 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
680 : "failed to inline (intrinsic), method not annotated";
681 }
682 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
683 } else {
684 // Root compile
685 tty->print("Did not generate intrinsic %s%s at bci:%d in",
686 vmIntrinsics::name_at(intrinsic_id()),
687 (is_virtual() ? " (virtual)" : ""), bci);
688 }
689 }
690 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
691 C->print_inlining_update(this);
692 return NULL;
693 }
694
695 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
696 LibraryCallKit kit(jvms, this);
697 Compile* C = kit.C;
698 int nodes = C->unique();
699 _last_predicate = predicate;
700 #ifndef PRODUCT
701 assert(is_predicated() && predicate < predicates_count(), "sanity");
702 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
703 char buf[1000];
704 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
705 tty->print_cr("Predicate for intrinsic %s", str);
706 }
707 #endif
708 ciMethod* callee = kit.callee();
709 const int bci = kit.bci();
710
711 Node* slow_ctl = kit.try_to_predicate(predicate);
712 if (!kit.failing()) {
713 if (C->print_intrinsics() || C->print_inlining()) {
714 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
715 }
716 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
717 if (C->log()) {
718 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
719 vmIntrinsics::name_at(intrinsic_id()),
720 (is_virtual() ? " virtual='1'" : ""),
721 C->unique() - nodes);
722 }
723 return slow_ctl; // Could be NULL if the check folds.
724 }
725
726 // The intrinsic bailed out
727 if (C->print_intrinsics() || C->print_inlining()) {
728 if (jvms->has_method()) {
729 // Not a root compile.
730 const char* msg = "failed to generate predicate for intrinsic";
731 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
732 } else {
733 // Root compile
734 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
735 vmIntrinsics::name_at(intrinsic_id()),
736 (is_virtual() ? " (virtual)" : ""), bci);
737 }
738 }
739 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
740 return NULL;
741 }
742
743 bool LibraryCallKit::try_to_inline(int predicate) {
744 // Handle symbolic names for otherwise undistinguished boolean switches:
745 const bool is_store = true;
746 const bool is_native_ptr = true;
747 const bool is_static = true;
748 const bool is_volatile = true;
749
750 if (!jvms()->has_method()) {
751 // Root JVMState has a null method.
752 assert(map()->memory()->Opcode() == Op_Parm, "");
753 // Insert the memory aliasing node
754 set_all_memory(reset_memory());
755 }
756 assert(merged_memory(), "");
757
758
759 switch (intrinsic_id()) {
760 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
761 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
762 case vmIntrinsics::_getClass: return inline_native_getClass();
763
764 case vmIntrinsics::_dsin:
765 case vmIntrinsics::_dcos:
766 case vmIntrinsics::_dtan:
767 case vmIntrinsics::_dabs:
768 case vmIntrinsics::_datan2:
769 case vmIntrinsics::_dsqrt:
770 case vmIntrinsics::_dexp:
771 case vmIntrinsics::_dlog:
772 case vmIntrinsics::_dlog10:
773 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
774
775 case vmIntrinsics::_min:
776 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
777
778 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
779 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
780 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
781 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
782 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
783 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
784 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
785 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
786 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
787 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
788 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
789 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
790
791 case vmIntrinsics::_arraycopy: return inline_arraycopy();
792
793 case vmIntrinsics::_compareTo: return inline_string_compareTo();
794 case vmIntrinsics::_indexOf: return inline_string_indexOf();
795 case vmIntrinsics::_equals: return inline_string_equals();
796
797 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
798 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
799 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
800 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
801 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
802 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
803 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
804 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
805 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
806
807 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
808 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
809 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
810 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
811 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
812 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
813 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
814 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
815 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
816
817 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
818 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
819 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
820 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
821 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
822 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
823 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
824 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
825
826 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
827 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
828 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
829 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
830 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
831 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
832 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
833 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
834
835 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
836 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
837 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
838 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
839 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
840 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
841 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
842 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
843 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
844
845 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
846 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
847 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
848 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
849 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
850 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
851 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
852 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
853 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
854
855 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
856 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
857 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
858 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
859
860 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
861 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
862 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
863 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
864
865 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
866 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
867 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
868
869 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
870 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
871 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
872
873 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
874 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
875 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
876 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
877 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
878
879 case vmIntrinsics::_loadFence:
880 case vmIntrinsics::_storeFence:
881 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
882
883 case vmIntrinsics::_currentThread: return inline_native_currentThread();
884 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
885
886 #ifdef TRACE_HAVE_INTRINSICS
887 case vmIntrinsics::_classID: return inline_native_classID();
888 case vmIntrinsics::_threadID: return inline_native_threadID();
889 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
890 #endif
891 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
892 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
893 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
894 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
895 case vmIntrinsics::_newArray: return inline_native_newArray();
896 case vmIntrinsics::_getLength: return inline_native_getLength();
897 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
898 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
899 case vmIntrinsics::_equalsC: return inline_array_equals();
900 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
901
902 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
903
904 case vmIntrinsics::_isInstance:
905 case vmIntrinsics::_getModifiers:
906 case vmIntrinsics::_isInterface:
907 case vmIntrinsics::_isArray:
908 case vmIntrinsics::_isPrimitive:
909 case vmIntrinsics::_getSuperclass:
910 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
911
912 case vmIntrinsics::_floatToRawIntBits:
913 case vmIntrinsics::_floatToIntBits:
914 case vmIntrinsics::_intBitsToFloat:
915 case vmIntrinsics::_doubleToRawLongBits:
916 case vmIntrinsics::_doubleToLongBits:
917 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
918
919 case vmIntrinsics::_numberOfLeadingZeros_i:
920 case vmIntrinsics::_numberOfLeadingZeros_l:
921 case vmIntrinsics::_numberOfTrailingZeros_i:
922 case vmIntrinsics::_numberOfTrailingZeros_l:
923 case vmIntrinsics::_bitCount_i:
924 case vmIntrinsics::_bitCount_l:
925 case vmIntrinsics::_reverseBytes_i:
926 case vmIntrinsics::_reverseBytes_l:
927 case vmIntrinsics::_reverseBytes_s:
928 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
929
930 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
931
932 case vmIntrinsics::_Reference_get: return inline_reference_get();
933
934 case vmIntrinsics::_Class_cast: return inline_Class_cast();
935
936 case vmIntrinsics::_aescrypt_encryptBlock:
937 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
938
939 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
940 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
941 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
942
943 case vmIntrinsics::_sha_implCompress:
944 case vmIntrinsics::_sha2_implCompress:
945 case vmIntrinsics::_sha5_implCompress:
946 return inline_sha_implCompress(intrinsic_id());
947
948 case vmIntrinsics::_digestBase_implCompressMB:
949 return inline_digestBase_implCompressMB(predicate);
950
951 case vmIntrinsics::_multiplyToLen:
952 return inline_multiplyToLen();
953
954 case vmIntrinsics::_squareToLen:
955 return inline_squareToLen();
956
957 case vmIntrinsics::_mulAdd:
958 return inline_mulAdd();
959
960 case vmIntrinsics::_montgomeryMultiply:
961 return inline_montgomeryMultiply();
962 case vmIntrinsics::_montgomerySquare:
963 return inline_montgomerySquare();
964
965 case vmIntrinsics::_ghash_processBlocks:
966 return inline_ghash_processBlocks();
967
968 case vmIntrinsics::_encodeISOArray:
969 return inline_encodeISOArray();
970
971 case vmIntrinsics::_updateCRC32:
972 return inline_updateCRC32();
973 case vmIntrinsics::_updateBytesCRC32:
974 return inline_updateBytesCRC32();
975 case vmIntrinsics::_updateByteBufferCRC32:
976 return inline_updateByteBufferCRC32();
977
978 case vmIntrinsics::_updateBytesCRC32C:
979 return inline_updateBytesCRC32C();
980 case vmIntrinsics::_updateDirectByteBufferCRC32C:
981 return inline_updateDirectByteBufferCRC32C();
982
983 case vmIntrinsics::_profileBoolean:
984 return inline_profileBoolean();
985 case vmIntrinsics::_isCompileConstant:
986 return inline_isCompileConstant();
987
988 default:
989 // If you get here, it may be that someone has added a new intrinsic
990 // to the list in vmSymbols.hpp without implementing it here.
991 #ifndef PRODUCT
992 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
993 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
994 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
995 }
996 #endif
997 return false;
998 }
999 }
1000
1001 Node* LibraryCallKit::try_to_predicate(int predicate) {
1002 if (!jvms()->has_method()) {
1003 // Root JVMState has a null method.
1004 assert(map()->memory()->Opcode() == Op_Parm, "");
1005 // Insert the memory aliasing node
1006 set_all_memory(reset_memory());
1007 }
1008 assert(merged_memory(), "");
1009
1010 switch (intrinsic_id()) {
1011 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1012 return inline_cipherBlockChaining_AESCrypt_predicate(false);
1013 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1014 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1015 case vmIntrinsics::_digestBase_implCompressMB:
1016 return inline_digestBase_implCompressMB_predicate(predicate);
1017
1018 default:
1019 // If you get here, it may be that someone has added a new intrinsic
1020 // to the list in vmSymbols.hpp without implementing it here.
1021 #ifndef PRODUCT
1022 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1023 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1024 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1025 }
1026 #endif
1027 Node* slow_ctl = control();
1028 set_control(top()); // No fast path instrinsic
1029 return slow_ctl;
1030 }
1031 }
1032
1033 //------------------------------set_result-------------------------------
1034 // Helper function for finishing intrinsics.
1035 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1036 record_for_igvn(region);
1037 set_control(_gvn.transform(region));
1038 set_result( _gvn.transform(value));
1039 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1040 }
1041
1042 //------------------------------generate_guard---------------------------
1043 // Helper function for generating guarded fast-slow graph structures.
1044 // The given 'test', if true, guards a slow path. If the test fails
1045 // then a fast path can be taken. (We generally hope it fails.)
1046 // In all cases, GraphKit::control() is updated to the fast path.
1047 // The returned value represents the control for the slow path.
1048 // The return value is never 'top'; it is either a valid control
1049 // or NULL if it is obvious that the slow path can never be taken.
1050 // Also, if region and the slow control are not NULL, the slow edge
1051 // is appended to the region.
1052 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1053 if (stopped()) {
1054 // Already short circuited.
1055 return NULL;
1056 }
1057
1058 // Build an if node and its projections.
1059 // If test is true we take the slow path, which we assume is uncommon.
1060 if (_gvn.type(test) == TypeInt::ZERO) {
1061 // The slow branch is never taken. No need to build this guard.
1062 return NULL;
1063 }
1064
1065 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1066
1067 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1068 if (if_slow == top()) {
1069 // The slow branch is never taken. No need to build this guard.
1070 return NULL;
1071 }
1072
1073 if (region != NULL)
1074 region->add_req(if_slow);
1075
1076 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1077 set_control(if_fast);
1078
1079 return if_slow;
1080 }
1081
1082 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1083 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1084 }
1085 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1086 return generate_guard(test, region, PROB_FAIR);
1087 }
1088
1089 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1090 Node* *pos_index) {
1091 if (stopped())
1092 return NULL; // already stopped
1093 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1094 return NULL; // index is already adequately typed
1095 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1096 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1097 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1098 if (is_neg != NULL && pos_index != NULL) {
1099 // Emulate effect of Parse::adjust_map_after_if.
1100 Node* ccast = new CastIINode(index, TypeInt::POS);
1101 ccast->set_req(0, control());
1102 (*pos_index) = _gvn.transform(ccast);
1103 }
1104 return is_neg;
1105 }
1106
1107 // Make sure that 'position' is a valid limit index, in [0..length].
1108 // There are two equivalent plans for checking this:
1109 // A. (offset + copyLength) unsigned<= arrayLength
1110 // B. offset <= (arrayLength - copyLength)
1111 // We require that all of the values above, except for the sum and
1112 // difference, are already known to be non-negative.
1113 // Plan A is robust in the face of overflow, if offset and copyLength
1114 // are both hugely positive.
1115 //
1116 // Plan B is less direct and intuitive, but it does not overflow at
1117 // all, since the difference of two non-negatives is always
1118 // representable. Whenever Java methods must perform the equivalent
1119 // check they generally use Plan B instead of Plan A.
1120 // For the moment we use Plan A.
1121 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1122 Node* subseq_length,
1123 Node* array_length,
1124 RegionNode* region) {
1125 if (stopped())
1126 return NULL; // already stopped
1127 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1128 if (zero_offset && subseq_length->eqv_uncast(array_length))
1129 return NULL; // common case of whole-array copy
1130 Node* last = subseq_length;
1131 if (!zero_offset) // last += offset
1132 last = _gvn.transform(new AddINode(last, offset));
1133 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1134 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1135 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1136 return is_over;
1137 }
1138
1139
1140 //--------------------------generate_current_thread--------------------
1141 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1142 ciKlass* thread_klass = env()->Thread_klass();
1143 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1144 Node* thread = _gvn.transform(new ThreadLocalNode());
1145 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1146 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1147 tls_output = thread;
1148 return threadObj;
1149 }
1150
1151
1152 //------------------------------make_string_method_node------------------------
1153 // Helper method for String intrinsic functions. This version is called
1154 // with str1 and str2 pointing to String object nodes.
1155 //
1156 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1157 Node* no_ctrl = NULL;
1158
1159 // Get start addr of string
1160 Node* str1_value = load_String_value(no_ctrl, str1);
1161 Node* str1_offset = load_String_offset(no_ctrl, str1);
1162 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1163
1164 // Get length of string 1
1165 Node* str1_len = load_String_length(no_ctrl, str1);
1166
1167 Node* str2_value = load_String_value(no_ctrl, str2);
1168 Node* str2_offset = load_String_offset(no_ctrl, str2);
1169 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1170
1171 Node* str2_len = NULL;
1172 Node* result = NULL;
1173
1174 switch (opcode) {
1175 case Op_StrIndexOf:
1176 // Get length of string 2
1177 str2_len = load_String_length(no_ctrl, str2);
1178
1179 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1180 str1_start, str1_len, str2_start, str2_len);
1181 break;
1182 case Op_StrComp:
1183 // Get length of string 2
1184 str2_len = load_String_length(no_ctrl, str2);
1185
1186 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1187 str1_start, str1_len, str2_start, str2_len);
1188 break;
1189 case Op_StrEquals:
1190 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1191 str1_start, str2_start, str1_len);
1192 break;
1193 default:
1194 ShouldNotReachHere();
1195 return NULL;
1196 }
1197
1198 // All these intrinsics have checks.
1199 C->set_has_split_ifs(true); // Has chance for split-if optimization
1200
1201 return _gvn.transform(result);
1202 }
1203
1204 // Helper method for String intrinsic functions. This version is called
1205 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1206 // to Int nodes containing the lenghts of str1 and str2.
1207 //
1208 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1209 Node* result = NULL;
1210 switch (opcode) {
1211 case Op_StrIndexOf:
1212 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1213 str1_start, cnt1, str2_start, cnt2);
1214 break;
1215 case Op_StrComp:
1216 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1217 str1_start, cnt1, str2_start, cnt2);
1218 break;
1219 case Op_StrEquals:
1220 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1221 str1_start, str2_start, cnt1);
1222 break;
1223 default:
1224 ShouldNotReachHere();
1225 return NULL;
1226 }
1227
1228 // All these intrinsics have checks.
1229 C->set_has_split_ifs(true); // Has chance for split-if optimization
1230
1231 return _gvn.transform(result);
1232 }
1233
1234 //------------------------------inline_string_compareTo------------------------
1235 // public int java.lang.String.compareTo(String anotherString);
1236 bool LibraryCallKit::inline_string_compareTo() {
1237 Node* receiver = null_check(argument(0));
1238 Node* arg = null_check(argument(1));
1239 if (stopped()) {
1240 return true;
1241 }
1242 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1243 return true;
1244 }
1245
1246 //------------------------------inline_string_equals------------------------
1247 bool LibraryCallKit::inline_string_equals() {
1248 Node* receiver = null_check_receiver();
1249 // NOTE: Do not null check argument for String.equals() because spec
1250 // allows to specify NULL as argument.
1251 Node* argument = this->argument(1);
1252 if (stopped()) {
1253 return true;
1254 }
1255
1256 // paths (plus control) merge
1257 RegionNode* region = new RegionNode(5);
1258 Node* phi = new PhiNode(region, TypeInt::BOOL);
1259
1260 // does source == target string?
1261 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1262 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1263
1264 Node* if_eq = generate_slow_guard(bol, NULL);
1265 if (if_eq != NULL) {
1266 // receiver == argument
1267 phi->init_req(2, intcon(1));
1268 region->init_req(2, if_eq);
1269 }
1270
1271 // get String klass for instanceOf
1272 ciInstanceKlass* klass = env()->String_klass();
1273
1274 if (!stopped()) {
1275 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1276 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1277 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1278
1279 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1280 //instanceOf == true, fallthrough
1281
1282 if (inst_false != NULL) {
1283 phi->init_req(3, intcon(0));
1284 region->init_req(3, inst_false);
1285 }
1286 }
1287
1288 if (!stopped()) {
1289 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1290
1291 // Properly cast the argument to String
1292 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1293 // This path is taken only when argument's type is String:NotNull.
1294 argument = cast_not_null(argument, false);
1295
1296 Node* no_ctrl = NULL;
1297
1298 // Get start addr of receiver
1299 Node* receiver_val = load_String_value(no_ctrl, receiver);
1300 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1301 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1302
1303 // Get length of receiver
1304 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1305
1306 // Get start addr of argument
1307 Node* argument_val = load_String_value(no_ctrl, argument);
1308 Node* argument_offset = load_String_offset(no_ctrl, argument);
1309 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1310
1311 // Get length of argument
1312 Node* argument_cnt = load_String_length(no_ctrl, argument);
1313
1314 // Check for receiver count != argument count
1315 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1316 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1317 Node* if_ne = generate_slow_guard(bol, NULL);
1318 if (if_ne != NULL) {
1319 phi->init_req(4, intcon(0));
1320 region->init_req(4, if_ne);
1321 }
1322
1323 // Check for count == 0 is done by assembler code for StrEquals.
1324
1325 if (!stopped()) {
1326 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1327 phi->init_req(1, equals);
1328 region->init_req(1, control());
1329 }
1330 }
1331
1332 // post merge
1333 set_control(_gvn.transform(region));
1334 record_for_igvn(region);
1335
1336 set_result(_gvn.transform(phi));
1337 return true;
1338 }
1339
1340 //------------------------------inline_array_equals----------------------------
1341 bool LibraryCallKit::inline_array_equals() {
1342 Node* arg1 = argument(0);
1343 Node* arg2 = argument(1);
1344 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1345 return true;
1346 }
1347
1348 // Java version of String.indexOf(constant string)
1349 // class StringDecl {
1350 // StringDecl(char[] ca) {
1351 // offset = 0;
1352 // count = ca.length;
1353 // value = ca;
1354 // }
1355 // int offset;
1356 // int count;
1357 // char[] value;
1358 // }
1359 //
1360 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1361 // int targetOffset, int cache_i, int md2) {
1362 // int cache = cache_i;
1363 // int sourceOffset = string_object.offset;
1364 // int sourceCount = string_object.count;
1365 // int targetCount = target_object.length;
1366 //
1367 // int targetCountLess1 = targetCount - 1;
1368 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1369 //
1370 // char[] source = string_object.value;
1371 // char[] target = target_object;
1372 // int lastChar = target[targetCountLess1];
1373 //
1374 // outer_loop:
1375 // for (int i = sourceOffset; i < sourceEnd; ) {
1376 // int src = source[i + targetCountLess1];
1377 // if (src == lastChar) {
1378 // // With random strings and a 4-character alphabet,
1379 // // reverse matching at this point sets up 0.8% fewer
1380 // // frames, but (paradoxically) makes 0.3% more probes.
1381 // // Since those probes are nearer the lastChar probe,
1382 // // there is may be a net D$ win with reverse matching.
1383 // // But, reversing loop inhibits unroll of inner loop
1384 // // for unknown reason. So, does running outer loop from
1385 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1386 // for (int j = 0; j < targetCountLess1; j++) {
1387 // if (target[targetOffset + j] != source[i+j]) {
1388 // if ((cache & (1 << source[i+j])) == 0) {
1389 // if (md2 < j+1) {
1390 // i += j+1;
1391 // continue outer_loop;
1392 // }
1393 // }
1394 // i += md2;
1395 // continue outer_loop;
1396 // }
1397 // }
1398 // return i - sourceOffset;
1399 // }
1400 // if ((cache & (1 << src)) == 0) {
1401 // i += targetCountLess1;
1402 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1403 // i++;
1404 // }
1405 // return -1;
1406 // }
1407
1408 //------------------------------string_indexOf------------------------
1409 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1410 jint cache_i, jint md2_i) {
1411
1412 Node* no_ctrl = NULL;
1413 float likely = PROB_LIKELY(0.9);
1414 float unlikely = PROB_UNLIKELY(0.9);
1415
1416 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1417
1418 Node* source = load_String_value(no_ctrl, string_object);
1419 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1420 Node* sourceCount = load_String_length(no_ctrl, string_object);
1421
1422 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1423 jint target_length = target_array->length();
1424 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1425 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1426
1427 // String.value field is known to be @Stable.
1428 if (UseImplicitStableValues) {
1429 target = cast_array_to_stable(target, target_type);
1430 }
1431
1432 IdealKit kit(this, false, true);
1433 #define __ kit.
1434 Node* zero = __ ConI(0);
1435 Node* one = __ ConI(1);
1436 Node* cache = __ ConI(cache_i);
1437 Node* md2 = __ ConI(md2_i);
1438 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1439 Node* targetCountLess1 = __ ConI(target_length - 1);
1440 Node* targetOffset = __ ConI(targetOffset_i);
1441 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1442
1443 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1444 Node* outer_loop = __ make_label(2 /* goto */);
1445 Node* return_ = __ make_label(1);
1446
1447 __ set(rtn,__ ConI(-1));
1448 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1449 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1450 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1451 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1452 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1453 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1454 Node* tpj = __ AddI(targetOffset, __ value(j));
1455 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1456 Node* ipj = __ AddI(__ value(i), __ value(j));
1457 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1458 __ if_then(targ, BoolTest::ne, src2); {
1459 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1460 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1461 __ increment(i, __ AddI(__ value(j), one));
1462 __ goto_(outer_loop);
1463 } __ end_if(); __ dead(j);
1464 }__ end_if(); __ dead(j);
1465 __ increment(i, md2);
1466 __ goto_(outer_loop);
1467 }__ end_if();
1468 __ increment(j, one);
1469 }__ end_loop(); __ dead(j);
1470 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1471 __ goto_(return_);
1472 }__ end_if();
1473 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1474 __ increment(i, targetCountLess1);
1475 }__ end_if();
1476 __ increment(i, one);
1477 __ bind(outer_loop);
1478 }__ end_loop(); __ dead(i);
1479 __ bind(return_);
1480
1481 // Final sync IdealKit and GraphKit.
1482 final_sync(kit);
1483 Node* result = __ value(rtn);
1484 #undef __
1485 C->set_has_loops(true);
1486 return result;
1487 }
1488
1489 //------------------------------inline_string_indexOf------------------------
1490 bool LibraryCallKit::inline_string_indexOf() {
1491 Node* receiver = argument(0);
1492 Node* arg = argument(1);
1493
1494 Node* result;
1495 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1496 UseSSE42Intrinsics) {
1497 // Generate SSE4.2 version of indexOf
1498 // We currently only have match rules that use SSE4.2
1499
1500 receiver = null_check(receiver);
1501 arg = null_check(arg);
1502 if (stopped()) {
1503 return true;
1504 }
1505
1506 // Make the merge point
1507 RegionNode* result_rgn = new RegionNode(4);
1508 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1509 Node* no_ctrl = NULL;
1510
1511 // Get start addr of source string
1512 Node* source = load_String_value(no_ctrl, receiver);
1513 Node* source_offset = load_String_offset(no_ctrl, receiver);
1514 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1515
1516 // Get length of source string
1517 Node* source_cnt = load_String_length(no_ctrl, receiver);
1518
1519 // Get start addr of substring
1520 Node* substr = load_String_value(no_ctrl, arg);
1521 Node* substr_offset = load_String_offset(no_ctrl, arg);
1522 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1523
1524 // Get length of source string
1525 Node* substr_cnt = load_String_length(no_ctrl, arg);
1526
1527 // Check for substr count > string count
1528 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1529 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1530 Node* if_gt = generate_slow_guard(bol, NULL);
1531 if (if_gt != NULL) {
1532 result_phi->init_req(2, intcon(-1));
1533 result_rgn->init_req(2, if_gt);
1534 }
1535
1536 if (!stopped()) {
1537 // Check for substr count == 0
1538 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1539 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1540 Node* if_zero = generate_slow_guard(bol, NULL);
1541 if (if_zero != NULL) {
1542 result_phi->init_req(3, intcon(0));
1543 result_rgn->init_req(3, if_zero);
1544 }
1545 }
1546
1547 if (!stopped()) {
1548 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1549 result_phi->init_req(1, result);
1550 result_rgn->init_req(1, control());
1551 }
1552 set_control(_gvn.transform(result_rgn));
1553 record_for_igvn(result_rgn);
1554 result = _gvn.transform(result_phi);
1555
1556 } else { // Use LibraryCallKit::string_indexOf
1557 // don't intrinsify if argument isn't a constant string.
1558 if (!arg->is_Con()) {
1559 return false;
1560 }
1561 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1562 if (str_type == NULL) {
1563 return false;
1564 }
1565 ciInstanceKlass* klass = env()->String_klass();
1566 ciObject* str_const = str_type->const_oop();
1567 if (str_const == NULL || str_const->klass() != klass) {
1568 return false;
1569 }
1570 ciInstance* str = str_const->as_instance();
1571 assert(str != NULL, "must be instance");
1572
1573 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1574 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1575
1576 int o;
1577 int c;
1578 if (java_lang_String::has_offset_field()) {
1579 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1580 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1581 } else {
1582 o = 0;
1583 c = pat->length();
1584 }
1585
1586 // constant strings have no offset and count == length which
1587 // simplifies the resulting code somewhat so lets optimize for that.
1588 if (o != 0 || c != pat->length()) {
1589 return false;
1590 }
1591
1592 receiver = null_check(receiver, T_OBJECT);
1593 // NOTE: No null check on the argument is needed since it's a constant String oop.
1594 if (stopped()) {
1595 return true;
1596 }
1597
1598 // The null string as a pattern always returns 0 (match at beginning of string)
1599 if (c == 0) {
1600 set_result(intcon(0));
1601 return true;
1602 }
1603
1604 // Generate default indexOf
1605 jchar lastChar = pat->char_at(o + (c - 1));
1606 int cache = 0;
1607 int i;
1608 for (i = 0; i < c - 1; i++) {
1609 assert(i < pat->length(), "out of range");
1610 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1611 }
1612
1613 int md2 = c;
1614 for (i = 0; i < c - 1; i++) {
1615 assert(i < pat->length(), "out of range");
1616 if (pat->char_at(o + i) == lastChar) {
1617 md2 = (c - 1) - i;
1618 }
1619 }
1620
1621 result = string_indexOf(receiver, pat, o, cache, md2);
1622 }
1623 set_result(result);
1624 return true;
1625 }
1626
1627 //--------------------------round_double_node--------------------------------
1628 // Round a double node if necessary.
1629 Node* LibraryCallKit::round_double_node(Node* n) {
1630 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1631 n = _gvn.transform(new RoundDoubleNode(0, n));
1632 return n;
1633 }
1634
1635 //------------------------------inline_math-----------------------------------
1636 // public static double Math.abs(double)
1637 // public static double Math.sqrt(double)
1638 // public static double Math.log(double)
1639 // public static double Math.log10(double)
1640 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1641 Node* arg = round_double_node(argument(0));
1642 Node* n;
1643 switch (id) {
1644 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1645 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1646 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1647 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1648 default: fatal_unexpected_iid(id); break;
1649 }
1650 set_result(_gvn.transform(n));
1651 return true;
1652 }
1653
1654 //------------------------------inline_trig----------------------------------
1655 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1656 // argument reduction which will turn into a fast/slow diamond.
1657 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1658 Node* arg = round_double_node(argument(0));
1659 Node* n = NULL;
1660
1661 switch (id) {
1662 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1663 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1664 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1665 default: fatal_unexpected_iid(id); break;
1666 }
1667 n = _gvn.transform(n);
1668
1669 // Rounding required? Check for argument reduction!
1670 if (Matcher::strict_fp_requires_explicit_rounding) {
1671 static const double pi_4 = 0.7853981633974483;
1672 static const double neg_pi_4 = -0.7853981633974483;
1673 // pi/2 in 80-bit extended precision
1674 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1675 // -pi/2 in 80-bit extended precision
1676 // 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};
1677 // Cutoff value for using this argument reduction technique
1678 //static const double pi_2_minus_epsilon = 1.564660403643354;
1679 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1680
1681 // Pseudocode for sin:
1682 // if (x <= Math.PI / 4.0) {
1683 // if (x >= -Math.PI / 4.0) return fsin(x);
1684 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1685 // } else {
1686 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1687 // }
1688 // return StrictMath.sin(x);
1689
1690 // Pseudocode for cos:
1691 // if (x <= Math.PI / 4.0) {
1692 // if (x >= -Math.PI / 4.0) return fcos(x);
1693 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1694 // } else {
1695 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1696 // }
1697 // return StrictMath.cos(x);
1698
1699 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1700 // requires a special machine instruction to load it. Instead we'll try
1701 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1702 // probably do the math inside the SIN encoding.
1703
1704 // Make the merge point
1705 RegionNode* r = new RegionNode(3);
1706 Node* phi = new PhiNode(r, Type::DOUBLE);
1707
1708 // Flatten arg so we need only 1 test
1709 Node *abs = _gvn.transform(new AbsDNode(arg));
1710 // Node for PI/4 constant
1711 Node *pi4 = makecon(TypeD::make(pi_4));
1712 // Check PI/4 : abs(arg)
1713 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1714 // Check: If PI/4 < abs(arg) then go slow
1715 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1716 // Branch either way
1717 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1718 set_control(opt_iff(r,iff));
1719
1720 // Set fast path result
1721 phi->init_req(2, n);
1722
1723 // Slow path - non-blocking leaf call
1724 Node* call = NULL;
1725 switch (id) {
1726 case vmIntrinsics::_dsin:
1727 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1728 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1729 "Sin", NULL, arg, top());
1730 break;
1731 case vmIntrinsics::_dcos:
1732 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1733 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1734 "Cos", NULL, arg, top());
1735 break;
1736 case vmIntrinsics::_dtan:
1737 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1738 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1739 "Tan", NULL, arg, top());
1740 break;
1741 }
1742 assert(control()->in(0) == call, "");
1743 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1744 r->init_req(1, control());
1745 phi->init_req(1, slow_result);
1746
1747 // Post-merge
1748 set_control(_gvn.transform(r));
1749 record_for_igvn(r);
1750 n = _gvn.transform(phi);
1751
1752 C->set_has_split_ifs(true); // Has chance for split-if optimization
1753 }
1754 set_result(n);
1755 return true;
1756 }
1757
1758 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1759 //-------------------
1760 //result=(result.isNaN())? funcAddr():result;
1761 // Check: If isNaN() by checking result!=result? then either trap
1762 // or go to runtime
1763 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1764 // Build the boolean node
1765 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1766
1767 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1768 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1769 // The pow or exp intrinsic returned a NaN, which requires a call
1770 // to the runtime. Recompile with the runtime call.
1771 uncommon_trap(Deoptimization::Reason_intrinsic,
1772 Deoptimization::Action_make_not_entrant);
1773 }
1774 return result;
1775 } else {
1776 // If this inlining ever returned NaN in the past, we compile a call
1777 // to the runtime to properly handle corner cases
1778
1779 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1780 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1781 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1782
1783 if (!if_slow->is_top()) {
1784 RegionNode* result_region = new RegionNode(3);
1785 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1786
1787 result_region->init_req(1, if_fast);
1788 result_val->init_req(1, result);
1789
1790 set_control(if_slow);
1791
1792 const TypePtr* no_memory_effects = NULL;
1793 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1794 no_memory_effects,
1795 x, top(), y, y ? top() : NULL);
1796 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1797 #ifdef ASSERT
1798 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1799 assert(value_top == top(), "second value must be top");
1800 #endif
1801
1802 result_region->init_req(2, control());
1803 result_val->init_req(2, value);
1804 set_control(_gvn.transform(result_region));
1805 return _gvn.transform(result_val);
1806 } else {
1807 return result;
1808 }
1809 }
1810 }
1811
1812 //------------------------------inline_pow-------------------------------------
1813 // Inline power instructions, if possible.
1814 bool LibraryCallKit::inline_pow() {
1815 // Pseudocode for pow
1816 // if (y == 2) {
1817 // return x * x;
1818 // } else {
1819 // if (x <= 0.0) {
1820 // long longy = (long)y;
1821 // if ((double)longy == y) { // if y is long
1822 // if (y + 1 == y) longy = 0; // huge number: even
1823 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1824 // } else {
1825 // result = NaN;
1826 // }
1827 // } else {
1828 // result = DPow(x,y);
1829 // }
1830 // if (result != result)? {
1831 // result = uncommon_trap() or runtime_call();
1832 // }
1833 // return result;
1834 // }
1835
1836 Node* x = round_double_node(argument(0));
1837 Node* y = round_double_node(argument(2));
1838
1839 Node* result = NULL;
1840
1841 Node* const_two_node = makecon(TypeD::make(2.0));
1842 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1843 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1844 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1845 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1846 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1847
1848 RegionNode* region_node = new RegionNode(3);
1849 region_node->init_req(1, if_true);
1850
1851 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1852 // special case for x^y where y == 2, we can convert it to x * x
1853 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1854
1855 // set control to if_false since we will now process the false branch
1856 set_control(if_false);
1857
1858 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1859 // Short form: skip the fancy tests and just check for NaN result.
1860 result = _gvn.transform(new PowDNode(C, control(), x, y));
1861 } else {
1862 // If this inlining ever returned NaN in the past, include all
1863 // checks + call to the runtime.
1864
1865 // Set the merge point for If node with condition of (x <= 0.0)
1866 // There are four possible paths to region node and phi node
1867 RegionNode *r = new RegionNode(4);
1868 Node *phi = new PhiNode(r, Type::DOUBLE);
1869
1870 // Build the first if node: if (x <= 0.0)
1871 // Node for 0 constant
1872 Node *zeronode = makecon(TypeD::ZERO);
1873 // Check x:0
1874 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1875 // Check: If (x<=0) then go complex path
1876 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1877 // Branch either way
1878 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1879 // Fast path taken; set region slot 3
1880 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1881 r->init_req(3,fast_taken); // Capture fast-control
1882
1883 // Fast path not-taken, i.e. slow path
1884 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1885
1886 // Set fast path result
1887 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1888 phi->init_req(3, fast_result);
1889
1890 // Complex path
1891 // Build the second if node (if y is long)
1892 // Node for (long)y
1893 Node *longy = _gvn.transform(new ConvD2LNode(y));
1894 // Node for (double)((long) y)
1895 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1896 // Check (double)((long) y) : y
1897 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1898 // Check if (y isn't long) then go to slow path
1899
1900 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1901 // Branch either way
1902 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1903 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1904
1905 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1906
1907 // Calculate DPow(abs(x), y)*(1 & (long)y)
1908 // Node for constant 1
1909 Node *conone = longcon(1);
1910 // 1& (long)y
1911 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1912
1913 // A huge number is always even. Detect a huge number by checking
1914 // if y + 1 == y and set integer to be tested for parity to 0.
1915 // Required for corner case:
1916 // (long)9.223372036854776E18 = max_jlong
1917 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1918 // max_jlong is odd but 9.223372036854776E18 is even
1919 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1920 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1921 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1922 Node* correctedsign = NULL;
1923 if (ConditionalMoveLimit != 0) {
1924 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1925 } else {
1926 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1927 RegionNode *r = new RegionNode(3);
1928 Node *phi = new PhiNode(r, TypeLong::LONG);
1929 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1930 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1931 phi->init_req(1, signnode);
1932 phi->init_req(2, longcon(0));
1933 correctedsign = _gvn.transform(phi);
1934 ylong_path = _gvn.transform(r);
1935 record_for_igvn(r);
1936 }
1937
1938 // zero node
1939 Node *conzero = longcon(0);
1940 // Check (1&(long)y)==0?
1941 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1942 // Check if (1&(long)y)!=0?, if so the result is negative
1943 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1944 // abs(x)
1945 Node *absx=_gvn.transform(new AbsDNode(x));
1946 // abs(x)^y
1947 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1948 // -abs(x)^y
1949 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1950 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1951 Node *signresult = NULL;
1952 if (ConditionalMoveLimit != 0) {
1953 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1954 } else {
1955 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1956 RegionNode *r = new RegionNode(3);
1957 Node *phi = new PhiNode(r, Type::DOUBLE);
1958 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1959 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1960 phi->init_req(1, absxpowy);
1961 phi->init_req(2, negabsxpowy);
1962 signresult = _gvn.transform(phi);
1963 ylong_path = _gvn.transform(r);
1964 record_for_igvn(r);
1965 }
1966 // Set complex path fast result
1967 r->init_req(2, ylong_path);
1968 phi->init_req(2, signresult);
1969
1970 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1971 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1972 r->init_req(1,slow_path);
1973 phi->init_req(1,slow_result);
1974
1975 // Post merge
1976 set_control(_gvn.transform(r));
1977 record_for_igvn(r);
1978 result = _gvn.transform(phi);
1979 }
1980
1981 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1982
1983 // control from finish_pow_exp is now input to the region node
1984 region_node->set_req(2, control());
1985 // the result from finish_pow_exp is now input to the phi node
1986 phi_node->init_req(2, result);
1987 set_control(_gvn.transform(region_node));
1988 record_for_igvn(region_node);
1989 set_result(_gvn.transform(phi_node));
1990
1991 C->set_has_split_ifs(true); // Has chance for split-if optimization
1992 return true;
1993 }
1994
1995 //------------------------------runtime_math-----------------------------
1996 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1997 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1998 "must be (DD)D or (D)D type");
1999
2000 // Inputs
2001 Node* a = round_double_node(argument(0));
2002 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2003
2004 const TypePtr* no_memory_effects = NULL;
2005 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2006 no_memory_effects,
2007 a, top(), b, b ? top() : NULL);
2008 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
2009 #ifdef ASSERT
2010 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
2011 assert(value_top == top(), "second value must be top");
2012 #endif
2013
2014 set_result(value);
2015 return true;
2016 }
2017
2018 //------------------------------inline_math_native-----------------------------
2019 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2020 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2021 switch (id) {
2022 // These intrinsics are not properly supported on all hardware
2023 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2024 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2025 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2026 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2027 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2028 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2029
2030 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2031 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2032 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2033 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2034
2035 // These intrinsics are supported on all hardware
2036 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2037 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2038
2039 case vmIntrinsics::_dexp:
2040 return VM_Version::supports_sse2() ? runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
2041 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2042 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2043 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2044 #undef FN_PTR
2045
2046 // These intrinsics are not yet correctly implemented
2047 case vmIntrinsics::_datan2:
2048 return false;
2049
2050 default:
2051 fatal_unexpected_iid(id);
2052 return false;
2053 }
2054 }
2055
2056 static bool is_simple_name(Node* n) {
2057 return (n->req() == 1 // constant
2058 || (n->is_Type() && n->as_Type()->type()->singleton())
2059 || n->is_Proj() // parameter or return value
2060 || n->is_Phi() // local of some sort
2061 );
2062 }
2063
2064 //----------------------------inline_min_max-----------------------------------
2065 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2066 set_result(generate_min_max(id, argument(0), argument(1)));
2067 return true;
2068 }
2069
2070 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2071 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2072 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2073 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2074 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2075
2076 {
2077 PreserveJVMState pjvms(this);
2078 PreserveReexecuteState preexecs(this);
2079 jvms()->set_should_reexecute(true);
2080
2081 set_control(slow_path);
2082 set_i_o(i_o());
2083
2084 uncommon_trap(Deoptimization::Reason_intrinsic,
2085 Deoptimization::Action_none);
2086 }
2087
2088 set_control(fast_path);
2089 set_result(math);
2090 }
2091
2092 template <typename OverflowOp>
2093 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2094 typedef typename OverflowOp::MathOp MathOp;
2095
2096 MathOp* mathOp = new MathOp(arg1, arg2);
2097 Node* operation = _gvn.transform( mathOp );
2098 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2099 inline_math_mathExact(operation, ofcheck);
2100 return true;
2101 }
2102
2103 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2104 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2105 }
2106
2107 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2108 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2109 }
2110
2111 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2112 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2113 }
2114
2115 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2116 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2117 }
2118
2119 bool LibraryCallKit::inline_math_negateExactI() {
2120 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2121 }
2122
2123 bool LibraryCallKit::inline_math_negateExactL() {
2124 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2125 }
2126
2127 bool LibraryCallKit::inline_math_multiplyExactI() {
2128 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2129 }
2130
2131 bool LibraryCallKit::inline_math_multiplyExactL() {
2132 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2133 }
2134
2135 Node*
2136 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2137 // These are the candidate return value:
2138 Node* xvalue = x0;
2139 Node* yvalue = y0;
2140
2141 if (xvalue == yvalue) {
2142 return xvalue;
2143 }
2144
2145 bool want_max = (id == vmIntrinsics::_max);
2146
2147 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2148 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2149 if (txvalue == NULL || tyvalue == NULL) return top();
2150 // This is not really necessary, but it is consistent with a
2151 // hypothetical MaxINode::Value method:
2152 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2153
2154 // %%% This folding logic should (ideally) be in a different place.
2155 // Some should be inside IfNode, and there to be a more reliable
2156 // transformation of ?: style patterns into cmoves. We also want
2157 // more powerful optimizations around cmove and min/max.
2158
2159 // Try to find a dominating comparison of these guys.
2160 // It can simplify the index computation for Arrays.copyOf
2161 // and similar uses of System.arraycopy.
2162 // First, compute the normalized version of CmpI(x, y).
2163 int cmp_op = Op_CmpI;
2164 Node* xkey = xvalue;
2165 Node* ykey = yvalue;
2166 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2167 if (ideal_cmpxy->is_Cmp()) {
2168 // E.g., if we have CmpI(length - offset, count),
2169 // it might idealize to CmpI(length, count + offset)
2170 cmp_op = ideal_cmpxy->Opcode();
2171 xkey = ideal_cmpxy->in(1);
2172 ykey = ideal_cmpxy->in(2);
2173 }
2174
2175 // Start by locating any relevant comparisons.
2176 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2177 Node* cmpxy = NULL;
2178 Node* cmpyx = NULL;
2179 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2180 Node* cmp = start_from->fast_out(k);
2181 if (cmp->outcnt() > 0 && // must have prior uses
2182 cmp->in(0) == NULL && // must be context-independent
2183 cmp->Opcode() == cmp_op) { // right kind of compare
2184 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2185 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2186 }
2187 }
2188
2189 const int NCMPS = 2;
2190 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2191 int cmpn;
2192 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2193 if (cmps[cmpn] != NULL) break; // find a result
2194 }
2195 if (cmpn < NCMPS) {
2196 // Look for a dominating test that tells us the min and max.
2197 int depth = 0; // Limit search depth for speed
2198 Node* dom = control();
2199 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2200 if (++depth >= 100) break;
2201 Node* ifproj = dom;
2202 if (!ifproj->is_Proj()) continue;
2203 Node* iff = ifproj->in(0);
2204 if (!iff->is_If()) continue;
2205 Node* bol = iff->in(1);
2206 if (!bol->is_Bool()) continue;
2207 Node* cmp = bol->in(1);
2208 if (cmp == NULL) continue;
2209 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2210 if (cmps[cmpn] == cmp) break;
2211 if (cmpn == NCMPS) continue;
2212 BoolTest::mask btest = bol->as_Bool()->_test._test;
2213 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2214 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2215 // At this point, we know that 'x btest y' is true.
2216 switch (btest) {
2217 case BoolTest::eq:
2218 // They are proven equal, so we can collapse the min/max.
2219 // Either value is the answer. Choose the simpler.
2220 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2221 return yvalue;
2222 return xvalue;
2223 case BoolTest::lt: // x < y
2224 case BoolTest::le: // x <= y
2225 return (want_max ? yvalue : xvalue);
2226 case BoolTest::gt: // x > y
2227 case BoolTest::ge: // x >= y
2228 return (want_max ? xvalue : yvalue);
2229 }
2230 }
2231 }
2232
2233 // We failed to find a dominating test.
2234 // Let's pick a test that might GVN with prior tests.
2235 Node* best_bol = NULL;
2236 BoolTest::mask best_btest = BoolTest::illegal;
2237 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2238 Node* cmp = cmps[cmpn];
2239 if (cmp == NULL) continue;
2240 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2241 Node* bol = cmp->fast_out(j);
2242 if (!bol->is_Bool()) continue;
2243 BoolTest::mask btest = bol->as_Bool()->_test._test;
2244 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2245 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2246 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2247 best_bol = bol->as_Bool();
2248 best_btest = btest;
2249 }
2250 }
2251 }
2252
2253 Node* answer_if_true = NULL;
2254 Node* answer_if_false = NULL;
2255 switch (best_btest) {
2256 default:
2257 if (cmpxy == NULL)
2258 cmpxy = ideal_cmpxy;
2259 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2260 // and fall through:
2261 case BoolTest::lt: // x < y
2262 case BoolTest::le: // x <= y
2263 answer_if_true = (want_max ? yvalue : xvalue);
2264 answer_if_false = (want_max ? xvalue : yvalue);
2265 break;
2266 case BoolTest::gt: // x > y
2267 case BoolTest::ge: // x >= y
2268 answer_if_true = (want_max ? xvalue : yvalue);
2269 answer_if_false = (want_max ? yvalue : xvalue);
2270 break;
2271 }
2272
2273 jint hi, lo;
2274 if (want_max) {
2275 // We can sharpen the minimum.
2276 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2277 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2278 } else {
2279 // We can sharpen the maximum.
2280 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2281 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2282 }
2283
2284 // Use a flow-free graph structure, to avoid creating excess control edges
2285 // which could hinder other optimizations.
2286 // Since Math.min/max is often used with arraycopy, we want
2287 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2288 Node* cmov = CMoveNode::make(NULL, best_bol,
2289 answer_if_false, answer_if_true,
2290 TypeInt::make(lo, hi, widen));
2291
2292 return _gvn.transform(cmov);
2293
2294 /*
2295 // This is not as desirable as it may seem, since Min and Max
2296 // nodes do not have a full set of optimizations.
2297 // And they would interfere, anyway, with 'if' optimizations
2298 // and with CMoveI canonical forms.
2299 switch (id) {
2300 case vmIntrinsics::_min:
2301 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2302 case vmIntrinsics::_max:
2303 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2304 default:
2305 ShouldNotReachHere();
2306 }
2307 */
2308 }
2309
2310 inline int
2311 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2312 const TypePtr* base_type = TypePtr::NULL_PTR;
2313 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2314 if (base_type == NULL) {
2315 // Unknown type.
2316 return Type::AnyPtr;
2317 } else if (base_type == TypePtr::NULL_PTR) {
2318 // Since this is a NULL+long form, we have to switch to a rawptr.
2319 base = _gvn.transform(new CastX2PNode(offset));
2320 offset = MakeConX(0);
2321 return Type::RawPtr;
2322 } else if (base_type->base() == Type::RawPtr) {
2323 return Type::RawPtr;
2324 } else if (base_type->isa_oopptr()) {
2325 // Base is never null => always a heap address.
2326 if (base_type->ptr() == TypePtr::NotNull) {
2327 return Type::OopPtr;
2328 }
2329 // Offset is small => always a heap address.
2330 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2331 if (offset_type != NULL &&
2332 base_type->offset() == 0 && // (should always be?)
2333 offset_type->_lo >= 0 &&
2334 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2335 return Type::OopPtr;
2336 }
2337 // Otherwise, it might either be oop+off or NULL+addr.
2338 return Type::AnyPtr;
2339 } else {
2340 // No information:
2341 return Type::AnyPtr;
2342 }
2343 }
2344
2345 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2346 int kind = classify_unsafe_addr(base, offset);
2347 if (kind == Type::RawPtr) {
2348 return basic_plus_adr(top(), base, offset);
2349 } else {
2350 return basic_plus_adr(base, offset);
2351 }
2352 }
2353
2354 //--------------------------inline_number_methods-----------------------------
2355 // inline int Integer.numberOfLeadingZeros(int)
2356 // inline int Long.numberOfLeadingZeros(long)
2357 //
2358 // inline int Integer.numberOfTrailingZeros(int)
2359 // inline int Long.numberOfTrailingZeros(long)
2360 //
2361 // inline int Integer.bitCount(int)
2362 // inline int Long.bitCount(long)
2363 //
2364 // inline char Character.reverseBytes(char)
2365 // inline short Short.reverseBytes(short)
2366 // inline int Integer.reverseBytes(int)
2367 // inline long Long.reverseBytes(long)
2368 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2369 Node* arg = argument(0);
2370 Node* n;
2371 switch (id) {
2372 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2373 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2374 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2375 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2376 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2377 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2378 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2379 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2380 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2381 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2382 default: fatal_unexpected_iid(id); break;
2383 }
2384 set_result(_gvn.transform(n));
2385 return true;
2386 }
2387
2388 //----------------------------inline_unsafe_access----------------------------
2389
2390 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2391
2392 // Helper that guards and inserts a pre-barrier.
2393 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2394 Node* pre_val, bool need_mem_bar) {
2395 // We could be accessing the referent field of a reference object. If so, when G1
2396 // is enabled, we need to log the value in the referent field in an SATB buffer.
2397 // This routine performs some compile time filters and generates suitable
2398 // runtime filters that guard the pre-barrier code.
2399 // Also add memory barrier for non volatile load from the referent field
2400 // to prevent commoning of loads across safepoint.
2401 if (!UseG1GC && !need_mem_bar)
2402 return;
2403
2404 // Some compile time checks.
2405
2406 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2407 const TypeX* otype = offset->find_intptr_t_type();
2408 if (otype != NULL && otype->is_con() &&
2409 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2410 // Constant offset but not the reference_offset so just return
2411 return;
2412 }
2413
2414 // We only need to generate the runtime guards for instances.
2415 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2416 if (btype != NULL) {
2417 if (btype->isa_aryptr()) {
2418 // Array type so nothing to do
2419 return;
2420 }
2421
2422 const TypeInstPtr* itype = btype->isa_instptr();
2423 if (itype != NULL) {
2424 // Can the klass of base_oop be statically determined to be
2425 // _not_ a sub-class of Reference and _not_ Object?
2426 ciKlass* klass = itype->klass();
2427 if ( klass->is_loaded() &&
2428 !klass->is_subtype_of(env()->Reference_klass()) &&
2429 !env()->Object_klass()->is_subtype_of(klass)) {
2430 return;
2431 }
2432 }
2433 }
2434
2435 // The compile time filters did not reject base_oop/offset so
2436 // we need to generate the following runtime filters
2437 //
2438 // if (offset == java_lang_ref_Reference::_reference_offset) {
2439 // if (instance_of(base, java.lang.ref.Reference)) {
2440 // pre_barrier(_, pre_val, ...);
2441 // }
2442 // }
2443
2444 float likely = PROB_LIKELY( 0.999);
2445 float unlikely = PROB_UNLIKELY(0.999);
2446
2447 IdealKit ideal(this);
2448 #define __ ideal.
2449
2450 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2451
2452 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2453 // Update graphKit memory and control from IdealKit.
2454 sync_kit(ideal);
2455
2456 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2457 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2458
2459 // Update IdealKit memory and control from graphKit.
2460 __ sync_kit(this);
2461
2462 Node* one = __ ConI(1);
2463 // is_instof == 0 if base_oop == NULL
2464 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2465
2466 // Update graphKit from IdeakKit.
2467 sync_kit(ideal);
2468
2469 // Use the pre-barrier to record the value in the referent field
2470 pre_barrier(false /* do_load */,
2471 __ ctrl(),
2472 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2473 pre_val /* pre_val */,
2474 T_OBJECT);
2475 if (need_mem_bar) {
2476 // Add memory barrier to prevent commoning reads from this field
2477 // across safepoint since GC can change its value.
2478 insert_mem_bar(Op_MemBarCPUOrder);
2479 }
2480 // Update IdealKit from graphKit.
2481 __ sync_kit(this);
2482
2483 } __ end_if(); // _ref_type != ref_none
2484 } __ end_if(); // offset == referent_offset
2485
2486 // Final sync IdealKit and GraphKit.
2487 final_sync(ideal);
2488 #undef __
2489 }
2490
2491
2492 // Interpret Unsafe.fieldOffset cookies correctly:
2493 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2494
2495 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2496 // Attempt to infer a sharper value type from the offset and base type.
2497 ciKlass* sharpened_klass = NULL;
2498
2499 // See if it is an instance field, with an object type.
2500 if (alias_type->field() != NULL) {
2501 assert(!is_native_ptr, "native pointer op cannot use a java address");
2502 if (alias_type->field()->type()->is_klass()) {
2503 sharpened_klass = alias_type->field()->type()->as_klass();
2504 }
2505 }
2506
2507 // See if it is a narrow oop array.
2508 if (adr_type->isa_aryptr()) {
2509 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2510 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2511 if (elem_type != NULL) {
2512 sharpened_klass = elem_type->klass();
2513 }
2514 }
2515 }
2516
2517 // The sharpened class might be unloaded if there is no class loader
2518 // contraint in place.
2519 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2520 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2521
2522 #ifndef PRODUCT
2523 if (C->print_intrinsics() || C->print_inlining()) {
2524 tty->print(" from base type: "); adr_type->dump();
2525 tty->print(" sharpened value: "); tjp->dump();
2526 }
2527 #endif
2528 // Sharpen the value type.
2529 return tjp;
2530 }
2531 return NULL;
2532 }
2533
2534 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2535 if (callee()->is_static()) return false; // caller must have the capability!
2536
2537 #ifndef PRODUCT
2538 {
2539 ResourceMark rm;
2540 // Check the signatures.
2541 ciSignature* sig = callee()->signature();
2542 #ifdef ASSERT
2543 if (!is_store) {
2544 // Object getObject(Object base, int/long offset), etc.
2545 BasicType rtype = sig->return_type()->basic_type();
2546 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2547 rtype = T_ADDRESS; // it is really a C void*
2548 assert(rtype == type, "getter must return the expected value");
2549 if (!is_native_ptr) {
2550 assert(sig->count() == 2, "oop getter has 2 arguments");
2551 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2552 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2553 } else {
2554 assert(sig->count() == 1, "native getter has 1 argument");
2555 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2556 }
2557 } else {
2558 // void putObject(Object base, int/long offset, Object x), etc.
2559 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2560 if (!is_native_ptr) {
2561 assert(sig->count() == 3, "oop putter has 3 arguments");
2562 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2563 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2564 } else {
2565 assert(sig->count() == 2, "native putter has 2 arguments");
2566 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2567 }
2568 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2569 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2570 vtype = T_ADDRESS; // it is really a C void*
2571 assert(vtype == type, "putter must accept the expected value");
2572 }
2573 #endif // ASSERT
2574 }
2575 #endif //PRODUCT
2576
2577 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2578
2579 Node* receiver = argument(0); // type: oop
2580
2581 // Build address expression.
2582 Node* adr;
2583 Node* heap_base_oop = top();
2584 Node* offset = top();
2585 Node* val;
2586
2587 if (!is_native_ptr) {
2588 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2589 Node* base = argument(1); // type: oop
2590 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2591 offset = argument(2); // type: long
2592 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2593 // to be plain byte offsets, which are also the same as those accepted
2594 // by oopDesc::field_base.
2595 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2596 "fieldOffset must be byte-scaled");
2597 // 32-bit machines ignore the high half!
2598 offset = ConvL2X(offset);
2599 adr = make_unsafe_address(base, offset);
2600 heap_base_oop = base;
2601 val = is_store ? argument(4) : NULL;
2602 } else {
2603 Node* ptr = argument(1); // type: long
2604 ptr = ConvL2X(ptr); // adjust Java long to machine word
2605 adr = make_unsafe_address(NULL, ptr);
2606 val = is_store ? argument(3) : NULL;
2607 }
2608
2609 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2610
2611 // First guess at the value type.
2612 const Type *value_type = Type::get_const_basic_type(type);
2613
2614 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2615 // there was not enough information to nail it down.
2616 Compile::AliasType* alias_type = C->alias_type(adr_type);
2617 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2618
2619 // We will need memory barriers unless we can determine a unique
2620 // alias category for this reference. (Note: If for some reason
2621 // the barriers get omitted and the unsafe reference begins to "pollute"
2622 // the alias analysis of the rest of the graph, either Compile::can_alias
2623 // or Compile::must_alias will throw a diagnostic assert.)
2624 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2625
2626 // If we are reading the value of the referent field of a Reference
2627 // object (either by using Unsafe directly or through reflection)
2628 // then, if G1 is enabled, we need to record the referent in an
2629 // SATB log buffer using the pre-barrier mechanism.
2630 // Also we need to add memory barrier to prevent commoning reads
2631 // from this field across safepoint since GC can change its value.
2632 bool need_read_barrier = !is_native_ptr && !is_store &&
2633 offset != top() && heap_base_oop != top();
2634
2635 if (!is_store && type == T_OBJECT) {
2636 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2637 if (tjp != NULL) {
2638 value_type = tjp;
2639 }
2640 }
2641
2642 receiver = null_check(receiver);
2643 if (stopped()) {
2644 return true;
2645 }
2646 // Heap pointers get a null-check from the interpreter,
2647 // as a courtesy. However, this is not guaranteed by Unsafe,
2648 // and it is not possible to fully distinguish unintended nulls
2649 // from intended ones in this API.
2650
2651 if (is_volatile) {
2652 // We need to emit leading and trailing CPU membars (see below) in
2653 // addition to memory membars when is_volatile. This is a little
2654 // too strong, but avoids the need to insert per-alias-type
2655 // volatile membars (for stores; compare Parse::do_put_xxx), which
2656 // we cannot do effectively here because we probably only have a
2657 // rough approximation of type.
2658 need_mem_bar = true;
2659 // For Stores, place a memory ordering barrier now.
2660 if (is_store) {
2661 insert_mem_bar(Op_MemBarRelease);
2662 } else {
2663 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2664 insert_mem_bar(Op_MemBarVolatile);
2665 }
2666 }
2667 }
2668
2669 // Memory barrier to prevent normal and 'unsafe' accesses from
2670 // bypassing each other. Happens after null checks, so the
2671 // exception paths do not take memory state from the memory barrier,
2672 // so there's no problems making a strong assert about mixing users
2673 // of safe & unsafe memory.
2674 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2675
2676 if (!is_store) {
2677 Node* p = NULL;
2678 // Try to constant fold a load from a constant field
2679 ciField* field = alias_type->field();
2680 if (heap_base_oop != top() &&
2681 field != NULL && field->is_constant() && field->layout_type() == type) {
2682 // final or stable field
2683 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2684 if (con_type != NULL) {
2685 p = makecon(con_type);
2686 }
2687 }
2688 if (p == NULL) {
2689 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2690 // To be valid, unsafe loads may depend on other conditions than
2691 // the one that guards them: pin the Load node
2692 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2693 // load value
2694 switch (type) {
2695 case T_BOOLEAN:
2696 case T_CHAR:
2697 case T_BYTE:
2698 case T_SHORT:
2699 case T_INT:
2700 case T_LONG:
2701 case T_FLOAT:
2702 case T_DOUBLE:
2703 break;
2704 case T_OBJECT:
2705 if (need_read_barrier) {
2706 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2707 }
2708 break;
2709 case T_ADDRESS:
2710 // Cast to an int type.
2711 p = _gvn.transform(new CastP2XNode(NULL, p));
2712 p = ConvX2UL(p);
2713 break;
2714 default:
2715 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2716 break;
2717 }
2718 }
2719 // The load node has the control of the preceding MemBarCPUOrder. All
2720 // following nodes will have the control of the MemBarCPUOrder inserted at
2721 // the end of this method. So, pushing the load onto the stack at a later
2722 // point is fine.
2723 set_result(p);
2724 } else {
2725 // place effect of store into memory
2726 switch (type) {
2727 case T_DOUBLE:
2728 val = dstore_rounding(val);
2729 break;
2730 case T_ADDRESS:
2731 // Repackage the long as a pointer.
2732 val = ConvL2X(val);
2733 val = _gvn.transform(new CastX2PNode(val));
2734 break;
2735 }
2736
2737 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2738 if (type != T_OBJECT ) {
2739 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2740 } else {
2741 // Possibly an oop being stored to Java heap or native memory
2742 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2743 // oop to Java heap.
2744 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2745 } else {
2746 // We can't tell at compile time if we are storing in the Java heap or outside
2747 // of it. So we need to emit code to conditionally do the proper type of
2748 // store.
2749
2750 IdealKit ideal(this);
2751 #define __ ideal.
2752 // QQQ who knows what probability is here??
2753 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2754 // Sync IdealKit and graphKit.
2755 sync_kit(ideal);
2756 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2757 // Update IdealKit memory.
2758 __ sync_kit(this);
2759 } __ else_(); {
2760 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2761 } __ end_if();
2762 // Final sync IdealKit and GraphKit.
2763 final_sync(ideal);
2764 #undef __
2765 }
2766 }
2767 }
2768
2769 if (is_volatile) {
2770 if (!is_store) {
2771 insert_mem_bar(Op_MemBarAcquire);
2772 } else {
2773 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2774 insert_mem_bar(Op_MemBarVolatile);
2775 }
2776 }
2777 }
2778
2779 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2780
2781 return true;
2782 }
2783
2784 //----------------------------inline_unsafe_load_store----------------------------
2785 // This method serves a couple of different customers (depending on LoadStoreKind):
2786 //
2787 // LS_cmpxchg:
2788 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2789 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2790 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2791 //
2792 // LS_xadd:
2793 // public int getAndAddInt( Object o, long offset, int delta)
2794 // public long getAndAddLong(Object o, long offset, long delta)
2795 //
2796 // LS_xchg:
2797 // int getAndSet(Object o, long offset, int newValue)
2798 // long getAndSet(Object o, long offset, long newValue)
2799 // Object getAndSet(Object o, long offset, Object newValue)
2800 //
2801 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2802 // This basic scheme here is the same as inline_unsafe_access, but
2803 // differs in enough details that combining them would make the code
2804 // overly confusing. (This is a true fact! I originally combined
2805 // them, but even I was confused by it!) As much code/comments as
2806 // possible are retained from inline_unsafe_access though to make
2807 // the correspondences clearer. - dl
2808
2809 if (callee()->is_static()) return false; // caller must have the capability!
2810
2811 #ifndef PRODUCT
2812 BasicType rtype;
2813 {
2814 ResourceMark rm;
2815 // Check the signatures.
2816 ciSignature* sig = callee()->signature();
2817 rtype = sig->return_type()->basic_type();
2818 if (kind == LS_xadd || kind == LS_xchg) {
2819 // Check the signatures.
2820 #ifdef ASSERT
2821 assert(rtype == type, "get and set must return the expected type");
2822 assert(sig->count() == 3, "get and set has 3 arguments");
2823 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2824 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2825 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2826 #endif // ASSERT
2827 } else if (kind == LS_cmpxchg) {
2828 // Check the signatures.
2829 #ifdef ASSERT
2830 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2831 assert(sig->count() == 4, "CAS has 4 arguments");
2832 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2833 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2834 #endif // ASSERT
2835 } else {
2836 ShouldNotReachHere();
2837 }
2838 }
2839 #endif //PRODUCT
2840
2841 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2842
2843 // Get arguments:
2844 Node* receiver = NULL;
2845 Node* base = NULL;
2846 Node* offset = NULL;
2847 Node* oldval = NULL;
2848 Node* newval = NULL;
2849 if (kind == LS_cmpxchg) {
2850 const bool two_slot_type = type2size[type] == 2;
2851 receiver = argument(0); // type: oop
2852 base = argument(1); // type: oop
2853 offset = argument(2); // type: long
2854 oldval = argument(4); // type: oop, int, or long
2855 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2856 } else if (kind == LS_xadd || kind == LS_xchg){
2857 receiver = argument(0); // type: oop
2858 base = argument(1); // type: oop
2859 offset = argument(2); // type: long
2860 oldval = NULL;
2861 newval = argument(4); // type: oop, int, or long
2862 }
2863
2864 // Null check receiver.
2865 receiver = null_check(receiver);
2866 if (stopped()) {
2867 return true;
2868 }
2869
2870 // Build field offset expression.
2871 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2872 // to be plain byte offsets, which are also the same as those accepted
2873 // by oopDesc::field_base.
2874 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2875 // 32-bit machines ignore the high half of long offsets
2876 offset = ConvL2X(offset);
2877 Node* adr = make_unsafe_address(base, offset);
2878 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2879
2880 // For CAS, unlike inline_unsafe_access, there seems no point in
2881 // trying to refine types. Just use the coarse types here.
2882 const Type *value_type = Type::get_const_basic_type(type);
2883 Compile::AliasType* alias_type = C->alias_type(adr_type);
2884 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2885
2886 if (kind == LS_xchg && type == T_OBJECT) {
2887 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2888 if (tjp != NULL) {
2889 value_type = tjp;
2890 }
2891 }
2892
2893 int alias_idx = C->get_alias_index(adr_type);
2894
2895 // Memory-model-wise, a LoadStore acts like a little synchronized
2896 // block, so needs barriers on each side. These don't translate
2897 // into actual barriers on most machines, but we still need rest of
2898 // compiler to respect ordering.
2899
2900 insert_mem_bar(Op_MemBarRelease);
2901 insert_mem_bar(Op_MemBarCPUOrder);
2902
2903 // 4984716: MemBars must be inserted before this
2904 // memory node in order to avoid a false
2905 // dependency which will confuse the scheduler.
2906 Node *mem = memory(alias_idx);
2907
2908 // For now, we handle only those cases that actually exist: ints,
2909 // longs, and Object. Adding others should be straightforward.
2910 Node* load_store;
2911 switch(type) {
2912 case T_INT:
2913 if (kind == LS_xadd) {
2914 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2915 } else if (kind == LS_xchg) {
2916 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2917 } else if (kind == LS_cmpxchg) {
2918 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2919 } else {
2920 ShouldNotReachHere();
2921 }
2922 break;
2923 case T_LONG:
2924 if (kind == LS_xadd) {
2925 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2926 } else if (kind == LS_xchg) {
2927 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2928 } else if (kind == LS_cmpxchg) {
2929 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2930 } else {
2931 ShouldNotReachHere();
2932 }
2933 break;
2934 case T_OBJECT:
2935 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2936 // could be delayed during Parse (for example, in adjust_map_after_if()).
2937 // Execute transformation here to avoid barrier generation in such case.
2938 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2939 newval = _gvn.makecon(TypePtr::NULL_PTR);
2940
2941 // Reference stores need a store barrier.
2942 if (kind == LS_xchg) {
2943 // If pre-barrier must execute before the oop store, old value will require do_load here.
2944 if (!can_move_pre_barrier()) {
2945 pre_barrier(true /* do_load*/,
2946 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2947 NULL /* pre_val*/,
2948 T_OBJECT);
2949 } // Else move pre_barrier to use load_store value, see below.
2950 } else if (kind == LS_cmpxchg) {
2951 // Same as for newval above:
2952 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2953 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2954 }
2955 // The only known value which might get overwritten is oldval.
2956 pre_barrier(false /* do_load */,
2957 control(), NULL, NULL, max_juint, NULL, NULL,
2958 oldval /* pre_val */,
2959 T_OBJECT);
2960 } else {
2961 ShouldNotReachHere();
2962 }
2963
2964 #ifdef _LP64
2965 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2966 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2967 if (kind == LS_xchg) {
2968 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2969 newval_enc, adr_type, value_type->make_narrowoop()));
2970 } else {
2971 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2972 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2973 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2974 newval_enc, oldval_enc));
2975 }
2976 } else
2977 #endif
2978 {
2979 if (kind == LS_xchg) {
2980 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2981 } else {
2982 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2983 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2984 }
2985 }
2986 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2987 break;
2988 default:
2989 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2990 break;
2991 }
2992
2993 // SCMemProjNodes represent the memory state of a LoadStore. Their
2994 // main role is to prevent LoadStore nodes from being optimized away
2995 // when their results aren't used.
2996 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2997 set_memory(proj, alias_idx);
2998
2999 if (type == T_OBJECT && kind == LS_xchg) {
3000 #ifdef _LP64
3001 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3002 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3003 }
3004 #endif
3005 if (can_move_pre_barrier()) {
3006 // Don't need to load pre_val. The old value is returned by load_store.
3007 // The pre_barrier can execute after the xchg as long as no safepoint
3008 // gets inserted between them.
3009 pre_barrier(false /* do_load */,
3010 control(), NULL, NULL, max_juint, NULL, NULL,
3011 load_store /* pre_val */,
3012 T_OBJECT);
3013 }
3014 }
3015
3016 // Add the trailing membar surrounding the access
3017 insert_mem_bar(Op_MemBarCPUOrder);
3018 insert_mem_bar(Op_MemBarAcquire);
3019
3020 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3021 set_result(load_store);
3022 return true;
3023 }
3024
3025 //----------------------------inline_unsafe_ordered_store----------------------
3026 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3027 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3028 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3029 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3030 // This is another variant of inline_unsafe_access, differing in
3031 // that it always issues store-store ("release") barrier and ensures
3032 // store-atomicity (which only matters for "long").
3033
3034 if (callee()->is_static()) return false; // caller must have the capability!
3035
3036 #ifndef PRODUCT
3037 {
3038 ResourceMark rm;
3039 // Check the signatures.
3040 ciSignature* sig = callee()->signature();
3041 #ifdef ASSERT
3042 BasicType rtype = sig->return_type()->basic_type();
3043 assert(rtype == T_VOID, "must return void");
3044 assert(sig->count() == 3, "has 3 arguments");
3045 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3046 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3047 #endif // ASSERT
3048 }
3049 #endif //PRODUCT
3050
3051 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3052
3053 // Get arguments:
3054 Node* receiver = argument(0); // type: oop
3055 Node* base = argument(1); // type: oop
3056 Node* offset = argument(2); // type: long
3057 Node* val = argument(4); // type: oop, int, or long
3058
3059 // Null check receiver.
3060 receiver = null_check(receiver);
3061 if (stopped()) {
3062 return true;
3063 }
3064
3065 // Build field offset expression.
3066 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3067 // 32-bit machines ignore the high half of long offsets
3068 offset = ConvL2X(offset);
3069 Node* adr = make_unsafe_address(base, offset);
3070 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3071 const Type *value_type = Type::get_const_basic_type(type);
3072 Compile::AliasType* alias_type = C->alias_type(adr_type);
3073
3074 insert_mem_bar(Op_MemBarRelease);
3075 insert_mem_bar(Op_MemBarCPUOrder);
3076 // Ensure that the store is atomic for longs:
3077 const bool require_atomic_access = true;
3078 Node* store;
3079 if (type == T_OBJECT) // reference stores need a store barrier.
3080 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3081 else {
3082 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3083 }
3084 insert_mem_bar(Op_MemBarCPUOrder);
3085 return true;
3086 }
3087
3088 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3089 // Regardless of form, don't allow previous ld/st to move down,
3090 // then issue acquire, release, or volatile mem_bar.
3091 insert_mem_bar(Op_MemBarCPUOrder);
3092 switch(id) {
3093 case vmIntrinsics::_loadFence:
3094 insert_mem_bar(Op_LoadFence);
3095 return true;
3096 case vmIntrinsics::_storeFence:
3097 insert_mem_bar(Op_StoreFence);
3098 return true;
3099 case vmIntrinsics::_fullFence:
3100 insert_mem_bar(Op_MemBarVolatile);
3101 return true;
3102 default:
3103 fatal_unexpected_iid(id);
3104 return false;
3105 }
3106 }
3107
3108 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3109 if (!kls->is_Con()) {
3110 return true;
3111 }
3112 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3113 if (klsptr == NULL) {
3114 return true;
3115 }
3116 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3117 // don't need a guard for a klass that is already initialized
3118 return !ik->is_initialized();
3119 }
3120
3121 //----------------------------inline_unsafe_allocate---------------------------
3122 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3123 bool LibraryCallKit::inline_unsafe_allocate() {
3124 if (callee()->is_static()) return false; // caller must have the capability!
3125
3126 null_check_receiver(); // null-check, then ignore
3127 Node* cls = null_check(argument(1));
3128 if (stopped()) return true;
3129
3130 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3131 kls = null_check(kls);
3132 if (stopped()) return true; // argument was like int.class
3133
3134 Node* test = NULL;
3135 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3136 // Note: The argument might still be an illegal value like
3137 // Serializable.class or Object[].class. The runtime will handle it.
3138 // But we must make an explicit check for initialization.
3139 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3140 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3141 // can generate code to load it as unsigned byte.
3142 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3143 Node* bits = intcon(InstanceKlass::fully_initialized);
3144 test = _gvn.transform(new SubINode(inst, bits));
3145 // The 'test' is non-zero if we need to take a slow path.
3146 }
3147
3148 Node* obj = new_instance(kls, test);
3149 set_result(obj);
3150 return true;
3151 }
3152
3153 #ifdef TRACE_HAVE_INTRINSICS
3154 /*
3155 * oop -> myklass
3156 * myklass->trace_id |= USED
3157 * return myklass->trace_id & ~0x3
3158 */
3159 bool LibraryCallKit::inline_native_classID() {
3160 null_check_receiver(); // null-check, then ignore
3161 Node* cls = null_check(argument(1), T_OBJECT);
3162 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3163 kls = null_check(kls, T_OBJECT);
3164 ByteSize offset = TRACE_ID_OFFSET;
3165 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3166 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3167 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3168 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3169 Node* clsused = longcon(0x01l); // set the class bit
3170 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3171
3172 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3173 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3174 set_result(andl);
3175 return true;
3176 }
3177
3178 bool LibraryCallKit::inline_native_threadID() {
3179 Node* tls_ptr = NULL;
3180 Node* cur_thr = generate_current_thread(tls_ptr);
3181 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3182 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3183 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3184
3185 Node* threadid = NULL;
3186 size_t thread_id_size = OSThread::thread_id_size();
3187 if (thread_id_size == (size_t) BytesPerLong) {
3188 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3189 } else if (thread_id_size == (size_t) BytesPerInt) {
3190 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3191 } else {
3192 ShouldNotReachHere();
3193 }
3194 set_result(threadid);
3195 return true;
3196 }
3197 #endif
3198
3199 //------------------------inline_native_time_funcs--------------
3200 // inline code for System.currentTimeMillis() and System.nanoTime()
3201 // these have the same type and signature
3202 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3203 const TypeFunc* tf = OptoRuntime::void_long_Type();
3204 const TypePtr* no_memory_effects = NULL;
3205 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3206 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3207 #ifdef ASSERT
3208 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3209 assert(value_top == top(), "second value must be top");
3210 #endif
3211 set_result(value);
3212 return true;
3213 }
3214
3215 //------------------------inline_native_currentThread------------------
3216 bool LibraryCallKit::inline_native_currentThread() {
3217 Node* junk = NULL;
3218 set_result(generate_current_thread(junk));
3219 return true;
3220 }
3221
3222 //------------------------inline_native_isInterrupted------------------
3223 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3224 bool LibraryCallKit::inline_native_isInterrupted() {
3225 // Add a fast path to t.isInterrupted(clear_int):
3226 // (t == Thread.current() &&
3227 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3228 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3229 // So, in the common case that the interrupt bit is false,
3230 // we avoid making a call into the VM. Even if the interrupt bit
3231 // is true, if the clear_int argument is false, we avoid the VM call.
3232 // However, if the receiver is not currentThread, we must call the VM,
3233 // because there must be some locking done around the operation.
3234
3235 // We only go to the fast case code if we pass two guards.
3236 // Paths which do not pass are accumulated in the slow_region.
3237
3238 enum {
3239 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3240 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3241 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3242 PATH_LIMIT
3243 };
3244
3245 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3246 // out of the function.
3247 insert_mem_bar(Op_MemBarCPUOrder);
3248
3249 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3250 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3251
3252 RegionNode* slow_region = new RegionNode(1);
3253 record_for_igvn(slow_region);
3254
3255 // (a) Receiving thread must be the current thread.
3256 Node* rec_thr = argument(0);
3257 Node* tls_ptr = NULL;
3258 Node* cur_thr = generate_current_thread(tls_ptr);
3259 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3260 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3261
3262 generate_slow_guard(bol_thr, slow_region);
3263
3264 // (b) Interrupt bit on TLS must be false.
3265 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3266 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3267 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3268
3269 // Set the control input on the field _interrupted read to prevent it floating up.
3270 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3271 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3272 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3273
3274 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3275
3276 // First fast path: if (!TLS._interrupted) return false;
3277 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3278 result_rgn->init_req(no_int_result_path, false_bit);
3279 result_val->init_req(no_int_result_path, intcon(0));
3280
3281 // drop through to next case
3282 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3283
3284 #ifndef TARGET_OS_FAMILY_windows
3285 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3286 Node* clr_arg = argument(1);
3287 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3288 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3289 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3290
3291 // Second fast path: ... else if (!clear_int) return true;
3292 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3293 result_rgn->init_req(no_clear_result_path, false_arg);
3294 result_val->init_req(no_clear_result_path, intcon(1));
3295
3296 // drop through to next case
3297 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3298 #else
3299 // To return true on Windows you must read the _interrupted field
3300 // and check the the event state i.e. take the slow path.
3301 #endif // TARGET_OS_FAMILY_windows
3302
3303 // (d) Otherwise, go to the slow path.
3304 slow_region->add_req(control());
3305 set_control( _gvn.transform(slow_region));
3306
3307 if (stopped()) {
3308 // There is no slow path.
3309 result_rgn->init_req(slow_result_path, top());
3310 result_val->init_req(slow_result_path, top());
3311 } else {
3312 // non-virtual because it is a private non-static
3313 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3314
3315 Node* slow_val = set_results_for_java_call(slow_call);
3316 // this->control() comes from set_results_for_java_call
3317
3318 Node* fast_io = slow_call->in(TypeFunc::I_O);
3319 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3320
3321 // These two phis are pre-filled with copies of of the fast IO and Memory
3322 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3323 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3324
3325 result_rgn->init_req(slow_result_path, control());
3326 result_io ->init_req(slow_result_path, i_o());
3327 result_mem->init_req(slow_result_path, reset_memory());
3328 result_val->init_req(slow_result_path, slow_val);
3329
3330 set_all_memory(_gvn.transform(result_mem));
3331 set_i_o( _gvn.transform(result_io));
3332 }
3333
3334 C->set_has_split_ifs(true); // Has chance for split-if optimization
3335 set_result(result_rgn, result_val);
3336 return true;
3337 }
3338
3339 //---------------------------load_mirror_from_klass----------------------------
3340 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3341 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3342 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3343 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3344 }
3345
3346 //-----------------------load_klass_from_mirror_common-------------------------
3347 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3348 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3349 // and branch to the given path on the region.
3350 // If never_see_null, take an uncommon trap on null, so we can optimistically
3351 // compile for the non-null case.
3352 // If the region is NULL, force never_see_null = true.
3353 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3354 bool never_see_null,
3355 RegionNode* region,
3356 int null_path,
3357 int offset) {
3358 if (region == NULL) never_see_null = true;
3359 Node* p = basic_plus_adr(mirror, offset);
3360 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3361 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3362 Node* null_ctl = top();
3363 kls = null_check_oop(kls, &null_ctl, never_see_null);
3364 if (region != NULL) {
3365 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3366 region->init_req(null_path, null_ctl);
3367 } else {
3368 assert(null_ctl == top(), "no loose ends");
3369 }
3370 return kls;
3371 }
3372
3373 //--------------------(inline_native_Class_query helpers)---------------------
3374 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3375 // Fall through if (mods & mask) == bits, take the guard otherwise.
3376 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3377 // Branch around if the given klass has the given modifier bit set.
3378 // Like generate_guard, adds a new path onto the region.
3379 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3380 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3381 Node* mask = intcon(modifier_mask);
3382 Node* bits = intcon(modifier_bits);
3383 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3384 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3385 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3386 return generate_fair_guard(bol, region);
3387 }
3388 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3389 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3390 }
3391
3392 //-------------------------inline_native_Class_query-------------------
3393 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3394 const Type* return_type = TypeInt::BOOL;
3395 Node* prim_return_value = top(); // what happens if it's a primitive class?
3396 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3397 bool expect_prim = false; // most of these guys expect to work on refs
3398
3399 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3400
3401 Node* mirror = argument(0);
3402 Node* obj = top();
3403
3404 switch (id) {
3405 case vmIntrinsics::_isInstance:
3406 // nothing is an instance of a primitive type
3407 prim_return_value = intcon(0);
3408 obj = argument(1);
3409 break;
3410 case vmIntrinsics::_getModifiers:
3411 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3412 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3413 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3414 break;
3415 case vmIntrinsics::_isInterface:
3416 prim_return_value = intcon(0);
3417 break;
3418 case vmIntrinsics::_isArray:
3419 prim_return_value = intcon(0);
3420 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3421 break;
3422 case vmIntrinsics::_isPrimitive:
3423 prim_return_value = intcon(1);
3424 expect_prim = true; // obviously
3425 break;
3426 case vmIntrinsics::_getSuperclass:
3427 prim_return_value = null();
3428 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3429 break;
3430 case vmIntrinsics::_getClassAccessFlags:
3431 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3432 return_type = TypeInt::INT; // not bool! 6297094
3433 break;
3434 default:
3435 fatal_unexpected_iid(id);
3436 break;
3437 }
3438
3439 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3440 if (mirror_con == NULL) return false; // cannot happen?
3441
3442 #ifndef PRODUCT
3443 if (C->print_intrinsics() || C->print_inlining()) {
3444 ciType* k = mirror_con->java_mirror_type();
3445 if (k) {
3446 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3447 k->print_name();
3448 tty->cr();
3449 }
3450 }
3451 #endif
3452
3453 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3454 RegionNode* region = new RegionNode(PATH_LIMIT);
3455 record_for_igvn(region);
3456 PhiNode* phi = new PhiNode(region, return_type);
3457
3458 // The mirror will never be null of Reflection.getClassAccessFlags, however
3459 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3460 // if it is. See bug 4774291.
3461
3462 // For Reflection.getClassAccessFlags(), the null check occurs in
3463 // the wrong place; see inline_unsafe_access(), above, for a similar
3464 // situation.
3465 mirror = null_check(mirror);
3466 // If mirror or obj is dead, only null-path is taken.
3467 if (stopped()) return true;
3468
3469 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3470
3471 // Now load the mirror's klass metaobject, and null-check it.
3472 // Side-effects region with the control path if the klass is null.
3473 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3474 // If kls is null, we have a primitive mirror.
3475 phi->init_req(_prim_path, prim_return_value);
3476 if (stopped()) { set_result(region, phi); return true; }
3477 bool safe_for_replace = (region->in(_prim_path) == top());
3478
3479 Node* p; // handy temp
3480 Node* null_ctl;
3481
3482 // Now that we have the non-null klass, we can perform the real query.
3483 // For constant classes, the query will constant-fold in LoadNode::Value.
3484 Node* query_value = top();
3485 switch (id) {
3486 case vmIntrinsics::_isInstance:
3487 // nothing is an instance of a primitive type
3488 query_value = gen_instanceof(obj, kls, safe_for_replace);
3489 break;
3490
3491 case vmIntrinsics::_getModifiers:
3492 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3493 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3494 break;
3495
3496 case vmIntrinsics::_isInterface:
3497 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3498 if (generate_interface_guard(kls, region) != NULL)
3499 // A guard was added. If the guard is taken, it was an interface.
3500 phi->add_req(intcon(1));
3501 // If we fall through, it's a plain class.
3502 query_value = intcon(0);
3503 break;
3504
3505 case vmIntrinsics::_isArray:
3506 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3507 if (generate_array_guard(kls, region) != NULL)
3508 // A guard was added. If the guard is taken, it was an array.
3509 phi->add_req(intcon(1));
3510 // If we fall through, it's a plain class.
3511 query_value = intcon(0);
3512 break;
3513
3514 case vmIntrinsics::_isPrimitive:
3515 query_value = intcon(0); // "normal" path produces false
3516 break;
3517
3518 case vmIntrinsics::_getSuperclass:
3519 // The rules here are somewhat unfortunate, but we can still do better
3520 // with random logic than with a JNI call.
3521 // Interfaces store null or Object as _super, but must report null.
3522 // Arrays store an intermediate super as _super, but must report Object.
3523 // Other types can report the actual _super.
3524 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3525 if (generate_interface_guard(kls, region) != NULL)
3526 // A guard was added. If the guard is taken, it was an interface.
3527 phi->add_req(null());
3528 if (generate_array_guard(kls, region) != NULL)
3529 // A guard was added. If the guard is taken, it was an array.
3530 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3531 // If we fall through, it's a plain class. Get its _super.
3532 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3533 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3534 null_ctl = top();
3535 kls = null_check_oop(kls, &null_ctl);
3536 if (null_ctl != top()) {
3537 // If the guard is taken, Object.superClass is null (both klass and mirror).
3538 region->add_req(null_ctl);
3539 phi ->add_req(null());
3540 }
3541 if (!stopped()) {
3542 query_value = load_mirror_from_klass(kls);
3543 }
3544 break;
3545
3546 case vmIntrinsics::_getClassAccessFlags:
3547 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3548 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3549 break;
3550
3551 default:
3552 fatal_unexpected_iid(id);
3553 break;
3554 }
3555
3556 // Fall-through is the normal case of a query to a real class.
3557 phi->init_req(1, query_value);
3558 region->init_req(1, control());
3559
3560 C->set_has_split_ifs(true); // Has chance for split-if optimization
3561 set_result(region, phi);
3562 return true;
3563 }
3564
3565 //-------------------------inline_Class_cast-------------------
3566 bool LibraryCallKit::inline_Class_cast() {
3567 Node* mirror = argument(0); // Class
3568 Node* obj = argument(1);
3569 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3570 if (mirror_con == NULL) {
3571 return false; // dead path (mirror->is_top()).
3572 }
3573 if (obj == NULL || obj->is_top()) {
3574 return false; // dead path
3575 }
3576 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3577
3578 // First, see if Class.cast() can be folded statically.
3579 // java_mirror_type() returns non-null for compile-time Class constants.
3580 ciType* tm = mirror_con->java_mirror_type();
3581 if (tm != NULL && tm->is_klass() &&
3582 tp != NULL && tp->klass() != NULL) {
3583 if (!tp->klass()->is_loaded()) {
3584 // Don't use intrinsic when class is not loaded.
3585 return false;
3586 } else {
3587 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3588 if (static_res == Compile::SSC_always_true) {
3589 // isInstance() is true - fold the code.
3590 set_result(obj);
3591 return true;
3592 } else if (static_res == Compile::SSC_always_false) {
3593 // Don't use intrinsic, have to throw ClassCastException.
3594 // If the reference is null, the non-intrinsic bytecode will
3595 // be optimized appropriately.
3596 return false;
3597 }
3598 }
3599 }
3600
3601 // Bailout intrinsic and do normal inlining if exception path is frequent.
3602 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3603 return false;
3604 }
3605
3606 // Generate dynamic checks.
3607 // Class.cast() is java implementation of _checkcast bytecode.
3608 // Do checkcast (Parse::do_checkcast()) optimizations here.
3609
3610 mirror = null_check(mirror);
3611 // If mirror is dead, only null-path is taken.
3612 if (stopped()) {
3613 return true;
3614 }
3615
3616 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3617 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3618 RegionNode* region = new RegionNode(PATH_LIMIT);
3619 record_for_igvn(region);
3620
3621 // Now load the mirror's klass metaobject, and null-check it.
3622 // If kls is null, we have a primitive mirror and
3623 // nothing is an instance of a primitive type.
3624 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3625
3626 Node* res = top();
3627 if (!stopped()) {
3628 Node* bad_type_ctrl = top();
3629 // Do checkcast optimizations.
3630 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3631 region->init_req(_bad_type_path, bad_type_ctrl);
3632 }
3633 if (region->in(_prim_path) != top() ||
3634 region->in(_bad_type_path) != top()) {
3635 // Let Interpreter throw ClassCastException.
3636 PreserveJVMState pjvms(this);
3637 set_control(_gvn.transform(region));
3638 uncommon_trap(Deoptimization::Reason_intrinsic,
3639 Deoptimization::Action_maybe_recompile);
3640 }
3641 if (!stopped()) {
3642 set_result(res);
3643 }
3644 return true;
3645 }
3646
3647
3648 //--------------------------inline_native_subtype_check------------------------
3649 // This intrinsic takes the JNI calls out of the heart of
3650 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3651 bool LibraryCallKit::inline_native_subtype_check() {
3652 // Pull both arguments off the stack.
3653 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3654 args[0] = argument(0);
3655 args[1] = argument(1);
3656 Node* klasses[2]; // corresponding Klasses: superk, subk
3657 klasses[0] = klasses[1] = top();
3658
3659 enum {
3660 // A full decision tree on {superc is prim, subc is prim}:
3661 _prim_0_path = 1, // {P,N} => false
3662 // {P,P} & superc!=subc => false
3663 _prim_same_path, // {P,P} & superc==subc => true
3664 _prim_1_path, // {N,P} => false
3665 _ref_subtype_path, // {N,N} & subtype check wins => true
3666 _both_ref_path, // {N,N} & subtype check loses => false
3667 PATH_LIMIT
3668 };
3669
3670 RegionNode* region = new RegionNode(PATH_LIMIT);
3671 Node* phi = new PhiNode(region, TypeInt::BOOL);
3672 record_for_igvn(region);
3673
3674 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3675 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3676 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3677
3678 // First null-check both mirrors and load each mirror's klass metaobject.
3679 int which_arg;
3680 for (which_arg = 0; which_arg <= 1; which_arg++) {
3681 Node* arg = args[which_arg];
3682 arg = null_check(arg);
3683 if (stopped()) break;
3684 args[which_arg] = arg;
3685
3686 Node* p = basic_plus_adr(arg, class_klass_offset);
3687 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3688 klasses[which_arg] = _gvn.transform(kls);
3689 }
3690
3691 // Having loaded both klasses, test each for null.
3692 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3693 for (which_arg = 0; which_arg <= 1; which_arg++) {
3694 Node* kls = klasses[which_arg];
3695 Node* null_ctl = top();
3696 kls = null_check_oop(kls, &null_ctl, never_see_null);
3697 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3698 region->init_req(prim_path, null_ctl);
3699 if (stopped()) break;
3700 klasses[which_arg] = kls;
3701 }
3702
3703 if (!stopped()) {
3704 // now we have two reference types, in klasses[0..1]
3705 Node* subk = klasses[1]; // the argument to isAssignableFrom
3706 Node* superk = klasses[0]; // the receiver
3707 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3708 // now we have a successful reference subtype check
3709 region->set_req(_ref_subtype_path, control());
3710 }
3711
3712 // If both operands are primitive (both klasses null), then
3713 // we must return true when they are identical primitives.
3714 // It is convenient to test this after the first null klass check.
3715 set_control(region->in(_prim_0_path)); // go back to first null check
3716 if (!stopped()) {
3717 // Since superc is primitive, make a guard for the superc==subc case.
3718 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3719 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3720 generate_guard(bol_eq, region, PROB_FAIR);
3721 if (region->req() == PATH_LIMIT+1) {
3722 // A guard was added. If the added guard is taken, superc==subc.
3723 region->swap_edges(PATH_LIMIT, _prim_same_path);
3724 region->del_req(PATH_LIMIT);
3725 }
3726 region->set_req(_prim_0_path, control()); // Not equal after all.
3727 }
3728
3729 // these are the only paths that produce 'true':
3730 phi->set_req(_prim_same_path, intcon(1));
3731 phi->set_req(_ref_subtype_path, intcon(1));
3732
3733 // pull together the cases:
3734 assert(region->req() == PATH_LIMIT, "sane region");
3735 for (uint i = 1; i < region->req(); i++) {
3736 Node* ctl = region->in(i);
3737 if (ctl == NULL || ctl == top()) {
3738 region->set_req(i, top());
3739 phi ->set_req(i, top());
3740 } else if (phi->in(i) == NULL) {
3741 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3742 }
3743 }
3744
3745 set_control(_gvn.transform(region));
3746 set_result(_gvn.transform(phi));
3747 return true;
3748 }
3749
3750 //---------------------generate_array_guard_common------------------------
3751 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3752 bool obj_array, bool not_array) {
3753
3754 if (stopped()) {
3755 return NULL;
3756 }
3757
3758 // If obj_array/non_array==false/false:
3759 // Branch around if the given klass is in fact an array (either obj or prim).
3760 // If obj_array/non_array==false/true:
3761 // Branch around if the given klass is not an array klass of any kind.
3762 // If obj_array/non_array==true/true:
3763 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3764 // If obj_array/non_array==true/false:
3765 // Branch around if the kls is an oop array (Object[] or subtype)
3766 //
3767 // Like generate_guard, adds a new path onto the region.
3768 jint layout_con = 0;
3769 Node* layout_val = get_layout_helper(kls, layout_con);
3770 if (layout_val == NULL) {
3771 bool query = (obj_array
3772 ? Klass::layout_helper_is_objArray(layout_con)
3773 : Klass::layout_helper_is_array(layout_con));
3774 if (query == not_array) {
3775 return NULL; // never a branch
3776 } else { // always a branch
3777 Node* always_branch = control();
3778 if (region != NULL)
3779 region->add_req(always_branch);
3780 set_control(top());
3781 return always_branch;
3782 }
3783 }
3784 // Now test the correct condition.
3785 jint nval = (obj_array
3786 ? ((jint)Klass::_lh_array_tag_type_value
3787 << Klass::_lh_array_tag_shift)
3788 : Klass::_lh_neutral_value);
3789 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3790 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3791 // invert the test if we are looking for a non-array
3792 if (not_array) btest = BoolTest(btest).negate();
3793 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3794 return generate_fair_guard(bol, region);
3795 }
3796
3797
3798 //-----------------------inline_native_newArray--------------------------
3799 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3800 bool LibraryCallKit::inline_native_newArray() {
3801 Node* mirror = argument(0);
3802 Node* count_val = argument(1);
3803
3804 mirror = null_check(mirror);
3805 // If mirror or obj is dead, only null-path is taken.
3806 if (stopped()) return true;
3807
3808 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3809 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3810 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3811 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3812 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3813
3814 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3815 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3816 result_reg, _slow_path);
3817 Node* normal_ctl = control();
3818 Node* no_array_ctl = result_reg->in(_slow_path);
3819
3820 // Generate code for the slow case. We make a call to newArray().
3821 set_control(no_array_ctl);
3822 if (!stopped()) {
3823 // Either the input type is void.class, or else the
3824 // array klass has not yet been cached. Either the
3825 // ensuing call will throw an exception, or else it
3826 // will cache the array klass for next time.
3827 PreserveJVMState pjvms(this);
3828 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3829 Node* slow_result = set_results_for_java_call(slow_call);
3830 // this->control() comes from set_results_for_java_call
3831 result_reg->set_req(_slow_path, control());
3832 result_val->set_req(_slow_path, slow_result);
3833 result_io ->set_req(_slow_path, i_o());
3834 result_mem->set_req(_slow_path, reset_memory());
3835 }
3836
3837 set_control(normal_ctl);
3838 if (!stopped()) {
3839 // Normal case: The array type has been cached in the java.lang.Class.
3840 // The following call works fine even if the array type is polymorphic.
3841 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3842 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3843 result_reg->init_req(_normal_path, control());
3844 result_val->init_req(_normal_path, obj);
3845 result_io ->init_req(_normal_path, i_o());
3846 result_mem->init_req(_normal_path, reset_memory());
3847 }
3848
3849 // Return the combined state.
3850 set_i_o( _gvn.transform(result_io) );
3851 set_all_memory( _gvn.transform(result_mem));
3852
3853 C->set_has_split_ifs(true); // Has chance for split-if optimization
3854 set_result(result_reg, result_val);
3855 return true;
3856 }
3857
3858 //----------------------inline_native_getLength--------------------------
3859 // public static native int java.lang.reflect.Array.getLength(Object array);
3860 bool LibraryCallKit::inline_native_getLength() {
3861 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3862
3863 Node* array = null_check(argument(0));
3864 // If array is dead, only null-path is taken.
3865 if (stopped()) return true;
3866
3867 // Deoptimize if it is a non-array.
3868 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3869
3870 if (non_array != NULL) {
3871 PreserveJVMState pjvms(this);
3872 set_control(non_array);
3873 uncommon_trap(Deoptimization::Reason_intrinsic,
3874 Deoptimization::Action_maybe_recompile);
3875 }
3876
3877 // If control is dead, only non-array-path is taken.
3878 if (stopped()) return true;
3879
3880 // The works fine even if the array type is polymorphic.
3881 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3882 Node* result = load_array_length(array);
3883
3884 C->set_has_split_ifs(true); // Has chance for split-if optimization
3885 set_result(result);
3886 return true;
3887 }
3888
3889 //------------------------inline_array_copyOf----------------------------
3890 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3891 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3892 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3893 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3894
3895 // Get the arguments.
3896 Node* original = argument(0);
3897 Node* start = is_copyOfRange? argument(1): intcon(0);
3898 Node* end = is_copyOfRange? argument(2): argument(1);
3899 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3900
3901 Node* newcopy;
3902
3903 // Set the original stack and the reexecute bit for the interpreter to reexecute
3904 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3905 { PreserveReexecuteState preexecs(this);
3906 jvms()->set_should_reexecute(true);
3907
3908 array_type_mirror = null_check(array_type_mirror);
3909 original = null_check(original);
3910
3911 // Check if a null path was taken unconditionally.
3912 if (stopped()) return true;
3913
3914 Node* orig_length = load_array_length(original);
3915
3916 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3917 klass_node = null_check(klass_node);
3918
3919 RegionNode* bailout = new RegionNode(1);
3920 record_for_igvn(bailout);
3921
3922 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3923 // Bail out if that is so.
3924 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3925 if (not_objArray != NULL) {
3926 // Improve the klass node's type from the new optimistic assumption:
3927 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3928 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3929 Node* cast = new CastPPNode(klass_node, akls);
3930 cast->init_req(0, control());
3931 klass_node = _gvn.transform(cast);
3932 }
3933
3934 // Bail out if either start or end is negative.
3935 generate_negative_guard(start, bailout, &start);
3936 generate_negative_guard(end, bailout, &end);
3937
3938 Node* length = end;
3939 if (_gvn.type(start) != TypeInt::ZERO) {
3940 length = _gvn.transform(new SubINode(end, start));
3941 }
3942
3943 // Bail out if length is negative.
3944 // Without this the new_array would throw
3945 // NegativeArraySizeException but IllegalArgumentException is what
3946 // should be thrown
3947 generate_negative_guard(length, bailout, &length);
3948
3949 if (bailout->req() > 1) {
3950 PreserveJVMState pjvms(this);
3951 set_control(_gvn.transform(bailout));
3952 uncommon_trap(Deoptimization::Reason_intrinsic,
3953 Deoptimization::Action_maybe_recompile);
3954 }
3955
3956 if (!stopped()) {
3957 // How many elements will we copy from the original?
3958 // The answer is MinI(orig_length - start, length).
3959 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3960 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3961
3962 // Generate a direct call to the right arraycopy function(s).
3963 // We know the copy is disjoint but we might not know if the
3964 // oop stores need checking.
3965 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3966 // This will fail a store-check if x contains any non-nulls.
3967
3968 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3969 // loads/stores but it is legal only if we're sure the
3970 // Arrays.copyOf would succeed. So we need all input arguments
3971 // to the copyOf to be validated, including that the copy to the
3972 // new array won't trigger an ArrayStoreException. That subtype
3973 // check can be optimized if we know something on the type of
3974 // the input array from type speculation.
3975 if (_gvn.type(klass_node)->singleton()) {
3976 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3977 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3978
3979 int test = C->static_subtype_check(superk, subk);
3980 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3981 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3982 if (t_original->speculative_type() != NULL) {
3983 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3984 }
3985 }
3986 }
3987
3988 bool validated = false;
3989 // Reason_class_check rather than Reason_intrinsic because we
3990 // want to intrinsify even if this traps.
3991 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3992 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3993 klass_node);
3994
3995 if (not_subtype_ctrl != top()) {
3996 PreserveJVMState pjvms(this);
3997 set_control(not_subtype_ctrl);
3998 uncommon_trap(Deoptimization::Reason_class_check,
3999 Deoptimization::Action_make_not_entrant);
4000 assert(stopped(), "Should be stopped");
4001 }
4002 validated = true;
4003 }
4004
4005 if (!stopped()) {
4006 newcopy = new_array(klass_node, length, 0); // no arguments to push
4007
4008 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
4009 load_object_klass(original), klass_node);
4010 if (!is_copyOfRange) {
4011 ac->set_copyof(validated);
4012 } else {
4013 ac->set_copyofrange(validated);
4014 }
4015 Node* n = _gvn.transform(ac);
4016 if (n == ac) {
4017 ac->connect_outputs(this);
4018 } else {
4019 assert(validated, "shouldn't transform if all arguments not validated");
4020 set_all_memory(n);
4021 }
4022 }
4023 }
4024 } // original reexecute is set back here
4025
4026 C->set_has_split_ifs(true); // Has chance for split-if optimization
4027 if (!stopped()) {
4028 set_result(newcopy);
4029 }
4030 return true;
4031 }
4032
4033
4034 //----------------------generate_virtual_guard---------------------------
4035 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4036 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4037 RegionNode* slow_region) {
4038 ciMethod* method = callee();
4039 int vtable_index = method->vtable_index();
4040 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4041 err_msg_res("bad index %d", vtable_index));
4042 // Get the Method* out of the appropriate vtable entry.
4043 int entry_offset = (InstanceKlass::vtable_start_offset() +
4044 vtable_index*vtableEntry::size()) * wordSize +
4045 vtableEntry::method_offset_in_bytes();
4046 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4047 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4048
4049 // Compare the target method with the expected method (e.g., Object.hashCode).
4050 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4051
4052 Node* native_call = makecon(native_call_addr);
4053 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4054 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4055
4056 return generate_slow_guard(test_native, slow_region);
4057 }
4058
4059 //-----------------------generate_method_call----------------------------
4060 // Use generate_method_call to make a slow-call to the real
4061 // method if the fast path fails. An alternative would be to
4062 // use a stub like OptoRuntime::slow_arraycopy_Java.
4063 // This only works for expanding the current library call,
4064 // not another intrinsic. (E.g., don't use this for making an
4065 // arraycopy call inside of the copyOf intrinsic.)
4066 CallJavaNode*
4067 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4068 // When compiling the intrinsic method itself, do not use this technique.
4069 guarantee(callee() != C->method(), "cannot make slow-call to self");
4070
4071 ciMethod* method = callee();
4072 // ensure the JVMS we have will be correct for this call
4073 guarantee(method_id == method->intrinsic_id(), "must match");
4074
4075 const TypeFunc* tf = TypeFunc::make(method);
4076 CallJavaNode* slow_call;
4077 if (is_static) {
4078 assert(!is_virtual, "");
4079 slow_call = new CallStaticJavaNode(C, tf,
4080 SharedRuntime::get_resolve_static_call_stub(),
4081 method, bci());
4082 } else if (is_virtual) {
4083 null_check_receiver();
4084 int vtable_index = Method::invalid_vtable_index;
4085 if (UseInlineCaches) {
4086 // Suppress the vtable call
4087 } else {
4088 // hashCode and clone are not a miranda methods,
4089 // so the vtable index is fixed.
4090 // No need to use the linkResolver to get it.
4091 vtable_index = method->vtable_index();
4092 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4093 err_msg_res("bad index %d", vtable_index));
4094 }
4095 slow_call = new CallDynamicJavaNode(tf,
4096 SharedRuntime::get_resolve_virtual_call_stub(),
4097 method, vtable_index, bci());
4098 } else { // neither virtual nor static: opt_virtual
4099 null_check_receiver();
4100 slow_call = new CallStaticJavaNode(C, tf,
4101 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4102 method, bci());
4103 slow_call->set_optimized_virtual(true);
4104 }
4105 set_arguments_for_java_call(slow_call);
4106 set_edges_for_java_call(slow_call);
4107 return slow_call;
4108 }
4109
4110
4111 /**
4112 * Build special case code for calls to hashCode on an object. This call may
4113 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4114 * slightly different code.
4115 */
4116 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4117 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4118 assert(!(is_virtual && is_static), "either virtual, special, or static");
4119
4120 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4121
4122 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4123 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4124 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4125 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4126 Node* obj = NULL;
4127 if (!is_static) {
4128 // Check for hashing null object
4129 obj = null_check_receiver();
4130 if (stopped()) return true; // unconditionally null
4131 result_reg->init_req(_null_path, top());
4132 result_val->init_req(_null_path, top());
4133 } else {
4134 // Do a null check, and return zero if null.
4135 // System.identityHashCode(null) == 0
4136 obj = argument(0);
4137 Node* null_ctl = top();
4138 obj = null_check_oop(obj, &null_ctl);
4139 result_reg->init_req(_null_path, null_ctl);
4140 result_val->init_req(_null_path, _gvn.intcon(0));
4141 }
4142
4143 // Unconditionally null? Then return right away.
4144 if (stopped()) {
4145 set_control( result_reg->in(_null_path));
4146 if (!stopped())
4147 set_result(result_val->in(_null_path));
4148 return true;
4149 }
4150
4151 // We only go to the fast case code if we pass a number of guards. The
4152 // paths which do not pass are accumulated in the slow_region.
4153 RegionNode* slow_region = new RegionNode(1);
4154 record_for_igvn(slow_region);
4155
4156 // If this is a virtual call, we generate a funny guard. We pull out
4157 // the vtable entry corresponding to hashCode() from the target object.
4158 // If the target method which we are calling happens to be the native
4159 // Object hashCode() method, we pass the guard. We do not need this
4160 // guard for non-virtual calls -- the caller is known to be the native
4161 // Object hashCode().
4162 if (is_virtual) {
4163 // After null check, get the object's klass.
4164 Node* obj_klass = load_object_klass(obj);
4165 generate_virtual_guard(obj_klass, slow_region);
4166 }
4167
4168 // Get the header out of the object, use LoadMarkNode when available
4169 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4170 // The control of the load must be NULL. Otherwise, the load can move before
4171 // the null check after castPP removal.
4172 Node* no_ctrl = NULL;
4173 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4174
4175 // Test the header to see if it is unlocked.
4176 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4177 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4178 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4179 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4180 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4181
4182 generate_slow_guard(test_unlocked, slow_region);
4183
4184 // Get the hash value and check to see that it has been properly assigned.
4185 // We depend on hash_mask being at most 32 bits and avoid the use of
4186 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4187 // vm: see markOop.hpp.
4188 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4189 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4190 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4191 // This hack lets the hash bits live anywhere in the mark object now, as long
4192 // as the shift drops the relevant bits into the low 32 bits. Note that
4193 // Java spec says that HashCode is an int so there's no point in capturing
4194 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4195 hshifted_header = ConvX2I(hshifted_header);
4196 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4197
4198 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4199 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4200 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4201
4202 generate_slow_guard(test_assigned, slow_region);
4203
4204 Node* init_mem = reset_memory();
4205 // fill in the rest of the null path:
4206 result_io ->init_req(_null_path, i_o());
4207 result_mem->init_req(_null_path, init_mem);
4208
4209 result_val->init_req(_fast_path, hash_val);
4210 result_reg->init_req(_fast_path, control());
4211 result_io ->init_req(_fast_path, i_o());
4212 result_mem->init_req(_fast_path, init_mem);
4213
4214 // Generate code for the slow case. We make a call to hashCode().
4215 set_control(_gvn.transform(slow_region));
4216 if (!stopped()) {
4217 // No need for PreserveJVMState, because we're using up the present state.
4218 set_all_memory(init_mem);
4219 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4220 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4221 Node* slow_result = set_results_for_java_call(slow_call);
4222 // this->control() comes from set_results_for_java_call
4223 result_reg->init_req(_slow_path, control());
4224 result_val->init_req(_slow_path, slow_result);
4225 result_io ->set_req(_slow_path, i_o());
4226 result_mem ->set_req(_slow_path, reset_memory());
4227 }
4228
4229 // Return the combined state.
4230 set_i_o( _gvn.transform(result_io) );
4231 set_all_memory( _gvn.transform(result_mem));
4232
4233 set_result(result_reg, result_val);
4234 return true;
4235 }
4236
4237 //---------------------------inline_native_getClass----------------------------
4238 // public final native Class<?> java.lang.Object.getClass();
4239 //
4240 // Build special case code for calls to getClass on an object.
4241 bool LibraryCallKit::inline_native_getClass() {
4242 Node* obj = null_check_receiver();
4243 if (stopped()) return true;
4244 set_result(load_mirror_from_klass(load_object_klass(obj)));
4245 return true;
4246 }
4247
4248 //-----------------inline_native_Reflection_getCallerClass---------------------
4249 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4250 //
4251 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4252 //
4253 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4254 // in that it must skip particular security frames and checks for
4255 // caller sensitive methods.
4256 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4257 #ifndef PRODUCT
4258 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4259 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4260 }
4261 #endif
4262
4263 if (!jvms()->has_method()) {
4264 #ifndef PRODUCT
4265 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4266 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4267 }
4268 #endif
4269 return false;
4270 }
4271
4272 // Walk back up the JVM state to find the caller at the required
4273 // depth.
4274 JVMState* caller_jvms = jvms();
4275
4276 // Cf. JVM_GetCallerClass
4277 // NOTE: Start the loop at depth 1 because the current JVM state does
4278 // not include the Reflection.getCallerClass() frame.
4279 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4280 ciMethod* m = caller_jvms->method();
4281 switch (n) {
4282 case 0:
4283 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4284 break;
4285 case 1:
4286 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4287 if (!m->caller_sensitive()) {
4288 #ifndef PRODUCT
4289 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4290 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4291 }
4292 #endif
4293 return false; // bail-out; let JVM_GetCallerClass do the work
4294 }
4295 break;
4296 default:
4297 if (!m->is_ignored_by_security_stack_walk()) {
4298 // We have reached the desired frame; return the holder class.
4299 // Acquire method holder as java.lang.Class and push as constant.
4300 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4301 ciInstance* caller_mirror = caller_klass->java_mirror();
4302 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4303
4304 #ifndef PRODUCT
4305 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4306 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());
4307 tty->print_cr(" JVM state at this point:");
4308 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4309 ciMethod* m = jvms()->of_depth(i)->method();
4310 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4311 }
4312 }
4313 #endif
4314 return true;
4315 }
4316 break;
4317 }
4318 }
4319
4320 #ifndef PRODUCT
4321 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4322 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4323 tty->print_cr(" JVM state at this point:");
4324 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4325 ciMethod* m = jvms()->of_depth(i)->method();
4326 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4327 }
4328 }
4329 #endif
4330
4331 return false; // bail-out; let JVM_GetCallerClass do the work
4332 }
4333
4334 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4335 Node* arg = argument(0);
4336 Node* result;
4337
4338 switch (id) {
4339 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4340 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4341 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4342 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4343
4344 case vmIntrinsics::_doubleToLongBits: {
4345 // two paths (plus control) merge in a wood
4346 RegionNode *r = new RegionNode(3);
4347 Node *phi = new PhiNode(r, TypeLong::LONG);
4348
4349 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4350 // Build the boolean node
4351 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4352
4353 // Branch either way.
4354 // NaN case is less traveled, which makes all the difference.
4355 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4356 Node *opt_isnan = _gvn.transform(ifisnan);
4357 assert( opt_isnan->is_If(), "Expect an IfNode");
4358 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4359 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4360
4361 set_control(iftrue);
4362
4363 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4364 Node *slow_result = longcon(nan_bits); // return NaN
4365 phi->init_req(1, _gvn.transform( slow_result ));
4366 r->init_req(1, iftrue);
4367
4368 // Else fall through
4369 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4370 set_control(iffalse);
4371
4372 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4373 r->init_req(2, iffalse);
4374
4375 // Post merge
4376 set_control(_gvn.transform(r));
4377 record_for_igvn(r);
4378
4379 C->set_has_split_ifs(true); // Has chance for split-if optimization
4380 result = phi;
4381 assert(result->bottom_type()->isa_long(), "must be");
4382 break;
4383 }
4384
4385 case vmIntrinsics::_floatToIntBits: {
4386 // two paths (plus control) merge in a wood
4387 RegionNode *r = new RegionNode(3);
4388 Node *phi = new PhiNode(r, TypeInt::INT);
4389
4390 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4391 // Build the boolean node
4392 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4393
4394 // Branch either way.
4395 // NaN case is less traveled, which makes all the difference.
4396 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4397 Node *opt_isnan = _gvn.transform(ifisnan);
4398 assert( opt_isnan->is_If(), "Expect an IfNode");
4399 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4400 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4401
4402 set_control(iftrue);
4403
4404 static const jint nan_bits = 0x7fc00000;
4405 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4406 phi->init_req(1, _gvn.transform( slow_result ));
4407 r->init_req(1, iftrue);
4408
4409 // Else fall through
4410 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4411 set_control(iffalse);
4412
4413 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4414 r->init_req(2, iffalse);
4415
4416 // Post merge
4417 set_control(_gvn.transform(r));
4418 record_for_igvn(r);
4419
4420 C->set_has_split_ifs(true); // Has chance for split-if optimization
4421 result = phi;
4422 assert(result->bottom_type()->isa_int(), "must be");
4423 break;
4424 }
4425
4426 default:
4427 fatal_unexpected_iid(id);
4428 break;
4429 }
4430 set_result(_gvn.transform(result));
4431 return true;
4432 }
4433
4434 #ifdef _LP64
4435 #define XTOP ,top() /*additional argument*/
4436 #else //_LP64
4437 #define XTOP /*no additional argument*/
4438 #endif //_LP64
4439
4440 //----------------------inline_unsafe_copyMemory-------------------------
4441 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4442 bool LibraryCallKit::inline_unsafe_copyMemory() {
4443 if (callee()->is_static()) return false; // caller must have the capability!
4444 null_check_receiver(); // null-check receiver
4445 if (stopped()) return true;
4446
4447 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4448
4449 Node* src_ptr = argument(1); // type: oop
4450 Node* src_off = ConvL2X(argument(2)); // type: long
4451 Node* dst_ptr = argument(4); // type: oop
4452 Node* dst_off = ConvL2X(argument(5)); // type: long
4453 Node* size = ConvL2X(argument(7)); // type: long
4454
4455 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4456 "fieldOffset must be byte-scaled");
4457
4458 Node* src = make_unsafe_address(src_ptr, src_off);
4459 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4460
4461 // Conservatively insert a memory barrier on all memory slices.
4462 // Do not let writes of the copy source or destination float below the copy.
4463 insert_mem_bar(Op_MemBarCPUOrder);
4464
4465 // Call it. Note that the length argument is not scaled.
4466 make_runtime_call(RC_LEAF|RC_NO_FP,
4467 OptoRuntime::fast_arraycopy_Type(),
4468 StubRoutines::unsafe_arraycopy(),
4469 "unsafe_arraycopy",
4470 TypeRawPtr::BOTTOM,
4471 src, dst, size XTOP);
4472
4473 // Do not let reads of the copy destination float above the copy.
4474 insert_mem_bar(Op_MemBarCPUOrder);
4475
4476 return true;
4477 }
4478
4479 //------------------------clone_coping-----------------------------------
4480 // Helper function for inline_native_clone.
4481 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4482 assert(obj_size != NULL, "");
4483 Node* raw_obj = alloc_obj->in(1);
4484 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4485
4486 AllocateNode* alloc = NULL;
4487 if (ReduceBulkZeroing) {
4488 // We will be completely responsible for initializing this object -
4489 // mark Initialize node as complete.
4490 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4491 // The object was just allocated - there should be no any stores!
4492 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4493 // Mark as complete_with_arraycopy so that on AllocateNode
4494 // expansion, we know this AllocateNode is initialized by an array
4495 // copy and a StoreStore barrier exists after the array copy.
4496 alloc->initialization()->set_complete_with_arraycopy();
4497 }
4498
4499 // Copy the fastest available way.
4500 // TODO: generate fields copies for small objects instead.
4501 Node* src = obj;
4502 Node* dest = alloc_obj;
4503 Node* size = _gvn.transform(obj_size);
4504
4505 // Exclude the header but include array length to copy by 8 bytes words.
4506 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4507 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4508 instanceOopDesc::base_offset_in_bytes();
4509 // base_off:
4510 // 8 - 32-bit VM
4511 // 12 - 64-bit VM, compressed klass
4512 // 16 - 64-bit VM, normal klass
4513 if (base_off % BytesPerLong != 0) {
4514 assert(UseCompressedClassPointers, "");
4515 if (is_array) {
4516 // Exclude length to copy by 8 bytes words.
4517 base_off += sizeof(int);
4518 } else {
4519 // Include klass to copy by 8 bytes words.
4520 base_off = instanceOopDesc::klass_offset_in_bytes();
4521 }
4522 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4523 }
4524 src = basic_plus_adr(src, base_off);
4525 dest = basic_plus_adr(dest, base_off);
4526
4527 // Compute the length also, if needed:
4528 Node* countx = size;
4529 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4530 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4531
4532 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4533
4534 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4535 ac->set_clonebasic();
4536 Node* n = _gvn.transform(ac);
4537 if (n == ac) {
4538 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4539 } else {
4540 set_all_memory(n);
4541 }
4542
4543 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4544 if (card_mark) {
4545 assert(!is_array, "");
4546 // Put in store barrier for any and all oops we are sticking
4547 // into this object. (We could avoid this if we could prove
4548 // that the object type contains no oop fields at all.)
4549 Node* no_particular_value = NULL;
4550 Node* no_particular_field = NULL;
4551 int raw_adr_idx = Compile::AliasIdxRaw;
4552 post_barrier(control(),
4553 memory(raw_adr_type),
4554 alloc_obj,
4555 no_particular_field,
4556 raw_adr_idx,
4557 no_particular_value,
4558 T_OBJECT,
4559 false);
4560 }
4561
4562 // Do not let reads from the cloned object float above the arraycopy.
4563 if (alloc != NULL) {
4564 // Do not let stores that initialize this object be reordered with
4565 // a subsequent store that would make this object accessible by
4566 // other threads.
4567 // Record what AllocateNode this StoreStore protects so that
4568 // escape analysis can go from the MemBarStoreStoreNode to the
4569 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4570 // based on the escape status of the AllocateNode.
4571 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4572 } else {
4573 insert_mem_bar(Op_MemBarCPUOrder);
4574 }
4575 }
4576
4577 //------------------------inline_native_clone----------------------------
4578 // protected native Object java.lang.Object.clone();
4579 //
4580 // Here are the simple edge cases:
4581 // null receiver => normal trap
4582 // virtual and clone was overridden => slow path to out-of-line clone
4583 // not cloneable or finalizer => slow path to out-of-line Object.clone
4584 //
4585 // The general case has two steps, allocation and copying.
4586 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4587 //
4588 // Copying also has two cases, oop arrays and everything else.
4589 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4590 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4591 //
4592 // These steps fold up nicely if and when the cloned object's klass
4593 // can be sharply typed as an object array, a type array, or an instance.
4594 //
4595 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4596 PhiNode* result_val;
4597
4598 // Set the reexecute bit for the interpreter to reexecute
4599 // the bytecode that invokes Object.clone if deoptimization happens.
4600 { PreserveReexecuteState preexecs(this);
4601 jvms()->set_should_reexecute(true);
4602
4603 Node* obj = null_check_receiver();
4604 if (stopped()) return true;
4605
4606 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4607
4608 // If we are going to clone an instance, we need its exact type to
4609 // know the number and types of fields to convert the clone to
4610 // loads/stores. Maybe a speculative type can help us.
4611 if (!obj_type->klass_is_exact() &&
4612 obj_type->speculative_type() != NULL &&
4613 obj_type->speculative_type()->is_instance_klass()) {
4614 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4615 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4616 !spec_ik->has_injected_fields()) {
4617 ciKlass* k = obj_type->klass();
4618 if (!k->is_instance_klass() ||
4619 k->as_instance_klass()->is_interface() ||
4620 k->as_instance_klass()->has_subklass()) {
4621 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4622 }
4623 }
4624 }
4625
4626 Node* obj_klass = load_object_klass(obj);
4627 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4628 const TypeOopPtr* toop = ((tklass != NULL)
4629 ? tklass->as_instance_type()
4630 : TypeInstPtr::NOTNULL);
4631
4632 // Conservatively insert a memory barrier on all memory slices.
4633 // Do not let writes into the original float below the clone.
4634 insert_mem_bar(Op_MemBarCPUOrder);
4635
4636 // paths into result_reg:
4637 enum {
4638 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4639 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4640 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4641 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4642 PATH_LIMIT
4643 };
4644 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4645 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4646 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4647 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4648 record_for_igvn(result_reg);
4649
4650 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4651 int raw_adr_idx = Compile::AliasIdxRaw;
4652
4653 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4654 if (array_ctl != NULL) {
4655 // It's an array.
4656 PreserveJVMState pjvms(this);
4657 set_control(array_ctl);
4658 Node* obj_length = load_array_length(obj);
4659 Node* obj_size = NULL;
4660 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4661
4662 if (!use_ReduceInitialCardMarks()) {
4663 // If it is an oop array, it requires very special treatment,
4664 // because card marking is required on each card of the array.
4665 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4666 if (is_obja != NULL) {
4667 PreserveJVMState pjvms2(this);
4668 set_control(is_obja);
4669 // Generate a direct call to the right arraycopy function(s).
4670 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4671 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4672 ac->set_cloneoop();
4673 Node* n = _gvn.transform(ac);
4674 assert(n == ac, "cannot disappear");
4675 ac->connect_outputs(this);
4676
4677 result_reg->init_req(_objArray_path, control());
4678 result_val->init_req(_objArray_path, alloc_obj);
4679 result_i_o ->set_req(_objArray_path, i_o());
4680 result_mem ->set_req(_objArray_path, reset_memory());
4681 }
4682 }
4683 // Otherwise, there are no card marks to worry about.
4684 // (We can dispense with card marks if we know the allocation
4685 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4686 // causes the non-eden paths to take compensating steps to
4687 // simulate a fresh allocation, so that no further
4688 // card marks are required in compiled code to initialize
4689 // the object.)
4690
4691 if (!stopped()) {
4692 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4693
4694 // Present the results of the copy.
4695 result_reg->init_req(_array_path, control());
4696 result_val->init_req(_array_path, alloc_obj);
4697 result_i_o ->set_req(_array_path, i_o());
4698 result_mem ->set_req(_array_path, reset_memory());
4699 }
4700 }
4701
4702 // We only go to the instance fast case code if we pass a number of guards.
4703 // The paths which do not pass are accumulated in the slow_region.
4704 RegionNode* slow_region = new RegionNode(1);
4705 record_for_igvn(slow_region);
4706 if (!stopped()) {
4707 // It's an instance (we did array above). Make the slow-path tests.
4708 // If this is a virtual call, we generate a funny guard. We grab
4709 // the vtable entry corresponding to clone() from the target object.
4710 // If the target method which we are calling happens to be the
4711 // Object clone() method, we pass the guard. We do not need this
4712 // guard for non-virtual calls; the caller is known to be the native
4713 // Object clone().
4714 if (is_virtual) {
4715 generate_virtual_guard(obj_klass, slow_region);
4716 }
4717
4718 // The object must be cloneable and must not have a finalizer.
4719 // Both of these conditions may be checked in a single test.
4720 // We could optimize the cloneable test further, but we don't care.
4721 generate_access_flags_guard(obj_klass,
4722 // Test both conditions:
4723 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4724 // Must be cloneable but not finalizer:
4725 JVM_ACC_IS_CLONEABLE,
4726 slow_region);
4727 }
4728
4729 if (!stopped()) {
4730 // It's an instance, and it passed the slow-path tests.
4731 PreserveJVMState pjvms(this);
4732 Node* obj_size = NULL;
4733 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4734 // is reexecuted if deoptimization occurs and there could be problems when merging
4735 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4736 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4737
4738 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4739
4740 // Present the results of the slow call.
4741 result_reg->init_req(_instance_path, control());
4742 result_val->init_req(_instance_path, alloc_obj);
4743 result_i_o ->set_req(_instance_path, i_o());
4744 result_mem ->set_req(_instance_path, reset_memory());
4745 }
4746
4747 // Generate code for the slow case. We make a call to clone().
4748 set_control(_gvn.transform(slow_region));
4749 if (!stopped()) {
4750 PreserveJVMState pjvms(this);
4751 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4752 Node* slow_result = set_results_for_java_call(slow_call);
4753 // this->control() comes from set_results_for_java_call
4754 result_reg->init_req(_slow_path, control());
4755 result_val->init_req(_slow_path, slow_result);
4756 result_i_o ->set_req(_slow_path, i_o());
4757 result_mem ->set_req(_slow_path, reset_memory());
4758 }
4759
4760 // Return the combined state.
4761 set_control( _gvn.transform(result_reg));
4762 set_i_o( _gvn.transform(result_i_o));
4763 set_all_memory( _gvn.transform(result_mem));
4764 } // original reexecute is set back here
4765
4766 set_result(_gvn.transform(result_val));
4767 return true;
4768 }
4769
4770 // If we have a tighly coupled allocation, the arraycopy may take care
4771 // of the array initialization. If one of the guards we insert between
4772 // the allocation and the arraycopy causes a deoptimization, an
4773 // unitialized array will escape the compiled method. To prevent that
4774 // we set the JVM state for uncommon traps between the allocation and
4775 // the arraycopy to the state before the allocation so, in case of
4776 // deoptimization, we'll reexecute the allocation and the
4777 // initialization.
4778 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4779 if (alloc != NULL) {
4780 ciMethod* trap_method = alloc->jvms()->method();
4781 int trap_bci = alloc->jvms()->bci();
4782
4783 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4784 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4785 // Make sure there's no store between the allocation and the
4786 // arraycopy otherwise visible side effects could be rexecuted
4787 // in case of deoptimization and cause incorrect execution.
4788 bool no_interfering_store = true;
4789 Node* mem = alloc->in(TypeFunc::Memory);
4790 if (mem->is_MergeMem()) {
4791 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4792 Node* n = mms.memory();
4793 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4794 assert(n->is_Store(), "what else?");
4795 no_interfering_store = false;
4796 break;
4797 }
4798 }
4799 } else {
4800 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4801 Node* n = mms.memory();
4802 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4803 assert(n->is_Store(), "what else?");
4804 no_interfering_store = false;
4805 break;
4806 }
4807 }
4808 }
4809
4810 if (no_interfering_store) {
4811 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4812 uint size = alloc->req();
4813 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4814 old_jvms->set_map(sfpt);
4815 for (uint i = 0; i < size; i++) {
4816 sfpt->init_req(i, alloc->in(i));
4817 }
4818 // re-push array length for deoptimization
4819 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4820 old_jvms->set_sp(old_jvms->sp()+1);
4821 old_jvms->set_monoff(old_jvms->monoff()+1);
4822 old_jvms->set_scloff(old_jvms->scloff()+1);
4823 old_jvms->set_endoff(old_jvms->endoff()+1);
4824 old_jvms->set_should_reexecute(true);
4825
4826 sfpt->set_i_o(map()->i_o());
4827 sfpt->set_memory(map()->memory());
4828 sfpt->set_control(map()->control());
4829
4830 JVMState* saved_jvms = jvms();
4831 saved_reexecute_sp = _reexecute_sp;
4832
4833 set_jvms(sfpt->jvms());
4834 _reexecute_sp = jvms()->sp();
4835
4836 return saved_jvms;
4837 }
4838 }
4839 }
4840 return NULL;
4841 }
4842
4843 // In case of a deoptimization, we restart execution at the
4844 // allocation, allocating a new array. We would leave an uninitialized
4845 // array in the heap that GCs wouldn't expect. Move the allocation
4846 // after the traps so we don't allocate the array if we
4847 // deoptimize. This is possible because tightly_coupled_allocation()
4848 // guarantees there's no observer of the allocated array at this point
4849 // and the control flow is simple enough.
4850 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4851 if (saved_jvms != NULL && !stopped()) {
4852 assert(alloc != NULL, "only with a tightly coupled allocation");
4853 // restore JVM state to the state at the arraycopy
4854 saved_jvms->map()->set_control(map()->control());
4855 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4856 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4857 // If we've improved the types of some nodes (null check) while
4858 // emitting the guards, propagate them to the current state
4859 map()->replaced_nodes().apply(saved_jvms->map());
4860 set_jvms(saved_jvms);
4861 _reexecute_sp = saved_reexecute_sp;
4862
4863 // Remove the allocation from above the guards
4864 CallProjections callprojs;
4865 alloc->extract_projections(&callprojs, true);
4866 InitializeNode* init = alloc->initialization();
4867 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4868 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4869 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4870 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4871
4872 // move the allocation here (after the guards)
4873 _gvn.hash_delete(alloc);
4874 alloc->set_req(TypeFunc::Control, control());
4875 alloc->set_req(TypeFunc::I_O, i_o());
4876 Node *mem = reset_memory();
4877 set_all_memory(mem);
4878 alloc->set_req(TypeFunc::Memory, mem);
4879 set_control(init->proj_out(TypeFunc::Control));
4880 set_i_o(callprojs.fallthrough_ioproj);
4881
4882 // Update memory as done in GraphKit::set_output_for_allocation()
4883 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4884 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4885 if (ary_type->isa_aryptr() && length_type != NULL) {
4886 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4887 }
4888 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4889 int elemidx = C->get_alias_index(telemref);
4890 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4891 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4892
4893 Node* allocx = _gvn.transform(alloc);
4894 assert(allocx == alloc, "where has the allocation gone?");
4895 assert(dest->is_CheckCastPP(), "not an allocation result?");
4896
4897 _gvn.hash_delete(dest);
4898 dest->set_req(0, control());
4899 Node* destx = _gvn.transform(dest);
4900 assert(destx == dest, "where has the allocation result gone?");
4901 }
4902 }
4903
4904
4905 //------------------------------inline_arraycopy-----------------------
4906 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4907 // Object dest, int destPos,
4908 // int length);
4909 bool LibraryCallKit::inline_arraycopy() {
4910 // Get the arguments.
4911 Node* src = argument(0); // type: oop
4912 Node* src_offset = argument(1); // type: int
4913 Node* dest = argument(2); // type: oop
4914 Node* dest_offset = argument(3); // type: int
4915 Node* length = argument(4); // type: int
4916
4917
4918 // Check for allocation before we add nodes that would confuse
4919 // tightly_coupled_allocation()
4920 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4921
4922 int saved_reexecute_sp = -1;
4923 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4924 // See arraycopy_restore_alloc_state() comment
4925 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4926 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4927 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4928 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4929
4930 // The following tests must be performed
4931 // (1) src and dest are arrays.
4932 // (2) src and dest arrays must have elements of the same BasicType
4933 // (3) src and dest must not be null.
4934 // (4) src_offset must not be negative.
4935 // (5) dest_offset must not be negative.
4936 // (6) length must not be negative.
4937 // (7) src_offset + length must not exceed length of src.
4938 // (8) dest_offset + length must not exceed length of dest.
4939 // (9) each element of an oop array must be assignable
4940
4941 // (3) src and dest must not be null.
4942 // always do this here because we need the JVM state for uncommon traps
4943 Node* null_ctl = top();
4944 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4945 assert(null_ctl->is_top(), "no null control here");
4946 dest = null_check(dest, T_ARRAY);
4947
4948 if (!can_emit_guards) {
4949 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4950 // guards but the arraycopy node could still take advantage of a
4951 // tightly allocated allocation. tightly_coupled_allocation() is
4952 // called again to make sure it takes the null check above into
4953 // account: the null check is mandatory and if it caused an
4954 // uncommon trap to be emitted then the allocation can't be
4955 // considered tightly coupled in this context.
4956 alloc = tightly_coupled_allocation(dest, NULL);
4957 }
4958
4959 bool validated = false;
4960
4961 const Type* src_type = _gvn.type(src);
4962 const Type* dest_type = _gvn.type(dest);
4963 const TypeAryPtr* top_src = src_type->isa_aryptr();
4964 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4965
4966 // Do we have the type of src?
4967 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4968 // Do we have the type of dest?
4969 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4970 // Is the type for src from speculation?
4971 bool src_spec = false;
4972 // Is the type for dest from speculation?
4973 bool dest_spec = false;
4974
4975 if ((!has_src || !has_dest) && can_emit_guards) {
4976 // We don't have sufficient type information, let's see if
4977 // speculative types can help. We need to have types for both src
4978 // and dest so that it pays off.
4979
4980 // Do we already have or could we have type information for src
4981 bool could_have_src = has_src;
4982 // Do we already have or could we have type information for dest
4983 bool could_have_dest = has_dest;
4984
4985 ciKlass* src_k = NULL;
4986 if (!has_src) {
4987 src_k = src_type->speculative_type_not_null();
4988 if (src_k != NULL && src_k->is_array_klass()) {
4989 could_have_src = true;
4990 }
4991 }
4992
4993 ciKlass* dest_k = NULL;
4994 if (!has_dest) {
4995 dest_k = dest_type->speculative_type_not_null();
4996 if (dest_k != NULL && dest_k->is_array_klass()) {
4997 could_have_dest = true;
4998 }
4999 }
5000
5001 if (could_have_src && could_have_dest) {
5002 // This is going to pay off so emit the required guards
5003 if (!has_src) {
5004 src = maybe_cast_profiled_obj(src, src_k, true);
5005 src_type = _gvn.type(src);
5006 top_src = src_type->isa_aryptr();
5007 has_src = (top_src != NULL && top_src->klass() != NULL);
5008 src_spec = true;
5009 }
5010 if (!has_dest) {
5011 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5012 dest_type = _gvn.type(dest);
5013 top_dest = dest_type->isa_aryptr();
5014 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5015 dest_spec = true;
5016 }
5017 }
5018 }
5019
5020 if (has_src && has_dest && can_emit_guards) {
5021 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5022 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5023 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5024 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5025
5026 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5027 // If both arrays are object arrays then having the exact types
5028 // for both will remove the need for a subtype check at runtime
5029 // before the call and may make it possible to pick a faster copy
5030 // routine (without a subtype check on every element)
5031 // Do we have the exact type of src?
5032 bool could_have_src = src_spec;
5033 // Do we have the exact type of dest?
5034 bool could_have_dest = dest_spec;
5035 ciKlass* src_k = top_src->klass();
5036 ciKlass* dest_k = top_dest->klass();
5037 if (!src_spec) {
5038 src_k = src_type->speculative_type_not_null();
5039 if (src_k != NULL && src_k->is_array_klass()) {
5040 could_have_src = true;
5041 }
5042 }
5043 if (!dest_spec) {
5044 dest_k = dest_type->speculative_type_not_null();
5045 if (dest_k != NULL && dest_k->is_array_klass()) {
5046 could_have_dest = true;
5047 }
5048 }
5049 if (could_have_src && could_have_dest) {
5050 // If we can have both exact types, emit the missing guards
5051 if (could_have_src && !src_spec) {
5052 src = maybe_cast_profiled_obj(src, src_k, true);
5053 }
5054 if (could_have_dest && !dest_spec) {
5055 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5056 }
5057 }
5058 }
5059 }
5060
5061 ciMethod* trap_method = method();
5062 int trap_bci = bci();
5063 if (saved_jvms != NULL) {
5064 trap_method = alloc->jvms()->method();
5065 trap_bci = alloc->jvms()->bci();
5066 }
5067
5068 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5069 can_emit_guards &&
5070 !src->is_top() && !dest->is_top()) {
5071 // validate arguments: enables transformation the ArrayCopyNode
5072 validated = true;
5073
5074 RegionNode* slow_region = new RegionNode(1);
5075 record_for_igvn(slow_region);
5076
5077 // (1) src and dest are arrays.
5078 generate_non_array_guard(load_object_klass(src), slow_region);
5079 generate_non_array_guard(load_object_klass(dest), slow_region);
5080
5081 // (2) src and dest arrays must have elements of the same BasicType
5082 // done at macro expansion or at Ideal transformation time
5083
5084 // (4) src_offset must not be negative.
5085 generate_negative_guard(src_offset, slow_region);
5086
5087 // (5) dest_offset must not be negative.
5088 generate_negative_guard(dest_offset, slow_region);
5089
5090 // (7) src_offset + length must not exceed length of src.
5091 generate_limit_guard(src_offset, length,
5092 load_array_length(src),
5093 slow_region);
5094
5095 // (8) dest_offset + length must not exceed length of dest.
5096 generate_limit_guard(dest_offset, length,
5097 load_array_length(dest),
5098 slow_region);
5099
5100 // (9) each element of an oop array must be assignable
5101 Node* src_klass = load_object_klass(src);
5102 Node* dest_klass = load_object_klass(dest);
5103 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5104
5105 if (not_subtype_ctrl != top()) {
5106 PreserveJVMState pjvms(this);
5107 set_control(not_subtype_ctrl);
5108 uncommon_trap(Deoptimization::Reason_intrinsic,
5109 Deoptimization::Action_make_not_entrant);
5110 assert(stopped(), "Should be stopped");
5111 }
5112 {
5113 PreserveJVMState pjvms(this);
5114 set_control(_gvn.transform(slow_region));
5115 uncommon_trap(Deoptimization::Reason_intrinsic,
5116 Deoptimization::Action_make_not_entrant);
5117 assert(stopped(), "Should be stopped");
5118 }
5119 }
5120
5121 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5122
5123 if (stopped()) {
5124 return true;
5125 }
5126
5127 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5128 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5129 // so the compiler has a chance to eliminate them: during macro expansion,
5130 // we have to set their control (CastPP nodes are eliminated).
5131 load_object_klass(src), load_object_klass(dest),
5132 load_array_length(src), load_array_length(dest));
5133
5134 ac->set_arraycopy(validated);
5135
5136 Node* n = _gvn.transform(ac);
5137 if (n == ac) {
5138 ac->connect_outputs(this);
5139 } else {
5140 assert(validated, "shouldn't transform if all arguments not validated");
5141 set_all_memory(n);
5142 }
5143
5144 return true;
5145 }
5146
5147
5148 // Helper function which determines if an arraycopy immediately follows
5149 // an allocation, with no intervening tests or other escapes for the object.
5150 AllocateArrayNode*
5151 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5152 RegionNode* slow_region) {
5153 if (stopped()) return NULL; // no fast path
5154 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5155
5156 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5157 if (alloc == NULL) return NULL;
5158
5159 Node* rawmem = memory(Compile::AliasIdxRaw);
5160 // Is the allocation's memory state untouched?
5161 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5162 // Bail out if there have been raw-memory effects since the allocation.
5163 // (Example: There might have been a call or safepoint.)
5164 return NULL;
5165 }
5166 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5167 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5168 return NULL;
5169 }
5170
5171 // There must be no unexpected observers of this allocation.
5172 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5173 Node* obs = ptr->fast_out(i);
5174 if (obs != this->map()) {
5175 return NULL;
5176 }
5177 }
5178
5179 // This arraycopy must unconditionally follow the allocation of the ptr.
5180 Node* alloc_ctl = ptr->in(0);
5181 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5182
5183 Node* ctl = control();
5184 while (ctl != alloc_ctl) {
5185 // There may be guards which feed into the slow_region.
5186 // Any other control flow means that we might not get a chance
5187 // to finish initializing the allocated object.
5188 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5189 IfNode* iff = ctl->in(0)->as_If();
5190 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5191 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5192 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5193 ctl = iff->in(0); // This test feeds the known slow_region.
5194 continue;
5195 }
5196 // One more try: Various low-level checks bottom out in
5197 // uncommon traps. If the debug-info of the trap omits
5198 // any reference to the allocation, as we've already
5199 // observed, then there can be no objection to the trap.
5200 bool found_trap = false;
5201 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5202 Node* obs = not_ctl->fast_out(j);
5203 if (obs->in(0) == not_ctl && obs->is_Call() &&
5204 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5205 found_trap = true; break;
5206 }
5207 }
5208 if (found_trap) {
5209 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5210 continue;
5211 }
5212 }
5213 return NULL;
5214 }
5215
5216 // If we get this far, we have an allocation which immediately
5217 // precedes the arraycopy, and we can take over zeroing the new object.
5218 // The arraycopy will finish the initialization, and provide
5219 // a new control state to which we will anchor the destination pointer.
5220
5221 return alloc;
5222 }
5223
5224 //-------------inline_encodeISOArray-----------------------------------
5225 // encode char[] to byte[] in ISO_8859_1
5226 bool LibraryCallKit::inline_encodeISOArray() {
5227 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5228 // no receiver since it is static method
5229 Node *src = argument(0);
5230 Node *src_offset = argument(1);
5231 Node *dst = argument(2);
5232 Node *dst_offset = argument(3);
5233 Node *length = argument(4);
5234
5235 const Type* src_type = src->Value(&_gvn);
5236 const Type* dst_type = dst->Value(&_gvn);
5237 const TypeAryPtr* top_src = src_type->isa_aryptr();
5238 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5239 if (top_src == NULL || top_src->klass() == NULL ||
5240 top_dest == NULL || top_dest->klass() == NULL) {
5241 // failed array check
5242 return false;
5243 }
5244
5245 // Figure out the size and type of the elements we will be copying.
5246 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5247 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5248 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5249 return false;
5250 }
5251 Node* src_start = array_element_address(src, src_offset, src_elem);
5252 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5253 // 'src_start' points to src array + scaled offset
5254 // 'dst_start' points to dst array + scaled offset
5255
5256 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5257 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5258 enc = _gvn.transform(enc);
5259 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5260 set_memory(res_mem, mtype);
5261 set_result(enc);
5262 return true;
5263 }
5264
5265 //-------------inline_multiplyToLen-----------------------------------
5266 bool LibraryCallKit::inline_multiplyToLen() {
5267 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5268
5269 address stubAddr = StubRoutines::multiplyToLen();
5270 if (stubAddr == NULL) {
5271 return false; // Intrinsic's stub is not implemented on this platform
5272 }
5273 const char* stubName = "multiplyToLen";
5274
5275 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5276
5277 // no receiver because it is a static method
5278 Node* x = argument(0);
5279 Node* xlen = argument(1);
5280 Node* y = argument(2);
5281 Node* ylen = argument(3);
5282 Node* z = argument(4);
5283
5284 const Type* x_type = x->Value(&_gvn);
5285 const Type* y_type = y->Value(&_gvn);
5286 const TypeAryPtr* top_x = x_type->isa_aryptr();
5287 const TypeAryPtr* top_y = y_type->isa_aryptr();
5288 if (top_x == NULL || top_x->klass() == NULL ||
5289 top_y == NULL || top_y->klass() == NULL) {
5290 // failed array check
5291 return false;
5292 }
5293
5294 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5295 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5296 if (x_elem != T_INT || y_elem != T_INT) {
5297 return false;
5298 }
5299
5300 // Set the original stack and the reexecute bit for the interpreter to reexecute
5301 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5302 // on the return from z array allocation in runtime.
5303 { PreserveReexecuteState preexecs(this);
5304 jvms()->set_should_reexecute(true);
5305
5306 Node* x_start = array_element_address(x, intcon(0), x_elem);
5307 Node* y_start = array_element_address(y, intcon(0), y_elem);
5308 // 'x_start' points to x array + scaled xlen
5309 // 'y_start' points to y array + scaled ylen
5310
5311 // Allocate the result array
5312 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5313 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5314 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5315
5316 IdealKit ideal(this);
5317
5318 #define __ ideal.
5319 Node* one = __ ConI(1);
5320 Node* zero = __ ConI(0);
5321 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5322 __ set(need_alloc, zero);
5323 __ set(z_alloc, z);
5324 __ if_then(z, BoolTest::eq, null()); {
5325 __ increment (need_alloc, one);
5326 } __ else_(); {
5327 // Update graphKit memory and control from IdealKit.
5328 sync_kit(ideal);
5329 Node* zlen_arg = load_array_length(z);
5330 // Update IdealKit memory and control from graphKit.
5331 __ sync_kit(this);
5332 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5333 __ increment (need_alloc, one);
5334 } __ end_if();
5335 } __ end_if();
5336
5337 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5338 // Update graphKit memory and control from IdealKit.
5339 sync_kit(ideal);
5340 Node * narr = new_array(klass_node, zlen, 1);
5341 // Update IdealKit memory and control from graphKit.
5342 __ sync_kit(this);
5343 __ set(z_alloc, narr);
5344 } __ end_if();
5345
5346 sync_kit(ideal);
5347 z = __ value(z_alloc);
5348 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5349 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5350 // Final sync IdealKit and GraphKit.
5351 final_sync(ideal);
5352 #undef __
5353
5354 Node* z_start = array_element_address(z, intcon(0), T_INT);
5355
5356 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5357 OptoRuntime::multiplyToLen_Type(),
5358 stubAddr, stubName, TypePtr::BOTTOM,
5359 x_start, xlen, y_start, ylen, z_start, zlen);
5360 } // original reexecute is set back here
5361
5362 C->set_has_split_ifs(true); // Has chance for split-if optimization
5363 set_result(z);
5364 return true;
5365 }
5366
5367 //-------------inline_squareToLen------------------------------------
5368 bool LibraryCallKit::inline_squareToLen() {
5369 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5370
5371 address stubAddr = StubRoutines::squareToLen();
5372 if (stubAddr == NULL) {
5373 return false; // Intrinsic's stub is not implemented on this platform
5374 }
5375 const char* stubName = "squareToLen";
5376
5377 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5378
5379 Node* x = argument(0);
5380 Node* len = argument(1);
5381 Node* z = argument(2);
5382 Node* zlen = argument(3);
5383
5384 const Type* x_type = x->Value(&_gvn);
5385 const Type* z_type = z->Value(&_gvn);
5386 const TypeAryPtr* top_x = x_type->isa_aryptr();
5387 const TypeAryPtr* top_z = z_type->isa_aryptr();
5388 if (top_x == NULL || top_x->klass() == NULL ||
5389 top_z == NULL || top_z->klass() == NULL) {
5390 // failed array check
5391 return false;
5392 }
5393
5394 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5395 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5396 if (x_elem != T_INT || z_elem != T_INT) {
5397 return false;
5398 }
5399
5400
5401 Node* x_start = array_element_address(x, intcon(0), x_elem);
5402 Node* z_start = array_element_address(z, intcon(0), z_elem);
5403
5404 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5405 OptoRuntime::squareToLen_Type(),
5406 stubAddr, stubName, TypePtr::BOTTOM,
5407 x_start, len, z_start, zlen);
5408
5409 set_result(z);
5410 return true;
5411 }
5412
5413 //-------------inline_mulAdd------------------------------------------
5414 bool LibraryCallKit::inline_mulAdd() {
5415 assert(UseMulAddIntrinsic, "not implementated on this platform");
5416
5417 address stubAddr = StubRoutines::mulAdd();
5418 if (stubAddr == NULL) {
5419 return false; // Intrinsic's stub is not implemented on this platform
5420 }
5421 const char* stubName = "mulAdd";
5422
5423 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5424
5425 Node* out = argument(0);
5426 Node* in = argument(1);
5427 Node* offset = argument(2);
5428 Node* len = argument(3);
5429 Node* k = argument(4);
5430
5431 const Type* out_type = out->Value(&_gvn);
5432 const Type* in_type = in->Value(&_gvn);
5433 const TypeAryPtr* top_out = out_type->isa_aryptr();
5434 const TypeAryPtr* top_in = in_type->isa_aryptr();
5435 if (top_out == NULL || top_out->klass() == NULL ||
5436 top_in == NULL || top_in->klass() == NULL) {
5437 // failed array check
5438 return false;
5439 }
5440
5441 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5442 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5443 if (out_elem != T_INT || in_elem != T_INT) {
5444 return false;
5445 }
5446
5447 Node* outlen = load_array_length(out);
5448 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5449 Node* out_start = array_element_address(out, intcon(0), out_elem);
5450 Node* in_start = array_element_address(in, intcon(0), in_elem);
5451
5452 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5453 OptoRuntime::mulAdd_Type(),
5454 stubAddr, stubName, TypePtr::BOTTOM,
5455 out_start,in_start, new_offset, len, k);
5456 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5457 set_result(result);
5458 return true;
5459 }
5460
5461 //-------------inline_montgomeryMultiply-----------------------------------
5462 bool LibraryCallKit::inline_montgomeryMultiply() {
5463 address stubAddr = StubRoutines::montgomeryMultiply();
5464 if (stubAddr == NULL) {
5465 return false; // Intrinsic's stub is not implemented on this platform
5466 }
5467
5468 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5469 const char* stubName = "montgomery_square";
5470
5471 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5472
5473 Node* a = argument(0);
5474 Node* b = argument(1);
5475 Node* n = argument(2);
5476 Node* len = argument(3);
5477 Node* inv = argument(4);
5478 Node* m = argument(6);
5479
5480 const Type* a_type = a->Value(&_gvn);
5481 const TypeAryPtr* top_a = a_type->isa_aryptr();
5482 const Type* b_type = b->Value(&_gvn);
5483 const TypeAryPtr* top_b = b_type->isa_aryptr();
5484 const Type* n_type = a->Value(&_gvn);
5485 const TypeAryPtr* top_n = n_type->isa_aryptr();
5486 const Type* m_type = a->Value(&_gvn);
5487 const TypeAryPtr* top_m = m_type->isa_aryptr();
5488 if (top_a == NULL || top_a->klass() == NULL ||
5489 top_b == NULL || top_b->klass() == NULL ||
5490 top_n == NULL || top_n->klass() == NULL ||
5491 top_m == NULL || top_m->klass() == NULL) {
5492 // failed array check
5493 return false;
5494 }
5495
5496 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5497 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5498 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5499 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5500 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5501 return false;
5502 }
5503
5504 // Make the call
5505 {
5506 Node* a_start = array_element_address(a, intcon(0), a_elem);
5507 Node* b_start = array_element_address(b, intcon(0), b_elem);
5508 Node* n_start = array_element_address(n, intcon(0), n_elem);
5509 Node* m_start = array_element_address(m, intcon(0), m_elem);
5510
5511 Node* call = make_runtime_call(RC_LEAF,
5512 OptoRuntime::montgomeryMultiply_Type(),
5513 stubAddr, stubName, TypePtr::BOTTOM,
5514 a_start, b_start, n_start, len, inv, top(),
5515 m_start);
5516 set_result(m);
5517 }
5518
5519 return true;
5520 }
5521
5522 bool LibraryCallKit::inline_montgomerySquare() {
5523 address stubAddr = StubRoutines::montgomerySquare();
5524 if (stubAddr == NULL) {
5525 return false; // Intrinsic's stub is not implemented on this platform
5526 }
5527
5528 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5529 const char* stubName = "montgomery_square";
5530
5531 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5532
5533 Node* a = argument(0);
5534 Node* n = argument(1);
5535 Node* len = argument(2);
5536 Node* inv = argument(3);
5537 Node* m = argument(5);
5538
5539 const Type* a_type = a->Value(&_gvn);
5540 const TypeAryPtr* top_a = a_type->isa_aryptr();
5541 const Type* n_type = a->Value(&_gvn);
5542 const TypeAryPtr* top_n = n_type->isa_aryptr();
5543 const Type* m_type = a->Value(&_gvn);
5544 const TypeAryPtr* top_m = m_type->isa_aryptr();
5545 if (top_a == NULL || top_a->klass() == NULL ||
5546 top_n == NULL || top_n->klass() == NULL ||
5547 top_m == NULL || top_m->klass() == NULL) {
5548 // failed array check
5549 return false;
5550 }
5551
5552 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5553 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5554 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5555 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5556 return false;
5557 }
5558
5559 // Make the call
5560 {
5561 Node* a_start = array_element_address(a, intcon(0), a_elem);
5562 Node* n_start = array_element_address(n, intcon(0), n_elem);
5563 Node* m_start = array_element_address(m, intcon(0), m_elem);
5564
5565 Node* call = make_runtime_call(RC_LEAF,
5566 OptoRuntime::montgomerySquare_Type(),
5567 stubAddr, stubName, TypePtr::BOTTOM,
5568 a_start, n_start, len, inv, top(),
5569 m_start);
5570 set_result(m);
5571 }
5572
5573 return true;
5574 }
5575
5576
5577 /**
5578 * Calculate CRC32 for byte.
5579 * int java.util.zip.CRC32.update(int crc, int b)
5580 */
5581 bool LibraryCallKit::inline_updateCRC32() {
5582 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5583 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5584 // no receiver since it is static method
5585 Node* crc = argument(0); // type: int
5586 Node* b = argument(1); // type: int
5587
5588 /*
5589 * int c = ~ crc;
5590 * b = timesXtoThe32[(b ^ c) & 0xFF];
5591 * b = b ^ (c >>> 8);
5592 * crc = ~b;
5593 */
5594
5595 Node* M1 = intcon(-1);
5596 crc = _gvn.transform(new XorINode(crc, M1));
5597 Node* result = _gvn.transform(new XorINode(crc, b));
5598 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5599
5600 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5601 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5602 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5603 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5604
5605 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5606 result = _gvn.transform(new XorINode(crc, result));
5607 result = _gvn.transform(new XorINode(result, M1));
5608 set_result(result);
5609 return true;
5610 }
5611
5612 /**
5613 * Calculate CRC32 for byte[] array.
5614 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5615 */
5616 bool LibraryCallKit::inline_updateBytesCRC32() {
5617 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5618 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5619 // no receiver since it is static method
5620 Node* crc = argument(0); // type: int
5621 Node* src = argument(1); // type: oop
5622 Node* offset = argument(2); // type: int
5623 Node* length = argument(3); // type: int
5624
5625 const Type* src_type = src->Value(&_gvn);
5626 const TypeAryPtr* top_src = src_type->isa_aryptr();
5627 if (top_src == NULL || top_src->klass() == NULL) {
5628 // failed array check
5629 return false;
5630 }
5631
5632 // Figure out the size and type of the elements we will be copying.
5633 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5634 if (src_elem != T_BYTE) {
5635 return false;
5636 }
5637
5638 // 'src_start' points to src array + scaled offset
5639 Node* src_start = array_element_address(src, offset, src_elem);
5640
5641 // We assume that range check is done by caller.
5642 // TODO: generate range check (offset+length < src.length) in debug VM.
5643
5644 // Call the stub.
5645 address stubAddr = StubRoutines::updateBytesCRC32();
5646 const char *stubName = "updateBytesCRC32";
5647
5648 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5649 stubAddr, stubName, TypePtr::BOTTOM,
5650 crc, src_start, length);
5651 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5652 set_result(result);
5653 return true;
5654 }
5655
5656 /**
5657 * Calculate CRC32 for ByteBuffer.
5658 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5659 */
5660 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5661 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5662 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5663 // no receiver since it is static method
5664 Node* crc = argument(0); // type: int
5665 Node* src = argument(1); // type: long
5666 Node* offset = argument(3); // type: int
5667 Node* length = argument(4); // type: int
5668
5669 src = ConvL2X(src); // adjust Java long to machine word
5670 Node* base = _gvn.transform(new CastX2PNode(src));
5671 offset = ConvI2X(offset);
5672
5673 // 'src_start' points to src array + scaled offset
5674 Node* src_start = basic_plus_adr(top(), base, offset);
5675
5676 // Call the stub.
5677 address stubAddr = StubRoutines::updateBytesCRC32();
5678 const char *stubName = "updateBytesCRC32";
5679
5680 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5681 stubAddr, stubName, TypePtr::BOTTOM,
5682 crc, src_start, length);
5683 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5684 set_result(result);
5685 return true;
5686 }
5687
5688 //------------------------------get_table_from_crc32c_class-----------------------
5689 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5690 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5691 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5692
5693 return table;
5694 }
5695
5696 //------------------------------inline_updateBytesCRC32C-----------------------
5697 //
5698 // Calculate CRC32C for byte[] array.
5699 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5700 //
5701 bool LibraryCallKit::inline_updateBytesCRC32C() {
5702 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5703 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5704 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5705 // no receiver since it is a static method
5706 Node* crc = argument(0); // type: int
5707 Node* src = argument(1); // type: oop
5708 Node* offset = argument(2); // type: int
5709 Node* end = argument(3); // type: int
5710
5711 Node* length = _gvn.transform(new SubINode(end, offset));
5712
5713 const Type* src_type = src->Value(&_gvn);
5714 const TypeAryPtr* top_src = src_type->isa_aryptr();
5715 if (top_src == NULL || top_src->klass() == NULL) {
5716 // failed array check
5717 return false;
5718 }
5719
5720 // Figure out the size and type of the elements we will be copying.
5721 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5722 if (src_elem != T_BYTE) {
5723 return false;
5724 }
5725
5726 // 'src_start' points to src array + scaled offset
5727 Node* src_start = array_element_address(src, offset, src_elem);
5728
5729 // static final int[] byteTable in class CRC32C
5730 Node* table = get_table_from_crc32c_class(callee()->holder());
5731 Node* table_start = array_element_address(table, intcon(0), T_INT);
5732
5733 // We assume that range check is done by caller.
5734 // TODO: generate range check (offset+length < src.length) in debug VM.
5735
5736 // Call the stub.
5737 address stubAddr = StubRoutines::updateBytesCRC32C();
5738 const char *stubName = "updateBytesCRC32C";
5739
5740 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5741 stubAddr, stubName, TypePtr::BOTTOM,
5742 crc, src_start, length, table_start);
5743 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5744 set_result(result);
5745 return true;
5746 }
5747
5748 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5749 //
5750 // Calculate CRC32C for DirectByteBuffer.
5751 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5752 //
5753 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5754 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5755 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5756 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5757 // no receiver since it is a static method
5758 Node* crc = argument(0); // type: int
5759 Node* src = argument(1); // type: long
5760 Node* offset = argument(3); // type: int
5761 Node* end = argument(4); // type: int
5762
5763 Node* length = _gvn.transform(new SubINode(end, offset));
5764
5765 src = ConvL2X(src); // adjust Java long to machine word
5766 Node* base = _gvn.transform(new CastX2PNode(src));
5767 offset = ConvI2X(offset);
5768
5769 // 'src_start' points to src array + scaled offset
5770 Node* src_start = basic_plus_adr(top(), base, offset);
5771
5772 // static final int[] byteTable in class CRC32C
5773 Node* table = get_table_from_crc32c_class(callee()->holder());
5774 Node* table_start = array_element_address(table, intcon(0), T_INT);
5775
5776 // Call the stub.
5777 address stubAddr = StubRoutines::updateBytesCRC32C();
5778 const char *stubName = "updateBytesCRC32C";
5779
5780 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5781 stubAddr, stubName, TypePtr::BOTTOM,
5782 crc, src_start, length, table_start);
5783 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5784 set_result(result);
5785 return true;
5786 }
5787
5788 //----------------------------inline_reference_get----------------------------
5789 // public T java.lang.ref.Reference.get();
5790 bool LibraryCallKit::inline_reference_get() {
5791 const int referent_offset = java_lang_ref_Reference::referent_offset;
5792 guarantee(referent_offset > 0, "should have already been set");
5793
5794 // Get the argument:
5795 Node* reference_obj = null_check_receiver();
5796 if (stopped()) return true;
5797
5798 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5799
5800 ciInstanceKlass* klass = env()->Object_klass();
5801 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5802
5803 Node* no_ctrl = NULL;
5804 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5805
5806 // Use the pre-barrier to record the value in the referent field
5807 pre_barrier(false /* do_load */,
5808 control(),
5809 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5810 result /* pre_val */,
5811 T_OBJECT);
5812
5813 // Add memory barrier to prevent commoning reads from this field
5814 // across safepoint since GC can change its value.
5815 insert_mem_bar(Op_MemBarCPUOrder);
5816
5817 set_result(result);
5818 return true;
5819 }
5820
5821
5822 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5823 bool is_exact=true, bool is_static=false,
5824 ciInstanceKlass * fromKls=NULL) {
5825 if (fromKls == NULL) {
5826 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5827 assert(tinst != NULL, "obj is null");
5828 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5829 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5830 fromKls = tinst->klass()->as_instance_klass();
5831 } else {
5832 assert(is_static, "only for static field access");
5833 }
5834 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5835 ciSymbol::make(fieldTypeString),
5836 is_static);
5837
5838 assert (field != NULL, "undefined field");
5839 if (field == NULL) return (Node *) NULL;
5840
5841 if (is_static) {
5842 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5843 fromObj = makecon(tip);
5844 }
5845
5846 // Next code copied from Parse::do_get_xxx():
5847
5848 // Compute address and memory type.
5849 int offset = field->offset_in_bytes();
5850 bool is_vol = field->is_volatile();
5851 ciType* field_klass = field->type();
5852 assert(field_klass->is_loaded(), "should be loaded");
5853 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5854 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5855 BasicType bt = field->layout_type();
5856
5857 // Build the resultant type of the load
5858 const Type *type;
5859 if (bt == T_OBJECT) {
5860 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5861 } else {
5862 type = Type::get_const_basic_type(bt);
5863 }
5864
5865 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5866 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5867 }
5868 // Build the load.
5869 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5870 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5871 // If reference is volatile, prevent following memory ops from
5872 // floating up past the volatile read. Also prevents commoning
5873 // another volatile read.
5874 if (is_vol) {
5875 // Memory barrier includes bogus read of value to force load BEFORE membar
5876 insert_mem_bar(Op_MemBarAcquire, loadedField);
5877 }
5878 return loadedField;
5879 }
5880
5881
5882 //------------------------------inline_aescrypt_Block-----------------------
5883 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5884 address stubAddr;
5885 const char *stubName;
5886 assert(UseAES, "need AES instruction support");
5887
5888 switch(id) {
5889 case vmIntrinsics::_aescrypt_encryptBlock:
5890 stubAddr = StubRoutines::aescrypt_encryptBlock();
5891 stubName = "aescrypt_encryptBlock";
5892 break;
5893 case vmIntrinsics::_aescrypt_decryptBlock:
5894 stubAddr = StubRoutines::aescrypt_decryptBlock();
5895 stubName = "aescrypt_decryptBlock";
5896 break;
5897 }
5898 if (stubAddr == NULL) return false;
5899
5900 Node* aescrypt_object = argument(0);
5901 Node* src = argument(1);
5902 Node* src_offset = argument(2);
5903 Node* dest = argument(3);
5904 Node* dest_offset = argument(4);
5905
5906 // (1) src and dest are arrays.
5907 const Type* src_type = src->Value(&_gvn);
5908 const Type* dest_type = dest->Value(&_gvn);
5909 const TypeAryPtr* top_src = src_type->isa_aryptr();
5910 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5911 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5912
5913 // for the quick and dirty code we will skip all the checks.
5914 // we are just trying to get the call to be generated.
5915 Node* src_start = src;
5916 Node* dest_start = dest;
5917 if (src_offset != NULL || dest_offset != NULL) {
5918 assert(src_offset != NULL && dest_offset != NULL, "");
5919 src_start = array_element_address(src, src_offset, T_BYTE);
5920 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5921 }
5922
5923 // now need to get the start of its expanded key array
5924 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5925 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5926 if (k_start == NULL) return false;
5927
5928 if (Matcher::pass_original_key_for_aes()) {
5929 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5930 // compatibility issues between Java key expansion and SPARC crypto instructions
5931 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5932 if (original_k_start == NULL) return false;
5933
5934 // Call the stub.
5935 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5936 stubAddr, stubName, TypePtr::BOTTOM,
5937 src_start, dest_start, k_start, original_k_start);
5938 } else {
5939 // Call the stub.
5940 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5941 stubAddr, stubName, TypePtr::BOTTOM,
5942 src_start, dest_start, k_start);
5943 }
5944
5945 return true;
5946 }
5947
5948 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5949 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5950 address stubAddr;
5951 const char *stubName;
5952
5953 assert(UseAES, "need AES instruction support");
5954
5955 switch(id) {
5956 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5957 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5958 stubName = "cipherBlockChaining_encryptAESCrypt";
5959 break;
5960 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5961 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5962 stubName = "cipherBlockChaining_decryptAESCrypt";
5963 break;
5964 }
5965 if (stubAddr == NULL) return false;
5966
5967 Node* cipherBlockChaining_object = argument(0);
5968 Node* src = argument(1);
5969 Node* src_offset = argument(2);
5970 Node* len = argument(3);
5971 Node* dest = argument(4);
5972 Node* dest_offset = argument(5);
5973
5974 // (1) src and dest are arrays.
5975 const Type* src_type = src->Value(&_gvn);
5976 const Type* dest_type = dest->Value(&_gvn);
5977 const TypeAryPtr* top_src = src_type->isa_aryptr();
5978 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5979 assert (top_src != NULL && top_src->klass() != NULL
5980 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5981
5982 // checks are the responsibility of the caller
5983 Node* src_start = src;
5984 Node* dest_start = dest;
5985 if (src_offset != NULL || dest_offset != NULL) {
5986 assert(src_offset != NULL && dest_offset != NULL, "");
5987 src_start = array_element_address(src, src_offset, T_BYTE);
5988 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5989 }
5990
5991 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5992 // (because of the predicated logic executed earlier).
5993 // so we cast it here safely.
5994 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5995
5996 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5997 if (embeddedCipherObj == NULL) return false;
5998
5999 // cast it to what we know it will be at runtime
6000 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6001 assert(tinst != NULL, "CBC obj is null");
6002 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6003 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6004 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6005
6006 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6007 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6008 const TypeOopPtr* xtype = aklass->as_instance_type();
6009 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6010 aescrypt_object = _gvn.transform(aescrypt_object);
6011
6012 // we need to get the start of the aescrypt_object's expanded key array
6013 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6014 if (k_start == NULL) return false;
6015
6016 // similarly, get the start address of the r vector
6017 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6018 if (objRvec == NULL) return false;
6019 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6020
6021 Node* cbcCrypt;
6022 if (Matcher::pass_original_key_for_aes()) {
6023 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6024 // compatibility issues between Java key expansion and SPARC crypto instructions
6025 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6026 if (original_k_start == NULL) return false;
6027
6028 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6029 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6030 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6031 stubAddr, stubName, TypePtr::BOTTOM,
6032 src_start, dest_start, k_start, r_start, len, original_k_start);
6033 } else {
6034 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6035 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6036 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6037 stubAddr, stubName, TypePtr::BOTTOM,
6038 src_start, dest_start, k_start, r_start, len);
6039 }
6040
6041 // return cipher length (int)
6042 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6043 set_result(retvalue);
6044 return true;
6045 }
6046
6047 //------------------------------get_key_start_from_aescrypt_object-----------------------
6048 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6049 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6050 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6051 if (objAESCryptKey == NULL) return (Node *) NULL;
6052
6053 // now have the array, need to get the start address of the K array
6054 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6055 return k_start;
6056 }
6057
6058 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6059 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6060 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6061 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6062 if (objAESCryptKey == NULL) return (Node *) NULL;
6063
6064 // now have the array, need to get the start address of the lastKey array
6065 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6066 return original_k_start;
6067 }
6068
6069 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6070 // Return node representing slow path of predicate check.
6071 // the pseudo code we want to emulate with this predicate is:
6072 // for encryption:
6073 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6074 // for decryption:
6075 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6076 // note cipher==plain is more conservative than the original java code but that's OK
6077 //
6078 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6079 // The receiver was checked for NULL already.
6080 Node* objCBC = argument(0);
6081
6082 // Load embeddedCipher field of CipherBlockChaining object.
6083 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6084
6085 // get AESCrypt klass for instanceOf check
6086 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6087 // will have same classloader as CipherBlockChaining object
6088 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6089 assert(tinst != NULL, "CBCobj is null");
6090 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6091
6092 // we want to do an instanceof comparison against the AESCrypt class
6093 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6094 if (!klass_AESCrypt->is_loaded()) {
6095 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6096 Node* ctrl = control();
6097 set_control(top()); // no regular fast path
6098 return ctrl;
6099 }
6100 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6101
6102 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6103 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6104 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6105
6106 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6107
6108 // for encryption, we are done
6109 if (!decrypting)
6110 return instof_false; // even if it is NULL
6111
6112 // for decryption, we need to add a further check to avoid
6113 // taking the intrinsic path when cipher and plain are the same
6114 // see the original java code for why.
6115 RegionNode* region = new RegionNode(3);
6116 region->init_req(1, instof_false);
6117 Node* src = argument(1);
6118 Node* dest = argument(4);
6119 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6120 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6121 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6122 region->init_req(2, src_dest_conjoint);
6123
6124 record_for_igvn(region);
6125 return _gvn.transform(region);
6126 }
6127
6128 //------------------------------inline_ghash_processBlocks
6129 bool LibraryCallKit::inline_ghash_processBlocks() {
6130 address stubAddr;
6131 const char *stubName;
6132 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6133
6134 stubAddr = StubRoutines::ghash_processBlocks();
6135 stubName = "ghash_processBlocks";
6136
6137 Node* data = argument(0);
6138 Node* offset = argument(1);
6139 Node* len = argument(2);
6140 Node* state = argument(3);
6141 Node* subkeyH = argument(4);
6142
6143 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6144 assert(state_start, "state is NULL");
6145 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6146 assert(subkeyH_start, "subkeyH is NULL");
6147 Node* data_start = array_element_address(data, offset, T_BYTE);
6148 assert(data_start, "data is NULL");
6149
6150 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6151 OptoRuntime::ghash_processBlocks_Type(),
6152 stubAddr, stubName, TypePtr::BOTTOM,
6153 state_start, subkeyH_start, data_start, len);
6154 return true;
6155 }
6156
6157 //------------------------------inline_sha_implCompress-----------------------
6158 //
6159 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6160 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6161 //
6162 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6163 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6164 //
6165 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6166 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6167 //
6168 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6169 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6170
6171 Node* sha_obj = argument(0);
6172 Node* src = argument(1); // type oop
6173 Node* ofs = argument(2); // type int
6174
6175 const Type* src_type = src->Value(&_gvn);
6176 const TypeAryPtr* top_src = src_type->isa_aryptr();
6177 if (top_src == NULL || top_src->klass() == NULL) {
6178 // failed array check
6179 return false;
6180 }
6181 // Figure out the size and type of the elements we will be copying.
6182 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6183 if (src_elem != T_BYTE) {
6184 return false;
6185 }
6186 // 'src_start' points to src array + offset
6187 Node* src_start = array_element_address(src, ofs, src_elem);
6188 Node* state = NULL;
6189 address stubAddr;
6190 const char *stubName;
6191
6192 switch(id) {
6193 case vmIntrinsics::_sha_implCompress:
6194 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6195 state = get_state_from_sha_object(sha_obj);
6196 stubAddr = StubRoutines::sha1_implCompress();
6197 stubName = "sha1_implCompress";
6198 break;
6199 case vmIntrinsics::_sha2_implCompress:
6200 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6201 state = get_state_from_sha_object(sha_obj);
6202 stubAddr = StubRoutines::sha256_implCompress();
6203 stubName = "sha256_implCompress";
6204 break;
6205 case vmIntrinsics::_sha5_implCompress:
6206 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6207 state = get_state_from_sha5_object(sha_obj);
6208 stubAddr = StubRoutines::sha512_implCompress();
6209 stubName = "sha512_implCompress";
6210 break;
6211 default:
6212 fatal_unexpected_iid(id);
6213 return false;
6214 }
6215 if (state == NULL) return false;
6216
6217 // Call the stub.
6218 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6219 stubAddr, stubName, TypePtr::BOTTOM,
6220 src_start, state);
6221
6222 return true;
6223 }
6224
6225 //------------------------------inline_digestBase_implCompressMB-----------------------
6226 //
6227 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6228 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6229 //
6230 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6231 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6232 "need SHA1/SHA256/SHA512 instruction support");
6233 assert((uint)predicate < 3, "sanity");
6234 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6235
6236 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6237 Node* src = argument(1); // byte[] array
6238 Node* ofs = argument(2); // type int
6239 Node* limit = argument(3); // type int
6240
6241 const Type* src_type = src->Value(&_gvn);
6242 const TypeAryPtr* top_src = src_type->isa_aryptr();
6243 if (top_src == NULL || top_src->klass() == NULL) {
6244 // failed array check
6245 return false;
6246 }
6247 // Figure out the size and type of the elements we will be copying.
6248 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6249 if (src_elem != T_BYTE) {
6250 return false;
6251 }
6252 // 'src_start' points to src array + offset
6253 Node* src_start = array_element_address(src, ofs, src_elem);
6254
6255 const char* klass_SHA_name = NULL;
6256 const char* stub_name = NULL;
6257 address stub_addr = NULL;
6258 bool long_state = false;
6259
6260 switch (predicate) {
6261 case 0:
6262 if (UseSHA1Intrinsics) {
6263 klass_SHA_name = "sun/security/provider/SHA";
6264 stub_name = "sha1_implCompressMB";
6265 stub_addr = StubRoutines::sha1_implCompressMB();
6266 }
6267 break;
6268 case 1:
6269 if (UseSHA256Intrinsics) {
6270 klass_SHA_name = "sun/security/provider/SHA2";
6271 stub_name = "sha256_implCompressMB";
6272 stub_addr = StubRoutines::sha256_implCompressMB();
6273 }
6274 break;
6275 case 2:
6276 if (UseSHA512Intrinsics) {
6277 klass_SHA_name = "sun/security/provider/SHA5";
6278 stub_name = "sha512_implCompressMB";
6279 stub_addr = StubRoutines::sha512_implCompressMB();
6280 long_state = true;
6281 }
6282 break;
6283 default:
6284 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6285 }
6286 if (klass_SHA_name != NULL) {
6287 // get DigestBase klass to lookup for SHA klass
6288 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6289 assert(tinst != NULL, "digestBase_obj is not instance???");
6290 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6291
6292 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6293 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6294 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6295 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6296 }
6297 return false;
6298 }
6299 //------------------------------inline_sha_implCompressMB-----------------------
6300 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6301 bool long_state, address stubAddr, const char *stubName,
6302 Node* src_start, Node* ofs, Node* limit) {
6303 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6304 const TypeOopPtr* xtype = aklass->as_instance_type();
6305 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6306 sha_obj = _gvn.transform(sha_obj);
6307
6308 Node* state;
6309 if (long_state) {
6310 state = get_state_from_sha5_object(sha_obj);
6311 } else {
6312 state = get_state_from_sha_object(sha_obj);
6313 }
6314 if (state == NULL) return false;
6315
6316 // Call the stub.
6317 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6318 OptoRuntime::digestBase_implCompressMB_Type(),
6319 stubAddr, stubName, TypePtr::BOTTOM,
6320 src_start, state, ofs, limit);
6321 // return ofs (int)
6322 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6323 set_result(result);
6324
6325 return true;
6326 }
6327
6328 //------------------------------get_state_from_sha_object-----------------------
6329 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6330 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6331 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6332 if (sha_state == NULL) return (Node *) NULL;
6333
6334 // now have the array, need to get the start address of the state array
6335 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6336 return state;
6337 }
6338
6339 //------------------------------get_state_from_sha5_object-----------------------
6340 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6341 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6342 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6343 if (sha_state == NULL) return (Node *) NULL;
6344
6345 // now have the array, need to get the start address of the state array
6346 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6347 return state;
6348 }
6349
6350 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6351 // Return node representing slow path of predicate check.
6352 // the pseudo code we want to emulate with this predicate is:
6353 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6354 //
6355 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6356 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6357 "need SHA1/SHA256/SHA512 instruction support");
6358 assert((uint)predicate < 3, "sanity");
6359
6360 // The receiver was checked for NULL already.
6361 Node* digestBaseObj = argument(0);
6362
6363 // get DigestBase klass for instanceOf check
6364 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6365 assert(tinst != NULL, "digestBaseObj is null");
6366 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6367
6368 const char* klass_SHA_name = NULL;
6369 switch (predicate) {
6370 case 0:
6371 if (UseSHA1Intrinsics) {
6372 // we want to do an instanceof comparison against the SHA class
6373 klass_SHA_name = "sun/security/provider/SHA";
6374 }
6375 break;
6376 case 1:
6377 if (UseSHA256Intrinsics) {
6378 // we want to do an instanceof comparison against the SHA2 class
6379 klass_SHA_name = "sun/security/provider/SHA2";
6380 }
6381 break;
6382 case 2:
6383 if (UseSHA512Intrinsics) {
6384 // we want to do an instanceof comparison against the SHA5 class
6385 klass_SHA_name = "sun/security/provider/SHA5";
6386 }
6387 break;
6388 default:
6389 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6390 }
6391
6392 ciKlass* klass_SHA = NULL;
6393 if (klass_SHA_name != NULL) {
6394 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6395 }
6396 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6397 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6398 Node* ctrl = control();
6399 set_control(top()); // no intrinsic path
6400 return ctrl;
6401 }
6402 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6403
6404 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6405 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6406 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6407 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6408
6409 return instof_false; // even if it is NULL
6410 }
6411
6412 bool LibraryCallKit::inline_profileBoolean() {
6413 Node* counts = argument(1);
6414 const TypeAryPtr* ary = NULL;
6415 ciArray* aobj = NULL;
6416 if (counts->is_Con()
6417 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6418 && (aobj = ary->const_oop()->as_array()) != NULL
6419 && (aobj->length() == 2)) {
6420 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6421 jint false_cnt = aobj->element_value(0).as_int();
6422 jint true_cnt = aobj->element_value(1).as_int();
6423
6424 if (C->log() != NULL) {
6425 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6426 false_cnt, true_cnt);
6427 }
6428
6429 if (false_cnt + true_cnt == 0) {
6430 // According to profile, never executed.
6431 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6432 Deoptimization::Action_reinterpret);
6433 return true;
6434 }
6435
6436 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6437 // is a number of each value occurrences.
6438 Node* result = argument(0);
6439 if (false_cnt == 0 || true_cnt == 0) {
6440 // According to profile, one value has been never seen.
6441 int expected_val = (false_cnt == 0) ? 1 : 0;
6442
6443 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6444 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6445
6446 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6447 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6448 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6449
6450 { // Slow path: uncommon trap for never seen value and then reexecute
6451 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6452 // the value has been seen at least once.
6453 PreserveJVMState pjvms(this);
6454 PreserveReexecuteState preexecs(this);
6455 jvms()->set_should_reexecute(true);
6456
6457 set_control(slow_path);
6458 set_i_o(i_o());
6459
6460 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6461 Deoptimization::Action_reinterpret);
6462 }
6463 // The guard for never seen value enables sharpening of the result and
6464 // returning a constant. It allows to eliminate branches on the same value
6465 // later on.
6466 set_control(fast_path);
6467 result = intcon(expected_val);
6468 }
6469 // Stop profiling.
6470 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6471 // By replacing method body with profile data (represented as ProfileBooleanNode
6472 // on IR level) we effectively disable profiling.
6473 // It enables full speed execution once optimized code is generated.
6474 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6475 C->record_for_igvn(profile);
6476 set_result(profile);
6477 return true;
6478 } else {
6479 // Continue profiling.
6480 // Profile data isn't available at the moment. So, execute method's bytecode version.
6481 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6482 // is compiled and counters aren't available since corresponding MethodHandle
6483 // isn't a compile-time constant.
6484 return false;
6485 }
6486 }
6487
6488 bool LibraryCallKit::inline_isCompileConstant() {
6489 Node* n = argument(0);
6490 set_result(n->is_Con() ? intcon(1) : intcon(0));
6491 return true;
6492 }