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_exp();
225 bool inline_pow();
226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
227 bool inline_min_max(vmIntrinsics::ID id);
228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
229 // This returns Type::AnyPtr, RawPtr, or OopPtr.
230 int classify_unsafe_addr(Node* &base, Node* &offset);
231 Node* make_unsafe_address(Node* base, Node* offset);
232 // Helper for inline_unsafe_access.
233 // Generates the guards that check whether the result of
234 // Unsafe.getObject should be recorded in an SATB log buffer.
235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
237 static bool klass_needs_init_guard(Node* kls);
238 bool inline_unsafe_allocate();
239 bool inline_unsafe_copyMemory();
240 bool inline_native_currentThread();
241 #ifdef TRACE_HAVE_INTRINSICS
242 bool inline_native_classID();
243 bool inline_native_threadID();
244 #endif
245 bool inline_native_time_funcs(address method, const char* funcName);
246 bool inline_native_isInterrupted();
247 bool inline_native_Class_query(vmIntrinsics::ID id);
248 bool inline_native_subtype_check();
249
250 bool inline_native_newArray();
251 bool inline_native_getLength();
252 bool inline_array_copyOf(bool is_copyOfRange);
253 bool inline_array_equals();
254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
255 bool inline_native_clone(bool is_virtual);
256 bool inline_native_Reflection_getCallerClass();
257 // Helper function for inlining native object hash method
258 bool inline_native_hashcode(bool is_virtual, bool is_static);
259 bool inline_native_getClass();
260
261 // Helper functions for inlining arraycopy
262 bool inline_arraycopy();
263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
264 RegionNode* slow_region);
265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
267
268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
270 bool inline_unsafe_ordered_store(BasicType type);
271 bool inline_unsafe_fence(vmIntrinsics::ID id);
272 bool inline_fp_conversions(vmIntrinsics::ID id);
273 bool inline_number_methods(vmIntrinsics::ID id);
274 bool inline_reference_get();
275 bool inline_Class_cast();
276 bool inline_aescrypt_Block(vmIntrinsics::ID id);
277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
281 bool inline_ghash_processBlocks();
282 bool inline_sha_implCompress(vmIntrinsics::ID id);
283 bool inline_digestBase_implCompressMB(int predicate);
284 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
285 bool long_state, address stubAddr, const char *stubName,
286 Node* src_start, Node* ofs, Node* limit);
287 Node* get_state_from_sha_object(Node *sha_object);
288 Node* get_state_from_sha5_object(Node *sha_object);
289 Node* inline_digestBase_implCompressMB_predicate(int predicate);
290 bool inline_encodeISOArray();
291 bool inline_updateCRC32();
292 bool inline_updateBytesCRC32();
293 bool inline_updateByteBufferCRC32();
294 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
295 bool inline_updateBytesCRC32C();
296 bool inline_updateDirectByteBufferCRC32C();
297 bool inline_multiplyToLen();
298 bool inline_squareToLen();
299 bool inline_mulAdd();
300 bool inline_montgomeryMultiply();
301 bool inline_montgomerySquare();
302
303 bool inline_profileBoolean();
304 bool inline_isCompileConstant();
305 };
306
307
308 //---------------------------make_vm_intrinsic----------------------------
309 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
310 vmIntrinsics::ID id = m->intrinsic_id();
311 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
312
313 ccstr disable_intr = NULL;
314
315 if ((DisableIntrinsic[0] != '\0'
316 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
317 (method_has_option_value("DisableIntrinsic", disable_intr)
318 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
319 // disabled by a user request on the command line:
320 // example: -XX:DisableIntrinsic=_hashCode,_getClass
321 return NULL;
322 }
323
324 if (!m->is_loaded()) {
325 // do not attempt to inline unloaded methods
326 return NULL;
327 }
328
329 // Only a few intrinsics implement a virtual dispatch.
330 // They are expensive calls which are also frequently overridden.
331 if (is_virtual) {
332 switch (id) {
333 case vmIntrinsics::_hashCode:
334 case vmIntrinsics::_clone:
335 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
336 break;
337 default:
338 return NULL;
339 }
340 }
341
342 // -XX:-InlineNatives disables nearly all intrinsics:
343 if (!InlineNatives) {
344 switch (id) {
345 case vmIntrinsics::_indexOf:
346 case vmIntrinsics::_compareTo:
347 case vmIntrinsics::_equals:
348 case vmIntrinsics::_equalsC:
349 case vmIntrinsics::_getAndAddInt:
350 case vmIntrinsics::_getAndAddLong:
351 case vmIntrinsics::_getAndSetInt:
352 case vmIntrinsics::_getAndSetLong:
353 case vmIntrinsics::_getAndSetObject:
354 case vmIntrinsics::_loadFence:
355 case vmIntrinsics::_storeFence:
356 case vmIntrinsics::_fullFence:
357 break; // InlineNatives does not control String.compareTo
358 case vmIntrinsics::_Reference_get:
359 break; // InlineNatives does not control Reference.get
360 default:
361 return NULL;
362 }
363 }
364
365 int predicates = 0;
366 bool does_virtual_dispatch = false;
367
368 switch (id) {
369 case vmIntrinsics::_compareTo:
370 if (!SpecialStringCompareTo) return NULL;
371 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
372 break;
373 case vmIntrinsics::_indexOf:
374 if (!SpecialStringIndexOf) return NULL;
375 break;
376 case vmIntrinsics::_equals:
377 if (!SpecialStringEquals) return NULL;
378 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
379 break;
380 case vmIntrinsics::_equalsC:
381 if (!SpecialArraysEquals) return NULL;
382 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
383 break;
384 case vmIntrinsics::_arraycopy:
385 if (!InlineArrayCopy) return NULL;
386 break;
387 case vmIntrinsics::_copyMemory:
388 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
389 if (!InlineArrayCopy) return NULL;
390 break;
391 case vmIntrinsics::_hashCode:
392 if (!InlineObjectHash) return NULL;
393 does_virtual_dispatch = true;
394 break;
395 case vmIntrinsics::_clone:
396 does_virtual_dispatch = true;
397 case vmIntrinsics::_copyOf:
398 case vmIntrinsics::_copyOfRange:
399 if (!InlineObjectCopy) return NULL;
400 // These also use the arraycopy intrinsic mechanism:
401 if (!InlineArrayCopy) return NULL;
402 break;
403 case vmIntrinsics::_encodeISOArray:
404 if (!SpecialEncodeISOArray) return NULL;
405 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
406 break;
407 case vmIntrinsics::_checkIndex:
408 // We do not intrinsify this. The optimizer does fine with it.
409 return NULL;
410
411 case vmIntrinsics::_getCallerClass:
412 if (!InlineReflectionGetCallerClass) return NULL;
413 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
414 break;
415
416 case vmIntrinsics::_bitCount_i:
417 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
418 break;
419
420 case vmIntrinsics::_bitCount_l:
421 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
422 break;
423
424 case vmIntrinsics::_numberOfLeadingZeros_i:
425 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
426 break;
427
428 case vmIntrinsics::_numberOfLeadingZeros_l:
429 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
430 break;
431
432 case vmIntrinsics::_numberOfTrailingZeros_i:
433 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
434 break;
435
436 case vmIntrinsics::_numberOfTrailingZeros_l:
437 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
438 break;
439
440 case vmIntrinsics::_reverseBytes_c:
441 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
442 break;
443 case vmIntrinsics::_reverseBytes_s:
444 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
445 break;
446 case vmIntrinsics::_reverseBytes_i:
447 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
448 break;
449 case vmIntrinsics::_reverseBytes_l:
450 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
451 break;
452
453 case vmIntrinsics::_Reference_get:
454 // Use the intrinsic version of Reference.get() so that the value in
455 // the referent field can be registered by the G1 pre-barrier code.
456 // Also add memory barrier to prevent commoning reads from this field
457 // across safepoint since GC can change it value.
458 break;
459
460 case vmIntrinsics::_compareAndSwapObject:
461 #ifdef _LP64
462 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
463 #endif
464 break;
465
466 case vmIntrinsics::_compareAndSwapLong:
467 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
468 break;
469
470 case vmIntrinsics::_getAndAddInt:
471 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
472 break;
473
474 case vmIntrinsics::_getAndAddLong:
475 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
476 break;
477
478 case vmIntrinsics::_getAndSetInt:
479 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndSetLong:
483 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
484 break;
485
486 case vmIntrinsics::_getAndSetObject:
487 #ifdef _LP64
488 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
489 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
490 break;
491 #else
492 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
493 break;
494 #endif
495
496 case vmIntrinsics::_aescrypt_encryptBlock:
497 case vmIntrinsics::_aescrypt_decryptBlock:
498 if (!UseAESIntrinsics) return NULL;
499 break;
500
501 case vmIntrinsics::_multiplyToLen:
502 if (!UseMultiplyToLenIntrinsic) return NULL;
503 break;
504
505 case vmIntrinsics::_squareToLen:
506 if (!UseSquareToLenIntrinsic) return NULL;
507 break;
508
509 case vmIntrinsics::_mulAdd:
510 if (!UseMulAddIntrinsic) return NULL;
511 break;
512
513 case vmIntrinsics::_montgomeryMultiply:
514 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
515 break;
516 case vmIntrinsics::_montgomerySquare:
517 if (!UseMontgomerySquareIntrinsic) return NULL;
518 break;
519
520 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
521 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
522 if (!UseAESIntrinsics) return NULL;
523 // these two require the predicated logic
524 predicates = 1;
525 break;
526
527 case vmIntrinsics::_sha_implCompress:
528 if (!UseSHA1Intrinsics) return NULL;
529 break;
530
531 case vmIntrinsics::_sha2_implCompress:
532 if (!UseSHA256Intrinsics) return NULL;
533 break;
534
535 case vmIntrinsics::_sha5_implCompress:
536 if (!UseSHA512Intrinsics) return NULL;
537 break;
538
539 case vmIntrinsics::_digestBase_implCompressMB:
540 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
541 predicates = 3;
542 break;
543
544 case vmIntrinsics::_ghash_processBlocks:
545 if (!UseGHASHIntrinsics) return NULL;
546 break;
547
548 case vmIntrinsics::_updateCRC32:
549 case vmIntrinsics::_updateBytesCRC32:
550 case vmIntrinsics::_updateByteBufferCRC32:
551 if (!UseCRC32Intrinsics) return NULL;
552 break;
553
554 case vmIntrinsics::_updateBytesCRC32C:
555 case vmIntrinsics::_updateDirectByteBufferCRC32C:
556 if (!UseCRC32CIntrinsics) return NULL;
557 break;
558
559 case vmIntrinsics::_incrementExactI:
560 case vmIntrinsics::_addExactI:
561 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
562 break;
563 case vmIntrinsics::_incrementExactL:
564 case vmIntrinsics::_addExactL:
565 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
566 break;
567 case vmIntrinsics::_decrementExactI:
568 case vmIntrinsics::_subtractExactI:
569 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
570 break;
571 case vmIntrinsics::_decrementExactL:
572 case vmIntrinsics::_subtractExactL:
573 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
574 break;
575 case vmIntrinsics::_negateExactI:
576 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
577 break;
578 case vmIntrinsics::_negateExactL:
579 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
580 break;
581 case vmIntrinsics::_multiplyExactI:
582 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
583 break;
584 case vmIntrinsics::_multiplyExactL:
585 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
586 break;
587
588 case vmIntrinsics::_getShortUnaligned:
589 case vmIntrinsics::_getCharUnaligned:
590 case vmIntrinsics::_getIntUnaligned:
591 case vmIntrinsics::_getLongUnaligned:
592 case vmIntrinsics::_putShortUnaligned:
593 case vmIntrinsics::_putCharUnaligned:
594 case vmIntrinsics::_putIntUnaligned:
595 case vmIntrinsics::_putLongUnaligned:
596 if (!UseUnalignedAccesses) return NULL;
597 break;
598
599 default:
600 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
601 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
602 break;
603 }
604
605 // -XX:-InlineClassNatives disables natives from the Class class.
606 // The flag applies to all reflective calls, notably Array.newArray
607 // (visible to Java programmers as Array.newInstance).
608 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
609 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
610 if (!InlineClassNatives) return NULL;
611 }
612
613 // -XX:-InlineThreadNatives disables natives from the Thread class.
614 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
615 if (!InlineThreadNatives) return NULL;
616 }
617
618 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
619 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
620 m->holder()->name() == ciSymbol::java_lang_Float() ||
621 m->holder()->name() == ciSymbol::java_lang_Double()) {
622 if (!InlineMathNatives) return NULL;
623 }
624
625 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
626 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
627 if (!InlineUnsafeOps) return NULL;
628 }
629
630 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
631 }
632
633 //----------------------register_library_intrinsics-----------------------
634 // Initialize this file's data structures, for each Compile instance.
635 void Compile::register_library_intrinsics() {
636 // Nothing to do here.
637 }
638
639 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
640 LibraryCallKit kit(jvms, this);
641 Compile* C = kit.C;
642 int nodes = C->unique();
643 #ifndef PRODUCT
644 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
645 char buf[1000];
646 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
647 tty->print_cr("Intrinsic %s", str);
648 }
649 #endif
650 ciMethod* callee = kit.callee();
651 const int bci = kit.bci();
652
653 // Try to inline the intrinsic.
654 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
655 kit.try_to_inline(_last_predicate)) {
656 if (C->print_intrinsics() || C->print_inlining()) {
657 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
658 }
659 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
660 if (C->log()) {
661 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
662 vmIntrinsics::name_at(intrinsic_id()),
663 (is_virtual() ? " virtual='1'" : ""),
664 C->unique() - nodes);
665 }
666 // Push the result from the inlined method onto the stack.
667 kit.push_result();
668 C->print_inlining_update(this);
669 return kit.transfer_exceptions_into_jvms();
670 }
671
672 // The intrinsic bailed out
673 if (C->print_intrinsics() || C->print_inlining()) {
674 if (jvms->has_method()) {
675 // Not a root compile.
676 const char* msg;
677 if (callee->intrinsic_candidate()) {
678 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
679 } else {
680 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
681 : "failed to inline (intrinsic), method not annotated";
682 }
683 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
684 } else {
685 // Root compile
686 tty->print("Did not generate intrinsic %s%s at bci:%d in",
687 vmIntrinsics::name_at(intrinsic_id()),
688 (is_virtual() ? " (virtual)" : ""), bci);
689 }
690 }
691 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
692 C->print_inlining_update(this);
693 return NULL;
694 }
695
696 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
697 LibraryCallKit kit(jvms, this);
698 Compile* C = kit.C;
699 int nodes = C->unique();
700 _last_predicate = predicate;
701 #ifndef PRODUCT
702 assert(is_predicated() && predicate < predicates_count(), "sanity");
703 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
704 char buf[1000];
705 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
706 tty->print_cr("Predicate for intrinsic %s", str);
707 }
708 #endif
709 ciMethod* callee = kit.callee();
710 const int bci = kit.bci();
711
712 Node* slow_ctl = kit.try_to_predicate(predicate);
713 if (!kit.failing()) {
714 if (C->print_intrinsics() || C->print_inlining()) {
715 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
716 }
717 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
718 if (C->log()) {
719 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
720 vmIntrinsics::name_at(intrinsic_id()),
721 (is_virtual() ? " virtual='1'" : ""),
722 C->unique() - nodes);
723 }
724 return slow_ctl; // Could be NULL if the check folds.
725 }
726
727 // The intrinsic bailed out
728 if (C->print_intrinsics() || C->print_inlining()) {
729 if (jvms->has_method()) {
730 // Not a root compile.
731 const char* msg = "failed to generate predicate for intrinsic";
732 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
733 } else {
734 // Root compile
735 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
736 vmIntrinsics::name_at(intrinsic_id()),
737 (is_virtual() ? " (virtual)" : ""), bci);
738 }
739 }
740 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
741 return NULL;
742 }
743
744 bool LibraryCallKit::try_to_inline(int predicate) {
745 // Handle symbolic names for otherwise undistinguished boolean switches:
746 const bool is_store = true;
747 const bool is_native_ptr = true;
748 const bool is_static = true;
749 const bool is_volatile = true;
750
751 if (!jvms()->has_method()) {
752 // Root JVMState has a null method.
753 assert(map()->memory()->Opcode() == Op_Parm, "");
754 // Insert the memory aliasing node
755 set_all_memory(reset_memory());
756 }
757 assert(merged_memory(), "");
758
759
760 switch (intrinsic_id()) {
761 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
762 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
763 case vmIntrinsics::_getClass: return inline_native_getClass();
764
765 case vmIntrinsics::_dsin:
766 case vmIntrinsics::_dcos:
767 case vmIntrinsics::_dtan:
768 case vmIntrinsics::_dabs:
769 case vmIntrinsics::_datan2:
770 case vmIntrinsics::_dsqrt:
771 case vmIntrinsics::_dexp:
772 case vmIntrinsics::_dlog:
773 case vmIntrinsics::_dlog10:
774 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
775
776 case vmIntrinsics::_min:
777 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
778
779 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
780 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
781 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
782 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
783 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
784 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
785 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
786 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
787 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
788 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
789 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
790 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
791
792 case vmIntrinsics::_arraycopy: return inline_arraycopy();
793
794 case vmIntrinsics::_compareTo: return inline_string_compareTo();
795 case vmIntrinsics::_indexOf: return inline_string_indexOf();
796 case vmIntrinsics::_equals: return inline_string_equals();
797
798 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
799 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
800 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
801 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
802 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
803 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
804 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
805 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
806 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
807
808 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
809 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
810 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
811 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
812 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
813 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
814 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
815 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
816 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
817
818 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
819 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
820 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
821 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
822 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
823 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
824 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
825 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
826
827 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
828 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
829 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
830 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
831 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
832 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
833 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
834 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
835
836 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
837 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
838 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
839 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
840 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
841 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
842 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
843 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
844 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
845
846 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
847 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
848 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
849 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
850 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
851 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
852 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
853 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
854 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
855
856 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
857 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
858 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
859 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
860
861 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
862 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
863 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
864 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
865
866 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
867 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
868 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
869
870 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
871 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
872 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
873
874 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
875 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
876 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
877 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
878 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
879
880 case vmIntrinsics::_loadFence:
881 case vmIntrinsics::_storeFence:
882 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
883
884 case vmIntrinsics::_currentThread: return inline_native_currentThread();
885 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
886
887 #ifdef TRACE_HAVE_INTRINSICS
888 case vmIntrinsics::_classID: return inline_native_classID();
889 case vmIntrinsics::_threadID: return inline_native_threadID();
890 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
891 #endif
892 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
893 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
894 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
895 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
896 case vmIntrinsics::_newArray: return inline_native_newArray();
897 case vmIntrinsics::_getLength: return inline_native_getLength();
898 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
899 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
900 case vmIntrinsics::_equalsC: return inline_array_equals();
901 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
902
903 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
904
905 case vmIntrinsics::_isInstance:
906 case vmIntrinsics::_getModifiers:
907 case vmIntrinsics::_isInterface:
908 case vmIntrinsics::_isArray:
909 case vmIntrinsics::_isPrimitive:
910 case vmIntrinsics::_getSuperclass:
911 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
912
913 case vmIntrinsics::_floatToRawIntBits:
914 case vmIntrinsics::_floatToIntBits:
915 case vmIntrinsics::_intBitsToFloat:
916 case vmIntrinsics::_doubleToRawLongBits:
917 case vmIntrinsics::_doubleToLongBits:
918 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
919
920 case vmIntrinsics::_numberOfLeadingZeros_i:
921 case vmIntrinsics::_numberOfLeadingZeros_l:
922 case vmIntrinsics::_numberOfTrailingZeros_i:
923 case vmIntrinsics::_numberOfTrailingZeros_l:
924 case vmIntrinsics::_bitCount_i:
925 case vmIntrinsics::_bitCount_l:
926 case vmIntrinsics::_reverseBytes_i:
927 case vmIntrinsics::_reverseBytes_l:
928 case vmIntrinsics::_reverseBytes_s:
929 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
930
931 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
932
933 case vmIntrinsics::_Reference_get: return inline_reference_get();
934
935 case vmIntrinsics::_Class_cast: return inline_Class_cast();
936
937 case vmIntrinsics::_aescrypt_encryptBlock:
938 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
939
940 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
941 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
942 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
943
944 case vmIntrinsics::_sha_implCompress:
945 case vmIntrinsics::_sha2_implCompress:
946 case vmIntrinsics::_sha5_implCompress:
947 return inline_sha_implCompress(intrinsic_id());
948
949 case vmIntrinsics::_digestBase_implCompressMB:
950 return inline_digestBase_implCompressMB(predicate);
951
952 case vmIntrinsics::_multiplyToLen:
953 return inline_multiplyToLen();
954
955 case vmIntrinsics::_squareToLen:
956 return inline_squareToLen();
957
958 case vmIntrinsics::_mulAdd:
959 return inline_mulAdd();
960
961 case vmIntrinsics::_montgomeryMultiply:
962 return inline_montgomeryMultiply();
963 case vmIntrinsics::_montgomerySquare:
964 return inline_montgomerySquare();
965
966 case vmIntrinsics::_ghash_processBlocks:
967 return inline_ghash_processBlocks();
968
969 case vmIntrinsics::_encodeISOArray:
970 return inline_encodeISOArray();
971
972 case vmIntrinsics::_updateCRC32:
973 return inline_updateCRC32();
974 case vmIntrinsics::_updateBytesCRC32:
975 return inline_updateBytesCRC32();
976 case vmIntrinsics::_updateByteBufferCRC32:
977 return inline_updateByteBufferCRC32();
978
979 case vmIntrinsics::_updateBytesCRC32C:
980 return inline_updateBytesCRC32C();
981 case vmIntrinsics::_updateDirectByteBufferCRC32C:
982 return inline_updateDirectByteBufferCRC32C();
983
984 case vmIntrinsics::_profileBoolean:
985 return inline_profileBoolean();
986 case vmIntrinsics::_isCompileConstant:
987 return inline_isCompileConstant();
988
989 default:
990 // If you get here, it may be that someone has added a new intrinsic
991 // to the list in vmSymbols.hpp without implementing it here.
992 #ifndef PRODUCT
993 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
994 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
995 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
996 }
997 #endif
998 return false;
999 }
1000 }
1001
1002 Node* LibraryCallKit::try_to_predicate(int predicate) {
1003 if (!jvms()->has_method()) {
1004 // Root JVMState has a null method.
1005 assert(map()->memory()->Opcode() == Op_Parm, "");
1006 // Insert the memory aliasing node
1007 set_all_memory(reset_memory());
1008 }
1009 assert(merged_memory(), "");
1010
1011 switch (intrinsic_id()) {
1012 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1013 return inline_cipherBlockChaining_AESCrypt_predicate(false);
1014 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1015 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1016 case vmIntrinsics::_digestBase_implCompressMB:
1017 return inline_digestBase_implCompressMB_predicate(predicate);
1018
1019 default:
1020 // If you get here, it may be that someone has added a new intrinsic
1021 // to the list in vmSymbols.hpp without implementing it here.
1022 #ifndef PRODUCT
1023 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1024 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1025 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1026 }
1027 #endif
1028 Node* slow_ctl = control();
1029 set_control(top()); // No fast path instrinsic
1030 return slow_ctl;
1031 }
1032 }
1033
1034 //------------------------------set_result-------------------------------
1035 // Helper function for finishing intrinsics.
1036 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1037 record_for_igvn(region);
1038 set_control(_gvn.transform(region));
1039 set_result( _gvn.transform(value));
1040 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1041 }
1042
1043 //------------------------------generate_guard---------------------------
1044 // Helper function for generating guarded fast-slow graph structures.
1045 // The given 'test', if true, guards a slow path. If the test fails
1046 // then a fast path can be taken. (We generally hope it fails.)
1047 // In all cases, GraphKit::control() is updated to the fast path.
1048 // The returned value represents the control for the slow path.
1049 // The return value is never 'top'; it is either a valid control
1050 // or NULL if it is obvious that the slow path can never be taken.
1051 // Also, if region and the slow control are not NULL, the slow edge
1052 // is appended to the region.
1053 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1054 if (stopped()) {
1055 // Already short circuited.
1056 return NULL;
1057 }
1058
1059 // Build an if node and its projections.
1060 // If test is true we take the slow path, which we assume is uncommon.
1061 if (_gvn.type(test) == TypeInt::ZERO) {
1062 // The slow branch is never taken. No need to build this guard.
1063 return NULL;
1064 }
1065
1066 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1067
1068 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1069 if (if_slow == top()) {
1070 // The slow branch is never taken. No need to build this guard.
1071 return NULL;
1072 }
1073
1074 if (region != NULL)
1075 region->add_req(if_slow);
1076
1077 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1078 set_control(if_fast);
1079
1080 return if_slow;
1081 }
1082
1083 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1084 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1085 }
1086 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1087 return generate_guard(test, region, PROB_FAIR);
1088 }
1089
1090 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1091 Node* *pos_index) {
1092 if (stopped())
1093 return NULL; // already stopped
1094 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1095 return NULL; // index is already adequately typed
1096 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1097 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1098 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1099 if (is_neg != NULL && pos_index != NULL) {
1100 // Emulate effect of Parse::adjust_map_after_if.
1101 Node* ccast = new CastIINode(index, TypeInt::POS);
1102 ccast->set_req(0, control());
1103 (*pos_index) = _gvn.transform(ccast);
1104 }
1105 return is_neg;
1106 }
1107
1108 // Make sure that 'position' is a valid limit index, in [0..length].
1109 // There are two equivalent plans for checking this:
1110 // A. (offset + copyLength) unsigned<= arrayLength
1111 // B. offset <= (arrayLength - copyLength)
1112 // We require that all of the values above, except for the sum and
1113 // difference, are already known to be non-negative.
1114 // Plan A is robust in the face of overflow, if offset and copyLength
1115 // are both hugely positive.
1116 //
1117 // Plan B is less direct and intuitive, but it does not overflow at
1118 // all, since the difference of two non-negatives is always
1119 // representable. Whenever Java methods must perform the equivalent
1120 // check they generally use Plan B instead of Plan A.
1121 // For the moment we use Plan A.
1122 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1123 Node* subseq_length,
1124 Node* array_length,
1125 RegionNode* region) {
1126 if (stopped())
1127 return NULL; // already stopped
1128 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1129 if (zero_offset && subseq_length->eqv_uncast(array_length))
1130 return NULL; // common case of whole-array copy
1131 Node* last = subseq_length;
1132 if (!zero_offset) // last += offset
1133 last = _gvn.transform(new AddINode(last, offset));
1134 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1135 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1136 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1137 return is_over;
1138 }
1139
1140
1141 //--------------------------generate_current_thread--------------------
1142 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1143 ciKlass* thread_klass = env()->Thread_klass();
1144 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1145 Node* thread = _gvn.transform(new ThreadLocalNode());
1146 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1147 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1148 tls_output = thread;
1149 return threadObj;
1150 }
1151
1152
1153 //------------------------------make_string_method_node------------------------
1154 // Helper method for String intrinsic functions. This version is called
1155 // with str1 and str2 pointing to String object nodes.
1156 //
1157 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1158 Node* no_ctrl = NULL;
1159
1160 // Get start addr of string
1161 Node* str1_value = load_String_value(no_ctrl, str1);
1162 Node* str1_offset = load_String_offset(no_ctrl, str1);
1163 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1164
1165 // Get length of string 1
1166 Node* str1_len = load_String_length(no_ctrl, str1);
1167
1168 Node* str2_value = load_String_value(no_ctrl, str2);
1169 Node* str2_offset = load_String_offset(no_ctrl, str2);
1170 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1171
1172 Node* str2_len = NULL;
1173 Node* result = NULL;
1174
1175 switch (opcode) {
1176 case Op_StrIndexOf:
1177 // Get length of string 2
1178 str2_len = load_String_length(no_ctrl, str2);
1179
1180 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1181 str1_start, str1_len, str2_start, str2_len);
1182 break;
1183 case Op_StrComp:
1184 // Get length of string 2
1185 str2_len = load_String_length(no_ctrl, str2);
1186
1187 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1188 str1_start, str1_len, str2_start, str2_len);
1189 break;
1190 case Op_StrEquals:
1191 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1192 str1_start, str2_start, str1_len);
1193 break;
1194 default:
1195 ShouldNotReachHere();
1196 return NULL;
1197 }
1198
1199 // All these intrinsics have checks.
1200 C->set_has_split_ifs(true); // Has chance for split-if optimization
1201
1202 return _gvn.transform(result);
1203 }
1204
1205 // Helper method for String intrinsic functions. This version is called
1206 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1207 // to Int nodes containing the lenghts of str1 and str2.
1208 //
1209 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1210 Node* result = NULL;
1211 switch (opcode) {
1212 case Op_StrIndexOf:
1213 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1214 str1_start, cnt1, str2_start, cnt2);
1215 break;
1216 case Op_StrComp:
1217 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1218 str1_start, cnt1, str2_start, cnt2);
1219 break;
1220 case Op_StrEquals:
1221 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1222 str1_start, str2_start, cnt1);
1223 break;
1224 default:
1225 ShouldNotReachHere();
1226 return NULL;
1227 }
1228
1229 // All these intrinsics have checks.
1230 C->set_has_split_ifs(true); // Has chance for split-if optimization
1231
1232 return _gvn.transform(result);
1233 }
1234
1235 //------------------------------inline_string_compareTo------------------------
1236 // public int java.lang.String.compareTo(String anotherString);
1237 bool LibraryCallKit::inline_string_compareTo() {
1238 Node* receiver = null_check(argument(0));
1239 Node* arg = null_check(argument(1));
1240 if (stopped()) {
1241 return true;
1242 }
1243 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1244 return true;
1245 }
1246
1247 //------------------------------inline_string_equals------------------------
1248 bool LibraryCallKit::inline_string_equals() {
1249 Node* receiver = null_check_receiver();
1250 // NOTE: Do not null check argument for String.equals() because spec
1251 // allows to specify NULL as argument.
1252 Node* argument = this->argument(1);
1253 if (stopped()) {
1254 return true;
1255 }
1256
1257 // paths (plus control) merge
1258 RegionNode* region = new RegionNode(5);
1259 Node* phi = new PhiNode(region, TypeInt::BOOL);
1260
1261 // does source == target string?
1262 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1263 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1264
1265 Node* if_eq = generate_slow_guard(bol, NULL);
1266 if (if_eq != NULL) {
1267 // receiver == argument
1268 phi->init_req(2, intcon(1));
1269 region->init_req(2, if_eq);
1270 }
1271
1272 // get String klass for instanceOf
1273 ciInstanceKlass* klass = env()->String_klass();
1274
1275 if (!stopped()) {
1276 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1277 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1278 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1279
1280 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1281 //instanceOf == true, fallthrough
1282
1283 if (inst_false != NULL) {
1284 phi->init_req(3, intcon(0));
1285 region->init_req(3, inst_false);
1286 }
1287 }
1288
1289 if (!stopped()) {
1290 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1291
1292 // Properly cast the argument to String
1293 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1294 // This path is taken only when argument's type is String:NotNull.
1295 argument = cast_not_null(argument, false);
1296
1297 Node* no_ctrl = NULL;
1298
1299 // Get start addr of receiver
1300 Node* receiver_val = load_String_value(no_ctrl, receiver);
1301 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1302 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1303
1304 // Get length of receiver
1305 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1306
1307 // Get start addr of argument
1308 Node* argument_val = load_String_value(no_ctrl, argument);
1309 Node* argument_offset = load_String_offset(no_ctrl, argument);
1310 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1311
1312 // Get length of argument
1313 Node* argument_cnt = load_String_length(no_ctrl, argument);
1314
1315 // Check for receiver count != argument count
1316 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1317 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1318 Node* if_ne = generate_slow_guard(bol, NULL);
1319 if (if_ne != NULL) {
1320 phi->init_req(4, intcon(0));
1321 region->init_req(4, if_ne);
1322 }
1323
1324 // Check for count == 0 is done by assembler code for StrEquals.
1325
1326 if (!stopped()) {
1327 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1328 phi->init_req(1, equals);
1329 region->init_req(1, control());
1330 }
1331 }
1332
1333 // post merge
1334 set_control(_gvn.transform(region));
1335 record_for_igvn(region);
1336
1337 set_result(_gvn.transform(phi));
1338 return true;
1339 }
1340
1341 //------------------------------inline_array_equals----------------------------
1342 bool LibraryCallKit::inline_array_equals() {
1343 Node* arg1 = argument(0);
1344 Node* arg2 = argument(1);
1345 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1346 return true;
1347 }
1348
1349 // Java version of String.indexOf(constant string)
1350 // class StringDecl {
1351 // StringDecl(char[] ca) {
1352 // offset = 0;
1353 // count = ca.length;
1354 // value = ca;
1355 // }
1356 // int offset;
1357 // int count;
1358 // char[] value;
1359 // }
1360 //
1361 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1362 // int targetOffset, int cache_i, int md2) {
1363 // int cache = cache_i;
1364 // int sourceOffset = string_object.offset;
1365 // int sourceCount = string_object.count;
1366 // int targetCount = target_object.length;
1367 //
1368 // int targetCountLess1 = targetCount - 1;
1369 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1370 //
1371 // char[] source = string_object.value;
1372 // char[] target = target_object;
1373 // int lastChar = target[targetCountLess1];
1374 //
1375 // outer_loop:
1376 // for (int i = sourceOffset; i < sourceEnd; ) {
1377 // int src = source[i + targetCountLess1];
1378 // if (src == lastChar) {
1379 // // With random strings and a 4-character alphabet,
1380 // // reverse matching at this point sets up 0.8% fewer
1381 // // frames, but (paradoxically) makes 0.3% more probes.
1382 // // Since those probes are nearer the lastChar probe,
1383 // // there is may be a net D$ win with reverse matching.
1384 // // But, reversing loop inhibits unroll of inner loop
1385 // // for unknown reason. So, does running outer loop from
1386 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1387 // for (int j = 0; j < targetCountLess1; j++) {
1388 // if (target[targetOffset + j] != source[i+j]) {
1389 // if ((cache & (1 << source[i+j])) == 0) {
1390 // if (md2 < j+1) {
1391 // i += j+1;
1392 // continue outer_loop;
1393 // }
1394 // }
1395 // i += md2;
1396 // continue outer_loop;
1397 // }
1398 // }
1399 // return i - sourceOffset;
1400 // }
1401 // if ((cache & (1 << src)) == 0) {
1402 // i += targetCountLess1;
1403 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1404 // i++;
1405 // }
1406 // return -1;
1407 // }
1408
1409 //------------------------------string_indexOf------------------------
1410 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1411 jint cache_i, jint md2_i) {
1412
1413 Node* no_ctrl = NULL;
1414 float likely = PROB_LIKELY(0.9);
1415 float unlikely = PROB_UNLIKELY(0.9);
1416
1417 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1418
1419 Node* source = load_String_value(no_ctrl, string_object);
1420 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1421 Node* sourceCount = load_String_length(no_ctrl, string_object);
1422
1423 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1424 jint target_length = target_array->length();
1425 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1426 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1427
1428 // String.value field is known to be @Stable.
1429 if (UseImplicitStableValues) {
1430 target = cast_array_to_stable(target, target_type);
1431 }
1432
1433 IdealKit kit(this, false, true);
1434 #define __ kit.
1435 Node* zero = __ ConI(0);
1436 Node* one = __ ConI(1);
1437 Node* cache = __ ConI(cache_i);
1438 Node* md2 = __ ConI(md2_i);
1439 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1440 Node* targetCountLess1 = __ ConI(target_length - 1);
1441 Node* targetOffset = __ ConI(targetOffset_i);
1442 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1443
1444 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1445 Node* outer_loop = __ make_label(2 /* goto */);
1446 Node* return_ = __ make_label(1);
1447
1448 __ set(rtn,__ ConI(-1));
1449 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1450 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1451 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1452 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1453 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1454 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1455 Node* tpj = __ AddI(targetOffset, __ value(j));
1456 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1457 Node* ipj = __ AddI(__ value(i), __ value(j));
1458 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1459 __ if_then(targ, BoolTest::ne, src2); {
1460 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1461 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1462 __ increment(i, __ AddI(__ value(j), one));
1463 __ goto_(outer_loop);
1464 } __ end_if(); __ dead(j);
1465 }__ end_if(); __ dead(j);
1466 __ increment(i, md2);
1467 __ goto_(outer_loop);
1468 }__ end_if();
1469 __ increment(j, one);
1470 }__ end_loop(); __ dead(j);
1471 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1472 __ goto_(return_);
1473 }__ end_if();
1474 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1475 __ increment(i, targetCountLess1);
1476 }__ end_if();
1477 __ increment(i, one);
1478 __ bind(outer_loop);
1479 }__ end_loop(); __ dead(i);
1480 __ bind(return_);
1481
1482 // Final sync IdealKit and GraphKit.
1483 final_sync(kit);
1484 Node* result = __ value(rtn);
1485 #undef __
1486 C->set_has_loops(true);
1487 return result;
1488 }
1489
1490 //------------------------------inline_string_indexOf------------------------
1491 bool LibraryCallKit::inline_string_indexOf() {
1492 Node* receiver = argument(0);
1493 Node* arg = argument(1);
1494
1495 Node* result;
1496 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1497 UseSSE42Intrinsics) {
1498 // Generate SSE4.2 version of indexOf
1499 // We currently only have match rules that use SSE4.2
1500
1501 receiver = null_check(receiver);
1502 arg = null_check(arg);
1503 if (stopped()) {
1504 return true;
1505 }
1506
1507 // Make the merge point
1508 RegionNode* result_rgn = new RegionNode(4);
1509 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1510 Node* no_ctrl = NULL;
1511
1512 // Get start addr of source string
1513 Node* source = load_String_value(no_ctrl, receiver);
1514 Node* source_offset = load_String_offset(no_ctrl, receiver);
1515 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1516
1517 // Get length of source string
1518 Node* source_cnt = load_String_length(no_ctrl, receiver);
1519
1520 // Get start addr of substring
1521 Node* substr = load_String_value(no_ctrl, arg);
1522 Node* substr_offset = load_String_offset(no_ctrl, arg);
1523 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1524
1525 // Get length of source string
1526 Node* substr_cnt = load_String_length(no_ctrl, arg);
1527
1528 // Check for substr count > string count
1529 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1530 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1531 Node* if_gt = generate_slow_guard(bol, NULL);
1532 if (if_gt != NULL) {
1533 result_phi->init_req(2, intcon(-1));
1534 result_rgn->init_req(2, if_gt);
1535 }
1536
1537 if (!stopped()) {
1538 // Check for substr count == 0
1539 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1540 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1541 Node* if_zero = generate_slow_guard(bol, NULL);
1542 if (if_zero != NULL) {
1543 result_phi->init_req(3, intcon(0));
1544 result_rgn->init_req(3, if_zero);
1545 }
1546 }
1547
1548 if (!stopped()) {
1549 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1550 result_phi->init_req(1, result);
1551 result_rgn->init_req(1, control());
1552 }
1553 set_control(_gvn.transform(result_rgn));
1554 record_for_igvn(result_rgn);
1555 result = _gvn.transform(result_phi);
1556
1557 } else { // Use LibraryCallKit::string_indexOf
1558 // don't intrinsify if argument isn't a constant string.
1559 if (!arg->is_Con()) {
1560 return false;
1561 }
1562 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1563 if (str_type == NULL) {
1564 return false;
1565 }
1566 ciInstanceKlass* klass = env()->String_klass();
1567 ciObject* str_const = str_type->const_oop();
1568 if (str_const == NULL || str_const->klass() != klass) {
1569 return false;
1570 }
1571 ciInstance* str = str_const->as_instance();
1572 assert(str != NULL, "must be instance");
1573
1574 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1575 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1576
1577 int o;
1578 int c;
1579 if (java_lang_String::has_offset_field()) {
1580 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1581 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1582 } else {
1583 o = 0;
1584 c = pat->length();
1585 }
1586
1587 // constant strings have no offset and count == length which
1588 // simplifies the resulting code somewhat so lets optimize for that.
1589 if (o != 0 || c != pat->length()) {
1590 return false;
1591 }
1592
1593 receiver = null_check(receiver, T_OBJECT);
1594 // NOTE: No null check on the argument is needed since it's a constant String oop.
1595 if (stopped()) {
1596 return true;
1597 }
1598
1599 // The null string as a pattern always returns 0 (match at beginning of string)
1600 if (c == 0) {
1601 set_result(intcon(0));
1602 return true;
1603 }
1604
1605 // Generate default indexOf
1606 jchar lastChar = pat->char_at(o + (c - 1));
1607 int cache = 0;
1608 int i;
1609 for (i = 0; i < c - 1; i++) {
1610 assert(i < pat->length(), "out of range");
1611 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1612 }
1613
1614 int md2 = c;
1615 for (i = 0; i < c - 1; i++) {
1616 assert(i < pat->length(), "out of range");
1617 if (pat->char_at(o + i) == lastChar) {
1618 md2 = (c - 1) - i;
1619 }
1620 }
1621
1622 result = string_indexOf(receiver, pat, o, cache, md2);
1623 }
1624 set_result(result);
1625 return true;
1626 }
1627
1628 //--------------------------round_double_node--------------------------------
1629 // Round a double node if necessary.
1630 Node* LibraryCallKit::round_double_node(Node* n) {
1631 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1632 n = _gvn.transform(new RoundDoubleNode(0, n));
1633 return n;
1634 }
1635
1636 //------------------------------inline_math-----------------------------------
1637 // public static double Math.abs(double)
1638 // public static double Math.sqrt(double)
1639 // public static double Math.log(double)
1640 // public static double Math.log10(double)
1641 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1642 Node* arg = round_double_node(argument(0));
1643 Node* n;
1644 switch (id) {
1645 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1646 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1647 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1648 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1649 default: fatal_unexpected_iid(id); break;
1650 }
1651 set_result(_gvn.transform(n));
1652 return true;
1653 }
1654
1655 //------------------------------inline_trig----------------------------------
1656 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1657 // argument reduction which will turn into a fast/slow diamond.
1658 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1659 Node* arg = round_double_node(argument(0));
1660 Node* n = NULL;
1661
1662 switch (id) {
1663 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1664 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1665 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1666 default: fatal_unexpected_iid(id); break;
1667 }
1668 n = _gvn.transform(n);
1669
1670 // Rounding required? Check for argument reduction!
1671 if (Matcher::strict_fp_requires_explicit_rounding) {
1672 static const double pi_4 = 0.7853981633974483;
1673 static const double neg_pi_4 = -0.7853981633974483;
1674 // pi/2 in 80-bit extended precision
1675 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1676 // -pi/2 in 80-bit extended precision
1677 // 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};
1678 // Cutoff value for using this argument reduction technique
1679 //static const double pi_2_minus_epsilon = 1.564660403643354;
1680 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1681
1682 // Pseudocode for sin:
1683 // if (x <= Math.PI / 4.0) {
1684 // if (x >= -Math.PI / 4.0) return fsin(x);
1685 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1686 // } else {
1687 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1688 // }
1689 // return StrictMath.sin(x);
1690
1691 // Pseudocode for cos:
1692 // if (x <= Math.PI / 4.0) {
1693 // if (x >= -Math.PI / 4.0) return fcos(x);
1694 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1695 // } else {
1696 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1697 // }
1698 // return StrictMath.cos(x);
1699
1700 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1701 // requires a special machine instruction to load it. Instead we'll try
1702 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1703 // probably do the math inside the SIN encoding.
1704
1705 // Make the merge point
1706 RegionNode* r = new RegionNode(3);
1707 Node* phi = new PhiNode(r, Type::DOUBLE);
1708
1709 // Flatten arg so we need only 1 test
1710 Node *abs = _gvn.transform(new AbsDNode(arg));
1711 // Node for PI/4 constant
1712 Node *pi4 = makecon(TypeD::make(pi_4));
1713 // Check PI/4 : abs(arg)
1714 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1715 // Check: If PI/4 < abs(arg) then go slow
1716 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1717 // Branch either way
1718 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1719 set_control(opt_iff(r,iff));
1720
1721 // Set fast path result
1722 phi->init_req(2, n);
1723
1724 // Slow path - non-blocking leaf call
1725 Node* call = NULL;
1726 switch (id) {
1727 case vmIntrinsics::_dsin:
1728 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1729 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1730 "Sin", NULL, arg, top());
1731 break;
1732 case vmIntrinsics::_dcos:
1733 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1734 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1735 "Cos", NULL, arg, top());
1736 break;
1737 case vmIntrinsics::_dtan:
1738 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1739 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1740 "Tan", NULL, arg, top());
1741 break;
1742 }
1743 assert(control()->in(0) == call, "");
1744 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1745 r->init_req(1, control());
1746 phi->init_req(1, slow_result);
1747
1748 // Post-merge
1749 set_control(_gvn.transform(r));
1750 record_for_igvn(r);
1751 n = _gvn.transform(phi);
1752
1753 C->set_has_split_ifs(true); // Has chance for split-if optimization
1754 }
1755 set_result(n);
1756 return true;
1757 }
1758
1759 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1760 //-------------------
1761 //result=(result.isNaN())? funcAddr():result;
1762 // Check: If isNaN() by checking result!=result? then either trap
1763 // or go to runtime
1764 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1765 // Build the boolean node
1766 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1767
1768 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1769 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1770 // The pow or exp intrinsic returned a NaN, which requires a call
1771 // to the runtime. Recompile with the runtime call.
1772 uncommon_trap(Deoptimization::Reason_intrinsic,
1773 Deoptimization::Action_make_not_entrant);
1774 }
1775 return result;
1776 } else {
1777 // If this inlining ever returned NaN in the past, we compile a call
1778 // to the runtime to properly handle corner cases
1779
1780 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1781 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1782 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1783
1784 if (!if_slow->is_top()) {
1785 RegionNode* result_region = new RegionNode(3);
1786 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1787
1788 result_region->init_req(1, if_fast);
1789 result_val->init_req(1, result);
1790
1791 set_control(if_slow);
1792
1793 const TypePtr* no_memory_effects = NULL;
1794 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1795 no_memory_effects,
1796 x, top(), y, y ? top() : NULL);
1797 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1798 #ifdef ASSERT
1799 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1800 assert(value_top == top(), "second value must be top");
1801 #endif
1802
1803 result_region->init_req(2, control());
1804 result_val->init_req(2, value);
1805 set_control(_gvn.transform(result_region));
1806 return _gvn.transform(result_val);
1807 } else {
1808 return result;
1809 }
1810 }
1811 }
1812
1813 //------------------------------inline_exp-------------------------------------
1814 // Inline exp instructions, if possible. The Intel hardware only misses
1815 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1816 bool LibraryCallKit::inline_exp() {
1817 Node* arg = round_double_node(argument(0));
1818 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1819
1820 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1821 set_result(n);
1822
1823 C->set_has_split_ifs(true); // Has chance for split-if optimization
1824 return true;
1825 }
1826
1827 //------------------------------inline_pow-------------------------------------
1828 // Inline power instructions, if possible.
1829 bool LibraryCallKit::inline_pow() {
1830 // Pseudocode for pow
1831 // if (y == 2) {
1832 // return x * x;
1833 // } else {
1834 // if (x <= 0.0) {
1835 // long longy = (long)y;
1836 // if ((double)longy == y) { // if y is long
1837 // if (y + 1 == y) longy = 0; // huge number: even
1838 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1839 // } else {
1840 // result = NaN;
1841 // }
1842 // } else {
1843 // result = DPow(x,y);
1844 // }
1845 // if (result != result)? {
1846 // result = uncommon_trap() or runtime_call();
1847 // }
1848 // return result;
1849 // }
1850
1851 Node* x = round_double_node(argument(0));
1852 Node* y = round_double_node(argument(2));
1853
1854 Node* result = NULL;
1855
1856 Node* const_two_node = makecon(TypeD::make(2.0));
1857 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1858 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1859 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1860 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1861 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1862
1863 RegionNode* region_node = new RegionNode(3);
1864 region_node->init_req(1, if_true);
1865
1866 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1867 // special case for x^y where y == 2, we can convert it to x * x
1868 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1869
1870 // set control to if_false since we will now process the false branch
1871 set_control(if_false);
1872
1873 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1874 // Short form: skip the fancy tests and just check for NaN result.
1875 result = _gvn.transform(new PowDNode(C, control(), x, y));
1876 } else {
1877 // If this inlining ever returned NaN in the past, include all
1878 // checks + call to the runtime.
1879
1880 // Set the merge point for If node with condition of (x <= 0.0)
1881 // There are four possible paths to region node and phi node
1882 RegionNode *r = new RegionNode(4);
1883 Node *phi = new PhiNode(r, Type::DOUBLE);
1884
1885 // Build the first if node: if (x <= 0.0)
1886 // Node for 0 constant
1887 Node *zeronode = makecon(TypeD::ZERO);
1888 // Check x:0
1889 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1890 // Check: If (x<=0) then go complex path
1891 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1892 // Branch either way
1893 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1894 // Fast path taken; set region slot 3
1895 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1896 r->init_req(3,fast_taken); // Capture fast-control
1897
1898 // Fast path not-taken, i.e. slow path
1899 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1900
1901 // Set fast path result
1902 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1903 phi->init_req(3, fast_result);
1904
1905 // Complex path
1906 // Build the second if node (if y is long)
1907 // Node for (long)y
1908 Node *longy = _gvn.transform(new ConvD2LNode(y));
1909 // Node for (double)((long) y)
1910 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1911 // Check (double)((long) y) : y
1912 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1913 // Check if (y isn't long) then go to slow path
1914
1915 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1916 // Branch either way
1917 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1918 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1919
1920 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1921
1922 // Calculate DPow(abs(x), y)*(1 & (long)y)
1923 // Node for constant 1
1924 Node *conone = longcon(1);
1925 // 1& (long)y
1926 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1927
1928 // A huge number is always even. Detect a huge number by checking
1929 // if y + 1 == y and set integer to be tested for parity to 0.
1930 // Required for corner case:
1931 // (long)9.223372036854776E18 = max_jlong
1932 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1933 // max_jlong is odd but 9.223372036854776E18 is even
1934 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1935 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1936 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1937 Node* correctedsign = NULL;
1938 if (ConditionalMoveLimit != 0) {
1939 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1940 } else {
1941 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1942 RegionNode *r = new RegionNode(3);
1943 Node *phi = new PhiNode(r, TypeLong::LONG);
1944 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1945 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1946 phi->init_req(1, signnode);
1947 phi->init_req(2, longcon(0));
1948 correctedsign = _gvn.transform(phi);
1949 ylong_path = _gvn.transform(r);
1950 record_for_igvn(r);
1951 }
1952
1953 // zero node
1954 Node *conzero = longcon(0);
1955 // Check (1&(long)y)==0?
1956 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1957 // Check if (1&(long)y)!=0?, if so the result is negative
1958 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1959 // abs(x)
1960 Node *absx=_gvn.transform(new AbsDNode(x));
1961 // abs(x)^y
1962 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1963 // -abs(x)^y
1964 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1965 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1966 Node *signresult = NULL;
1967 if (ConditionalMoveLimit != 0) {
1968 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1969 } else {
1970 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1971 RegionNode *r = new RegionNode(3);
1972 Node *phi = new PhiNode(r, Type::DOUBLE);
1973 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1974 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1975 phi->init_req(1, absxpowy);
1976 phi->init_req(2, negabsxpowy);
1977 signresult = _gvn.transform(phi);
1978 ylong_path = _gvn.transform(r);
1979 record_for_igvn(r);
1980 }
1981 // Set complex path fast result
1982 r->init_req(2, ylong_path);
1983 phi->init_req(2, signresult);
1984
1985 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1986 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1987 r->init_req(1,slow_path);
1988 phi->init_req(1,slow_result);
1989
1990 // Post merge
1991 set_control(_gvn.transform(r));
1992 record_for_igvn(r);
1993 result = _gvn.transform(phi);
1994 }
1995
1996 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1997
1998 // control from finish_pow_exp is now input to the region node
1999 region_node->set_req(2, control());
2000 // the result from finish_pow_exp is now input to the phi node
2001 phi_node->init_req(2, result);
2002 set_control(_gvn.transform(region_node));
2003 record_for_igvn(region_node);
2004 set_result(_gvn.transform(phi_node));
2005
2006 C->set_has_split_ifs(true); // Has chance for split-if optimization
2007 return true;
2008 }
2009
2010 //------------------------------runtime_math-----------------------------
2011 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2012 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2013 "must be (DD)D or (D)D type");
2014
2015 // Inputs
2016 Node* a = round_double_node(argument(0));
2017 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2018
2019 const TypePtr* no_memory_effects = NULL;
2020 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2021 no_memory_effects,
2022 a, top(), b, b ? top() : NULL);
2023 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
2024 #ifdef ASSERT
2025 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
2026 assert(value_top == top(), "second value must be top");
2027 #endif
2028
2029 set_result(value);
2030 return true;
2031 }
2032
2033 //------------------------------inline_math_native-----------------------------
2034 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2035 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2036 switch (id) {
2037 // These intrinsics are not properly supported on all hardware
2038 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2039 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2040 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2041 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2042 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2043 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2044
2045 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2046 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2047 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2048 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2049
2050 // These intrinsics are supported on all hardware
2051 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2052 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2053
2054 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2055 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2056 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2057 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2058 #undef FN_PTR
2059
2060 // These intrinsics are not yet correctly implemented
2061 case vmIntrinsics::_datan2:
2062 return false;
2063
2064 default:
2065 fatal_unexpected_iid(id);
2066 return false;
2067 }
2068 }
2069
2070 static bool is_simple_name(Node* n) {
2071 return (n->req() == 1 // constant
2072 || (n->is_Type() && n->as_Type()->type()->singleton())
2073 || n->is_Proj() // parameter or return value
2074 || n->is_Phi() // local of some sort
2075 );
2076 }
2077
2078 //----------------------------inline_min_max-----------------------------------
2079 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2080 set_result(generate_min_max(id, argument(0), argument(1)));
2081 return true;
2082 }
2083
2084 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2085 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2086 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2087 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2088 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2089
2090 {
2091 PreserveJVMState pjvms(this);
2092 PreserveReexecuteState preexecs(this);
2093 jvms()->set_should_reexecute(true);
2094
2095 set_control(slow_path);
2096 set_i_o(i_o());
2097
2098 uncommon_trap(Deoptimization::Reason_intrinsic,
2099 Deoptimization::Action_none);
2100 }
2101
2102 set_control(fast_path);
2103 set_result(math);
2104 }
2105
2106 template <typename OverflowOp>
2107 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2108 typedef typename OverflowOp::MathOp MathOp;
2109
2110 MathOp* mathOp = new MathOp(arg1, arg2);
2111 Node* operation = _gvn.transform( mathOp );
2112 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2113 inline_math_mathExact(operation, ofcheck);
2114 return true;
2115 }
2116
2117 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2118 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2119 }
2120
2121 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2122 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2123 }
2124
2125 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2126 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2127 }
2128
2129 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2130 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2131 }
2132
2133 bool LibraryCallKit::inline_math_negateExactI() {
2134 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2135 }
2136
2137 bool LibraryCallKit::inline_math_negateExactL() {
2138 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2139 }
2140
2141 bool LibraryCallKit::inline_math_multiplyExactI() {
2142 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2143 }
2144
2145 bool LibraryCallKit::inline_math_multiplyExactL() {
2146 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2147 }
2148
2149 Node*
2150 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2151 // These are the candidate return value:
2152 Node* xvalue = x0;
2153 Node* yvalue = y0;
2154
2155 if (xvalue == yvalue) {
2156 return xvalue;
2157 }
2158
2159 bool want_max = (id == vmIntrinsics::_max);
2160
2161 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2162 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2163 if (txvalue == NULL || tyvalue == NULL) return top();
2164 // This is not really necessary, but it is consistent with a
2165 // hypothetical MaxINode::Value method:
2166 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2167
2168 // %%% This folding logic should (ideally) be in a different place.
2169 // Some should be inside IfNode, and there to be a more reliable
2170 // transformation of ?: style patterns into cmoves. We also want
2171 // more powerful optimizations around cmove and min/max.
2172
2173 // Try to find a dominating comparison of these guys.
2174 // It can simplify the index computation for Arrays.copyOf
2175 // and similar uses of System.arraycopy.
2176 // First, compute the normalized version of CmpI(x, y).
2177 int cmp_op = Op_CmpI;
2178 Node* xkey = xvalue;
2179 Node* ykey = yvalue;
2180 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2181 if (ideal_cmpxy->is_Cmp()) {
2182 // E.g., if we have CmpI(length - offset, count),
2183 // it might idealize to CmpI(length, count + offset)
2184 cmp_op = ideal_cmpxy->Opcode();
2185 xkey = ideal_cmpxy->in(1);
2186 ykey = ideal_cmpxy->in(2);
2187 }
2188
2189 // Start by locating any relevant comparisons.
2190 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2191 Node* cmpxy = NULL;
2192 Node* cmpyx = NULL;
2193 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2194 Node* cmp = start_from->fast_out(k);
2195 if (cmp->outcnt() > 0 && // must have prior uses
2196 cmp->in(0) == NULL && // must be context-independent
2197 cmp->Opcode() == cmp_op) { // right kind of compare
2198 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2199 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2200 }
2201 }
2202
2203 const int NCMPS = 2;
2204 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2205 int cmpn;
2206 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2207 if (cmps[cmpn] != NULL) break; // find a result
2208 }
2209 if (cmpn < NCMPS) {
2210 // Look for a dominating test that tells us the min and max.
2211 int depth = 0; // Limit search depth for speed
2212 Node* dom = control();
2213 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2214 if (++depth >= 100) break;
2215 Node* ifproj = dom;
2216 if (!ifproj->is_Proj()) continue;
2217 Node* iff = ifproj->in(0);
2218 if (!iff->is_If()) continue;
2219 Node* bol = iff->in(1);
2220 if (!bol->is_Bool()) continue;
2221 Node* cmp = bol->in(1);
2222 if (cmp == NULL) continue;
2223 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2224 if (cmps[cmpn] == cmp) break;
2225 if (cmpn == NCMPS) continue;
2226 BoolTest::mask btest = bol->as_Bool()->_test._test;
2227 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2228 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2229 // At this point, we know that 'x btest y' is true.
2230 switch (btest) {
2231 case BoolTest::eq:
2232 // They are proven equal, so we can collapse the min/max.
2233 // Either value is the answer. Choose the simpler.
2234 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2235 return yvalue;
2236 return xvalue;
2237 case BoolTest::lt: // x < y
2238 case BoolTest::le: // x <= y
2239 return (want_max ? yvalue : xvalue);
2240 case BoolTest::gt: // x > y
2241 case BoolTest::ge: // x >= y
2242 return (want_max ? xvalue : yvalue);
2243 }
2244 }
2245 }
2246
2247 // We failed to find a dominating test.
2248 // Let's pick a test that might GVN with prior tests.
2249 Node* best_bol = NULL;
2250 BoolTest::mask best_btest = BoolTest::illegal;
2251 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2252 Node* cmp = cmps[cmpn];
2253 if (cmp == NULL) continue;
2254 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2255 Node* bol = cmp->fast_out(j);
2256 if (!bol->is_Bool()) continue;
2257 BoolTest::mask btest = bol->as_Bool()->_test._test;
2258 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2259 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2260 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2261 best_bol = bol->as_Bool();
2262 best_btest = btest;
2263 }
2264 }
2265 }
2266
2267 Node* answer_if_true = NULL;
2268 Node* answer_if_false = NULL;
2269 switch (best_btest) {
2270 default:
2271 if (cmpxy == NULL)
2272 cmpxy = ideal_cmpxy;
2273 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2274 // and fall through:
2275 case BoolTest::lt: // x < y
2276 case BoolTest::le: // x <= y
2277 answer_if_true = (want_max ? yvalue : xvalue);
2278 answer_if_false = (want_max ? xvalue : yvalue);
2279 break;
2280 case BoolTest::gt: // x > y
2281 case BoolTest::ge: // x >= y
2282 answer_if_true = (want_max ? xvalue : yvalue);
2283 answer_if_false = (want_max ? yvalue : xvalue);
2284 break;
2285 }
2286
2287 jint hi, lo;
2288 if (want_max) {
2289 // We can sharpen the minimum.
2290 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2291 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2292 } else {
2293 // We can sharpen the maximum.
2294 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2295 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2296 }
2297
2298 // Use a flow-free graph structure, to avoid creating excess control edges
2299 // which could hinder other optimizations.
2300 // Since Math.min/max is often used with arraycopy, we want
2301 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2302 Node* cmov = CMoveNode::make(NULL, best_bol,
2303 answer_if_false, answer_if_true,
2304 TypeInt::make(lo, hi, widen));
2305
2306 return _gvn.transform(cmov);
2307
2308 /*
2309 // This is not as desirable as it may seem, since Min and Max
2310 // nodes do not have a full set of optimizations.
2311 // And they would interfere, anyway, with 'if' optimizations
2312 // and with CMoveI canonical forms.
2313 switch (id) {
2314 case vmIntrinsics::_min:
2315 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2316 case vmIntrinsics::_max:
2317 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2318 default:
2319 ShouldNotReachHere();
2320 }
2321 */
2322 }
2323
2324 inline int
2325 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2326 const TypePtr* base_type = TypePtr::NULL_PTR;
2327 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2328 if (base_type == NULL) {
2329 // Unknown type.
2330 return Type::AnyPtr;
2331 } else if (base_type == TypePtr::NULL_PTR) {
2332 // Since this is a NULL+long form, we have to switch to a rawptr.
2333 base = _gvn.transform(new CastX2PNode(offset));
2334 offset = MakeConX(0);
2335 return Type::RawPtr;
2336 } else if (base_type->base() == Type::RawPtr) {
2337 return Type::RawPtr;
2338 } else if (base_type->isa_oopptr()) {
2339 // Base is never null => always a heap address.
2340 if (base_type->ptr() == TypePtr::NotNull) {
2341 return Type::OopPtr;
2342 }
2343 // Offset is small => always a heap address.
2344 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2345 if (offset_type != NULL &&
2346 base_type->offset() == 0 && // (should always be?)
2347 offset_type->_lo >= 0 &&
2348 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2349 return Type::OopPtr;
2350 }
2351 // Otherwise, it might either be oop+off or NULL+addr.
2352 return Type::AnyPtr;
2353 } else {
2354 // No information:
2355 return Type::AnyPtr;
2356 }
2357 }
2358
2359 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2360 int kind = classify_unsafe_addr(base, offset);
2361 if (kind == Type::RawPtr) {
2362 return basic_plus_adr(top(), base, offset);
2363 } else {
2364 return basic_plus_adr(base, offset);
2365 }
2366 }
2367
2368 //--------------------------inline_number_methods-----------------------------
2369 // inline int Integer.numberOfLeadingZeros(int)
2370 // inline int Long.numberOfLeadingZeros(long)
2371 //
2372 // inline int Integer.numberOfTrailingZeros(int)
2373 // inline int Long.numberOfTrailingZeros(long)
2374 //
2375 // inline int Integer.bitCount(int)
2376 // inline int Long.bitCount(long)
2377 //
2378 // inline char Character.reverseBytes(char)
2379 // inline short Short.reverseBytes(short)
2380 // inline int Integer.reverseBytes(int)
2381 // inline long Long.reverseBytes(long)
2382 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2383 Node* arg = argument(0);
2384 Node* n;
2385 switch (id) {
2386 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2387 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2388 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2389 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2390 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2391 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2392 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2393 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2394 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2395 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2396 default: fatal_unexpected_iid(id); break;
2397 }
2398 set_result(_gvn.transform(n));
2399 return true;
2400 }
2401
2402 //----------------------------inline_unsafe_access----------------------------
2403
2404 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2405
2406 // Helper that guards and inserts a pre-barrier.
2407 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2408 Node* pre_val, bool need_mem_bar) {
2409 // We could be accessing the referent field of a reference object. If so, when G1
2410 // is enabled, we need to log the value in the referent field in an SATB buffer.
2411 // This routine performs some compile time filters and generates suitable
2412 // runtime filters that guard the pre-barrier code.
2413 // Also add memory barrier for non volatile load from the referent field
2414 // to prevent commoning of loads across safepoint.
2415 if (!UseG1GC && !need_mem_bar)
2416 return;
2417
2418 // Some compile time checks.
2419
2420 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2421 const TypeX* otype = offset->find_intptr_t_type();
2422 if (otype != NULL && otype->is_con() &&
2423 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2424 // Constant offset but not the reference_offset so just return
2425 return;
2426 }
2427
2428 // We only need to generate the runtime guards for instances.
2429 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2430 if (btype != NULL) {
2431 if (btype->isa_aryptr()) {
2432 // Array type so nothing to do
2433 return;
2434 }
2435
2436 const TypeInstPtr* itype = btype->isa_instptr();
2437 if (itype != NULL) {
2438 // Can the klass of base_oop be statically determined to be
2439 // _not_ a sub-class of Reference and _not_ Object?
2440 ciKlass* klass = itype->klass();
2441 if ( klass->is_loaded() &&
2442 !klass->is_subtype_of(env()->Reference_klass()) &&
2443 !env()->Object_klass()->is_subtype_of(klass)) {
2444 return;
2445 }
2446 }
2447 }
2448
2449 // The compile time filters did not reject base_oop/offset so
2450 // we need to generate the following runtime filters
2451 //
2452 // if (offset == java_lang_ref_Reference::_reference_offset) {
2453 // if (instance_of(base, java.lang.ref.Reference)) {
2454 // pre_barrier(_, pre_val, ...);
2455 // }
2456 // }
2457
2458 float likely = PROB_LIKELY( 0.999);
2459 float unlikely = PROB_UNLIKELY(0.999);
2460
2461 IdealKit ideal(this);
2462 #define __ ideal.
2463
2464 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2465
2466 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2467 // Update graphKit memory and control from IdealKit.
2468 sync_kit(ideal);
2469
2470 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2471 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2472
2473 // Update IdealKit memory and control from graphKit.
2474 __ sync_kit(this);
2475
2476 Node* one = __ ConI(1);
2477 // is_instof == 0 if base_oop == NULL
2478 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2479
2480 // Update graphKit from IdeakKit.
2481 sync_kit(ideal);
2482
2483 // Use the pre-barrier to record the value in the referent field
2484 pre_barrier(false /* do_load */,
2485 __ ctrl(),
2486 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2487 pre_val /* pre_val */,
2488 T_OBJECT);
2489 if (need_mem_bar) {
2490 // Add memory barrier to prevent commoning reads from this field
2491 // across safepoint since GC can change its value.
2492 insert_mem_bar(Op_MemBarCPUOrder);
2493 }
2494 // Update IdealKit from graphKit.
2495 __ sync_kit(this);
2496
2497 } __ end_if(); // _ref_type != ref_none
2498 } __ end_if(); // offset == referent_offset
2499
2500 // Final sync IdealKit and GraphKit.
2501 final_sync(ideal);
2502 #undef __
2503 }
2504
2505
2506 // Interpret Unsafe.fieldOffset cookies correctly:
2507 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2508
2509 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2510 // Attempt to infer a sharper value type from the offset and base type.
2511 ciKlass* sharpened_klass = NULL;
2512
2513 // See if it is an instance field, with an object type.
2514 if (alias_type->field() != NULL) {
2515 assert(!is_native_ptr, "native pointer op cannot use a java address");
2516 if (alias_type->field()->type()->is_klass()) {
2517 sharpened_klass = alias_type->field()->type()->as_klass();
2518 }
2519 }
2520
2521 // See if it is a narrow oop array.
2522 if (adr_type->isa_aryptr()) {
2523 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2524 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2525 if (elem_type != NULL) {
2526 sharpened_klass = elem_type->klass();
2527 }
2528 }
2529 }
2530
2531 // The sharpened class might be unloaded if there is no class loader
2532 // contraint in place.
2533 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2534 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2535
2536 #ifndef PRODUCT
2537 if (C->print_intrinsics() || C->print_inlining()) {
2538 tty->print(" from base type: "); adr_type->dump();
2539 tty->print(" sharpened value: "); tjp->dump();
2540 }
2541 #endif
2542 // Sharpen the value type.
2543 return tjp;
2544 }
2545 return NULL;
2546 }
2547
2548 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2549 if (callee()->is_static()) return false; // caller must have the capability!
2550
2551 #ifndef PRODUCT
2552 {
2553 ResourceMark rm;
2554 // Check the signatures.
2555 ciSignature* sig = callee()->signature();
2556 #ifdef ASSERT
2557 if (!is_store) {
2558 // Object getObject(Object base, int/long offset), etc.
2559 BasicType rtype = sig->return_type()->basic_type();
2560 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2561 rtype = T_ADDRESS; // it is really a C void*
2562 assert(rtype == type, "getter must return the expected value");
2563 if (!is_native_ptr) {
2564 assert(sig->count() == 2, "oop getter has 2 arguments");
2565 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2566 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2567 } else {
2568 assert(sig->count() == 1, "native getter has 1 argument");
2569 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2570 }
2571 } else {
2572 // void putObject(Object base, int/long offset, Object x), etc.
2573 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2574 if (!is_native_ptr) {
2575 assert(sig->count() == 3, "oop putter has 3 arguments");
2576 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2577 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2578 } else {
2579 assert(sig->count() == 2, "native putter has 2 arguments");
2580 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2581 }
2582 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2583 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2584 vtype = T_ADDRESS; // it is really a C void*
2585 assert(vtype == type, "putter must accept the expected value");
2586 }
2587 #endif // ASSERT
2588 }
2589 #endif //PRODUCT
2590
2591 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2592
2593 Node* receiver = argument(0); // type: oop
2594
2595 // Build address expression.
2596 Node* adr;
2597 Node* heap_base_oop = top();
2598 Node* offset = top();
2599 Node* val;
2600
2601 if (!is_native_ptr) {
2602 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2603 Node* base = argument(1); // type: oop
2604 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2605 offset = argument(2); // type: long
2606 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2607 // to be plain byte offsets, which are also the same as those accepted
2608 // by oopDesc::field_base.
2609 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2610 "fieldOffset must be byte-scaled");
2611 // 32-bit machines ignore the high half!
2612 offset = ConvL2X(offset);
2613 adr = make_unsafe_address(base, offset);
2614 heap_base_oop = base;
2615 val = is_store ? argument(4) : NULL;
2616 } else {
2617 Node* ptr = argument(1); // type: long
2618 ptr = ConvL2X(ptr); // adjust Java long to machine word
2619 adr = make_unsafe_address(NULL, ptr);
2620 val = is_store ? argument(3) : NULL;
2621 }
2622
2623 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2624
2625 // First guess at the value type.
2626 const Type *value_type = Type::get_const_basic_type(type);
2627
2628 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2629 // there was not enough information to nail it down.
2630 Compile::AliasType* alias_type = C->alias_type(adr_type);
2631 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2632
2633 // We will need memory barriers unless we can determine a unique
2634 // alias category for this reference. (Note: If for some reason
2635 // the barriers get omitted and the unsafe reference begins to "pollute"
2636 // the alias analysis of the rest of the graph, either Compile::can_alias
2637 // or Compile::must_alias will throw a diagnostic assert.)
2638 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2639
2640 // If we are reading the value of the referent field of a Reference
2641 // object (either by using Unsafe directly or through reflection)
2642 // then, if G1 is enabled, we need to record the referent in an
2643 // SATB log buffer using the pre-barrier mechanism.
2644 // Also we need to add memory barrier to prevent commoning reads
2645 // from this field across safepoint since GC can change its value.
2646 bool need_read_barrier = !is_native_ptr && !is_store &&
2647 offset != top() && heap_base_oop != top();
2648
2649 if (!is_store && type == T_OBJECT) {
2650 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2651 if (tjp != NULL) {
2652 value_type = tjp;
2653 }
2654 }
2655
2656 receiver = null_check(receiver);
2657 if (stopped()) {
2658 return true;
2659 }
2660 // Heap pointers get a null-check from the interpreter,
2661 // as a courtesy. However, this is not guaranteed by Unsafe,
2662 // and it is not possible to fully distinguish unintended nulls
2663 // from intended ones in this API.
2664
2665 if (is_volatile) {
2666 // We need to emit leading and trailing CPU membars (see below) in
2667 // addition to memory membars when is_volatile. This is a little
2668 // too strong, but avoids the need to insert per-alias-type
2669 // volatile membars (for stores; compare Parse::do_put_xxx), which
2670 // we cannot do effectively here because we probably only have a
2671 // rough approximation of type.
2672 need_mem_bar = true;
2673 // For Stores, place a memory ordering barrier now.
2674 if (is_store) {
2675 insert_mem_bar(Op_MemBarRelease);
2676 } else {
2677 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2678 insert_mem_bar(Op_MemBarVolatile);
2679 }
2680 }
2681 }
2682
2683 // Memory barrier to prevent normal and 'unsafe' accesses from
2684 // bypassing each other. Happens after null checks, so the
2685 // exception paths do not take memory state from the memory barrier,
2686 // so there's no problems making a strong assert about mixing users
2687 // of safe & unsafe memory.
2688 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2689
2690 if (!is_store) {
2691 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2692 // To be valid, unsafe loads may depend on other conditions than
2693 // the one that guards them: pin the Load node
2694 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2695 // load value
2696 switch (type) {
2697 case T_BOOLEAN:
2698 case T_CHAR:
2699 case T_BYTE:
2700 case T_SHORT:
2701 case T_INT:
2702 case T_LONG:
2703 case T_FLOAT:
2704 case T_DOUBLE:
2705 break;
2706 case T_OBJECT:
2707 if (need_read_barrier) {
2708 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2709 }
2710 break;
2711 case T_ADDRESS:
2712 // Cast to an int type.
2713 p = _gvn.transform(new CastP2XNode(NULL, p));
2714 p = ConvX2UL(p);
2715 break;
2716 default:
2717 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2718 break;
2719 }
2720 // The load node has the control of the preceding MemBarCPUOrder. All
2721 // following nodes will have the control of the MemBarCPUOrder inserted at
2722 // the end of this method. So, pushing the load onto the stack at a later
2723 // point is fine.
2724 set_result(p);
2725 } else {
2726 // place effect of store into memory
2727 switch (type) {
2728 case T_DOUBLE:
2729 val = dstore_rounding(val);
2730 break;
2731 case T_ADDRESS:
2732 // Repackage the long as a pointer.
2733 val = ConvL2X(val);
2734 val = _gvn.transform(new CastX2PNode(val));
2735 break;
2736 }
2737
2738 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2739 if (type != T_OBJECT ) {
2740 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2741 } else {
2742 // Possibly an oop being stored to Java heap or native memory
2743 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2744 // oop to Java heap.
2745 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2746 } else {
2747 // We can't tell at compile time if we are storing in the Java heap or outside
2748 // of it. So we need to emit code to conditionally do the proper type of
2749 // store.
2750
2751 IdealKit ideal(this);
2752 #define __ ideal.
2753 // QQQ who knows what probability is here??
2754 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2755 // Sync IdealKit and graphKit.
2756 sync_kit(ideal);
2757 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2758 // Update IdealKit memory.
2759 __ sync_kit(this);
2760 } __ else_(); {
2761 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2762 } __ end_if();
2763 // Final sync IdealKit and GraphKit.
2764 final_sync(ideal);
2765 #undef __
2766 }
2767 }
2768 }
2769
2770 if (is_volatile) {
2771 if (!is_store) {
2772 insert_mem_bar(Op_MemBarAcquire);
2773 } else {
2774 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2775 insert_mem_bar(Op_MemBarVolatile);
2776 }
2777 }
2778 }
2779
2780 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2781
2782 return true;
2783 }
2784
2785 //----------------------------inline_unsafe_load_store----------------------------
2786 // This method serves a couple of different customers (depending on LoadStoreKind):
2787 //
2788 // LS_cmpxchg:
2789 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2790 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2791 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2792 //
2793 // LS_xadd:
2794 // public int getAndAddInt( Object o, long offset, int delta)
2795 // public long getAndAddLong(Object o, long offset, long delta)
2796 //
2797 // LS_xchg:
2798 // int getAndSet(Object o, long offset, int newValue)
2799 // long getAndSet(Object o, long offset, long newValue)
2800 // Object getAndSet(Object o, long offset, Object newValue)
2801 //
2802 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2803 // This basic scheme here is the same as inline_unsafe_access, but
2804 // differs in enough details that combining them would make the code
2805 // overly confusing. (This is a true fact! I originally combined
2806 // them, but even I was confused by it!) As much code/comments as
2807 // possible are retained from inline_unsafe_access though to make
2808 // the correspondences clearer. - dl
2809
2810 if (callee()->is_static()) return false; // caller must have the capability!
2811
2812 #ifndef PRODUCT
2813 BasicType rtype;
2814 {
2815 ResourceMark rm;
2816 // Check the signatures.
2817 ciSignature* sig = callee()->signature();
2818 rtype = sig->return_type()->basic_type();
2819 if (kind == LS_xadd || kind == LS_xchg) {
2820 // Check the signatures.
2821 #ifdef ASSERT
2822 assert(rtype == type, "get and set must return the expected type");
2823 assert(sig->count() == 3, "get and set has 3 arguments");
2824 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2825 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2826 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2827 #endif // ASSERT
2828 } else if (kind == LS_cmpxchg) {
2829 // Check the signatures.
2830 #ifdef ASSERT
2831 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2832 assert(sig->count() == 4, "CAS has 4 arguments");
2833 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2834 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2835 #endif // ASSERT
2836 } else {
2837 ShouldNotReachHere();
2838 }
2839 }
2840 #endif //PRODUCT
2841
2842 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2843
2844 // Get arguments:
2845 Node* receiver = NULL;
2846 Node* base = NULL;
2847 Node* offset = NULL;
2848 Node* oldval = NULL;
2849 Node* newval = NULL;
2850 if (kind == LS_cmpxchg) {
2851 const bool two_slot_type = type2size[type] == 2;
2852 receiver = argument(0); // type: oop
2853 base = argument(1); // type: oop
2854 offset = argument(2); // type: long
2855 oldval = argument(4); // type: oop, int, or long
2856 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2857 } else if (kind == LS_xadd || kind == LS_xchg){
2858 receiver = argument(0); // type: oop
2859 base = argument(1); // type: oop
2860 offset = argument(2); // type: long
2861 oldval = NULL;
2862 newval = argument(4); // type: oop, int, or long
2863 }
2864
2865 // Null check receiver.
2866 receiver = null_check(receiver);
2867 if (stopped()) {
2868 return true;
2869 }
2870
2871 // Build field offset expression.
2872 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2873 // to be plain byte offsets, which are also the same as those accepted
2874 // by oopDesc::field_base.
2875 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2876 // 32-bit machines ignore the high half of long offsets
2877 offset = ConvL2X(offset);
2878 Node* adr = make_unsafe_address(base, offset);
2879 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2880
2881 // For CAS, unlike inline_unsafe_access, there seems no point in
2882 // trying to refine types. Just use the coarse types here.
2883 const Type *value_type = Type::get_const_basic_type(type);
2884 Compile::AliasType* alias_type = C->alias_type(adr_type);
2885 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2886
2887 if (kind == LS_xchg && type == T_OBJECT) {
2888 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2889 if (tjp != NULL) {
2890 value_type = tjp;
2891 }
2892 }
2893
2894 int alias_idx = C->get_alias_index(adr_type);
2895
2896 // Memory-model-wise, a LoadStore acts like a little synchronized
2897 // block, so needs barriers on each side. These don't translate
2898 // into actual barriers on most machines, but we still need rest of
2899 // compiler to respect ordering.
2900
2901 insert_mem_bar(Op_MemBarRelease);
2902 insert_mem_bar(Op_MemBarCPUOrder);
2903
2904 // 4984716: MemBars must be inserted before this
2905 // memory node in order to avoid a false
2906 // dependency which will confuse the scheduler.
2907 Node *mem = memory(alias_idx);
2908
2909 // For now, we handle only those cases that actually exist: ints,
2910 // longs, and Object. Adding others should be straightforward.
2911 Node* load_store;
2912 switch(type) {
2913 case T_INT:
2914 if (kind == LS_xadd) {
2915 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2916 } else if (kind == LS_xchg) {
2917 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2918 } else if (kind == LS_cmpxchg) {
2919 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2920 } else {
2921 ShouldNotReachHere();
2922 }
2923 break;
2924 case T_LONG:
2925 if (kind == LS_xadd) {
2926 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2927 } else if (kind == LS_xchg) {
2928 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2929 } else if (kind == LS_cmpxchg) {
2930 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2931 } else {
2932 ShouldNotReachHere();
2933 }
2934 break;
2935 case T_OBJECT:
2936 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2937 // could be delayed during Parse (for example, in adjust_map_after_if()).
2938 // Execute transformation here to avoid barrier generation in such case.
2939 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2940 newval = _gvn.makecon(TypePtr::NULL_PTR);
2941
2942 // Reference stores need a store barrier.
2943 if (kind == LS_xchg) {
2944 // If pre-barrier must execute before the oop store, old value will require do_load here.
2945 if (!can_move_pre_barrier()) {
2946 pre_barrier(true /* do_load*/,
2947 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2948 NULL /* pre_val*/,
2949 T_OBJECT);
2950 } // Else move pre_barrier to use load_store value, see below.
2951 } else if (kind == LS_cmpxchg) {
2952 // Same as for newval above:
2953 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2954 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2955 }
2956 // The only known value which might get overwritten is oldval.
2957 pre_barrier(false /* do_load */,
2958 control(), NULL, NULL, max_juint, NULL, NULL,
2959 oldval /* pre_val */,
2960 T_OBJECT);
2961 } else {
2962 ShouldNotReachHere();
2963 }
2964
2965 #ifdef _LP64
2966 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2967 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2968 if (kind == LS_xchg) {
2969 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2970 newval_enc, adr_type, value_type->make_narrowoop()));
2971 } else {
2972 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2973 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2974 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2975 newval_enc, oldval_enc));
2976 }
2977 } else
2978 #endif
2979 {
2980 if (kind == LS_xchg) {
2981 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2982 } else {
2983 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2984 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2985 }
2986 }
2987 if (kind == LS_cmpxchg) {
2988 // Emit the post barrier only when the actual store happened.
2989 // This makes sense to check only for compareAndSet that can fail to set the value.
2990 // CAS success path is marked more likely since we anticipate this is a performance
2991 // critical path, while CAS failure path can use the penalty for going through unlikely
2992 // path as backoff. Which is still better than doing a store barrier there.
2993 IdealKit ideal(this);
2994 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2995 sync_kit(ideal);
2996 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2997 ideal.sync_kit(this);
2998 } ideal.end_if();
2999 final_sync(ideal);
3000 } else {
3001 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3002 }
3003 break;
3004 default:
3005 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3006 break;
3007 }
3008
3009 // SCMemProjNodes represent the memory state of a LoadStore. Their
3010 // main role is to prevent LoadStore nodes from being optimized away
3011 // when their results aren't used.
3012 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3013 set_memory(proj, alias_idx);
3014
3015 if (type == T_OBJECT && kind == LS_xchg) {
3016 #ifdef _LP64
3017 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3018 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3019 }
3020 #endif
3021 if (can_move_pre_barrier()) {
3022 // Don't need to load pre_val. The old value is returned by load_store.
3023 // The pre_barrier can execute after the xchg as long as no safepoint
3024 // gets inserted between them.
3025 pre_barrier(false /* do_load */,
3026 control(), NULL, NULL, max_juint, NULL, NULL,
3027 load_store /* pre_val */,
3028 T_OBJECT);
3029 }
3030 }
3031
3032 // Add the trailing membar surrounding the access
3033 insert_mem_bar(Op_MemBarCPUOrder);
3034 insert_mem_bar(Op_MemBarAcquire);
3035
3036 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3037 set_result(load_store);
3038 return true;
3039 }
3040
3041 //----------------------------inline_unsafe_ordered_store----------------------
3042 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3043 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3044 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3045 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3046 // This is another variant of inline_unsafe_access, differing in
3047 // that it always issues store-store ("release") barrier and ensures
3048 // store-atomicity (which only matters for "long").
3049
3050 if (callee()->is_static()) return false; // caller must have the capability!
3051
3052 #ifndef PRODUCT
3053 {
3054 ResourceMark rm;
3055 // Check the signatures.
3056 ciSignature* sig = callee()->signature();
3057 #ifdef ASSERT
3058 BasicType rtype = sig->return_type()->basic_type();
3059 assert(rtype == T_VOID, "must return void");
3060 assert(sig->count() == 3, "has 3 arguments");
3061 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3062 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3063 #endif // ASSERT
3064 }
3065 #endif //PRODUCT
3066
3067 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3068
3069 // Get arguments:
3070 Node* receiver = argument(0); // type: oop
3071 Node* base = argument(1); // type: oop
3072 Node* offset = argument(2); // type: long
3073 Node* val = argument(4); // type: oop, int, or long
3074
3075 // Null check receiver.
3076 receiver = null_check(receiver);
3077 if (stopped()) {
3078 return true;
3079 }
3080
3081 // Build field offset expression.
3082 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3083 // 32-bit machines ignore the high half of long offsets
3084 offset = ConvL2X(offset);
3085 Node* adr = make_unsafe_address(base, offset);
3086 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3087 const Type *value_type = Type::get_const_basic_type(type);
3088 Compile::AliasType* alias_type = C->alias_type(adr_type);
3089
3090 insert_mem_bar(Op_MemBarRelease);
3091 insert_mem_bar(Op_MemBarCPUOrder);
3092 // Ensure that the store is atomic for longs:
3093 const bool require_atomic_access = true;
3094 Node* store;
3095 if (type == T_OBJECT) // reference stores need a store barrier.
3096 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3097 else {
3098 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3099 }
3100 insert_mem_bar(Op_MemBarCPUOrder);
3101 return true;
3102 }
3103
3104 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3105 // Regardless of form, don't allow previous ld/st to move down,
3106 // then issue acquire, release, or volatile mem_bar.
3107 insert_mem_bar(Op_MemBarCPUOrder);
3108 switch(id) {
3109 case vmIntrinsics::_loadFence:
3110 insert_mem_bar(Op_LoadFence);
3111 return true;
3112 case vmIntrinsics::_storeFence:
3113 insert_mem_bar(Op_StoreFence);
3114 return true;
3115 case vmIntrinsics::_fullFence:
3116 insert_mem_bar(Op_MemBarVolatile);
3117 return true;
3118 default:
3119 fatal_unexpected_iid(id);
3120 return false;
3121 }
3122 }
3123
3124 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3125 if (!kls->is_Con()) {
3126 return true;
3127 }
3128 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3129 if (klsptr == NULL) {
3130 return true;
3131 }
3132 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3133 // don't need a guard for a klass that is already initialized
3134 return !ik->is_initialized();
3135 }
3136
3137 //----------------------------inline_unsafe_allocate---------------------------
3138 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3139 bool LibraryCallKit::inline_unsafe_allocate() {
3140 if (callee()->is_static()) return false; // caller must have the capability!
3141
3142 null_check_receiver(); // null-check, then ignore
3143 Node* cls = null_check(argument(1));
3144 if (stopped()) return true;
3145
3146 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3147 kls = null_check(kls);
3148 if (stopped()) return true; // argument was like int.class
3149
3150 Node* test = NULL;
3151 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3152 // Note: The argument might still be an illegal value like
3153 // Serializable.class or Object[].class. The runtime will handle it.
3154 // But we must make an explicit check for initialization.
3155 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3156 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3157 // can generate code to load it as unsigned byte.
3158 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3159 Node* bits = intcon(InstanceKlass::fully_initialized);
3160 test = _gvn.transform(new SubINode(inst, bits));
3161 // The 'test' is non-zero if we need to take a slow path.
3162 }
3163
3164 Node* obj = new_instance(kls, test);
3165 set_result(obj);
3166 return true;
3167 }
3168
3169 #ifdef TRACE_HAVE_INTRINSICS
3170 /*
3171 * oop -> myklass
3172 * myklass->trace_id |= USED
3173 * return myklass->trace_id & ~0x3
3174 */
3175 bool LibraryCallKit::inline_native_classID() {
3176 null_check_receiver(); // null-check, then ignore
3177 Node* cls = null_check(argument(1), T_OBJECT);
3178 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3179 kls = null_check(kls, T_OBJECT);
3180 ByteSize offset = TRACE_ID_OFFSET;
3181 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3182 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3183 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3184 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3185 Node* clsused = longcon(0x01l); // set the class bit
3186 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3187
3188 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3189 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3190 set_result(andl);
3191 return true;
3192 }
3193
3194 bool LibraryCallKit::inline_native_threadID() {
3195 Node* tls_ptr = NULL;
3196 Node* cur_thr = generate_current_thread(tls_ptr);
3197 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3198 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3199 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3200
3201 Node* threadid = NULL;
3202 size_t thread_id_size = OSThread::thread_id_size();
3203 if (thread_id_size == (size_t) BytesPerLong) {
3204 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3205 } else if (thread_id_size == (size_t) BytesPerInt) {
3206 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3207 } else {
3208 ShouldNotReachHere();
3209 }
3210 set_result(threadid);
3211 return true;
3212 }
3213 #endif
3214
3215 //------------------------inline_native_time_funcs--------------
3216 // inline code for System.currentTimeMillis() and System.nanoTime()
3217 // these have the same type and signature
3218 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3219 const TypeFunc* tf = OptoRuntime::void_long_Type();
3220 const TypePtr* no_memory_effects = NULL;
3221 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3222 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3223 #ifdef ASSERT
3224 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3225 assert(value_top == top(), "second value must be top");
3226 #endif
3227 set_result(value);
3228 return true;
3229 }
3230
3231 //------------------------inline_native_currentThread------------------
3232 bool LibraryCallKit::inline_native_currentThread() {
3233 Node* junk = NULL;
3234 set_result(generate_current_thread(junk));
3235 return true;
3236 }
3237
3238 //------------------------inline_native_isInterrupted------------------
3239 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3240 bool LibraryCallKit::inline_native_isInterrupted() {
3241 // Add a fast path to t.isInterrupted(clear_int):
3242 // (t == Thread.current() &&
3243 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3244 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3245 // So, in the common case that the interrupt bit is false,
3246 // we avoid making a call into the VM. Even if the interrupt bit
3247 // is true, if the clear_int argument is false, we avoid the VM call.
3248 // However, if the receiver is not currentThread, we must call the VM,
3249 // because there must be some locking done around the operation.
3250
3251 // We only go to the fast case code if we pass two guards.
3252 // Paths which do not pass are accumulated in the slow_region.
3253
3254 enum {
3255 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3256 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3257 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3258 PATH_LIMIT
3259 };
3260
3261 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3262 // out of the function.
3263 insert_mem_bar(Op_MemBarCPUOrder);
3264
3265 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3266 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3267
3268 RegionNode* slow_region = new RegionNode(1);
3269 record_for_igvn(slow_region);
3270
3271 // (a) Receiving thread must be the current thread.
3272 Node* rec_thr = argument(0);
3273 Node* tls_ptr = NULL;
3274 Node* cur_thr = generate_current_thread(tls_ptr);
3275 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3276 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3277
3278 generate_slow_guard(bol_thr, slow_region);
3279
3280 // (b) Interrupt bit on TLS must be false.
3281 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3282 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3283 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3284
3285 // Set the control input on the field _interrupted read to prevent it floating up.
3286 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3287 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3288 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3289
3290 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3291
3292 // First fast path: if (!TLS._interrupted) return false;
3293 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3294 result_rgn->init_req(no_int_result_path, false_bit);
3295 result_val->init_req(no_int_result_path, intcon(0));
3296
3297 // drop through to next case
3298 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3299
3300 #ifndef TARGET_OS_FAMILY_windows
3301 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3302 Node* clr_arg = argument(1);
3303 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3304 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3305 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3306
3307 // Second fast path: ... else if (!clear_int) return true;
3308 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3309 result_rgn->init_req(no_clear_result_path, false_arg);
3310 result_val->init_req(no_clear_result_path, intcon(1));
3311
3312 // drop through to next case
3313 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3314 #else
3315 // To return true on Windows you must read the _interrupted field
3316 // and check the the event state i.e. take the slow path.
3317 #endif // TARGET_OS_FAMILY_windows
3318
3319 // (d) Otherwise, go to the slow path.
3320 slow_region->add_req(control());
3321 set_control( _gvn.transform(slow_region));
3322
3323 if (stopped()) {
3324 // There is no slow path.
3325 result_rgn->init_req(slow_result_path, top());
3326 result_val->init_req(slow_result_path, top());
3327 } else {
3328 // non-virtual because it is a private non-static
3329 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3330
3331 Node* slow_val = set_results_for_java_call(slow_call);
3332 // this->control() comes from set_results_for_java_call
3333
3334 Node* fast_io = slow_call->in(TypeFunc::I_O);
3335 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3336
3337 // These two phis are pre-filled with copies of of the fast IO and Memory
3338 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3339 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3340
3341 result_rgn->init_req(slow_result_path, control());
3342 result_io ->init_req(slow_result_path, i_o());
3343 result_mem->init_req(slow_result_path, reset_memory());
3344 result_val->init_req(slow_result_path, slow_val);
3345
3346 set_all_memory(_gvn.transform(result_mem));
3347 set_i_o( _gvn.transform(result_io));
3348 }
3349
3350 C->set_has_split_ifs(true); // Has chance for split-if optimization
3351 set_result(result_rgn, result_val);
3352 return true;
3353 }
3354
3355 //---------------------------load_mirror_from_klass----------------------------
3356 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3357 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3358 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3359 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3360 }
3361
3362 //-----------------------load_klass_from_mirror_common-------------------------
3363 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3364 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3365 // and branch to the given path on the region.
3366 // If never_see_null, take an uncommon trap on null, so we can optimistically
3367 // compile for the non-null case.
3368 // If the region is NULL, force never_see_null = true.
3369 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3370 bool never_see_null,
3371 RegionNode* region,
3372 int null_path,
3373 int offset) {
3374 if (region == NULL) never_see_null = true;
3375 Node* p = basic_plus_adr(mirror, offset);
3376 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3377 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3378 Node* null_ctl = top();
3379 kls = null_check_oop(kls, &null_ctl, never_see_null);
3380 if (region != NULL) {
3381 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3382 region->init_req(null_path, null_ctl);
3383 } else {
3384 assert(null_ctl == top(), "no loose ends");
3385 }
3386 return kls;
3387 }
3388
3389 //--------------------(inline_native_Class_query helpers)---------------------
3390 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3391 // Fall through if (mods & mask) == bits, take the guard otherwise.
3392 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3393 // Branch around if the given klass has the given modifier bit set.
3394 // Like generate_guard, adds a new path onto the region.
3395 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3396 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3397 Node* mask = intcon(modifier_mask);
3398 Node* bits = intcon(modifier_bits);
3399 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3400 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3401 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3402 return generate_fair_guard(bol, region);
3403 }
3404 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3405 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3406 }
3407
3408 //-------------------------inline_native_Class_query-------------------
3409 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3410 const Type* return_type = TypeInt::BOOL;
3411 Node* prim_return_value = top(); // what happens if it's a primitive class?
3412 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3413 bool expect_prim = false; // most of these guys expect to work on refs
3414
3415 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3416
3417 Node* mirror = argument(0);
3418 Node* obj = top();
3419
3420 switch (id) {
3421 case vmIntrinsics::_isInstance:
3422 // nothing is an instance of a primitive type
3423 prim_return_value = intcon(0);
3424 obj = argument(1);
3425 break;
3426 case vmIntrinsics::_getModifiers:
3427 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3428 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3429 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3430 break;
3431 case vmIntrinsics::_isInterface:
3432 prim_return_value = intcon(0);
3433 break;
3434 case vmIntrinsics::_isArray:
3435 prim_return_value = intcon(0);
3436 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3437 break;
3438 case vmIntrinsics::_isPrimitive:
3439 prim_return_value = intcon(1);
3440 expect_prim = true; // obviously
3441 break;
3442 case vmIntrinsics::_getSuperclass:
3443 prim_return_value = null();
3444 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3445 break;
3446 case vmIntrinsics::_getClassAccessFlags:
3447 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3448 return_type = TypeInt::INT; // not bool! 6297094
3449 break;
3450 default:
3451 fatal_unexpected_iid(id);
3452 break;
3453 }
3454
3455 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3456 if (mirror_con == NULL) return false; // cannot happen?
3457
3458 #ifndef PRODUCT
3459 if (C->print_intrinsics() || C->print_inlining()) {
3460 ciType* k = mirror_con->java_mirror_type();
3461 if (k) {
3462 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3463 k->print_name();
3464 tty->cr();
3465 }
3466 }
3467 #endif
3468
3469 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3470 RegionNode* region = new RegionNode(PATH_LIMIT);
3471 record_for_igvn(region);
3472 PhiNode* phi = new PhiNode(region, return_type);
3473
3474 // The mirror will never be null of Reflection.getClassAccessFlags, however
3475 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3476 // if it is. See bug 4774291.
3477
3478 // For Reflection.getClassAccessFlags(), the null check occurs in
3479 // the wrong place; see inline_unsafe_access(), above, for a similar
3480 // situation.
3481 mirror = null_check(mirror);
3482 // If mirror or obj is dead, only null-path is taken.
3483 if (stopped()) return true;
3484
3485 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3486
3487 // Now load the mirror's klass metaobject, and null-check it.
3488 // Side-effects region with the control path if the klass is null.
3489 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3490 // If kls is null, we have a primitive mirror.
3491 phi->init_req(_prim_path, prim_return_value);
3492 if (stopped()) { set_result(region, phi); return true; }
3493 bool safe_for_replace = (region->in(_prim_path) == top());
3494
3495 Node* p; // handy temp
3496 Node* null_ctl;
3497
3498 // Now that we have the non-null klass, we can perform the real query.
3499 // For constant classes, the query will constant-fold in LoadNode::Value.
3500 Node* query_value = top();
3501 switch (id) {
3502 case vmIntrinsics::_isInstance:
3503 // nothing is an instance of a primitive type
3504 query_value = gen_instanceof(obj, kls, safe_for_replace);
3505 break;
3506
3507 case vmIntrinsics::_getModifiers:
3508 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3509 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3510 break;
3511
3512 case vmIntrinsics::_isInterface:
3513 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3514 if (generate_interface_guard(kls, region) != NULL)
3515 // A guard was added. If the guard is taken, it was an interface.
3516 phi->add_req(intcon(1));
3517 // If we fall through, it's a plain class.
3518 query_value = intcon(0);
3519 break;
3520
3521 case vmIntrinsics::_isArray:
3522 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3523 if (generate_array_guard(kls, region) != NULL)
3524 // A guard was added. If the guard is taken, it was an array.
3525 phi->add_req(intcon(1));
3526 // If we fall through, it's a plain class.
3527 query_value = intcon(0);
3528 break;
3529
3530 case vmIntrinsics::_isPrimitive:
3531 query_value = intcon(0); // "normal" path produces false
3532 break;
3533
3534 case vmIntrinsics::_getSuperclass:
3535 // The rules here are somewhat unfortunate, but we can still do better
3536 // with random logic than with a JNI call.
3537 // Interfaces store null or Object as _super, but must report null.
3538 // Arrays store an intermediate super as _super, but must report Object.
3539 // Other types can report the actual _super.
3540 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3541 if (generate_interface_guard(kls, region) != NULL)
3542 // A guard was added. If the guard is taken, it was an interface.
3543 phi->add_req(null());
3544 if (generate_array_guard(kls, region) != NULL)
3545 // A guard was added. If the guard is taken, it was an array.
3546 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3547 // If we fall through, it's a plain class. Get its _super.
3548 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3549 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3550 null_ctl = top();
3551 kls = null_check_oop(kls, &null_ctl);
3552 if (null_ctl != top()) {
3553 // If the guard is taken, Object.superClass is null (both klass and mirror).
3554 region->add_req(null_ctl);
3555 phi ->add_req(null());
3556 }
3557 if (!stopped()) {
3558 query_value = load_mirror_from_klass(kls);
3559 }
3560 break;
3561
3562 case vmIntrinsics::_getClassAccessFlags:
3563 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3564 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3565 break;
3566
3567 default:
3568 fatal_unexpected_iid(id);
3569 break;
3570 }
3571
3572 // Fall-through is the normal case of a query to a real class.
3573 phi->init_req(1, query_value);
3574 region->init_req(1, control());
3575
3576 C->set_has_split_ifs(true); // Has chance for split-if optimization
3577 set_result(region, phi);
3578 return true;
3579 }
3580
3581 //-------------------------inline_Class_cast-------------------
3582 bool LibraryCallKit::inline_Class_cast() {
3583 Node* mirror = argument(0); // Class
3584 Node* obj = argument(1);
3585 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3586 if (mirror_con == NULL) {
3587 return false; // dead path (mirror->is_top()).
3588 }
3589 if (obj == NULL || obj->is_top()) {
3590 return false; // dead path
3591 }
3592 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3593
3594 // First, see if Class.cast() can be folded statically.
3595 // java_mirror_type() returns non-null for compile-time Class constants.
3596 ciType* tm = mirror_con->java_mirror_type();
3597 if (tm != NULL && tm->is_klass() &&
3598 tp != NULL && tp->klass() != NULL) {
3599 if (!tp->klass()->is_loaded()) {
3600 // Don't use intrinsic when class is not loaded.
3601 return false;
3602 } else {
3603 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3604 if (static_res == Compile::SSC_always_true) {
3605 // isInstance() is true - fold the code.
3606 set_result(obj);
3607 return true;
3608 } else if (static_res == Compile::SSC_always_false) {
3609 // Don't use intrinsic, have to throw ClassCastException.
3610 // If the reference is null, the non-intrinsic bytecode will
3611 // be optimized appropriately.
3612 return false;
3613 }
3614 }
3615 }
3616
3617 // Bailout intrinsic and do normal inlining if exception path is frequent.
3618 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3619 return false;
3620 }
3621
3622 // Generate dynamic checks.
3623 // Class.cast() is java implementation of _checkcast bytecode.
3624 // Do checkcast (Parse::do_checkcast()) optimizations here.
3625
3626 mirror = null_check(mirror);
3627 // If mirror is dead, only null-path is taken.
3628 if (stopped()) {
3629 return true;
3630 }
3631
3632 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3633 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3634 RegionNode* region = new RegionNode(PATH_LIMIT);
3635 record_for_igvn(region);
3636
3637 // Now load the mirror's klass metaobject, and null-check it.
3638 // If kls is null, we have a primitive mirror and
3639 // nothing is an instance of a primitive type.
3640 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3641
3642 Node* res = top();
3643 if (!stopped()) {
3644 Node* bad_type_ctrl = top();
3645 // Do checkcast optimizations.
3646 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3647 region->init_req(_bad_type_path, bad_type_ctrl);
3648 }
3649 if (region->in(_prim_path) != top() ||
3650 region->in(_bad_type_path) != top()) {
3651 // Let Interpreter throw ClassCastException.
3652 PreserveJVMState pjvms(this);
3653 set_control(_gvn.transform(region));
3654 uncommon_trap(Deoptimization::Reason_intrinsic,
3655 Deoptimization::Action_maybe_recompile);
3656 }
3657 if (!stopped()) {
3658 set_result(res);
3659 }
3660 return true;
3661 }
3662
3663
3664 //--------------------------inline_native_subtype_check------------------------
3665 // This intrinsic takes the JNI calls out of the heart of
3666 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3667 bool LibraryCallKit::inline_native_subtype_check() {
3668 // Pull both arguments off the stack.
3669 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3670 args[0] = argument(0);
3671 args[1] = argument(1);
3672 Node* klasses[2]; // corresponding Klasses: superk, subk
3673 klasses[0] = klasses[1] = top();
3674
3675 enum {
3676 // A full decision tree on {superc is prim, subc is prim}:
3677 _prim_0_path = 1, // {P,N} => false
3678 // {P,P} & superc!=subc => false
3679 _prim_same_path, // {P,P} & superc==subc => true
3680 _prim_1_path, // {N,P} => false
3681 _ref_subtype_path, // {N,N} & subtype check wins => true
3682 _both_ref_path, // {N,N} & subtype check loses => false
3683 PATH_LIMIT
3684 };
3685
3686 RegionNode* region = new RegionNode(PATH_LIMIT);
3687 Node* phi = new PhiNode(region, TypeInt::BOOL);
3688 record_for_igvn(region);
3689
3690 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3691 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3692 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3693
3694 // First null-check both mirrors and load each mirror's klass metaobject.
3695 int which_arg;
3696 for (which_arg = 0; which_arg <= 1; which_arg++) {
3697 Node* arg = args[which_arg];
3698 arg = null_check(arg);
3699 if (stopped()) break;
3700 args[which_arg] = arg;
3701
3702 Node* p = basic_plus_adr(arg, class_klass_offset);
3703 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3704 klasses[which_arg] = _gvn.transform(kls);
3705 }
3706
3707 // Having loaded both klasses, test each for null.
3708 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3709 for (which_arg = 0; which_arg <= 1; which_arg++) {
3710 Node* kls = klasses[which_arg];
3711 Node* null_ctl = top();
3712 kls = null_check_oop(kls, &null_ctl, never_see_null);
3713 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3714 region->init_req(prim_path, null_ctl);
3715 if (stopped()) break;
3716 klasses[which_arg] = kls;
3717 }
3718
3719 if (!stopped()) {
3720 // now we have two reference types, in klasses[0..1]
3721 Node* subk = klasses[1]; // the argument to isAssignableFrom
3722 Node* superk = klasses[0]; // the receiver
3723 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3724 // now we have a successful reference subtype check
3725 region->set_req(_ref_subtype_path, control());
3726 }
3727
3728 // If both operands are primitive (both klasses null), then
3729 // we must return true when they are identical primitives.
3730 // It is convenient to test this after the first null klass check.
3731 set_control(region->in(_prim_0_path)); // go back to first null check
3732 if (!stopped()) {
3733 // Since superc is primitive, make a guard for the superc==subc case.
3734 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3735 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3736 generate_guard(bol_eq, region, PROB_FAIR);
3737 if (region->req() == PATH_LIMIT+1) {
3738 // A guard was added. If the added guard is taken, superc==subc.
3739 region->swap_edges(PATH_LIMIT, _prim_same_path);
3740 region->del_req(PATH_LIMIT);
3741 }
3742 region->set_req(_prim_0_path, control()); // Not equal after all.
3743 }
3744
3745 // these are the only paths that produce 'true':
3746 phi->set_req(_prim_same_path, intcon(1));
3747 phi->set_req(_ref_subtype_path, intcon(1));
3748
3749 // pull together the cases:
3750 assert(region->req() == PATH_LIMIT, "sane region");
3751 for (uint i = 1; i < region->req(); i++) {
3752 Node* ctl = region->in(i);
3753 if (ctl == NULL || ctl == top()) {
3754 region->set_req(i, top());
3755 phi ->set_req(i, top());
3756 } else if (phi->in(i) == NULL) {
3757 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3758 }
3759 }
3760
3761 set_control(_gvn.transform(region));
3762 set_result(_gvn.transform(phi));
3763 return true;
3764 }
3765
3766 //---------------------generate_array_guard_common------------------------
3767 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3768 bool obj_array, bool not_array) {
3769
3770 if (stopped()) {
3771 return NULL;
3772 }
3773
3774 // If obj_array/non_array==false/false:
3775 // Branch around if the given klass is in fact an array (either obj or prim).
3776 // If obj_array/non_array==false/true:
3777 // Branch around if the given klass is not an array klass of any kind.
3778 // If obj_array/non_array==true/true:
3779 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3780 // If obj_array/non_array==true/false:
3781 // Branch around if the kls is an oop array (Object[] or subtype)
3782 //
3783 // Like generate_guard, adds a new path onto the region.
3784 jint layout_con = 0;
3785 Node* layout_val = get_layout_helper(kls, layout_con);
3786 if (layout_val == NULL) {
3787 bool query = (obj_array
3788 ? Klass::layout_helper_is_objArray(layout_con)
3789 : Klass::layout_helper_is_array(layout_con));
3790 if (query == not_array) {
3791 return NULL; // never a branch
3792 } else { // always a branch
3793 Node* always_branch = control();
3794 if (region != NULL)
3795 region->add_req(always_branch);
3796 set_control(top());
3797 return always_branch;
3798 }
3799 }
3800 // Now test the correct condition.
3801 jint nval = (obj_array
3802 ? ((jint)Klass::_lh_array_tag_type_value
3803 << Klass::_lh_array_tag_shift)
3804 : Klass::_lh_neutral_value);
3805 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3806 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3807 // invert the test if we are looking for a non-array
3808 if (not_array) btest = BoolTest(btest).negate();
3809 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3810 return generate_fair_guard(bol, region);
3811 }
3812
3813
3814 //-----------------------inline_native_newArray--------------------------
3815 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3816 bool LibraryCallKit::inline_native_newArray() {
3817 Node* mirror = argument(0);
3818 Node* count_val = argument(1);
3819
3820 mirror = null_check(mirror);
3821 // If mirror or obj is dead, only null-path is taken.
3822 if (stopped()) return true;
3823
3824 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3825 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3826 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3827 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3828 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3829
3830 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3831 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3832 result_reg, _slow_path);
3833 Node* normal_ctl = control();
3834 Node* no_array_ctl = result_reg->in(_slow_path);
3835
3836 // Generate code for the slow case. We make a call to newArray().
3837 set_control(no_array_ctl);
3838 if (!stopped()) {
3839 // Either the input type is void.class, or else the
3840 // array klass has not yet been cached. Either the
3841 // ensuing call will throw an exception, or else it
3842 // will cache the array klass for next time.
3843 PreserveJVMState pjvms(this);
3844 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3845 Node* slow_result = set_results_for_java_call(slow_call);
3846 // this->control() comes from set_results_for_java_call
3847 result_reg->set_req(_slow_path, control());
3848 result_val->set_req(_slow_path, slow_result);
3849 result_io ->set_req(_slow_path, i_o());
3850 result_mem->set_req(_slow_path, reset_memory());
3851 }
3852
3853 set_control(normal_ctl);
3854 if (!stopped()) {
3855 // Normal case: The array type has been cached in the java.lang.Class.
3856 // The following call works fine even if the array type is polymorphic.
3857 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3858 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3859 result_reg->init_req(_normal_path, control());
3860 result_val->init_req(_normal_path, obj);
3861 result_io ->init_req(_normal_path, i_o());
3862 result_mem->init_req(_normal_path, reset_memory());
3863 }
3864
3865 // Return the combined state.
3866 set_i_o( _gvn.transform(result_io) );
3867 set_all_memory( _gvn.transform(result_mem));
3868
3869 C->set_has_split_ifs(true); // Has chance for split-if optimization
3870 set_result(result_reg, result_val);
3871 return true;
3872 }
3873
3874 //----------------------inline_native_getLength--------------------------
3875 // public static native int java.lang.reflect.Array.getLength(Object array);
3876 bool LibraryCallKit::inline_native_getLength() {
3877 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3878
3879 Node* array = null_check(argument(0));
3880 // If array is dead, only null-path is taken.
3881 if (stopped()) return true;
3882
3883 // Deoptimize if it is a non-array.
3884 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3885
3886 if (non_array != NULL) {
3887 PreserveJVMState pjvms(this);
3888 set_control(non_array);
3889 uncommon_trap(Deoptimization::Reason_intrinsic,
3890 Deoptimization::Action_maybe_recompile);
3891 }
3892
3893 // If control is dead, only non-array-path is taken.
3894 if (stopped()) return true;
3895
3896 // The works fine even if the array type is polymorphic.
3897 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3898 Node* result = load_array_length(array);
3899
3900 C->set_has_split_ifs(true); // Has chance for split-if optimization
3901 set_result(result);
3902 return true;
3903 }
3904
3905 //------------------------inline_array_copyOf----------------------------
3906 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3907 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3908 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3909 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3910
3911 // Get the arguments.
3912 Node* original = argument(0);
3913 Node* start = is_copyOfRange? argument(1): intcon(0);
3914 Node* end = is_copyOfRange? argument(2): argument(1);
3915 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3916
3917 Node* newcopy;
3918
3919 // Set the original stack and the reexecute bit for the interpreter to reexecute
3920 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3921 { PreserveReexecuteState preexecs(this);
3922 jvms()->set_should_reexecute(true);
3923
3924 array_type_mirror = null_check(array_type_mirror);
3925 original = null_check(original);
3926
3927 // Check if a null path was taken unconditionally.
3928 if (stopped()) return true;
3929
3930 Node* orig_length = load_array_length(original);
3931
3932 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3933 klass_node = null_check(klass_node);
3934
3935 RegionNode* bailout = new RegionNode(1);
3936 record_for_igvn(bailout);
3937
3938 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3939 // Bail out if that is so.
3940 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3941 if (not_objArray != NULL) {
3942 // Improve the klass node's type from the new optimistic assumption:
3943 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3944 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3945 Node* cast = new CastPPNode(klass_node, akls);
3946 cast->init_req(0, control());
3947 klass_node = _gvn.transform(cast);
3948 }
3949
3950 // Bail out if either start or end is negative.
3951 generate_negative_guard(start, bailout, &start);
3952 generate_negative_guard(end, bailout, &end);
3953
3954 Node* length = end;
3955 if (_gvn.type(start) != TypeInt::ZERO) {
3956 length = _gvn.transform(new SubINode(end, start));
3957 }
3958
3959 // Bail out if length is negative.
3960 // Without this the new_array would throw
3961 // NegativeArraySizeException but IllegalArgumentException is what
3962 // should be thrown
3963 generate_negative_guard(length, bailout, &length);
3964
3965 if (bailout->req() > 1) {
3966 PreserveJVMState pjvms(this);
3967 set_control(_gvn.transform(bailout));
3968 uncommon_trap(Deoptimization::Reason_intrinsic,
3969 Deoptimization::Action_maybe_recompile);
3970 }
3971
3972 if (!stopped()) {
3973 // How many elements will we copy from the original?
3974 // The answer is MinI(orig_length - start, length).
3975 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3976 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3977
3978 // Generate a direct call to the right arraycopy function(s).
3979 // We know the copy is disjoint but we might not know if the
3980 // oop stores need checking.
3981 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3982 // This will fail a store-check if x contains any non-nulls.
3983
3984 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3985 // loads/stores but it is legal only if we're sure the
3986 // Arrays.copyOf would succeed. So we need all input arguments
3987 // to the copyOf to be validated, including that the copy to the
3988 // new array won't trigger an ArrayStoreException. That subtype
3989 // check can be optimized if we know something on the type of
3990 // the input array from type speculation.
3991 if (_gvn.type(klass_node)->singleton()) {
3992 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3993 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3994
3995 int test = C->static_subtype_check(superk, subk);
3996 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3997 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3998 if (t_original->speculative_type() != NULL) {
3999 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4000 }
4001 }
4002 }
4003
4004 bool validated = false;
4005 // Reason_class_check rather than Reason_intrinsic because we
4006 // want to intrinsify even if this traps.
4007 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4008 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4009 klass_node);
4010
4011 if (not_subtype_ctrl != top()) {
4012 PreserveJVMState pjvms(this);
4013 set_control(not_subtype_ctrl);
4014 uncommon_trap(Deoptimization::Reason_class_check,
4015 Deoptimization::Action_make_not_entrant);
4016 assert(stopped(), "Should be stopped");
4017 }
4018 validated = true;
4019 }
4020
4021 if (!stopped()) {
4022 newcopy = new_array(klass_node, length, 0); // no arguments to push
4023
4024 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
4025 load_object_klass(original), klass_node);
4026 if (!is_copyOfRange) {
4027 ac->set_copyof(validated);
4028 } else {
4029 ac->set_copyofrange(validated);
4030 }
4031 Node* n = _gvn.transform(ac);
4032 if (n == ac) {
4033 ac->connect_outputs(this);
4034 } else {
4035 assert(validated, "shouldn't transform if all arguments not validated");
4036 set_all_memory(n);
4037 }
4038 }
4039 }
4040 } // original reexecute is set back here
4041
4042 C->set_has_split_ifs(true); // Has chance for split-if optimization
4043 if (!stopped()) {
4044 set_result(newcopy);
4045 }
4046 return true;
4047 }
4048
4049
4050 //----------------------generate_virtual_guard---------------------------
4051 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4052 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4053 RegionNode* slow_region) {
4054 ciMethod* method = callee();
4055 int vtable_index = method->vtable_index();
4056 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4057 err_msg_res("bad index %d", vtable_index));
4058 // Get the Method* out of the appropriate vtable entry.
4059 int entry_offset = (InstanceKlass::vtable_start_offset() +
4060 vtable_index*vtableEntry::size()) * wordSize +
4061 vtableEntry::method_offset_in_bytes();
4062 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4063 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4064
4065 // Compare the target method with the expected method (e.g., Object.hashCode).
4066 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4067
4068 Node* native_call = makecon(native_call_addr);
4069 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4070 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4071
4072 return generate_slow_guard(test_native, slow_region);
4073 }
4074
4075 //-----------------------generate_method_call----------------------------
4076 // Use generate_method_call to make a slow-call to the real
4077 // method if the fast path fails. An alternative would be to
4078 // use a stub like OptoRuntime::slow_arraycopy_Java.
4079 // This only works for expanding the current library call,
4080 // not another intrinsic. (E.g., don't use this for making an
4081 // arraycopy call inside of the copyOf intrinsic.)
4082 CallJavaNode*
4083 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4084 // When compiling the intrinsic method itself, do not use this technique.
4085 guarantee(callee() != C->method(), "cannot make slow-call to self");
4086
4087 ciMethod* method = callee();
4088 // ensure the JVMS we have will be correct for this call
4089 guarantee(method_id == method->intrinsic_id(), "must match");
4090
4091 const TypeFunc* tf = TypeFunc::make(method);
4092 CallJavaNode* slow_call;
4093 if (is_static) {
4094 assert(!is_virtual, "");
4095 slow_call = new CallStaticJavaNode(C, tf,
4096 SharedRuntime::get_resolve_static_call_stub(),
4097 method, bci());
4098 } else if (is_virtual) {
4099 null_check_receiver();
4100 int vtable_index = Method::invalid_vtable_index;
4101 if (UseInlineCaches) {
4102 // Suppress the vtable call
4103 } else {
4104 // hashCode and clone are not a miranda methods,
4105 // so the vtable index is fixed.
4106 // No need to use the linkResolver to get it.
4107 vtable_index = method->vtable_index();
4108 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4109 err_msg_res("bad index %d", vtable_index));
4110 }
4111 slow_call = new CallDynamicJavaNode(tf,
4112 SharedRuntime::get_resolve_virtual_call_stub(),
4113 method, vtable_index, bci());
4114 } else { // neither virtual nor static: opt_virtual
4115 null_check_receiver();
4116 slow_call = new CallStaticJavaNode(C, tf,
4117 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4118 method, bci());
4119 slow_call->set_optimized_virtual(true);
4120 }
4121 set_arguments_for_java_call(slow_call);
4122 set_edges_for_java_call(slow_call);
4123 return slow_call;
4124 }
4125
4126
4127 /**
4128 * Build special case code for calls to hashCode on an object. This call may
4129 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4130 * slightly different code.
4131 */
4132 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4133 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4134 assert(!(is_virtual && is_static), "either virtual, special, or static");
4135
4136 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4137
4138 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4139 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4140 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4141 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4142 Node* obj = NULL;
4143 if (!is_static) {
4144 // Check for hashing null object
4145 obj = null_check_receiver();
4146 if (stopped()) return true; // unconditionally null
4147 result_reg->init_req(_null_path, top());
4148 result_val->init_req(_null_path, top());
4149 } else {
4150 // Do a null check, and return zero if null.
4151 // System.identityHashCode(null) == 0
4152 obj = argument(0);
4153 Node* null_ctl = top();
4154 obj = null_check_oop(obj, &null_ctl);
4155 result_reg->init_req(_null_path, null_ctl);
4156 result_val->init_req(_null_path, _gvn.intcon(0));
4157 }
4158
4159 // Unconditionally null? Then return right away.
4160 if (stopped()) {
4161 set_control( result_reg->in(_null_path));
4162 if (!stopped())
4163 set_result(result_val->in(_null_path));
4164 return true;
4165 }
4166
4167 // We only go to the fast case code if we pass a number of guards. The
4168 // paths which do not pass are accumulated in the slow_region.
4169 RegionNode* slow_region = new RegionNode(1);
4170 record_for_igvn(slow_region);
4171
4172 // If this is a virtual call, we generate a funny guard. We pull out
4173 // the vtable entry corresponding to hashCode() from the target object.
4174 // If the target method which we are calling happens to be the native
4175 // Object hashCode() method, we pass the guard. We do not need this
4176 // guard for non-virtual calls -- the caller is known to be the native
4177 // Object hashCode().
4178 if (is_virtual) {
4179 // After null check, get the object's klass.
4180 Node* obj_klass = load_object_klass(obj);
4181 generate_virtual_guard(obj_klass, slow_region);
4182 }
4183
4184 // Get the header out of the object, use LoadMarkNode when available
4185 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4186 // The control of the load must be NULL. Otherwise, the load can move before
4187 // the null check after castPP removal.
4188 Node* no_ctrl = NULL;
4189 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4190
4191 // Test the header to see if it is unlocked.
4192 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4193 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4194 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4195 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4196 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4197
4198 generate_slow_guard(test_unlocked, slow_region);
4199
4200 // Get the hash value and check to see that it has been properly assigned.
4201 // We depend on hash_mask being at most 32 bits and avoid the use of
4202 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4203 // vm: see markOop.hpp.
4204 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4205 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4206 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4207 // This hack lets the hash bits live anywhere in the mark object now, as long
4208 // as the shift drops the relevant bits into the low 32 bits. Note that
4209 // Java spec says that HashCode is an int so there's no point in capturing
4210 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4211 hshifted_header = ConvX2I(hshifted_header);
4212 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4213
4214 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4215 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4216 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4217
4218 generate_slow_guard(test_assigned, slow_region);
4219
4220 Node* init_mem = reset_memory();
4221 // fill in the rest of the null path:
4222 result_io ->init_req(_null_path, i_o());
4223 result_mem->init_req(_null_path, init_mem);
4224
4225 result_val->init_req(_fast_path, hash_val);
4226 result_reg->init_req(_fast_path, control());
4227 result_io ->init_req(_fast_path, i_o());
4228 result_mem->init_req(_fast_path, init_mem);
4229
4230 // Generate code for the slow case. We make a call to hashCode().
4231 set_control(_gvn.transform(slow_region));
4232 if (!stopped()) {
4233 // No need for PreserveJVMState, because we're using up the present state.
4234 set_all_memory(init_mem);
4235 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4236 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4237 Node* slow_result = set_results_for_java_call(slow_call);
4238 // this->control() comes from set_results_for_java_call
4239 result_reg->init_req(_slow_path, control());
4240 result_val->init_req(_slow_path, slow_result);
4241 result_io ->set_req(_slow_path, i_o());
4242 result_mem ->set_req(_slow_path, reset_memory());
4243 }
4244
4245 // Return the combined state.
4246 set_i_o( _gvn.transform(result_io) );
4247 set_all_memory( _gvn.transform(result_mem));
4248
4249 set_result(result_reg, result_val);
4250 return true;
4251 }
4252
4253 //---------------------------inline_native_getClass----------------------------
4254 // public final native Class<?> java.lang.Object.getClass();
4255 //
4256 // Build special case code for calls to getClass on an object.
4257 bool LibraryCallKit::inline_native_getClass() {
4258 Node* obj = null_check_receiver();
4259 if (stopped()) return true;
4260 set_result(load_mirror_from_klass(load_object_klass(obj)));
4261 return true;
4262 }
4263
4264 //-----------------inline_native_Reflection_getCallerClass---------------------
4265 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4266 //
4267 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4268 //
4269 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4270 // in that it must skip particular security frames and checks for
4271 // caller sensitive methods.
4272 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4273 #ifndef PRODUCT
4274 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4275 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4276 }
4277 #endif
4278
4279 if (!jvms()->has_method()) {
4280 #ifndef PRODUCT
4281 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4282 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4283 }
4284 #endif
4285 return false;
4286 }
4287
4288 // Walk back up the JVM state to find the caller at the required
4289 // depth.
4290 JVMState* caller_jvms = jvms();
4291
4292 // Cf. JVM_GetCallerClass
4293 // NOTE: Start the loop at depth 1 because the current JVM state does
4294 // not include the Reflection.getCallerClass() frame.
4295 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4296 ciMethod* m = caller_jvms->method();
4297 switch (n) {
4298 case 0:
4299 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4300 break;
4301 case 1:
4302 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4303 if (!m->caller_sensitive()) {
4304 #ifndef PRODUCT
4305 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4306 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4307 }
4308 #endif
4309 return false; // bail-out; let JVM_GetCallerClass do the work
4310 }
4311 break;
4312 default:
4313 if (!m->is_ignored_by_security_stack_walk()) {
4314 // We have reached the desired frame; return the holder class.
4315 // Acquire method holder as java.lang.Class and push as constant.
4316 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4317 ciInstance* caller_mirror = caller_klass->java_mirror();
4318 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4319
4320 #ifndef PRODUCT
4321 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4322 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());
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 return true;
4331 }
4332 break;
4333 }
4334 }
4335
4336 #ifndef PRODUCT
4337 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4338 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4339 tty->print_cr(" JVM state at this point:");
4340 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4341 ciMethod* m = jvms()->of_depth(i)->method();
4342 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4343 }
4344 }
4345 #endif
4346
4347 return false; // bail-out; let JVM_GetCallerClass do the work
4348 }
4349
4350 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4351 Node* arg = argument(0);
4352 Node* result;
4353
4354 switch (id) {
4355 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4356 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4357 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4358 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4359
4360 case vmIntrinsics::_doubleToLongBits: {
4361 // two paths (plus control) merge in a wood
4362 RegionNode *r = new RegionNode(3);
4363 Node *phi = new PhiNode(r, TypeLong::LONG);
4364
4365 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4366 // Build the boolean node
4367 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4368
4369 // Branch either way.
4370 // NaN case is less traveled, which makes all the difference.
4371 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4372 Node *opt_isnan = _gvn.transform(ifisnan);
4373 assert( opt_isnan->is_If(), "Expect an IfNode");
4374 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4375 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4376
4377 set_control(iftrue);
4378
4379 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4380 Node *slow_result = longcon(nan_bits); // return NaN
4381 phi->init_req(1, _gvn.transform( slow_result ));
4382 r->init_req(1, iftrue);
4383
4384 // Else fall through
4385 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4386 set_control(iffalse);
4387
4388 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4389 r->init_req(2, iffalse);
4390
4391 // Post merge
4392 set_control(_gvn.transform(r));
4393 record_for_igvn(r);
4394
4395 C->set_has_split_ifs(true); // Has chance for split-if optimization
4396 result = phi;
4397 assert(result->bottom_type()->isa_long(), "must be");
4398 break;
4399 }
4400
4401 case vmIntrinsics::_floatToIntBits: {
4402 // two paths (plus control) merge in a wood
4403 RegionNode *r = new RegionNode(3);
4404 Node *phi = new PhiNode(r, TypeInt::INT);
4405
4406 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4407 // Build the boolean node
4408 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4409
4410 // Branch either way.
4411 // NaN case is less traveled, which makes all the difference.
4412 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4413 Node *opt_isnan = _gvn.transform(ifisnan);
4414 assert( opt_isnan->is_If(), "Expect an IfNode");
4415 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4416 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4417
4418 set_control(iftrue);
4419
4420 static const jint nan_bits = 0x7fc00000;
4421 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4422 phi->init_req(1, _gvn.transform( slow_result ));
4423 r->init_req(1, iftrue);
4424
4425 // Else fall through
4426 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4427 set_control(iffalse);
4428
4429 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4430 r->init_req(2, iffalse);
4431
4432 // Post merge
4433 set_control(_gvn.transform(r));
4434 record_for_igvn(r);
4435
4436 C->set_has_split_ifs(true); // Has chance for split-if optimization
4437 result = phi;
4438 assert(result->bottom_type()->isa_int(), "must be");
4439 break;
4440 }
4441
4442 default:
4443 fatal_unexpected_iid(id);
4444 break;
4445 }
4446 set_result(_gvn.transform(result));
4447 return true;
4448 }
4449
4450 #ifdef _LP64
4451 #define XTOP ,top() /*additional argument*/
4452 #else //_LP64
4453 #define XTOP /*no additional argument*/
4454 #endif //_LP64
4455
4456 //----------------------inline_unsafe_copyMemory-------------------------
4457 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4458 bool LibraryCallKit::inline_unsafe_copyMemory() {
4459 if (callee()->is_static()) return false; // caller must have the capability!
4460 null_check_receiver(); // null-check receiver
4461 if (stopped()) return true;
4462
4463 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4464
4465 Node* src_ptr = argument(1); // type: oop
4466 Node* src_off = ConvL2X(argument(2)); // type: long
4467 Node* dst_ptr = argument(4); // type: oop
4468 Node* dst_off = ConvL2X(argument(5)); // type: long
4469 Node* size = ConvL2X(argument(7)); // type: long
4470
4471 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4472 "fieldOffset must be byte-scaled");
4473
4474 Node* src = make_unsafe_address(src_ptr, src_off);
4475 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4476
4477 // Conservatively insert a memory barrier on all memory slices.
4478 // Do not let writes of the copy source or destination float below the copy.
4479 insert_mem_bar(Op_MemBarCPUOrder);
4480
4481 // Call it. Note that the length argument is not scaled.
4482 make_runtime_call(RC_LEAF|RC_NO_FP,
4483 OptoRuntime::fast_arraycopy_Type(),
4484 StubRoutines::unsafe_arraycopy(),
4485 "unsafe_arraycopy",
4486 TypeRawPtr::BOTTOM,
4487 src, dst, size XTOP);
4488
4489 // Do not let reads of the copy destination float above the copy.
4490 insert_mem_bar(Op_MemBarCPUOrder);
4491
4492 return true;
4493 }
4494
4495 //------------------------clone_coping-----------------------------------
4496 // Helper function for inline_native_clone.
4497 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4498 assert(obj_size != NULL, "");
4499 Node* raw_obj = alloc_obj->in(1);
4500 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4501
4502 AllocateNode* alloc = NULL;
4503 if (ReduceBulkZeroing) {
4504 // We will be completely responsible for initializing this object -
4505 // mark Initialize node as complete.
4506 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4507 // The object was just allocated - there should be no any stores!
4508 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4509 // Mark as complete_with_arraycopy so that on AllocateNode
4510 // expansion, we know this AllocateNode is initialized by an array
4511 // copy and a StoreStore barrier exists after the array copy.
4512 alloc->initialization()->set_complete_with_arraycopy();
4513 }
4514
4515 // Copy the fastest available way.
4516 // TODO: generate fields copies for small objects instead.
4517 Node* src = obj;
4518 Node* dest = alloc_obj;
4519 Node* size = _gvn.transform(obj_size);
4520
4521 // Exclude the header but include array length to copy by 8 bytes words.
4522 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4523 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4524 instanceOopDesc::base_offset_in_bytes();
4525 // base_off:
4526 // 8 - 32-bit VM
4527 // 12 - 64-bit VM, compressed klass
4528 // 16 - 64-bit VM, normal klass
4529 if (base_off % BytesPerLong != 0) {
4530 assert(UseCompressedClassPointers, "");
4531 if (is_array) {
4532 // Exclude length to copy by 8 bytes words.
4533 base_off += sizeof(int);
4534 } else {
4535 // Include klass to copy by 8 bytes words.
4536 base_off = instanceOopDesc::klass_offset_in_bytes();
4537 }
4538 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4539 }
4540 src = basic_plus_adr(src, base_off);
4541 dest = basic_plus_adr(dest, base_off);
4542
4543 // Compute the length also, if needed:
4544 Node* countx = size;
4545 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4546 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4547
4548 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4549
4550 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4551 ac->set_clonebasic();
4552 Node* n = _gvn.transform(ac);
4553 if (n == ac) {
4554 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4555 } else {
4556 set_all_memory(n);
4557 }
4558
4559 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4560 if (card_mark) {
4561 assert(!is_array, "");
4562 // Put in store barrier for any and all oops we are sticking
4563 // into this object. (We could avoid this if we could prove
4564 // that the object type contains no oop fields at all.)
4565 Node* no_particular_value = NULL;
4566 Node* no_particular_field = NULL;
4567 int raw_adr_idx = Compile::AliasIdxRaw;
4568 post_barrier(control(),
4569 memory(raw_adr_type),
4570 alloc_obj,
4571 no_particular_field,
4572 raw_adr_idx,
4573 no_particular_value,
4574 T_OBJECT,
4575 false);
4576 }
4577
4578 // Do not let reads from the cloned object float above the arraycopy.
4579 if (alloc != NULL) {
4580 // Do not let stores that initialize this object be reordered with
4581 // a subsequent store that would make this object accessible by
4582 // other threads.
4583 // Record what AllocateNode this StoreStore protects so that
4584 // escape analysis can go from the MemBarStoreStoreNode to the
4585 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4586 // based on the escape status of the AllocateNode.
4587 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4588 } else {
4589 insert_mem_bar(Op_MemBarCPUOrder);
4590 }
4591 }
4592
4593 //------------------------inline_native_clone----------------------------
4594 // protected native Object java.lang.Object.clone();
4595 //
4596 // Here are the simple edge cases:
4597 // null receiver => normal trap
4598 // virtual and clone was overridden => slow path to out-of-line clone
4599 // not cloneable or finalizer => slow path to out-of-line Object.clone
4600 //
4601 // The general case has two steps, allocation and copying.
4602 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4603 //
4604 // Copying also has two cases, oop arrays and everything else.
4605 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4606 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4607 //
4608 // These steps fold up nicely if and when the cloned object's klass
4609 // can be sharply typed as an object array, a type array, or an instance.
4610 //
4611 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4612 PhiNode* result_val;
4613
4614 // Set the reexecute bit for the interpreter to reexecute
4615 // the bytecode that invokes Object.clone if deoptimization happens.
4616 { PreserveReexecuteState preexecs(this);
4617 jvms()->set_should_reexecute(true);
4618
4619 Node* obj = null_check_receiver();
4620 if (stopped()) return true;
4621
4622 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4623
4624 // If we are going to clone an instance, we need its exact type to
4625 // know the number and types of fields to convert the clone to
4626 // loads/stores. Maybe a speculative type can help us.
4627 if (!obj_type->klass_is_exact() &&
4628 obj_type->speculative_type() != NULL &&
4629 obj_type->speculative_type()->is_instance_klass()) {
4630 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4631 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4632 !spec_ik->has_injected_fields()) {
4633 ciKlass* k = obj_type->klass();
4634 if (!k->is_instance_klass() ||
4635 k->as_instance_klass()->is_interface() ||
4636 k->as_instance_klass()->has_subklass()) {
4637 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4638 }
4639 }
4640 }
4641
4642 Node* obj_klass = load_object_klass(obj);
4643 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4644 const TypeOopPtr* toop = ((tklass != NULL)
4645 ? tklass->as_instance_type()
4646 : TypeInstPtr::NOTNULL);
4647
4648 // Conservatively insert a memory barrier on all memory slices.
4649 // Do not let writes into the original float below the clone.
4650 insert_mem_bar(Op_MemBarCPUOrder);
4651
4652 // paths into result_reg:
4653 enum {
4654 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4655 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4656 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4657 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4658 PATH_LIMIT
4659 };
4660 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4661 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4662 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4663 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4664 record_for_igvn(result_reg);
4665
4666 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4667 int raw_adr_idx = Compile::AliasIdxRaw;
4668
4669 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4670 if (array_ctl != NULL) {
4671 // It's an array.
4672 PreserveJVMState pjvms(this);
4673 set_control(array_ctl);
4674 Node* obj_length = load_array_length(obj);
4675 Node* obj_size = NULL;
4676 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4677
4678 if (!use_ReduceInitialCardMarks()) {
4679 // If it is an oop array, it requires very special treatment,
4680 // because card marking is required on each card of the array.
4681 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4682 if (is_obja != NULL) {
4683 PreserveJVMState pjvms2(this);
4684 set_control(is_obja);
4685 // Generate a direct call to the right arraycopy function(s).
4686 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4687 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4688 ac->set_cloneoop();
4689 Node* n = _gvn.transform(ac);
4690 assert(n == ac, "cannot disappear");
4691 ac->connect_outputs(this);
4692
4693 result_reg->init_req(_objArray_path, control());
4694 result_val->init_req(_objArray_path, alloc_obj);
4695 result_i_o ->set_req(_objArray_path, i_o());
4696 result_mem ->set_req(_objArray_path, reset_memory());
4697 }
4698 }
4699 // Otherwise, there are no card marks to worry about.
4700 // (We can dispense with card marks if we know the allocation
4701 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4702 // causes the non-eden paths to take compensating steps to
4703 // simulate a fresh allocation, so that no further
4704 // card marks are required in compiled code to initialize
4705 // the object.)
4706
4707 if (!stopped()) {
4708 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4709
4710 // Present the results of the copy.
4711 result_reg->init_req(_array_path, control());
4712 result_val->init_req(_array_path, alloc_obj);
4713 result_i_o ->set_req(_array_path, i_o());
4714 result_mem ->set_req(_array_path, reset_memory());
4715 }
4716 }
4717
4718 // We only go to the instance fast case code if we pass a number of guards.
4719 // The paths which do not pass are accumulated in the slow_region.
4720 RegionNode* slow_region = new RegionNode(1);
4721 record_for_igvn(slow_region);
4722 if (!stopped()) {
4723 // It's an instance (we did array above). Make the slow-path tests.
4724 // If this is a virtual call, we generate a funny guard. We grab
4725 // the vtable entry corresponding to clone() from the target object.
4726 // If the target method which we are calling happens to be the
4727 // Object clone() method, we pass the guard. We do not need this
4728 // guard for non-virtual calls; the caller is known to be the native
4729 // Object clone().
4730 if (is_virtual) {
4731 generate_virtual_guard(obj_klass, slow_region);
4732 }
4733
4734 // The object must be cloneable and must not have a finalizer.
4735 // Both of these conditions may be checked in a single test.
4736 // We could optimize the cloneable test further, but we don't care.
4737 generate_access_flags_guard(obj_klass,
4738 // Test both conditions:
4739 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4740 // Must be cloneable but not finalizer:
4741 JVM_ACC_IS_CLONEABLE,
4742 slow_region);
4743 }
4744
4745 if (!stopped()) {
4746 // It's an instance, and it passed the slow-path tests.
4747 PreserveJVMState pjvms(this);
4748 Node* obj_size = NULL;
4749 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4750 // is reexecuted if deoptimization occurs and there could be problems when merging
4751 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4752 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4753
4754 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4755
4756 // Present the results of the slow call.
4757 result_reg->init_req(_instance_path, control());
4758 result_val->init_req(_instance_path, alloc_obj);
4759 result_i_o ->set_req(_instance_path, i_o());
4760 result_mem ->set_req(_instance_path, reset_memory());
4761 }
4762
4763 // Generate code for the slow case. We make a call to clone().
4764 set_control(_gvn.transform(slow_region));
4765 if (!stopped()) {
4766 PreserveJVMState pjvms(this);
4767 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4768 Node* slow_result = set_results_for_java_call(slow_call);
4769 // this->control() comes from set_results_for_java_call
4770 result_reg->init_req(_slow_path, control());
4771 result_val->init_req(_slow_path, slow_result);
4772 result_i_o ->set_req(_slow_path, i_o());
4773 result_mem ->set_req(_slow_path, reset_memory());
4774 }
4775
4776 // Return the combined state.
4777 set_control( _gvn.transform(result_reg));
4778 set_i_o( _gvn.transform(result_i_o));
4779 set_all_memory( _gvn.transform(result_mem));
4780 } // original reexecute is set back here
4781
4782 set_result(_gvn.transform(result_val));
4783 return true;
4784 }
4785
4786 // If we have a tighly coupled allocation, the arraycopy may take care
4787 // of the array initialization. If one of the guards we insert between
4788 // the allocation and the arraycopy causes a deoptimization, an
4789 // unitialized array will escape the compiled method. To prevent that
4790 // we set the JVM state for uncommon traps between the allocation and
4791 // the arraycopy to the state before the allocation so, in case of
4792 // deoptimization, we'll reexecute the allocation and the
4793 // initialization.
4794 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4795 if (alloc != NULL) {
4796 ciMethod* trap_method = alloc->jvms()->method();
4797 int trap_bci = alloc->jvms()->bci();
4798
4799 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4800 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4801 // Make sure there's no store between the allocation and the
4802 // arraycopy otherwise visible side effects could be rexecuted
4803 // in case of deoptimization and cause incorrect execution.
4804 bool no_interfering_store = true;
4805 Node* mem = alloc->in(TypeFunc::Memory);
4806 if (mem->is_MergeMem()) {
4807 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4808 Node* n = mms.memory();
4809 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4810 assert(n->is_Store(), "what else?");
4811 no_interfering_store = false;
4812 break;
4813 }
4814 }
4815 } else {
4816 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4817 Node* n = mms.memory();
4818 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4819 assert(n->is_Store(), "what else?");
4820 no_interfering_store = false;
4821 break;
4822 }
4823 }
4824 }
4825
4826 if (no_interfering_store) {
4827 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4828 uint size = alloc->req();
4829 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4830 old_jvms->set_map(sfpt);
4831 for (uint i = 0; i < size; i++) {
4832 sfpt->init_req(i, alloc->in(i));
4833 }
4834 // re-push array length for deoptimization
4835 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4836 old_jvms->set_sp(old_jvms->sp()+1);
4837 old_jvms->set_monoff(old_jvms->monoff()+1);
4838 old_jvms->set_scloff(old_jvms->scloff()+1);
4839 old_jvms->set_endoff(old_jvms->endoff()+1);
4840 old_jvms->set_should_reexecute(true);
4841
4842 sfpt->set_i_o(map()->i_o());
4843 sfpt->set_memory(map()->memory());
4844 sfpt->set_control(map()->control());
4845
4846 JVMState* saved_jvms = jvms();
4847 saved_reexecute_sp = _reexecute_sp;
4848
4849 set_jvms(sfpt->jvms());
4850 _reexecute_sp = jvms()->sp();
4851
4852 return saved_jvms;
4853 }
4854 }
4855 }
4856 return NULL;
4857 }
4858
4859 // In case of a deoptimization, we restart execution at the
4860 // allocation, allocating a new array. We would leave an uninitialized
4861 // array in the heap that GCs wouldn't expect. Move the allocation
4862 // after the traps so we don't allocate the array if we
4863 // deoptimize. This is possible because tightly_coupled_allocation()
4864 // guarantees there's no observer of the allocated array at this point
4865 // and the control flow is simple enough.
4866 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4867 if (saved_jvms != NULL && !stopped()) {
4868 assert(alloc != NULL, "only with a tightly coupled allocation");
4869 // restore JVM state to the state at the arraycopy
4870 saved_jvms->map()->set_control(map()->control());
4871 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4872 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4873 // If we've improved the types of some nodes (null check) while
4874 // emitting the guards, propagate them to the current state
4875 map()->replaced_nodes().apply(saved_jvms->map());
4876 set_jvms(saved_jvms);
4877 _reexecute_sp = saved_reexecute_sp;
4878
4879 // Remove the allocation from above the guards
4880 CallProjections callprojs;
4881 alloc->extract_projections(&callprojs, true);
4882 InitializeNode* init = alloc->initialization();
4883 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4884 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4885 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4886 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4887
4888 // move the allocation here (after the guards)
4889 _gvn.hash_delete(alloc);
4890 alloc->set_req(TypeFunc::Control, control());
4891 alloc->set_req(TypeFunc::I_O, i_o());
4892 Node *mem = reset_memory();
4893 set_all_memory(mem);
4894 alloc->set_req(TypeFunc::Memory, mem);
4895 set_control(init->proj_out(TypeFunc::Control));
4896 set_i_o(callprojs.fallthrough_ioproj);
4897
4898 // Update memory as done in GraphKit::set_output_for_allocation()
4899 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4900 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4901 if (ary_type->isa_aryptr() && length_type != NULL) {
4902 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4903 }
4904 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4905 int elemidx = C->get_alias_index(telemref);
4906 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4907 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4908
4909 Node* allocx = _gvn.transform(alloc);
4910 assert(allocx == alloc, "where has the allocation gone?");
4911 assert(dest->is_CheckCastPP(), "not an allocation result?");
4912
4913 _gvn.hash_delete(dest);
4914 dest->set_req(0, control());
4915 Node* destx = _gvn.transform(dest);
4916 assert(destx == dest, "where has the allocation result gone?");
4917 }
4918 }
4919
4920
4921 //------------------------------inline_arraycopy-----------------------
4922 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4923 // Object dest, int destPos,
4924 // int length);
4925 bool LibraryCallKit::inline_arraycopy() {
4926 // Get the arguments.
4927 Node* src = argument(0); // type: oop
4928 Node* src_offset = argument(1); // type: int
4929 Node* dest = argument(2); // type: oop
4930 Node* dest_offset = argument(3); // type: int
4931 Node* length = argument(4); // type: int
4932
4933
4934 // Check for allocation before we add nodes that would confuse
4935 // tightly_coupled_allocation()
4936 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4937
4938 int saved_reexecute_sp = -1;
4939 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4940 // See arraycopy_restore_alloc_state() comment
4941 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4942 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4943 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4944 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4945
4946 // The following tests must be performed
4947 // (1) src and dest are arrays.
4948 // (2) src and dest arrays must have elements of the same BasicType
4949 // (3) src and dest must not be null.
4950 // (4) src_offset must not be negative.
4951 // (5) dest_offset must not be negative.
4952 // (6) length must not be negative.
4953 // (7) src_offset + length must not exceed length of src.
4954 // (8) dest_offset + length must not exceed length of dest.
4955 // (9) each element of an oop array must be assignable
4956
4957 // (3) src and dest must not be null.
4958 // always do this here because we need the JVM state for uncommon traps
4959 Node* null_ctl = top();
4960 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4961 assert(null_ctl->is_top(), "no null control here");
4962 dest = null_check(dest, T_ARRAY);
4963
4964 if (!can_emit_guards) {
4965 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4966 // guards but the arraycopy node could still take advantage of a
4967 // tightly allocated allocation. tightly_coupled_allocation() is
4968 // called again to make sure it takes the null check above into
4969 // account: the null check is mandatory and if it caused an
4970 // uncommon trap to be emitted then the allocation can't be
4971 // considered tightly coupled in this context.
4972 alloc = tightly_coupled_allocation(dest, NULL);
4973 }
4974
4975 bool validated = false;
4976
4977 const Type* src_type = _gvn.type(src);
4978 const Type* dest_type = _gvn.type(dest);
4979 const TypeAryPtr* top_src = src_type->isa_aryptr();
4980 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4981
4982 // Do we have the type of src?
4983 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4984 // Do we have the type of dest?
4985 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4986 // Is the type for src from speculation?
4987 bool src_spec = false;
4988 // Is the type for dest from speculation?
4989 bool dest_spec = false;
4990
4991 if ((!has_src || !has_dest) && can_emit_guards) {
4992 // We don't have sufficient type information, let's see if
4993 // speculative types can help. We need to have types for both src
4994 // and dest so that it pays off.
4995
4996 // Do we already have or could we have type information for src
4997 bool could_have_src = has_src;
4998 // Do we already have or could we have type information for dest
4999 bool could_have_dest = has_dest;
5000
5001 ciKlass* src_k = NULL;
5002 if (!has_src) {
5003 src_k = src_type->speculative_type_not_null();
5004 if (src_k != NULL && src_k->is_array_klass()) {
5005 could_have_src = true;
5006 }
5007 }
5008
5009 ciKlass* dest_k = NULL;
5010 if (!has_dest) {
5011 dest_k = dest_type->speculative_type_not_null();
5012 if (dest_k != NULL && dest_k->is_array_klass()) {
5013 could_have_dest = true;
5014 }
5015 }
5016
5017 if (could_have_src && could_have_dest) {
5018 // This is going to pay off so emit the required guards
5019 if (!has_src) {
5020 src = maybe_cast_profiled_obj(src, src_k, true);
5021 src_type = _gvn.type(src);
5022 top_src = src_type->isa_aryptr();
5023 has_src = (top_src != NULL && top_src->klass() != NULL);
5024 src_spec = true;
5025 }
5026 if (!has_dest) {
5027 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5028 dest_type = _gvn.type(dest);
5029 top_dest = dest_type->isa_aryptr();
5030 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5031 dest_spec = true;
5032 }
5033 }
5034 }
5035
5036 if (has_src && has_dest && can_emit_guards) {
5037 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5038 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5039 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5040 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5041
5042 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5043 // If both arrays are object arrays then having the exact types
5044 // for both will remove the need for a subtype check at runtime
5045 // before the call and may make it possible to pick a faster copy
5046 // routine (without a subtype check on every element)
5047 // Do we have the exact type of src?
5048 bool could_have_src = src_spec;
5049 // Do we have the exact type of dest?
5050 bool could_have_dest = dest_spec;
5051 ciKlass* src_k = top_src->klass();
5052 ciKlass* dest_k = top_dest->klass();
5053 if (!src_spec) {
5054 src_k = src_type->speculative_type_not_null();
5055 if (src_k != NULL && src_k->is_array_klass()) {
5056 could_have_src = true;
5057 }
5058 }
5059 if (!dest_spec) {
5060 dest_k = dest_type->speculative_type_not_null();
5061 if (dest_k != NULL && dest_k->is_array_klass()) {
5062 could_have_dest = true;
5063 }
5064 }
5065 if (could_have_src && could_have_dest) {
5066 // If we can have both exact types, emit the missing guards
5067 if (could_have_src && !src_spec) {
5068 src = maybe_cast_profiled_obj(src, src_k, true);
5069 }
5070 if (could_have_dest && !dest_spec) {
5071 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5072 }
5073 }
5074 }
5075 }
5076
5077 ciMethod* trap_method = method();
5078 int trap_bci = bci();
5079 if (saved_jvms != NULL) {
5080 trap_method = alloc->jvms()->method();
5081 trap_bci = alloc->jvms()->bci();
5082 }
5083
5084 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5085 can_emit_guards &&
5086 !src->is_top() && !dest->is_top()) {
5087 // validate arguments: enables transformation the ArrayCopyNode
5088 validated = true;
5089
5090 RegionNode* slow_region = new RegionNode(1);
5091 record_for_igvn(slow_region);
5092
5093 // (1) src and dest are arrays.
5094 generate_non_array_guard(load_object_klass(src), slow_region);
5095 generate_non_array_guard(load_object_klass(dest), slow_region);
5096
5097 // (2) src and dest arrays must have elements of the same BasicType
5098 // done at macro expansion or at Ideal transformation time
5099
5100 // (4) src_offset must not be negative.
5101 generate_negative_guard(src_offset, slow_region);
5102
5103 // (5) dest_offset must not be negative.
5104 generate_negative_guard(dest_offset, slow_region);
5105
5106 // (7) src_offset + length must not exceed length of src.
5107 generate_limit_guard(src_offset, length,
5108 load_array_length(src),
5109 slow_region);
5110
5111 // (8) dest_offset + length must not exceed length of dest.
5112 generate_limit_guard(dest_offset, length,
5113 load_array_length(dest),
5114 slow_region);
5115
5116 // (9) each element of an oop array must be assignable
5117 Node* src_klass = load_object_klass(src);
5118 Node* dest_klass = load_object_klass(dest);
5119 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5120
5121 if (not_subtype_ctrl != top()) {
5122 PreserveJVMState pjvms(this);
5123 set_control(not_subtype_ctrl);
5124 uncommon_trap(Deoptimization::Reason_intrinsic,
5125 Deoptimization::Action_make_not_entrant);
5126 assert(stopped(), "Should be stopped");
5127 }
5128 {
5129 PreserveJVMState pjvms(this);
5130 set_control(_gvn.transform(slow_region));
5131 uncommon_trap(Deoptimization::Reason_intrinsic,
5132 Deoptimization::Action_make_not_entrant);
5133 assert(stopped(), "Should be stopped");
5134 }
5135 }
5136
5137 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5138
5139 if (stopped()) {
5140 return true;
5141 }
5142
5143 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5144 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5145 // so the compiler has a chance to eliminate them: during macro expansion,
5146 // we have to set their control (CastPP nodes are eliminated).
5147 load_object_klass(src), load_object_klass(dest),
5148 load_array_length(src), load_array_length(dest));
5149
5150 ac->set_arraycopy(validated);
5151
5152 Node* n = _gvn.transform(ac);
5153 if (n == ac) {
5154 ac->connect_outputs(this);
5155 } else {
5156 assert(validated, "shouldn't transform if all arguments not validated");
5157 set_all_memory(n);
5158 }
5159
5160 return true;
5161 }
5162
5163
5164 // Helper function which determines if an arraycopy immediately follows
5165 // an allocation, with no intervening tests or other escapes for the object.
5166 AllocateArrayNode*
5167 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5168 RegionNode* slow_region) {
5169 if (stopped()) return NULL; // no fast path
5170 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5171
5172 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5173 if (alloc == NULL) return NULL;
5174
5175 Node* rawmem = memory(Compile::AliasIdxRaw);
5176 // Is the allocation's memory state untouched?
5177 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5178 // Bail out if there have been raw-memory effects since the allocation.
5179 // (Example: There might have been a call or safepoint.)
5180 return NULL;
5181 }
5182 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5183 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5184 return NULL;
5185 }
5186
5187 // There must be no unexpected observers of this allocation.
5188 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5189 Node* obs = ptr->fast_out(i);
5190 if (obs != this->map()) {
5191 return NULL;
5192 }
5193 }
5194
5195 // This arraycopy must unconditionally follow the allocation of the ptr.
5196 Node* alloc_ctl = ptr->in(0);
5197 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5198
5199 Node* ctl = control();
5200 while (ctl != alloc_ctl) {
5201 // There may be guards which feed into the slow_region.
5202 // Any other control flow means that we might not get a chance
5203 // to finish initializing the allocated object.
5204 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5205 IfNode* iff = ctl->in(0)->as_If();
5206 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5207 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5208 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5209 ctl = iff->in(0); // This test feeds the known slow_region.
5210 continue;
5211 }
5212 // One more try: Various low-level checks bottom out in
5213 // uncommon traps. If the debug-info of the trap omits
5214 // any reference to the allocation, as we've already
5215 // observed, then there can be no objection to the trap.
5216 bool found_trap = false;
5217 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5218 Node* obs = not_ctl->fast_out(j);
5219 if (obs->in(0) == not_ctl && obs->is_Call() &&
5220 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5221 found_trap = true; break;
5222 }
5223 }
5224 if (found_trap) {
5225 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5226 continue;
5227 }
5228 }
5229 return NULL;
5230 }
5231
5232 // If we get this far, we have an allocation which immediately
5233 // precedes the arraycopy, and we can take over zeroing the new object.
5234 // The arraycopy will finish the initialization, and provide
5235 // a new control state to which we will anchor the destination pointer.
5236
5237 return alloc;
5238 }
5239
5240 //-------------inline_encodeISOArray-----------------------------------
5241 // encode char[] to byte[] in ISO_8859_1
5242 bool LibraryCallKit::inline_encodeISOArray() {
5243 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5244 // no receiver since it is static method
5245 Node *src = argument(0);
5246 Node *src_offset = argument(1);
5247 Node *dst = argument(2);
5248 Node *dst_offset = argument(3);
5249 Node *length = argument(4);
5250
5251 const Type* src_type = src->Value(&_gvn);
5252 const Type* dst_type = dst->Value(&_gvn);
5253 const TypeAryPtr* top_src = src_type->isa_aryptr();
5254 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5255 if (top_src == NULL || top_src->klass() == NULL ||
5256 top_dest == NULL || top_dest->klass() == NULL) {
5257 // failed array check
5258 return false;
5259 }
5260
5261 // Figure out the size and type of the elements we will be copying.
5262 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5263 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5264 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5265 return false;
5266 }
5267 Node* src_start = array_element_address(src, src_offset, src_elem);
5268 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5269 // 'src_start' points to src array + scaled offset
5270 // 'dst_start' points to dst array + scaled offset
5271
5272 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5273 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5274 enc = _gvn.transform(enc);
5275 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5276 set_memory(res_mem, mtype);
5277 set_result(enc);
5278 return true;
5279 }
5280
5281 //-------------inline_multiplyToLen-----------------------------------
5282 bool LibraryCallKit::inline_multiplyToLen() {
5283 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5284
5285 address stubAddr = StubRoutines::multiplyToLen();
5286 if (stubAddr == NULL) {
5287 return false; // Intrinsic's stub is not implemented on this platform
5288 }
5289 const char* stubName = "multiplyToLen";
5290
5291 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5292
5293 // no receiver because it is a static method
5294 Node* x = argument(0);
5295 Node* xlen = argument(1);
5296 Node* y = argument(2);
5297 Node* ylen = argument(3);
5298 Node* z = argument(4);
5299
5300 const Type* x_type = x->Value(&_gvn);
5301 const Type* y_type = y->Value(&_gvn);
5302 const TypeAryPtr* top_x = x_type->isa_aryptr();
5303 const TypeAryPtr* top_y = y_type->isa_aryptr();
5304 if (top_x == NULL || top_x->klass() == NULL ||
5305 top_y == NULL || top_y->klass() == NULL) {
5306 // failed array check
5307 return false;
5308 }
5309
5310 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5311 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5312 if (x_elem != T_INT || y_elem != T_INT) {
5313 return false;
5314 }
5315
5316 // Set the original stack and the reexecute bit for the interpreter to reexecute
5317 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5318 // on the return from z array allocation in runtime.
5319 { PreserveReexecuteState preexecs(this);
5320 jvms()->set_should_reexecute(true);
5321
5322 Node* x_start = array_element_address(x, intcon(0), x_elem);
5323 Node* y_start = array_element_address(y, intcon(0), y_elem);
5324 // 'x_start' points to x array + scaled xlen
5325 // 'y_start' points to y array + scaled ylen
5326
5327 // Allocate the result array
5328 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5329 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5330 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5331
5332 IdealKit ideal(this);
5333
5334 #define __ ideal.
5335 Node* one = __ ConI(1);
5336 Node* zero = __ ConI(0);
5337 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5338 __ set(need_alloc, zero);
5339 __ set(z_alloc, z);
5340 __ if_then(z, BoolTest::eq, null()); {
5341 __ increment (need_alloc, one);
5342 } __ else_(); {
5343 // Update graphKit memory and control from IdealKit.
5344 sync_kit(ideal);
5345 Node* zlen_arg = load_array_length(z);
5346 // Update IdealKit memory and control from graphKit.
5347 __ sync_kit(this);
5348 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5349 __ increment (need_alloc, one);
5350 } __ end_if();
5351 } __ end_if();
5352
5353 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5354 // Update graphKit memory and control from IdealKit.
5355 sync_kit(ideal);
5356 Node * narr = new_array(klass_node, zlen, 1);
5357 // Update IdealKit memory and control from graphKit.
5358 __ sync_kit(this);
5359 __ set(z_alloc, narr);
5360 } __ end_if();
5361
5362 sync_kit(ideal);
5363 z = __ value(z_alloc);
5364 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5365 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5366 // Final sync IdealKit and GraphKit.
5367 final_sync(ideal);
5368 #undef __
5369
5370 Node* z_start = array_element_address(z, intcon(0), T_INT);
5371
5372 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5373 OptoRuntime::multiplyToLen_Type(),
5374 stubAddr, stubName, TypePtr::BOTTOM,
5375 x_start, xlen, y_start, ylen, z_start, zlen);
5376 } // original reexecute is set back here
5377
5378 C->set_has_split_ifs(true); // Has chance for split-if optimization
5379 set_result(z);
5380 return true;
5381 }
5382
5383 //-------------inline_squareToLen------------------------------------
5384 bool LibraryCallKit::inline_squareToLen() {
5385 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5386
5387 address stubAddr = StubRoutines::squareToLen();
5388 if (stubAddr == NULL) {
5389 return false; // Intrinsic's stub is not implemented on this platform
5390 }
5391 const char* stubName = "squareToLen";
5392
5393 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5394
5395 Node* x = argument(0);
5396 Node* len = argument(1);
5397 Node* z = argument(2);
5398 Node* zlen = argument(3);
5399
5400 const Type* x_type = x->Value(&_gvn);
5401 const Type* z_type = z->Value(&_gvn);
5402 const TypeAryPtr* top_x = x_type->isa_aryptr();
5403 const TypeAryPtr* top_z = z_type->isa_aryptr();
5404 if (top_x == NULL || top_x->klass() == NULL ||
5405 top_z == NULL || top_z->klass() == NULL) {
5406 // failed array check
5407 return false;
5408 }
5409
5410 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5411 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5412 if (x_elem != T_INT || z_elem != T_INT) {
5413 return false;
5414 }
5415
5416
5417 Node* x_start = array_element_address(x, intcon(0), x_elem);
5418 Node* z_start = array_element_address(z, intcon(0), z_elem);
5419
5420 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5421 OptoRuntime::squareToLen_Type(),
5422 stubAddr, stubName, TypePtr::BOTTOM,
5423 x_start, len, z_start, zlen);
5424
5425 set_result(z);
5426 return true;
5427 }
5428
5429 //-------------inline_mulAdd------------------------------------------
5430 bool LibraryCallKit::inline_mulAdd() {
5431 assert(UseMulAddIntrinsic, "not implementated on this platform");
5432
5433 address stubAddr = StubRoutines::mulAdd();
5434 if (stubAddr == NULL) {
5435 return false; // Intrinsic's stub is not implemented on this platform
5436 }
5437 const char* stubName = "mulAdd";
5438
5439 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5440
5441 Node* out = argument(0);
5442 Node* in = argument(1);
5443 Node* offset = argument(2);
5444 Node* len = argument(3);
5445 Node* k = argument(4);
5446
5447 const Type* out_type = out->Value(&_gvn);
5448 const Type* in_type = in->Value(&_gvn);
5449 const TypeAryPtr* top_out = out_type->isa_aryptr();
5450 const TypeAryPtr* top_in = in_type->isa_aryptr();
5451 if (top_out == NULL || top_out->klass() == NULL ||
5452 top_in == NULL || top_in->klass() == NULL) {
5453 // failed array check
5454 return false;
5455 }
5456
5457 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5458 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5459 if (out_elem != T_INT || in_elem != T_INT) {
5460 return false;
5461 }
5462
5463 Node* outlen = load_array_length(out);
5464 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5465 Node* out_start = array_element_address(out, intcon(0), out_elem);
5466 Node* in_start = array_element_address(in, intcon(0), in_elem);
5467
5468 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5469 OptoRuntime::mulAdd_Type(),
5470 stubAddr, stubName, TypePtr::BOTTOM,
5471 out_start,in_start, new_offset, len, k);
5472 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5473 set_result(result);
5474 return true;
5475 }
5476
5477 //-------------inline_montgomeryMultiply-----------------------------------
5478 bool LibraryCallKit::inline_montgomeryMultiply() {
5479 address stubAddr = StubRoutines::montgomeryMultiply();
5480 if (stubAddr == NULL) {
5481 return false; // Intrinsic's stub is not implemented on this platform
5482 }
5483
5484 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5485 const char* stubName = "montgomery_square";
5486
5487 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5488
5489 Node* a = argument(0);
5490 Node* b = argument(1);
5491 Node* n = argument(2);
5492 Node* len = argument(3);
5493 Node* inv = argument(4);
5494 Node* m = argument(6);
5495
5496 const Type* a_type = a->Value(&_gvn);
5497 const TypeAryPtr* top_a = a_type->isa_aryptr();
5498 const Type* b_type = b->Value(&_gvn);
5499 const TypeAryPtr* top_b = b_type->isa_aryptr();
5500 const Type* n_type = a->Value(&_gvn);
5501 const TypeAryPtr* top_n = n_type->isa_aryptr();
5502 const Type* m_type = a->Value(&_gvn);
5503 const TypeAryPtr* top_m = m_type->isa_aryptr();
5504 if (top_a == NULL || top_a->klass() == NULL ||
5505 top_b == NULL || top_b->klass() == NULL ||
5506 top_n == NULL || top_n->klass() == NULL ||
5507 top_m == NULL || top_m->klass() == NULL) {
5508 // failed array check
5509 return false;
5510 }
5511
5512 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5513 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5514 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5515 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5516 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5517 return false;
5518 }
5519
5520 // Make the call
5521 {
5522 Node* a_start = array_element_address(a, intcon(0), a_elem);
5523 Node* b_start = array_element_address(b, intcon(0), b_elem);
5524 Node* n_start = array_element_address(n, intcon(0), n_elem);
5525 Node* m_start = array_element_address(m, intcon(0), m_elem);
5526
5527 Node* call = make_runtime_call(RC_LEAF,
5528 OptoRuntime::montgomeryMultiply_Type(),
5529 stubAddr, stubName, TypePtr::BOTTOM,
5530 a_start, b_start, n_start, len, inv, top(),
5531 m_start);
5532 set_result(m);
5533 }
5534
5535 return true;
5536 }
5537
5538 bool LibraryCallKit::inline_montgomerySquare() {
5539 address stubAddr = StubRoutines::montgomerySquare();
5540 if (stubAddr == NULL) {
5541 return false; // Intrinsic's stub is not implemented on this platform
5542 }
5543
5544 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5545 const char* stubName = "montgomery_square";
5546
5547 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5548
5549 Node* a = argument(0);
5550 Node* n = argument(1);
5551 Node* len = argument(2);
5552 Node* inv = argument(3);
5553 Node* m = argument(5);
5554
5555 const Type* a_type = a->Value(&_gvn);
5556 const TypeAryPtr* top_a = a_type->isa_aryptr();
5557 const Type* n_type = a->Value(&_gvn);
5558 const TypeAryPtr* top_n = n_type->isa_aryptr();
5559 const Type* m_type = a->Value(&_gvn);
5560 const TypeAryPtr* top_m = m_type->isa_aryptr();
5561 if (top_a == NULL || top_a->klass() == NULL ||
5562 top_n == NULL || top_n->klass() == NULL ||
5563 top_m == NULL || top_m->klass() == NULL) {
5564 // failed array check
5565 return false;
5566 }
5567
5568 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5569 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5570 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5571 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5572 return false;
5573 }
5574
5575 // Make the call
5576 {
5577 Node* a_start = array_element_address(a, intcon(0), a_elem);
5578 Node* n_start = array_element_address(n, intcon(0), n_elem);
5579 Node* m_start = array_element_address(m, intcon(0), m_elem);
5580
5581 Node* call = make_runtime_call(RC_LEAF,
5582 OptoRuntime::montgomerySquare_Type(),
5583 stubAddr, stubName, TypePtr::BOTTOM,
5584 a_start, n_start, len, inv, top(),
5585 m_start);
5586 set_result(m);
5587 }
5588
5589 return true;
5590 }
5591
5592
5593 /**
5594 * Calculate CRC32 for byte.
5595 * int java.util.zip.CRC32.update(int crc, int b)
5596 */
5597 bool LibraryCallKit::inline_updateCRC32() {
5598 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5599 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5600 // no receiver since it is static method
5601 Node* crc = argument(0); // type: int
5602 Node* b = argument(1); // type: int
5603
5604 /*
5605 * int c = ~ crc;
5606 * b = timesXtoThe32[(b ^ c) & 0xFF];
5607 * b = b ^ (c >>> 8);
5608 * crc = ~b;
5609 */
5610
5611 Node* M1 = intcon(-1);
5612 crc = _gvn.transform(new XorINode(crc, M1));
5613 Node* result = _gvn.transform(new XorINode(crc, b));
5614 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5615
5616 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5617 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5618 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5619 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5620
5621 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5622 result = _gvn.transform(new XorINode(crc, result));
5623 result = _gvn.transform(new XorINode(result, M1));
5624 set_result(result);
5625 return true;
5626 }
5627
5628 /**
5629 * Calculate CRC32 for byte[] array.
5630 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5631 */
5632 bool LibraryCallKit::inline_updateBytesCRC32() {
5633 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5634 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5635 // no receiver since it is static method
5636 Node* crc = argument(0); // type: int
5637 Node* src = argument(1); // type: oop
5638 Node* offset = argument(2); // type: int
5639 Node* length = argument(3); // type: int
5640
5641 const Type* src_type = src->Value(&_gvn);
5642 const TypeAryPtr* top_src = src_type->isa_aryptr();
5643 if (top_src == NULL || top_src->klass() == NULL) {
5644 // failed array check
5645 return false;
5646 }
5647
5648 // Figure out the size and type of the elements we will be copying.
5649 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5650 if (src_elem != T_BYTE) {
5651 return false;
5652 }
5653
5654 // 'src_start' points to src array + scaled offset
5655 Node* src_start = array_element_address(src, offset, src_elem);
5656
5657 // We assume that range check is done by caller.
5658 // TODO: generate range check (offset+length < src.length) in debug VM.
5659
5660 // Call the stub.
5661 address stubAddr = StubRoutines::updateBytesCRC32();
5662 const char *stubName = "updateBytesCRC32";
5663
5664 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5665 stubAddr, stubName, TypePtr::BOTTOM,
5666 crc, src_start, length);
5667 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5668 set_result(result);
5669 return true;
5670 }
5671
5672 /**
5673 * Calculate CRC32 for ByteBuffer.
5674 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5675 */
5676 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5677 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5678 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5679 // no receiver since it is static method
5680 Node* crc = argument(0); // type: int
5681 Node* src = argument(1); // type: long
5682 Node* offset = argument(3); // type: int
5683 Node* length = argument(4); // type: int
5684
5685 src = ConvL2X(src); // adjust Java long to machine word
5686 Node* base = _gvn.transform(new CastX2PNode(src));
5687 offset = ConvI2X(offset);
5688
5689 // 'src_start' points to src array + scaled offset
5690 Node* src_start = basic_plus_adr(top(), base, offset);
5691
5692 // Call the stub.
5693 address stubAddr = StubRoutines::updateBytesCRC32();
5694 const char *stubName = "updateBytesCRC32";
5695
5696 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5697 stubAddr, stubName, TypePtr::BOTTOM,
5698 crc, src_start, length);
5699 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5700 set_result(result);
5701 return true;
5702 }
5703
5704 //------------------------------get_table_from_crc32c_class-----------------------
5705 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5706 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5707 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5708
5709 return table;
5710 }
5711
5712 //------------------------------inline_updateBytesCRC32C-----------------------
5713 //
5714 // Calculate CRC32C for byte[] array.
5715 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5716 //
5717 bool LibraryCallKit::inline_updateBytesCRC32C() {
5718 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5719 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5720 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5721 // no receiver since it is a static method
5722 Node* crc = argument(0); // type: int
5723 Node* src = argument(1); // type: oop
5724 Node* offset = argument(2); // type: int
5725 Node* end = argument(3); // type: int
5726
5727 Node* length = _gvn.transform(new SubINode(end, offset));
5728
5729 const Type* src_type = src->Value(&_gvn);
5730 const TypeAryPtr* top_src = src_type->isa_aryptr();
5731 if (top_src == NULL || top_src->klass() == NULL) {
5732 // failed array check
5733 return false;
5734 }
5735
5736 // Figure out the size and type of the elements we will be copying.
5737 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5738 if (src_elem != T_BYTE) {
5739 return false;
5740 }
5741
5742 // 'src_start' points to src array + scaled offset
5743 Node* src_start = array_element_address(src, offset, src_elem);
5744
5745 // static final int[] byteTable in class CRC32C
5746 Node* table = get_table_from_crc32c_class(callee()->holder());
5747 Node* table_start = array_element_address(table, intcon(0), T_INT);
5748
5749 // We assume that range check is done by caller.
5750 // TODO: generate range check (offset+length < src.length) in debug VM.
5751
5752 // Call the stub.
5753 address stubAddr = StubRoutines::updateBytesCRC32C();
5754 const char *stubName = "updateBytesCRC32C";
5755
5756 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5757 stubAddr, stubName, TypePtr::BOTTOM,
5758 crc, src_start, length, table_start);
5759 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5760 set_result(result);
5761 return true;
5762 }
5763
5764 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5765 //
5766 // Calculate CRC32C for DirectByteBuffer.
5767 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5768 //
5769 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5770 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5771 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5772 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5773 // no receiver since it is a static method
5774 Node* crc = argument(0); // type: int
5775 Node* src = argument(1); // type: long
5776 Node* offset = argument(3); // type: int
5777 Node* end = argument(4); // type: int
5778
5779 Node* length = _gvn.transform(new SubINode(end, offset));
5780
5781 src = ConvL2X(src); // adjust Java long to machine word
5782 Node* base = _gvn.transform(new CastX2PNode(src));
5783 offset = ConvI2X(offset);
5784
5785 // 'src_start' points to src array + scaled offset
5786 Node* src_start = basic_plus_adr(top(), base, offset);
5787
5788 // static final int[] byteTable in class CRC32C
5789 Node* table = get_table_from_crc32c_class(callee()->holder());
5790 Node* table_start = array_element_address(table, intcon(0), T_INT);
5791
5792 // Call the stub.
5793 address stubAddr = StubRoutines::updateBytesCRC32C();
5794 const char *stubName = "updateBytesCRC32C";
5795
5796 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5797 stubAddr, stubName, TypePtr::BOTTOM,
5798 crc, src_start, length, table_start);
5799 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5800 set_result(result);
5801 return true;
5802 }
5803
5804 //----------------------------inline_reference_get----------------------------
5805 // public T java.lang.ref.Reference.get();
5806 bool LibraryCallKit::inline_reference_get() {
5807 const int referent_offset = java_lang_ref_Reference::referent_offset;
5808 guarantee(referent_offset > 0, "should have already been set");
5809
5810 // Get the argument:
5811 Node* reference_obj = null_check_receiver();
5812 if (stopped()) return true;
5813
5814 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5815
5816 ciInstanceKlass* klass = env()->Object_klass();
5817 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5818
5819 Node* no_ctrl = NULL;
5820 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5821
5822 // Use the pre-barrier to record the value in the referent field
5823 pre_barrier(false /* do_load */,
5824 control(),
5825 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5826 result /* pre_val */,
5827 T_OBJECT);
5828
5829 // Add memory barrier to prevent commoning reads from this field
5830 // across safepoint since GC can change its value.
5831 insert_mem_bar(Op_MemBarCPUOrder);
5832
5833 set_result(result);
5834 return true;
5835 }
5836
5837
5838 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5839 bool is_exact=true, bool is_static=false,
5840 ciInstanceKlass * fromKls=NULL) {
5841 if (fromKls == NULL) {
5842 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5843 assert(tinst != NULL, "obj is null");
5844 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5845 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5846 fromKls = tinst->klass()->as_instance_klass();
5847 } else {
5848 assert(is_static, "only for static field access");
5849 }
5850 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5851 ciSymbol::make(fieldTypeString),
5852 is_static);
5853
5854 assert (field != NULL, "undefined field");
5855 if (field == NULL) return (Node *) NULL;
5856
5857 if (is_static) {
5858 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5859 fromObj = makecon(tip);
5860 }
5861
5862 // Next code copied from Parse::do_get_xxx():
5863
5864 // Compute address and memory type.
5865 int offset = field->offset_in_bytes();
5866 bool is_vol = field->is_volatile();
5867 ciType* field_klass = field->type();
5868 assert(field_klass->is_loaded(), "should be loaded");
5869 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5870 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5871 BasicType bt = field->layout_type();
5872
5873 // Build the resultant type of the load
5874 const Type *type;
5875 if (bt == T_OBJECT) {
5876 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5877 } else {
5878 type = Type::get_const_basic_type(bt);
5879 }
5880
5881 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5882 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5883 }
5884 // Build the load.
5885 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5886 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5887 // If reference is volatile, prevent following memory ops from
5888 // floating up past the volatile read. Also prevents commoning
5889 // another volatile read.
5890 if (is_vol) {
5891 // Memory barrier includes bogus read of value to force load BEFORE membar
5892 insert_mem_bar(Op_MemBarAcquire, loadedField);
5893 }
5894 return loadedField;
5895 }
5896
5897
5898 //------------------------------inline_aescrypt_Block-----------------------
5899 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5900 address stubAddr;
5901 const char *stubName;
5902 assert(UseAES, "need AES instruction support");
5903
5904 switch(id) {
5905 case vmIntrinsics::_aescrypt_encryptBlock:
5906 stubAddr = StubRoutines::aescrypt_encryptBlock();
5907 stubName = "aescrypt_encryptBlock";
5908 break;
5909 case vmIntrinsics::_aescrypt_decryptBlock:
5910 stubAddr = StubRoutines::aescrypt_decryptBlock();
5911 stubName = "aescrypt_decryptBlock";
5912 break;
5913 }
5914 if (stubAddr == NULL) return false;
5915
5916 Node* aescrypt_object = argument(0);
5917 Node* src = argument(1);
5918 Node* src_offset = argument(2);
5919 Node* dest = argument(3);
5920 Node* dest_offset = argument(4);
5921
5922 // (1) src and dest are arrays.
5923 const Type* src_type = src->Value(&_gvn);
5924 const Type* dest_type = dest->Value(&_gvn);
5925 const TypeAryPtr* top_src = src_type->isa_aryptr();
5926 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5927 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5928
5929 // for the quick and dirty code we will skip all the checks.
5930 // we are just trying to get the call to be generated.
5931 Node* src_start = src;
5932 Node* dest_start = dest;
5933 if (src_offset != NULL || dest_offset != NULL) {
5934 assert(src_offset != NULL && dest_offset != NULL, "");
5935 src_start = array_element_address(src, src_offset, T_BYTE);
5936 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5937 }
5938
5939 // now need to get the start of its expanded key array
5940 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5941 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5942 if (k_start == NULL) return false;
5943
5944 if (Matcher::pass_original_key_for_aes()) {
5945 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5946 // compatibility issues between Java key expansion and SPARC crypto instructions
5947 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5948 if (original_k_start == NULL) return false;
5949
5950 // Call the stub.
5951 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5952 stubAddr, stubName, TypePtr::BOTTOM,
5953 src_start, dest_start, k_start, original_k_start);
5954 } else {
5955 // Call the stub.
5956 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5957 stubAddr, stubName, TypePtr::BOTTOM,
5958 src_start, dest_start, k_start);
5959 }
5960
5961 return true;
5962 }
5963
5964 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5965 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5966 address stubAddr;
5967 const char *stubName;
5968
5969 assert(UseAES, "need AES instruction support");
5970
5971 switch(id) {
5972 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5973 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5974 stubName = "cipherBlockChaining_encryptAESCrypt";
5975 break;
5976 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5977 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5978 stubName = "cipherBlockChaining_decryptAESCrypt";
5979 break;
5980 }
5981 if (stubAddr == NULL) return false;
5982
5983 Node* cipherBlockChaining_object = argument(0);
5984 Node* src = argument(1);
5985 Node* src_offset = argument(2);
5986 Node* len = argument(3);
5987 Node* dest = argument(4);
5988 Node* dest_offset = argument(5);
5989
5990 // (1) src and dest are arrays.
5991 const Type* src_type = src->Value(&_gvn);
5992 const Type* dest_type = dest->Value(&_gvn);
5993 const TypeAryPtr* top_src = src_type->isa_aryptr();
5994 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5995 assert (top_src != NULL && top_src->klass() != NULL
5996 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5997
5998 // checks are the responsibility of the caller
5999 Node* src_start = src;
6000 Node* dest_start = dest;
6001 if (src_offset != NULL || dest_offset != NULL) {
6002 assert(src_offset != NULL && dest_offset != NULL, "");
6003 src_start = array_element_address(src, src_offset, T_BYTE);
6004 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6005 }
6006
6007 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6008 // (because of the predicated logic executed earlier).
6009 // so we cast it here safely.
6010 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6011
6012 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6013 if (embeddedCipherObj == NULL) return false;
6014
6015 // cast it to what we know it will be at runtime
6016 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6017 assert(tinst != NULL, "CBC obj is null");
6018 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6019 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6020 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6021
6022 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6023 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6024 const TypeOopPtr* xtype = aklass->as_instance_type();
6025 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6026 aescrypt_object = _gvn.transform(aescrypt_object);
6027
6028 // we need to get the start of the aescrypt_object's expanded key array
6029 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6030 if (k_start == NULL) return false;
6031
6032 // similarly, get the start address of the r vector
6033 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6034 if (objRvec == NULL) return false;
6035 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6036
6037 Node* cbcCrypt;
6038 if (Matcher::pass_original_key_for_aes()) {
6039 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6040 // compatibility issues between Java key expansion and SPARC crypto instructions
6041 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6042 if (original_k_start == NULL) return false;
6043
6044 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6045 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6046 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6047 stubAddr, stubName, TypePtr::BOTTOM,
6048 src_start, dest_start, k_start, r_start, len, original_k_start);
6049 } else {
6050 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6051 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6052 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6053 stubAddr, stubName, TypePtr::BOTTOM,
6054 src_start, dest_start, k_start, r_start, len);
6055 }
6056
6057 // return cipher length (int)
6058 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6059 set_result(retvalue);
6060 return true;
6061 }
6062
6063 //------------------------------get_key_start_from_aescrypt_object-----------------------
6064 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6065 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6066 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6067 if (objAESCryptKey == NULL) return (Node *) NULL;
6068
6069 // now have the array, need to get the start address of the K array
6070 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6071 return k_start;
6072 }
6073
6074 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6075 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6076 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6077 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6078 if (objAESCryptKey == NULL) return (Node *) NULL;
6079
6080 // now have the array, need to get the start address of the lastKey array
6081 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6082 return original_k_start;
6083 }
6084
6085 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6086 // Return node representing slow path of predicate check.
6087 // the pseudo code we want to emulate with this predicate is:
6088 // for encryption:
6089 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6090 // for decryption:
6091 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6092 // note cipher==plain is more conservative than the original java code but that's OK
6093 //
6094 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6095 // The receiver was checked for NULL already.
6096 Node* objCBC = argument(0);
6097
6098 // Load embeddedCipher field of CipherBlockChaining object.
6099 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6100
6101 // get AESCrypt klass for instanceOf check
6102 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6103 // will have same classloader as CipherBlockChaining object
6104 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6105 assert(tinst != NULL, "CBCobj is null");
6106 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6107
6108 // we want to do an instanceof comparison against the AESCrypt class
6109 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6110 if (!klass_AESCrypt->is_loaded()) {
6111 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6112 Node* ctrl = control();
6113 set_control(top()); // no regular fast path
6114 return ctrl;
6115 }
6116 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6117
6118 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6119 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6120 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6121
6122 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6123
6124 // for encryption, we are done
6125 if (!decrypting)
6126 return instof_false; // even if it is NULL
6127
6128 // for decryption, we need to add a further check to avoid
6129 // taking the intrinsic path when cipher and plain are the same
6130 // see the original java code for why.
6131 RegionNode* region = new RegionNode(3);
6132 region->init_req(1, instof_false);
6133 Node* src = argument(1);
6134 Node* dest = argument(4);
6135 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6136 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6137 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6138 region->init_req(2, src_dest_conjoint);
6139
6140 record_for_igvn(region);
6141 return _gvn.transform(region);
6142 }
6143
6144 //------------------------------inline_ghash_processBlocks
6145 bool LibraryCallKit::inline_ghash_processBlocks() {
6146 address stubAddr;
6147 const char *stubName;
6148 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6149
6150 stubAddr = StubRoutines::ghash_processBlocks();
6151 stubName = "ghash_processBlocks";
6152
6153 Node* data = argument(0);
6154 Node* offset = argument(1);
6155 Node* len = argument(2);
6156 Node* state = argument(3);
6157 Node* subkeyH = argument(4);
6158
6159 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6160 assert(state_start, "state is NULL");
6161 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6162 assert(subkeyH_start, "subkeyH is NULL");
6163 Node* data_start = array_element_address(data, offset, T_BYTE);
6164 assert(data_start, "data is NULL");
6165
6166 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6167 OptoRuntime::ghash_processBlocks_Type(),
6168 stubAddr, stubName, TypePtr::BOTTOM,
6169 state_start, subkeyH_start, data_start, len);
6170 return true;
6171 }
6172
6173 //------------------------------inline_sha_implCompress-----------------------
6174 //
6175 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6176 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6177 //
6178 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6179 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6180 //
6181 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6182 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6183 //
6184 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6185 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6186
6187 Node* sha_obj = argument(0);
6188 Node* src = argument(1); // type oop
6189 Node* ofs = argument(2); // type int
6190
6191 const Type* src_type = src->Value(&_gvn);
6192 const TypeAryPtr* top_src = src_type->isa_aryptr();
6193 if (top_src == NULL || top_src->klass() == NULL) {
6194 // failed array check
6195 return false;
6196 }
6197 // Figure out the size and type of the elements we will be copying.
6198 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6199 if (src_elem != T_BYTE) {
6200 return false;
6201 }
6202 // 'src_start' points to src array + offset
6203 Node* src_start = array_element_address(src, ofs, src_elem);
6204 Node* state = NULL;
6205 address stubAddr;
6206 const char *stubName;
6207
6208 switch(id) {
6209 case vmIntrinsics::_sha_implCompress:
6210 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6211 state = get_state_from_sha_object(sha_obj);
6212 stubAddr = StubRoutines::sha1_implCompress();
6213 stubName = "sha1_implCompress";
6214 break;
6215 case vmIntrinsics::_sha2_implCompress:
6216 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6217 state = get_state_from_sha_object(sha_obj);
6218 stubAddr = StubRoutines::sha256_implCompress();
6219 stubName = "sha256_implCompress";
6220 break;
6221 case vmIntrinsics::_sha5_implCompress:
6222 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6223 state = get_state_from_sha5_object(sha_obj);
6224 stubAddr = StubRoutines::sha512_implCompress();
6225 stubName = "sha512_implCompress";
6226 break;
6227 default:
6228 fatal_unexpected_iid(id);
6229 return false;
6230 }
6231 if (state == NULL) return false;
6232
6233 // Call the stub.
6234 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6235 stubAddr, stubName, TypePtr::BOTTOM,
6236 src_start, state);
6237
6238 return true;
6239 }
6240
6241 //------------------------------inline_digestBase_implCompressMB-----------------------
6242 //
6243 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6244 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6245 //
6246 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6247 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6248 "need SHA1/SHA256/SHA512 instruction support");
6249 assert((uint)predicate < 3, "sanity");
6250 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6251
6252 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6253 Node* src = argument(1); // byte[] array
6254 Node* ofs = argument(2); // type int
6255 Node* limit = argument(3); // type int
6256
6257 const Type* src_type = src->Value(&_gvn);
6258 const TypeAryPtr* top_src = src_type->isa_aryptr();
6259 if (top_src == NULL || top_src->klass() == NULL) {
6260 // failed array check
6261 return false;
6262 }
6263 // Figure out the size and type of the elements we will be copying.
6264 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6265 if (src_elem != T_BYTE) {
6266 return false;
6267 }
6268 // 'src_start' points to src array + offset
6269 Node* src_start = array_element_address(src, ofs, src_elem);
6270
6271 const char* klass_SHA_name = NULL;
6272 const char* stub_name = NULL;
6273 address stub_addr = NULL;
6274 bool long_state = false;
6275
6276 switch (predicate) {
6277 case 0:
6278 if (UseSHA1Intrinsics) {
6279 klass_SHA_name = "sun/security/provider/SHA";
6280 stub_name = "sha1_implCompressMB";
6281 stub_addr = StubRoutines::sha1_implCompressMB();
6282 }
6283 break;
6284 case 1:
6285 if (UseSHA256Intrinsics) {
6286 klass_SHA_name = "sun/security/provider/SHA2";
6287 stub_name = "sha256_implCompressMB";
6288 stub_addr = StubRoutines::sha256_implCompressMB();
6289 }
6290 break;
6291 case 2:
6292 if (UseSHA512Intrinsics) {
6293 klass_SHA_name = "sun/security/provider/SHA5";
6294 stub_name = "sha512_implCompressMB";
6295 stub_addr = StubRoutines::sha512_implCompressMB();
6296 long_state = true;
6297 }
6298 break;
6299 default:
6300 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6301 }
6302 if (klass_SHA_name != NULL) {
6303 // get DigestBase klass to lookup for SHA klass
6304 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6305 assert(tinst != NULL, "digestBase_obj is not instance???");
6306 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6307
6308 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6309 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6310 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6311 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6312 }
6313 return false;
6314 }
6315 //------------------------------inline_sha_implCompressMB-----------------------
6316 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6317 bool long_state, address stubAddr, const char *stubName,
6318 Node* src_start, Node* ofs, Node* limit) {
6319 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6320 const TypeOopPtr* xtype = aklass->as_instance_type();
6321 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6322 sha_obj = _gvn.transform(sha_obj);
6323
6324 Node* state;
6325 if (long_state) {
6326 state = get_state_from_sha5_object(sha_obj);
6327 } else {
6328 state = get_state_from_sha_object(sha_obj);
6329 }
6330 if (state == NULL) return false;
6331
6332 // Call the stub.
6333 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6334 OptoRuntime::digestBase_implCompressMB_Type(),
6335 stubAddr, stubName, TypePtr::BOTTOM,
6336 src_start, state, ofs, limit);
6337 // return ofs (int)
6338 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6339 set_result(result);
6340
6341 return true;
6342 }
6343
6344 //------------------------------get_state_from_sha_object-----------------------
6345 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6346 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6347 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6348 if (sha_state == NULL) return (Node *) NULL;
6349
6350 // now have the array, need to get the start address of the state array
6351 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6352 return state;
6353 }
6354
6355 //------------------------------get_state_from_sha5_object-----------------------
6356 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6357 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6358 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6359 if (sha_state == NULL) return (Node *) NULL;
6360
6361 // now have the array, need to get the start address of the state array
6362 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6363 return state;
6364 }
6365
6366 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6367 // Return node representing slow path of predicate check.
6368 // the pseudo code we want to emulate with this predicate is:
6369 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6370 //
6371 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6372 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6373 "need SHA1/SHA256/SHA512 instruction support");
6374 assert((uint)predicate < 3, "sanity");
6375
6376 // The receiver was checked for NULL already.
6377 Node* digestBaseObj = argument(0);
6378
6379 // get DigestBase klass for instanceOf check
6380 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6381 assert(tinst != NULL, "digestBaseObj is null");
6382 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6383
6384 const char* klass_SHA_name = NULL;
6385 switch (predicate) {
6386 case 0:
6387 if (UseSHA1Intrinsics) {
6388 // we want to do an instanceof comparison against the SHA class
6389 klass_SHA_name = "sun/security/provider/SHA";
6390 }
6391 break;
6392 case 1:
6393 if (UseSHA256Intrinsics) {
6394 // we want to do an instanceof comparison against the SHA2 class
6395 klass_SHA_name = "sun/security/provider/SHA2";
6396 }
6397 break;
6398 case 2:
6399 if (UseSHA512Intrinsics) {
6400 // we want to do an instanceof comparison against the SHA5 class
6401 klass_SHA_name = "sun/security/provider/SHA5";
6402 }
6403 break;
6404 default:
6405 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6406 }
6407
6408 ciKlass* klass_SHA = NULL;
6409 if (klass_SHA_name != NULL) {
6410 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6411 }
6412 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6413 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6414 Node* ctrl = control();
6415 set_control(top()); // no intrinsic path
6416 return ctrl;
6417 }
6418 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6419
6420 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6421 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6422 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6423 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6424
6425 return instof_false; // even if it is NULL
6426 }
6427
6428 bool LibraryCallKit::inline_profileBoolean() {
6429 Node* counts = argument(1);
6430 const TypeAryPtr* ary = NULL;
6431 ciArray* aobj = NULL;
6432 if (counts->is_Con()
6433 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6434 && (aobj = ary->const_oop()->as_array()) != NULL
6435 && (aobj->length() == 2)) {
6436 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6437 jint false_cnt = aobj->element_value(0).as_int();
6438 jint true_cnt = aobj->element_value(1).as_int();
6439
6440 if (C->log() != NULL) {
6441 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6442 false_cnt, true_cnt);
6443 }
6444
6445 if (false_cnt + true_cnt == 0) {
6446 // According to profile, never executed.
6447 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6448 Deoptimization::Action_reinterpret);
6449 return true;
6450 }
6451
6452 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6453 // is a number of each value occurrences.
6454 Node* result = argument(0);
6455 if (false_cnt == 0 || true_cnt == 0) {
6456 // According to profile, one value has been never seen.
6457 int expected_val = (false_cnt == 0) ? 1 : 0;
6458
6459 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6460 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6461
6462 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6463 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6464 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6465
6466 { // Slow path: uncommon trap for never seen value and then reexecute
6467 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6468 // the value has been seen at least once.
6469 PreserveJVMState pjvms(this);
6470 PreserveReexecuteState preexecs(this);
6471 jvms()->set_should_reexecute(true);
6472
6473 set_control(slow_path);
6474 set_i_o(i_o());
6475
6476 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6477 Deoptimization::Action_reinterpret);
6478 }
6479 // The guard for never seen value enables sharpening of the result and
6480 // returning a constant. It allows to eliminate branches on the same value
6481 // later on.
6482 set_control(fast_path);
6483 result = intcon(expected_val);
6484 }
6485 // Stop profiling.
6486 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6487 // By replacing method body with profile data (represented as ProfileBooleanNode
6488 // on IR level) we effectively disable profiling.
6489 // It enables full speed execution once optimized code is generated.
6490 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6491 C->record_for_igvn(profile);
6492 set_result(profile);
6493 return true;
6494 } else {
6495 // Continue profiling.
6496 // Profile data isn't available at the moment. So, execute method's bytecode version.
6497 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6498 // is compiled and counters aren't available since corresponding MethodHandle
6499 // isn't a compile-time constant.
6500 return false;
6501 }
6502 }
6503
6504 bool LibraryCallKit::inline_isCompileConstant() {
6505 Node* n = argument(0);
6506 set_result(n->is_Con() ? intcon(1) : intcon(0));
6507 return true;
6508 }