1 /*
2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callGenerator.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/countbitsnode.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/mathexactnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/mulnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/parse.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "prims/nativeLookup.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "trace/traceMacros.hpp"
51
52 class LibraryIntrinsic : public InlineCallGenerator {
53 // Extend the set of intrinsics known to the runtime:
54 public:
55 private:
56 bool _is_virtual;
57 bool _does_virtual_dispatch;
58 int8_t _predicates_count; // Intrinsic is predicated by several conditions
59 int8_t _last_predicate; // Last generated predicate
60 vmIntrinsics::ID _intrinsic_id;
61
62 public:
63 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
64 : InlineCallGenerator(m),
65 _is_virtual(is_virtual),
66 _does_virtual_dispatch(does_virtual_dispatch),
67 _predicates_count((int8_t)predicates_count),
68 _last_predicate((int8_t)-1),
69 _intrinsic_id(id)
70 {
71 }
72 virtual bool is_intrinsic() const { return true; }
73 virtual bool is_virtual() const { return _is_virtual; }
74 virtual bool is_predicated() const { return _predicates_count > 0; }
75 virtual int predicates_count() const { return _predicates_count; }
76 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
77 virtual JVMState* generate(JVMState* jvms);
78 virtual Node* generate_predicate(JVMState* jvms, int predicate);
79 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
80 };
81
82
83 // Local helper class for LibraryIntrinsic:
84 class LibraryCallKit : public GraphKit {
85 private:
86 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
87 Node* _result; // the result node, if any
88 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
89
90 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
91
92 public:
93 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
94 : GraphKit(jvms),
95 _intrinsic(intrinsic),
96 _result(NULL)
97 {
98 // Check if this is a root compile. In that case we don't have a caller.
99 if (!jvms->has_method()) {
100 _reexecute_sp = sp();
101 } else {
102 // Find out how many arguments the interpreter needs when deoptimizing
103 // and save the stack pointer value so it can used by uncommon_trap.
104 // We find the argument count by looking at the declared signature.
105 bool ignored_will_link;
106 ciSignature* declared_signature = NULL;
107 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
108 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
109 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
110 }
111 }
112
113 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
114
115 ciMethod* caller() const { return jvms()->method(); }
116 int bci() const { return jvms()->bci(); }
117 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
118 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
119 ciMethod* callee() const { return _intrinsic->method(); }
120
121 bool try_to_inline(int predicate);
122 Node* try_to_predicate(int predicate);
123
124 void push_result() {
125 // Push the result onto the stack.
126 if (!stopped() && result() != NULL) {
127 BasicType bt = result()->bottom_type()->basic_type();
128 push_node(bt, result());
129 }
130 }
131
132 private:
133 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
134 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
135 }
136
137 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
138 void set_result(RegionNode* region, PhiNode* value);
139 Node* result() { return _result; }
140
141 virtual int reexecute_sp() { return _reexecute_sp; }
142
143 // Helper functions to inline natives
144 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
145 Node* generate_slow_guard(Node* test, RegionNode* region);
146 Node* generate_fair_guard(Node* test, RegionNode* region);
147 Node* generate_negative_guard(Node* index, RegionNode* region,
148 // resulting CastII of index:
149 Node* *pos_index = NULL);
150 Node* generate_limit_guard(Node* offset, Node* subseq_length,
151 Node* array_length,
152 RegionNode* region);
153 Node* generate_current_thread(Node* &tls_output);
154 Node* load_mirror_from_klass(Node* klass);
155 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
156 RegionNode* region, int null_path,
157 int offset);
158 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
159 RegionNode* region, int null_path) {
160 int offset = java_lang_Class::klass_offset_in_bytes();
161 return load_klass_from_mirror_common(mirror, never_see_null,
162 region, null_path,
163 offset);
164 }
165 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
166 RegionNode* region, int null_path) {
167 int offset = java_lang_Class::array_klass_offset_in_bytes();
168 return load_klass_from_mirror_common(mirror, never_see_null,
169 region, null_path,
170 offset);
171 }
172 Node* generate_access_flags_guard(Node* kls,
173 int modifier_mask, int modifier_bits,
174 RegionNode* region);
175 Node* generate_interface_guard(Node* kls, RegionNode* region);
176 Node* generate_array_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, false, false);
178 }
179 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, false, true);
181 }
182 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
183 return generate_array_guard_common(kls, region, true, false);
184 }
185 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, true, true);
187 }
188 Node* generate_array_guard_common(Node* kls, RegionNode* region,
189 bool obj_array, bool not_array);
190 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
191 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
192 bool is_virtual = false, bool is_static = false);
193 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
194 return generate_method_call(method_id, false, true);
195 }
196 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
197 return generate_method_call(method_id, true, false);
198 }
199 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
200
201 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
202 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
203 bool inline_string_compareTo();
204 bool inline_string_indexOf();
205 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
206 bool inline_string_equals();
207 Node* round_double_node(Node* n);
208 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
209 bool inline_math_native(vmIntrinsics::ID id);
210 bool inline_trig(vmIntrinsics::ID id);
211 bool inline_math(vmIntrinsics::ID id);
212 template <typename OverflowOp>
213 bool inline_math_overflow(Node* arg1, Node* arg2);
214 void inline_math_mathExact(Node* math, Node* test);
215 bool inline_math_addExactI(bool is_increment);
216 bool inline_math_addExactL(bool is_increment);
217 bool inline_math_multiplyExactI();
218 bool inline_math_multiplyExactL();
219 bool inline_math_negateExactI();
220 bool inline_math_negateExactL();
221 bool inline_math_subtractExactI(bool is_decrement);
222 bool inline_math_subtractExactL(bool is_decrement);
223 bool inline_exp();
224 bool inline_pow();
225 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
226 bool inline_min_max(vmIntrinsics::ID id);
227 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
228 // This returns Type::AnyPtr, RawPtr, or OopPtr.
229 int classify_unsafe_addr(Node* &base, Node* &offset);
230 Node* make_unsafe_address(Node* base, Node* offset);
231 // Helper for inline_unsafe_access.
232 // Generates the guards that check whether the result of
233 // Unsafe.getObject should be recorded in an SATB log buffer.
234 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
235 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
236 static bool klass_needs_init_guard(Node* kls);
237 bool inline_unsafe_allocate();
238 bool inline_unsafe_copyMemory();
239 bool inline_native_currentThread();
240 #ifdef TRACE_HAVE_INTRINSICS
241 bool inline_native_classID();
242 bool inline_native_threadID();
243 #endif
244 bool inline_native_time_funcs(address method, const char* funcName);
245 bool inline_native_isInterrupted();
246 bool inline_native_Class_query(vmIntrinsics::ID id);
247 bool inline_native_subtype_check();
248
249 bool inline_native_newArray();
250 bool inline_native_getLength();
251 bool inline_array_copyOf(bool is_copyOfRange);
252 bool inline_array_equals();
253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
254 bool inline_native_clone(bool is_virtual);
255 bool inline_native_Reflection_getCallerClass();
256 // Helper function for inlining native object hash method
257 bool inline_native_hashcode(bool is_virtual, bool is_static);
258 bool inline_native_getClass();
259
260 // Helper functions for inlining arraycopy
261 bool inline_arraycopy();
262 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
263 RegionNode* slow_region);
264 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
265 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
266 bool inline_unsafe_ordered_store(BasicType type);
267 bool inline_unsafe_fence(vmIntrinsics::ID id);
268 bool inline_fp_conversions(vmIntrinsics::ID id);
269 bool inline_number_methods(vmIntrinsics::ID id);
270 bool inline_reference_get();
271 bool inline_Class_cast();
272 bool inline_aescrypt_Block(vmIntrinsics::ID id);
273 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
274 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
275 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
276 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
277 bool inline_sha_implCompress(vmIntrinsics::ID id);
278 bool inline_digestBase_implCompressMB(int predicate);
279 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
280 bool long_state, address stubAddr, const char *stubName,
281 Node* src_start, Node* ofs, Node* limit);
282 Node* get_state_from_sha_object(Node *sha_object);
283 Node* get_state_from_sha5_object(Node *sha_object);
284 Node* inline_digestBase_implCompressMB_predicate(int predicate);
285 bool inline_encodeISOArray();
286 bool inline_updateCRC32();
287 bool inline_updateBytesCRC32();
288 bool inline_updateByteBufferCRC32();
289 bool inline_multiplyToLen();
290
291 bool inline_profileBoolean();
292 };
293
294
295 //---------------------------make_vm_intrinsic----------------------------
296 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
297 vmIntrinsics::ID id = m->intrinsic_id();
298 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
299
300 ccstr disable_intr = NULL;
301
302 if ((DisableIntrinsic[0] != '\0'
303 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
304 (method_has_option_value("DisableIntrinsic", disable_intr)
305 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
306 // disabled by a user request on the command line:
307 // example: -XX:DisableIntrinsic=_hashCode,_getClass
308 return NULL;
309 }
310
311 if (!m->is_loaded()) {
312 // do not attempt to inline unloaded methods
313 return NULL;
314 }
315
316 // Only a few intrinsics implement a virtual dispatch.
317 // They are expensive calls which are also frequently overridden.
318 if (is_virtual) {
319 switch (id) {
320 case vmIntrinsics::_hashCode:
321 case vmIntrinsics::_clone:
322 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
323 break;
324 default:
325 return NULL;
326 }
327 }
328
329 // -XX:-InlineNatives disables nearly all intrinsics:
330 if (!InlineNatives) {
331 switch (id) {
332 case vmIntrinsics::_indexOf:
333 case vmIntrinsics::_compareTo:
334 case vmIntrinsics::_equals:
335 case vmIntrinsics::_equalsC:
336 case vmIntrinsics::_getAndAddInt:
337 case vmIntrinsics::_getAndAddLong:
338 case vmIntrinsics::_getAndSetInt:
339 case vmIntrinsics::_getAndSetLong:
340 case vmIntrinsics::_getAndSetObject:
341 case vmIntrinsics::_loadFence:
342 case vmIntrinsics::_storeFence:
343 case vmIntrinsics::_fullFence:
344 break; // InlineNatives does not control String.compareTo
345 case vmIntrinsics::_Reference_get:
346 break; // InlineNatives does not control Reference.get
347 default:
348 return NULL;
349 }
350 }
351
352 int predicates = 0;
353 bool does_virtual_dispatch = false;
354
355 switch (id) {
356 case vmIntrinsics::_compareTo:
357 if (!SpecialStringCompareTo) return NULL;
358 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
359 break;
360 case vmIntrinsics::_indexOf:
361 if (!SpecialStringIndexOf) return NULL;
362 break;
363 case vmIntrinsics::_equals:
364 if (!SpecialStringEquals) return NULL;
365 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
366 break;
367 case vmIntrinsics::_equalsC:
368 if (!SpecialArraysEquals) return NULL;
369 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
370 break;
371 case vmIntrinsics::_arraycopy:
372 if (!InlineArrayCopy) return NULL;
373 break;
374 case vmIntrinsics::_copyMemory:
375 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
376 if (!InlineArrayCopy) return NULL;
377 break;
378 case vmIntrinsics::_hashCode:
379 if (!InlineObjectHash) return NULL;
380 does_virtual_dispatch = true;
381 break;
382 case vmIntrinsics::_clone:
383 does_virtual_dispatch = true;
384 case vmIntrinsics::_copyOf:
385 case vmIntrinsics::_copyOfRange:
386 if (!InlineObjectCopy) return NULL;
387 // These also use the arraycopy intrinsic mechanism:
388 if (!InlineArrayCopy) return NULL;
389 break;
390 case vmIntrinsics::_encodeISOArray:
391 if (!SpecialEncodeISOArray) return NULL;
392 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
393 break;
394 case vmIntrinsics::_checkIndex:
395 // We do not intrinsify this. The optimizer does fine with it.
396 return NULL;
397
398 case vmIntrinsics::_getCallerClass:
399 if (!InlineReflectionGetCallerClass) return NULL;
400 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
401 break;
402
403 case vmIntrinsics::_bitCount_i:
404 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
405 break;
406
407 case vmIntrinsics::_bitCount_l:
408 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
409 break;
410
411 case vmIntrinsics::_numberOfLeadingZeros_i:
412 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
413 break;
414
415 case vmIntrinsics::_numberOfLeadingZeros_l:
416 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
417 break;
418
419 case vmIntrinsics::_numberOfTrailingZeros_i:
420 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
421 break;
422
423 case vmIntrinsics::_numberOfTrailingZeros_l:
424 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
425 break;
426
427 case vmIntrinsics::_reverseBytes_c:
428 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
429 break;
430 case vmIntrinsics::_reverseBytes_s:
431 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
432 break;
433 case vmIntrinsics::_reverseBytes_i:
434 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
435 break;
436 case vmIntrinsics::_reverseBytes_l:
437 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
438 break;
439
440 case vmIntrinsics::_Reference_get:
441 // Use the intrinsic version of Reference.get() so that the value in
442 // the referent field can be registered by the G1 pre-barrier code.
443 // Also add memory barrier to prevent commoning reads from this field
444 // across safepoint since GC can change it value.
445 break;
446
447 case vmIntrinsics::_compareAndSwapObject:
448 #ifdef _LP64
449 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
450 #endif
451 break;
452
453 case vmIntrinsics::_compareAndSwapLong:
454 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
455 break;
456
457 case vmIntrinsics::_getAndAddInt:
458 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
459 break;
460
461 case vmIntrinsics::_getAndAddLong:
462 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
463 break;
464
465 case vmIntrinsics::_getAndSetInt:
466 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
467 break;
468
469 case vmIntrinsics::_getAndSetLong:
470 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
471 break;
472
473 case vmIntrinsics::_getAndSetObject:
474 #ifdef _LP64
475 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
476 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
477 break;
478 #else
479 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
480 break;
481 #endif
482
483 case vmIntrinsics::_aescrypt_encryptBlock:
484 case vmIntrinsics::_aescrypt_decryptBlock:
485 if (!UseAESIntrinsics) return NULL;
486 break;
487
488 case vmIntrinsics::_multiplyToLen:
489 if (!UseMultiplyToLenIntrinsic) return NULL;
490 break;
491
492 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
493 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
494 if (!UseAESIntrinsics) return NULL;
495 // these two require the predicated logic
496 predicates = 1;
497 break;
498
499 case vmIntrinsics::_sha_implCompress:
500 if (!UseSHA1Intrinsics) return NULL;
501 break;
502
503 case vmIntrinsics::_sha2_implCompress:
504 if (!UseSHA256Intrinsics) return NULL;
505 break;
506
507 case vmIntrinsics::_sha5_implCompress:
508 if (!UseSHA512Intrinsics) return NULL;
509 break;
510
511 case vmIntrinsics::_digestBase_implCompressMB:
512 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
513 predicates = 3;
514 break;
515
516 case vmIntrinsics::_updateCRC32:
517 case vmIntrinsics::_updateBytesCRC32:
518 case vmIntrinsics::_updateByteBufferCRC32:
519 if (!UseCRC32Intrinsics) return NULL;
520 break;
521
522 case vmIntrinsics::_incrementExactI:
523 case vmIntrinsics::_addExactI:
524 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
525 break;
526 case vmIntrinsics::_incrementExactL:
527 case vmIntrinsics::_addExactL:
528 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
529 break;
530 case vmIntrinsics::_decrementExactI:
531 case vmIntrinsics::_subtractExactI:
532 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
533 break;
534 case vmIntrinsics::_decrementExactL:
535 case vmIntrinsics::_subtractExactL:
536 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
537 break;
538 case vmIntrinsics::_negateExactI:
539 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
540 break;
541 case vmIntrinsics::_negateExactL:
542 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
543 break;
544 case vmIntrinsics::_multiplyExactI:
545 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
546 break;
547 case vmIntrinsics::_multiplyExactL:
548 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
549 break;
550
551 default:
552 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
553 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
554 break;
555 }
556
557 // -XX:-InlineClassNatives disables natives from the Class class.
558 // The flag applies to all reflective calls, notably Array.newArray
559 // (visible to Java programmers as Array.newInstance).
560 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
561 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
562 if (!InlineClassNatives) return NULL;
563 }
564
565 // -XX:-InlineThreadNatives disables natives from the Thread class.
566 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
567 if (!InlineThreadNatives) return NULL;
568 }
569
570 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
571 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
572 m->holder()->name() == ciSymbol::java_lang_Float() ||
573 m->holder()->name() == ciSymbol::java_lang_Double()) {
574 if (!InlineMathNatives) return NULL;
575 }
576
577 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
578 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
579 if (!InlineUnsafeOps) return NULL;
580 }
581
582 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
583 }
584
585 //----------------------register_library_intrinsics-----------------------
586 // Initialize this file's data structures, for each Compile instance.
587 void Compile::register_library_intrinsics() {
588 // Nothing to do here.
589 }
590
591 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
592 LibraryCallKit kit(jvms, this);
593 Compile* C = kit.C;
594 int nodes = C->unique();
595 #ifndef PRODUCT
596 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
597 char buf[1000];
598 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
599 tty->print_cr("Intrinsic %s", str);
600 }
601 #endif
602 ciMethod* callee = kit.callee();
603 const int bci = kit.bci();
604
605 // Try to inline the intrinsic.
606 if (kit.try_to_inline(_last_predicate)) {
607 if (C->print_intrinsics() || C->print_inlining()) {
608 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
609 }
610 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
611 if (C->log()) {
612 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
613 vmIntrinsics::name_at(intrinsic_id()),
614 (is_virtual() ? " virtual='1'" : ""),
615 C->unique() - nodes);
616 }
617 // Push the result from the inlined method onto the stack.
618 kit.push_result();
619 C->print_inlining_update(this);
620 return kit.transfer_exceptions_into_jvms();
621 }
622
623 // The intrinsic bailed out
624 if (C->print_intrinsics() || C->print_inlining()) {
625 if (jvms->has_method()) {
626 // Not a root compile.
627 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
628 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
629 } else {
630 // Root compile
631 tty->print("Did not generate intrinsic %s%s at bci:%d in",
632 vmIntrinsics::name_at(intrinsic_id()),
633 (is_virtual() ? " (virtual)" : ""), bci);
634 }
635 }
636 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
637 C->print_inlining_update(this);
638 return NULL;
639 }
640
641 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
642 LibraryCallKit kit(jvms, this);
643 Compile* C = kit.C;
644 int nodes = C->unique();
645 _last_predicate = predicate;
646 #ifndef PRODUCT
647 assert(is_predicated() && predicate < predicates_count(), "sanity");
648 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
649 char buf[1000];
650 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
651 tty->print_cr("Predicate for intrinsic %s", str);
652 }
653 #endif
654 ciMethod* callee = kit.callee();
655 const int bci = kit.bci();
656
657 Node* slow_ctl = kit.try_to_predicate(predicate);
658 if (!kit.failing()) {
659 if (C->print_intrinsics() || C->print_inlining()) {
660 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
661 }
662 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
663 if (C->log()) {
664 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
665 vmIntrinsics::name_at(intrinsic_id()),
666 (is_virtual() ? " virtual='1'" : ""),
667 C->unique() - nodes);
668 }
669 return slow_ctl; // Could be NULL if the check folds.
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 = "failed to generate predicate for intrinsic";
677 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
678 } else {
679 // Root compile
680 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
681 vmIntrinsics::name_at(intrinsic_id()),
682 (is_virtual() ? " (virtual)" : ""), bci);
683 }
684 }
685 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
686 return NULL;
687 }
688
689 bool LibraryCallKit::try_to_inline(int predicate) {
690 // Handle symbolic names for otherwise undistinguished boolean switches:
691 const bool is_store = true;
692 const bool is_native_ptr = true;
693 const bool is_static = true;
694 const bool is_volatile = true;
695
696 if (!jvms()->has_method()) {
697 // Root JVMState has a null method.
698 assert(map()->memory()->Opcode() == Op_Parm, "");
699 // Insert the memory aliasing node
700 set_all_memory(reset_memory());
701 }
702 assert(merged_memory(), "");
703
704
705 switch (intrinsic_id()) {
706 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
707 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
708 case vmIntrinsics::_getClass: return inline_native_getClass();
709
710 case vmIntrinsics::_dsin:
711 case vmIntrinsics::_dcos:
712 case vmIntrinsics::_dtan:
713 case vmIntrinsics::_dabs:
714 case vmIntrinsics::_datan2:
715 case vmIntrinsics::_dsqrt:
716 case vmIntrinsics::_dexp:
717 case vmIntrinsics::_dlog:
718 case vmIntrinsics::_dlog10:
719 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
720
721 case vmIntrinsics::_min:
722 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
723
724 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
725 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
726 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
727 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
728 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
729 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
730 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
731 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
732 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
733 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
734 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
735 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
736
737 case vmIntrinsics::_arraycopy: return inline_arraycopy();
738
739 case vmIntrinsics::_compareTo: return inline_string_compareTo();
740 case vmIntrinsics::_indexOf: return inline_string_indexOf();
741 case vmIntrinsics::_equals: return inline_string_equals();
742
743 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
744 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
745 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
746 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
747 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
748 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
749 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
750 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
751 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
752
753 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
754 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
755 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
756 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
757 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
758 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
759 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
760 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
761 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
762
763 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
764 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
765 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
766 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
767 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
768 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
769 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
770 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
771
772 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
773 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
774 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
775 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
776 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
777 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
778 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
779 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
780
781 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
782 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
783 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
784 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
785 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
786 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
787 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
788 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
789 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
790
791 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
792 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
793 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
794 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
795 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
796 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
797 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
798 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
799 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
800
801 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
802 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
803 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
804
805 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
806 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
807 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
808
809 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
810 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
811 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
812 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
813 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
814
815 case vmIntrinsics::_loadFence:
816 case vmIntrinsics::_storeFence:
817 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
818
819 case vmIntrinsics::_currentThread: return inline_native_currentThread();
820 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
821
822 #ifdef TRACE_HAVE_INTRINSICS
823 case vmIntrinsics::_classID: return inline_native_classID();
824 case vmIntrinsics::_threadID: return inline_native_threadID();
825 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
826 #endif
827 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
828 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
829 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
830 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
831 case vmIntrinsics::_newArray: return inline_native_newArray();
832 case vmIntrinsics::_getLength: return inline_native_getLength();
833 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
834 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
835 case vmIntrinsics::_equalsC: return inline_array_equals();
836 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
837
838 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
839
840 case vmIntrinsics::_isInstance:
841 case vmIntrinsics::_getModifiers:
842 case vmIntrinsics::_isInterface:
843 case vmIntrinsics::_isArray:
844 case vmIntrinsics::_isPrimitive:
845 case vmIntrinsics::_getSuperclass:
846 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
847
848 case vmIntrinsics::_floatToRawIntBits:
849 case vmIntrinsics::_floatToIntBits:
850 case vmIntrinsics::_intBitsToFloat:
851 case vmIntrinsics::_doubleToRawLongBits:
852 case vmIntrinsics::_doubleToLongBits:
853 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
854
855 case vmIntrinsics::_numberOfLeadingZeros_i:
856 case vmIntrinsics::_numberOfLeadingZeros_l:
857 case vmIntrinsics::_numberOfTrailingZeros_i:
858 case vmIntrinsics::_numberOfTrailingZeros_l:
859 case vmIntrinsics::_bitCount_i:
860 case vmIntrinsics::_bitCount_l:
861 case vmIntrinsics::_reverseBytes_i:
862 case vmIntrinsics::_reverseBytes_l:
863 case vmIntrinsics::_reverseBytes_s:
864 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
865
866 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
867
868 case vmIntrinsics::_Reference_get: return inline_reference_get();
869
870 case vmIntrinsics::_Class_cast: return inline_Class_cast();
871
872 case vmIntrinsics::_aescrypt_encryptBlock:
873 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
874
875 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
876 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
877 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
878
879 case vmIntrinsics::_sha_implCompress:
880 case vmIntrinsics::_sha2_implCompress:
881 case vmIntrinsics::_sha5_implCompress:
882 return inline_sha_implCompress(intrinsic_id());
883
884 case vmIntrinsics::_digestBase_implCompressMB:
885 return inline_digestBase_implCompressMB(predicate);
886
887 case vmIntrinsics::_multiplyToLen:
888 return inline_multiplyToLen();
889
890 case vmIntrinsics::_encodeISOArray:
891 return inline_encodeISOArray();
892
893 case vmIntrinsics::_updateCRC32:
894 return inline_updateCRC32();
895 case vmIntrinsics::_updateBytesCRC32:
896 return inline_updateBytesCRC32();
897 case vmIntrinsics::_updateByteBufferCRC32:
898 return inline_updateByteBufferCRC32();
899
900 case vmIntrinsics::_profileBoolean:
901 return inline_profileBoolean();
902
903 default:
904 // If you get here, it may be that someone has added a new intrinsic
905 // to the list in vmSymbols.hpp without implementing it here.
906 #ifndef PRODUCT
907 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
908 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
909 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
910 }
911 #endif
912 return false;
913 }
914 }
915
916 Node* LibraryCallKit::try_to_predicate(int predicate) {
917 if (!jvms()->has_method()) {
918 // Root JVMState has a null method.
919 assert(map()->memory()->Opcode() == Op_Parm, "");
920 // Insert the memory aliasing node
921 set_all_memory(reset_memory());
922 }
923 assert(merged_memory(), "");
924
925 switch (intrinsic_id()) {
926 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
927 return inline_cipherBlockChaining_AESCrypt_predicate(false);
928 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
929 return inline_cipherBlockChaining_AESCrypt_predicate(true);
930 case vmIntrinsics::_digestBase_implCompressMB:
931 return inline_digestBase_implCompressMB_predicate(predicate);
932
933 default:
934 // If you get here, it may be that someone has added a new intrinsic
935 // to the list in vmSymbols.hpp without implementing it here.
936 #ifndef PRODUCT
937 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
938 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
939 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
940 }
941 #endif
942 Node* slow_ctl = control();
943 set_control(top()); // No fast path instrinsic
944 return slow_ctl;
945 }
946 }
947
948 //------------------------------set_result-------------------------------
949 // Helper function for finishing intrinsics.
950 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
951 record_for_igvn(region);
952 set_control(_gvn.transform(region));
953 set_result( _gvn.transform(value));
954 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
955 }
956
957 //------------------------------generate_guard---------------------------
958 // Helper function for generating guarded fast-slow graph structures.
959 // The given 'test', if true, guards a slow path. If the test fails
960 // then a fast path can be taken. (We generally hope it fails.)
961 // In all cases, GraphKit::control() is updated to the fast path.
962 // The returned value represents the control for the slow path.
963 // The return value is never 'top'; it is either a valid control
964 // or NULL if it is obvious that the slow path can never be taken.
965 // Also, if region and the slow control are not NULL, the slow edge
966 // is appended to the region.
967 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
968 if (stopped()) {
969 // Already short circuited.
970 return NULL;
971 }
972
973 // Build an if node and its projections.
974 // If test is true we take the slow path, which we assume is uncommon.
975 if (_gvn.type(test) == TypeInt::ZERO) {
976 // The slow branch is never taken. No need to build this guard.
977 return NULL;
978 }
979
980 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
981
982 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
983 if (if_slow == top()) {
984 // The slow branch is never taken. No need to build this guard.
985 return NULL;
986 }
987
988 if (region != NULL)
989 region->add_req(if_slow);
990
991 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
992 set_control(if_fast);
993
994 return if_slow;
995 }
996
997 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
998 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
999 }
1000 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1001 return generate_guard(test, region, PROB_FAIR);
1002 }
1003
1004 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1005 Node* *pos_index) {
1006 if (stopped())
1007 return NULL; // already stopped
1008 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1009 return NULL; // index is already adequately typed
1010 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1011 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1012 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1013 if (is_neg != NULL && pos_index != NULL) {
1014 // Emulate effect of Parse::adjust_map_after_if.
1015 Node* ccast = new CastIINode(index, TypeInt::POS);
1016 ccast->set_req(0, control());
1017 (*pos_index) = _gvn.transform(ccast);
1018 }
1019 return is_neg;
1020 }
1021
1022 // Make sure that 'position' is a valid limit index, in [0..length].
1023 // There are two equivalent plans for checking this:
1024 // A. (offset + copyLength) unsigned<= arrayLength
1025 // B. offset <= (arrayLength - copyLength)
1026 // We require that all of the values above, except for the sum and
1027 // difference, are already known to be non-negative.
1028 // Plan A is robust in the face of overflow, if offset and copyLength
1029 // are both hugely positive.
1030 //
1031 // Plan B is less direct and intuitive, but it does not overflow at
1032 // all, since the difference of two non-negatives is always
1033 // representable. Whenever Java methods must perform the equivalent
1034 // check they generally use Plan B instead of Plan A.
1035 // For the moment we use Plan A.
1036 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1037 Node* subseq_length,
1038 Node* array_length,
1039 RegionNode* region) {
1040 if (stopped())
1041 return NULL; // already stopped
1042 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1043 if (zero_offset && subseq_length->eqv_uncast(array_length))
1044 return NULL; // common case of whole-array copy
1045 Node* last = subseq_length;
1046 if (!zero_offset) // last += offset
1047 last = _gvn.transform(new AddINode(last, offset));
1048 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1049 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1050 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1051 return is_over;
1052 }
1053
1054
1055 //--------------------------generate_current_thread--------------------
1056 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1057 ciKlass* thread_klass = env()->Thread_klass();
1058 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1059 Node* thread = _gvn.transform(new ThreadLocalNode());
1060 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1061 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1062 tls_output = thread;
1063 return threadObj;
1064 }
1065
1066
1067 //------------------------------make_string_method_node------------------------
1068 // Helper method for String intrinsic functions. This version is called
1069 // with str1 and str2 pointing to String object nodes.
1070 //
1071 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1072 Node* no_ctrl = NULL;
1073
1074 // Get start addr of string
1075 Node* str1_value = load_String_value(no_ctrl, str1);
1076 Node* str1_offset = load_String_offset(no_ctrl, str1);
1077 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1078
1079 // Get length of string 1
1080 Node* str1_len = load_String_length(no_ctrl, str1);
1081
1082 Node* str2_value = load_String_value(no_ctrl, str2);
1083 Node* str2_offset = load_String_offset(no_ctrl, str2);
1084 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1085
1086 Node* str2_len = NULL;
1087 Node* result = NULL;
1088
1089 switch (opcode) {
1090 case Op_StrIndexOf:
1091 // Get length of string 2
1092 str2_len = load_String_length(no_ctrl, str2);
1093
1094 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1095 str1_start, str1_len, str2_start, str2_len);
1096 break;
1097 case Op_StrComp:
1098 // Get length of string 2
1099 str2_len = load_String_length(no_ctrl, str2);
1100
1101 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1102 str1_start, str1_len, str2_start, str2_len);
1103 break;
1104 case Op_StrEquals:
1105 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1106 str1_start, str2_start, str1_len);
1107 break;
1108 default:
1109 ShouldNotReachHere();
1110 return NULL;
1111 }
1112
1113 // All these intrinsics have checks.
1114 C->set_has_split_ifs(true); // Has chance for split-if optimization
1115
1116 return _gvn.transform(result);
1117 }
1118
1119 // Helper method for String intrinsic functions. This version is called
1120 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1121 // to Int nodes containing the lenghts of str1 and str2.
1122 //
1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1124 Node* result = NULL;
1125 switch (opcode) {
1126 case Op_StrIndexOf:
1127 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1128 str1_start, cnt1, str2_start, cnt2);
1129 break;
1130 case Op_StrComp:
1131 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1132 str1_start, cnt1, str2_start, cnt2);
1133 break;
1134 case Op_StrEquals:
1135 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1136 str1_start, str2_start, cnt1);
1137 break;
1138 default:
1139 ShouldNotReachHere();
1140 return NULL;
1141 }
1142
1143 // All these intrinsics have checks.
1144 C->set_has_split_ifs(true); // Has chance for split-if optimization
1145
1146 return _gvn.transform(result);
1147 }
1148
1149 //------------------------------inline_string_compareTo------------------------
1150 // public int java.lang.String.compareTo(String anotherString);
1151 bool LibraryCallKit::inline_string_compareTo() {
1152 Node* receiver = null_check(argument(0));
1153 Node* arg = null_check(argument(1));
1154 if (stopped()) {
1155 return true;
1156 }
1157 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1158 return true;
1159 }
1160
1161 //------------------------------inline_string_equals------------------------
1162 bool LibraryCallKit::inline_string_equals() {
1163 Node* receiver = null_check_receiver();
1164 // NOTE: Do not null check argument for String.equals() because spec
1165 // allows to specify NULL as argument.
1166 Node* argument = this->argument(1);
1167 if (stopped()) {
1168 return true;
1169 }
1170
1171 // paths (plus control) merge
1172 RegionNode* region = new RegionNode(5);
1173 Node* phi = new PhiNode(region, TypeInt::BOOL);
1174
1175 // does source == target string?
1176 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1177 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1178
1179 Node* if_eq = generate_slow_guard(bol, NULL);
1180 if (if_eq != NULL) {
1181 // receiver == argument
1182 phi->init_req(2, intcon(1));
1183 region->init_req(2, if_eq);
1184 }
1185
1186 // get String klass for instanceOf
1187 ciInstanceKlass* klass = env()->String_klass();
1188
1189 if (!stopped()) {
1190 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1191 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1192 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1193
1194 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1195 //instanceOf == true, fallthrough
1196
1197 if (inst_false != NULL) {
1198 phi->init_req(3, intcon(0));
1199 region->init_req(3, inst_false);
1200 }
1201 }
1202
1203 if (!stopped()) {
1204 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1205
1206 // Properly cast the argument to String
1207 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1208 // This path is taken only when argument's type is String:NotNull.
1209 argument = cast_not_null(argument, false);
1210
1211 Node* no_ctrl = NULL;
1212
1213 // Get start addr of receiver
1214 Node* receiver_val = load_String_value(no_ctrl, receiver);
1215 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1216 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1217
1218 // Get length of receiver
1219 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1220
1221 // Get start addr of argument
1222 Node* argument_val = load_String_value(no_ctrl, argument);
1223 Node* argument_offset = load_String_offset(no_ctrl, argument);
1224 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1225
1226 // Get length of argument
1227 Node* argument_cnt = load_String_length(no_ctrl, argument);
1228
1229 // Check for receiver count != argument count
1230 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1231 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1232 Node* if_ne = generate_slow_guard(bol, NULL);
1233 if (if_ne != NULL) {
1234 phi->init_req(4, intcon(0));
1235 region->init_req(4, if_ne);
1236 }
1237
1238 // Check for count == 0 is done by assembler code for StrEquals.
1239
1240 if (!stopped()) {
1241 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1242 phi->init_req(1, equals);
1243 region->init_req(1, control());
1244 }
1245 }
1246
1247 // post merge
1248 set_control(_gvn.transform(region));
1249 record_for_igvn(region);
1250
1251 set_result(_gvn.transform(phi));
1252 return true;
1253 }
1254
1255 //------------------------------inline_array_equals----------------------------
1256 bool LibraryCallKit::inline_array_equals() {
1257 Node* arg1 = argument(0);
1258 Node* arg2 = argument(1);
1259 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1260 return true;
1261 }
1262
1263 // Java version of String.indexOf(constant string)
1264 // class StringDecl {
1265 // StringDecl(char[] ca) {
1266 // offset = 0;
1267 // count = ca.length;
1268 // value = ca;
1269 // }
1270 // int offset;
1271 // int count;
1272 // char[] value;
1273 // }
1274 //
1275 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1276 // int targetOffset, int cache_i, int md2) {
1277 // int cache = cache_i;
1278 // int sourceOffset = string_object.offset;
1279 // int sourceCount = string_object.count;
1280 // int targetCount = target_object.length;
1281 //
1282 // int targetCountLess1 = targetCount - 1;
1283 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1284 //
1285 // char[] source = string_object.value;
1286 // char[] target = target_object;
1287 // int lastChar = target[targetCountLess1];
1288 //
1289 // outer_loop:
1290 // for (int i = sourceOffset; i < sourceEnd; ) {
1291 // int src = source[i + targetCountLess1];
1292 // if (src == lastChar) {
1293 // // With random strings and a 4-character alphabet,
1294 // // reverse matching at this point sets up 0.8% fewer
1295 // // frames, but (paradoxically) makes 0.3% more probes.
1296 // // Since those probes are nearer the lastChar probe,
1297 // // there is may be a net D$ win with reverse matching.
1298 // // But, reversing loop inhibits unroll of inner loop
1299 // // for unknown reason. So, does running outer loop from
1300 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1301 // for (int j = 0; j < targetCountLess1; j++) {
1302 // if (target[targetOffset + j] != source[i+j]) {
1303 // if ((cache & (1 << source[i+j])) == 0) {
1304 // if (md2 < j+1) {
1305 // i += j+1;
1306 // continue outer_loop;
1307 // }
1308 // }
1309 // i += md2;
1310 // continue outer_loop;
1311 // }
1312 // }
1313 // return i - sourceOffset;
1314 // }
1315 // if ((cache & (1 << src)) == 0) {
1316 // i += targetCountLess1;
1317 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1318 // i++;
1319 // }
1320 // return -1;
1321 // }
1322
1323 //------------------------------string_indexOf------------------------
1324 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1325 jint cache_i, jint md2_i) {
1326
1327 Node* no_ctrl = NULL;
1328 float likely = PROB_LIKELY(0.9);
1329 float unlikely = PROB_UNLIKELY(0.9);
1330
1331 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1332
1333 Node* source = load_String_value(no_ctrl, string_object);
1334 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1335 Node* sourceCount = load_String_length(no_ctrl, string_object);
1336
1337 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1338 jint target_length = target_array->length();
1339 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1340 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1341
1342 // String.value field is known to be @Stable.
1343 if (UseImplicitStableValues) {
1344 target = cast_array_to_stable(target, target_type);
1345 }
1346
1347 IdealKit kit(this, false, true);
1348 #define __ kit.
1349 Node* zero = __ ConI(0);
1350 Node* one = __ ConI(1);
1351 Node* cache = __ ConI(cache_i);
1352 Node* md2 = __ ConI(md2_i);
1353 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1354 Node* targetCountLess1 = __ ConI(target_length - 1);
1355 Node* targetOffset = __ ConI(targetOffset_i);
1356 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1357
1358 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1359 Node* outer_loop = __ make_label(2 /* goto */);
1360 Node* return_ = __ make_label(1);
1361
1362 __ set(rtn,__ ConI(-1));
1363 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1364 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1365 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1366 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1367 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1368 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1369 Node* tpj = __ AddI(targetOffset, __ value(j));
1370 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1371 Node* ipj = __ AddI(__ value(i), __ value(j));
1372 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1373 __ if_then(targ, BoolTest::ne, src2); {
1374 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1375 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1376 __ increment(i, __ AddI(__ value(j), one));
1377 __ goto_(outer_loop);
1378 } __ end_if(); __ dead(j);
1379 }__ end_if(); __ dead(j);
1380 __ increment(i, md2);
1381 __ goto_(outer_loop);
1382 }__ end_if();
1383 __ increment(j, one);
1384 }__ end_loop(); __ dead(j);
1385 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1386 __ goto_(return_);
1387 }__ end_if();
1388 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1389 __ increment(i, targetCountLess1);
1390 }__ end_if();
1391 __ increment(i, one);
1392 __ bind(outer_loop);
1393 }__ end_loop(); __ dead(i);
1394 __ bind(return_);
1395
1396 // Final sync IdealKit and GraphKit.
1397 final_sync(kit);
1398 Node* result = __ value(rtn);
1399 #undef __
1400 C->set_has_loops(true);
1401 return result;
1402 }
1403
1404 //------------------------------inline_string_indexOf------------------------
1405 bool LibraryCallKit::inline_string_indexOf() {
1406 Node* receiver = argument(0);
1407 Node* arg = argument(1);
1408
1409 Node* result;
1410 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1411 UseSSE42Intrinsics) {
1412 // Generate SSE4.2 version of indexOf
1413 // We currently only have match rules that use SSE4.2
1414
1415 receiver = null_check(receiver);
1416 arg = null_check(arg);
1417 if (stopped()) {
1418 return true;
1419 }
1420
1421 // Make the merge point
1422 RegionNode* result_rgn = new RegionNode(4);
1423 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1424 Node* no_ctrl = NULL;
1425
1426 // Get start addr of source string
1427 Node* source = load_String_value(no_ctrl, receiver);
1428 Node* source_offset = load_String_offset(no_ctrl, receiver);
1429 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1430
1431 // Get length of source string
1432 Node* source_cnt = load_String_length(no_ctrl, receiver);
1433
1434 // Get start addr of substring
1435 Node* substr = load_String_value(no_ctrl, arg);
1436 Node* substr_offset = load_String_offset(no_ctrl, arg);
1437 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1438
1439 // Get length of source string
1440 Node* substr_cnt = load_String_length(no_ctrl, arg);
1441
1442 // Check for substr count > string count
1443 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1444 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1445 Node* if_gt = generate_slow_guard(bol, NULL);
1446 if (if_gt != NULL) {
1447 result_phi->init_req(2, intcon(-1));
1448 result_rgn->init_req(2, if_gt);
1449 }
1450
1451 if (!stopped()) {
1452 // Check for substr count == 0
1453 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1454 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1455 Node* if_zero = generate_slow_guard(bol, NULL);
1456 if (if_zero != NULL) {
1457 result_phi->init_req(3, intcon(0));
1458 result_rgn->init_req(3, if_zero);
1459 }
1460 }
1461
1462 if (!stopped()) {
1463 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1464 result_phi->init_req(1, result);
1465 result_rgn->init_req(1, control());
1466 }
1467 set_control(_gvn.transform(result_rgn));
1468 record_for_igvn(result_rgn);
1469 result = _gvn.transform(result_phi);
1470
1471 } else { // Use LibraryCallKit::string_indexOf
1472 // don't intrinsify if argument isn't a constant string.
1473 if (!arg->is_Con()) {
1474 return false;
1475 }
1476 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1477 if (str_type == NULL) {
1478 return false;
1479 }
1480 ciInstanceKlass* klass = env()->String_klass();
1481 ciObject* str_const = str_type->const_oop();
1482 if (str_const == NULL || str_const->klass() != klass) {
1483 return false;
1484 }
1485 ciInstance* str = str_const->as_instance();
1486 assert(str != NULL, "must be instance");
1487
1488 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1489 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1490
1491 int o;
1492 int c;
1493 if (java_lang_String::has_offset_field()) {
1494 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1495 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1496 } else {
1497 o = 0;
1498 c = pat->length();
1499 }
1500
1501 // constant strings have no offset and count == length which
1502 // simplifies the resulting code somewhat so lets optimize for that.
1503 if (o != 0 || c != pat->length()) {
1504 return false;
1505 }
1506
1507 receiver = null_check(receiver, T_OBJECT);
1508 // NOTE: No null check on the argument is needed since it's a constant String oop.
1509 if (stopped()) {
1510 return true;
1511 }
1512
1513 // The null string as a pattern always returns 0 (match at beginning of string)
1514 if (c == 0) {
1515 set_result(intcon(0));
1516 return true;
1517 }
1518
1519 // Generate default indexOf
1520 jchar lastChar = pat->char_at(o + (c - 1));
1521 int cache = 0;
1522 int i;
1523 for (i = 0; i < c - 1; i++) {
1524 assert(i < pat->length(), "out of range");
1525 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1526 }
1527
1528 int md2 = c;
1529 for (i = 0; i < c - 1; i++) {
1530 assert(i < pat->length(), "out of range");
1531 if (pat->char_at(o + i) == lastChar) {
1532 md2 = (c - 1) - i;
1533 }
1534 }
1535
1536 result = string_indexOf(receiver, pat, o, cache, md2);
1537 }
1538 set_result(result);
1539 return true;
1540 }
1541
1542 //--------------------------round_double_node--------------------------------
1543 // Round a double node if necessary.
1544 Node* LibraryCallKit::round_double_node(Node* n) {
1545 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1546 n = _gvn.transform(new RoundDoubleNode(0, n));
1547 return n;
1548 }
1549
1550 //------------------------------inline_math-----------------------------------
1551 // public static double Math.abs(double)
1552 // public static double Math.sqrt(double)
1553 // public static double Math.log(double)
1554 // public static double Math.log10(double)
1555 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1556 Node* arg = round_double_node(argument(0));
1557 Node* n;
1558 switch (id) {
1559 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1560 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1561 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1562 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1563 default: fatal_unexpected_iid(id); break;
1564 }
1565 set_result(_gvn.transform(n));
1566 return true;
1567 }
1568
1569 //------------------------------inline_trig----------------------------------
1570 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1571 // argument reduction which will turn into a fast/slow diamond.
1572 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1573 Node* arg = round_double_node(argument(0));
1574 Node* n = NULL;
1575
1576 switch (id) {
1577 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1578 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1579 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1580 default: fatal_unexpected_iid(id); break;
1581 }
1582 n = _gvn.transform(n);
1583
1584 // Rounding required? Check for argument reduction!
1585 if (Matcher::strict_fp_requires_explicit_rounding) {
1586 static const double pi_4 = 0.7853981633974483;
1587 static const double neg_pi_4 = -0.7853981633974483;
1588 // pi/2 in 80-bit extended precision
1589 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1590 // -pi/2 in 80-bit extended precision
1591 // 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};
1592 // Cutoff value for using this argument reduction technique
1593 //static const double pi_2_minus_epsilon = 1.564660403643354;
1594 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1595
1596 // Pseudocode for sin:
1597 // if (x <= Math.PI / 4.0) {
1598 // if (x >= -Math.PI / 4.0) return fsin(x);
1599 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1600 // } else {
1601 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1602 // }
1603 // return StrictMath.sin(x);
1604
1605 // Pseudocode for cos:
1606 // if (x <= Math.PI / 4.0) {
1607 // if (x >= -Math.PI / 4.0) return fcos(x);
1608 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1609 // } else {
1610 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1611 // }
1612 // return StrictMath.cos(x);
1613
1614 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1615 // requires a special machine instruction to load it. Instead we'll try
1616 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1617 // probably do the math inside the SIN encoding.
1618
1619 // Make the merge point
1620 RegionNode* r = new RegionNode(3);
1621 Node* phi = new PhiNode(r, Type::DOUBLE);
1622
1623 // Flatten arg so we need only 1 test
1624 Node *abs = _gvn.transform(new AbsDNode(arg));
1625 // Node for PI/4 constant
1626 Node *pi4 = makecon(TypeD::make(pi_4));
1627 // Check PI/4 : abs(arg)
1628 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1629 // Check: If PI/4 < abs(arg) then go slow
1630 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1631 // Branch either way
1632 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1633 set_control(opt_iff(r,iff));
1634
1635 // Set fast path result
1636 phi->init_req(2, n);
1637
1638 // Slow path - non-blocking leaf call
1639 Node* call = NULL;
1640 switch (id) {
1641 case vmIntrinsics::_dsin:
1642 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1643 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1644 "Sin", NULL, arg, top());
1645 break;
1646 case vmIntrinsics::_dcos:
1647 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1648 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1649 "Cos", NULL, arg, top());
1650 break;
1651 case vmIntrinsics::_dtan:
1652 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1653 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1654 "Tan", NULL, arg, top());
1655 break;
1656 }
1657 assert(control()->in(0) == call, "");
1658 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1659 r->init_req(1, control());
1660 phi->init_req(1, slow_result);
1661
1662 // Post-merge
1663 set_control(_gvn.transform(r));
1664 record_for_igvn(r);
1665 n = _gvn.transform(phi);
1666
1667 C->set_has_split_ifs(true); // Has chance for split-if optimization
1668 }
1669 set_result(n);
1670 return true;
1671 }
1672
1673 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1674 //-------------------
1675 //result=(result.isNaN())? funcAddr():result;
1676 // Check: If isNaN() by checking result!=result? then either trap
1677 // or go to runtime
1678 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1679 // Build the boolean node
1680 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1681
1682 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1683 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1684 // The pow or exp intrinsic returned a NaN, which requires a call
1685 // to the runtime. Recompile with the runtime call.
1686 uncommon_trap(Deoptimization::Reason_intrinsic,
1687 Deoptimization::Action_make_not_entrant);
1688 }
1689 return result;
1690 } else {
1691 // If this inlining ever returned NaN in the past, we compile a call
1692 // to the runtime to properly handle corner cases
1693
1694 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1695 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1696 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1697
1698 if (!if_slow->is_top()) {
1699 RegionNode* result_region = new RegionNode(3);
1700 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1701
1702 result_region->init_req(1, if_fast);
1703 result_val->init_req(1, result);
1704
1705 set_control(if_slow);
1706
1707 const TypePtr* no_memory_effects = NULL;
1708 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1709 no_memory_effects,
1710 x, top(), y, y ? top() : NULL);
1711 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1712 #ifdef ASSERT
1713 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1714 assert(value_top == top(), "second value must be top");
1715 #endif
1716
1717 result_region->init_req(2, control());
1718 result_val->init_req(2, value);
1719 set_control(_gvn.transform(result_region));
1720 return _gvn.transform(result_val);
1721 } else {
1722 return result;
1723 }
1724 }
1725 }
1726
1727 //------------------------------inline_exp-------------------------------------
1728 // Inline exp instructions, if possible. The Intel hardware only misses
1729 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1730 bool LibraryCallKit::inline_exp() {
1731 Node* arg = round_double_node(argument(0));
1732 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1733
1734 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1735 set_result(n);
1736
1737 C->set_has_split_ifs(true); // Has chance for split-if optimization
1738 return true;
1739 }
1740
1741 //------------------------------inline_pow-------------------------------------
1742 // Inline power instructions, if possible.
1743 bool LibraryCallKit::inline_pow() {
1744 // Pseudocode for pow
1745 // if (y == 2) {
1746 // return x * x;
1747 // } else {
1748 // if (x <= 0.0) {
1749 // long longy = (long)y;
1750 // if ((double)longy == y) { // if y is long
1751 // if (y + 1 == y) longy = 0; // huge number: even
1752 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1753 // } else {
1754 // result = NaN;
1755 // }
1756 // } else {
1757 // result = DPow(x,y);
1758 // }
1759 // if (result != result)? {
1760 // result = uncommon_trap() or runtime_call();
1761 // }
1762 // return result;
1763 // }
1764
1765 Node* x = round_double_node(argument(0));
1766 Node* y = round_double_node(argument(2));
1767
1768 Node* result = NULL;
1769
1770 Node* const_two_node = makecon(TypeD::make(2.0));
1771 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1772 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1773 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1774 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1775 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1776
1777 RegionNode* region_node = new RegionNode(3);
1778 region_node->init_req(1, if_true);
1779
1780 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1781 // special case for x^y where y == 2, we can convert it to x * x
1782 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1783
1784 // set control to if_false since we will now process the false branch
1785 set_control(if_false);
1786
1787 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1788 // Short form: skip the fancy tests and just check for NaN result.
1789 result = _gvn.transform(new PowDNode(C, control(), x, y));
1790 } else {
1791 // If this inlining ever returned NaN in the past, include all
1792 // checks + call to the runtime.
1793
1794 // Set the merge point for If node with condition of (x <= 0.0)
1795 // There are four possible paths to region node and phi node
1796 RegionNode *r = new RegionNode(4);
1797 Node *phi = new PhiNode(r, Type::DOUBLE);
1798
1799 // Build the first if node: if (x <= 0.0)
1800 // Node for 0 constant
1801 Node *zeronode = makecon(TypeD::ZERO);
1802 // Check x:0
1803 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1804 // Check: If (x<=0) then go complex path
1805 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1806 // Branch either way
1807 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1808 // Fast path taken; set region slot 3
1809 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1810 r->init_req(3,fast_taken); // Capture fast-control
1811
1812 // Fast path not-taken, i.e. slow path
1813 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1814
1815 // Set fast path result
1816 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1817 phi->init_req(3, fast_result);
1818
1819 // Complex path
1820 // Build the second if node (if y is long)
1821 // Node for (long)y
1822 Node *longy = _gvn.transform(new ConvD2LNode(y));
1823 // Node for (double)((long) y)
1824 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1825 // Check (double)((long) y) : y
1826 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1827 // Check if (y isn't long) then go to slow path
1828
1829 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1830 // Branch either way
1831 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1832 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1833
1834 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1835
1836 // Calculate DPow(abs(x), y)*(1 & (long)y)
1837 // Node for constant 1
1838 Node *conone = longcon(1);
1839 // 1& (long)y
1840 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1841
1842 // A huge number is always even. Detect a huge number by checking
1843 // if y + 1 == y and set integer to be tested for parity to 0.
1844 // Required for corner case:
1845 // (long)9.223372036854776E18 = max_jlong
1846 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1847 // max_jlong is odd but 9.223372036854776E18 is even
1848 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1849 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1850 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1851 Node* correctedsign = NULL;
1852 if (ConditionalMoveLimit != 0) {
1853 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1854 } else {
1855 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1856 RegionNode *r = new RegionNode(3);
1857 Node *phi = new PhiNode(r, TypeLong::LONG);
1858 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1859 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1860 phi->init_req(1, signnode);
1861 phi->init_req(2, longcon(0));
1862 correctedsign = _gvn.transform(phi);
1863 ylong_path = _gvn.transform(r);
1864 record_for_igvn(r);
1865 }
1866
1867 // zero node
1868 Node *conzero = longcon(0);
1869 // Check (1&(long)y)==0?
1870 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1871 // Check if (1&(long)y)!=0?, if so the result is negative
1872 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1873 // abs(x)
1874 Node *absx=_gvn.transform(new AbsDNode(x));
1875 // abs(x)^y
1876 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1877 // -abs(x)^y
1878 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1879 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1880 Node *signresult = NULL;
1881 if (ConditionalMoveLimit != 0) {
1882 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1883 } else {
1884 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1885 RegionNode *r = new RegionNode(3);
1886 Node *phi = new PhiNode(r, Type::DOUBLE);
1887 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1888 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1889 phi->init_req(1, absxpowy);
1890 phi->init_req(2, negabsxpowy);
1891 signresult = _gvn.transform(phi);
1892 ylong_path = _gvn.transform(r);
1893 record_for_igvn(r);
1894 }
1895 // Set complex path fast result
1896 r->init_req(2, ylong_path);
1897 phi->init_req(2, signresult);
1898
1899 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1900 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1901 r->init_req(1,slow_path);
1902 phi->init_req(1,slow_result);
1903
1904 // Post merge
1905 set_control(_gvn.transform(r));
1906 record_for_igvn(r);
1907 result = _gvn.transform(phi);
1908 }
1909
1910 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1911
1912 // control from finish_pow_exp is now input to the region node
1913 region_node->set_req(2, control());
1914 // the result from finish_pow_exp is now input to the phi node
1915 phi_node->init_req(2, result);
1916 set_control(_gvn.transform(region_node));
1917 record_for_igvn(region_node);
1918 set_result(_gvn.transform(phi_node));
1919
1920 C->set_has_split_ifs(true); // Has chance for split-if optimization
1921 return true;
1922 }
1923
1924 //------------------------------runtime_math-----------------------------
1925 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1926 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1927 "must be (DD)D or (D)D type");
1928
1929 // Inputs
1930 Node* a = round_double_node(argument(0));
1931 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1932
1933 const TypePtr* no_memory_effects = NULL;
1934 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1935 no_memory_effects,
1936 a, top(), b, b ? top() : NULL);
1937 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1938 #ifdef ASSERT
1939 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1940 assert(value_top == top(), "second value must be top");
1941 #endif
1942
1943 set_result(value);
1944 return true;
1945 }
1946
1947 //------------------------------inline_math_native-----------------------------
1948 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1949 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1950 switch (id) {
1951 // These intrinsics are not properly supported on all hardware
1952 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1953 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1954 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1955 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1956 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1957 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1958
1959 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1960 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1961 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1962 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1963
1964 // These intrinsics are supported on all hardware
1965 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1966 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1967
1968 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1969 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1970 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1971 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1972 #undef FN_PTR
1973
1974 // These intrinsics are not yet correctly implemented
1975 case vmIntrinsics::_datan2:
1976 return false;
1977
1978 default:
1979 fatal_unexpected_iid(id);
1980 return false;
1981 }
1982 }
1983
1984 static bool is_simple_name(Node* n) {
1985 return (n->req() == 1 // constant
1986 || (n->is_Type() && n->as_Type()->type()->singleton())
1987 || n->is_Proj() // parameter or return value
1988 || n->is_Phi() // local of some sort
1989 );
1990 }
1991
1992 //----------------------------inline_min_max-----------------------------------
1993 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1994 set_result(generate_min_max(id, argument(0), argument(1)));
1995 return true;
1996 }
1997
1998 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1999 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2000 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2001 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2002 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2003
2004 {
2005 PreserveJVMState pjvms(this);
2006 PreserveReexecuteState preexecs(this);
2007 jvms()->set_should_reexecute(true);
2008
2009 set_control(slow_path);
2010 set_i_o(i_o());
2011
2012 uncommon_trap(Deoptimization::Reason_intrinsic,
2013 Deoptimization::Action_none);
2014 }
2015
2016 set_control(fast_path);
2017 set_result(math);
2018 }
2019
2020 template <typename OverflowOp>
2021 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2022 typedef typename OverflowOp::MathOp MathOp;
2023
2024 MathOp* mathOp = new MathOp(arg1, arg2);
2025 Node* operation = _gvn.transform( mathOp );
2026 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2027 inline_math_mathExact(operation, ofcheck);
2028 return true;
2029 }
2030
2031 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2032 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2033 }
2034
2035 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2036 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2037 }
2038
2039 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2040 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2041 }
2042
2043 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2044 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2045 }
2046
2047 bool LibraryCallKit::inline_math_negateExactI() {
2048 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2049 }
2050
2051 bool LibraryCallKit::inline_math_negateExactL() {
2052 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2053 }
2054
2055 bool LibraryCallKit::inline_math_multiplyExactI() {
2056 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2057 }
2058
2059 bool LibraryCallKit::inline_math_multiplyExactL() {
2060 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2061 }
2062
2063 Node*
2064 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2065 // These are the candidate return value:
2066 Node* xvalue = x0;
2067 Node* yvalue = y0;
2068
2069 if (xvalue == yvalue) {
2070 return xvalue;
2071 }
2072
2073 bool want_max = (id == vmIntrinsics::_max);
2074
2075 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2076 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2077 if (txvalue == NULL || tyvalue == NULL) return top();
2078 // This is not really necessary, but it is consistent with a
2079 // hypothetical MaxINode::Value method:
2080 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2081
2082 // %%% This folding logic should (ideally) be in a different place.
2083 // Some should be inside IfNode, and there to be a more reliable
2084 // transformation of ?: style patterns into cmoves. We also want
2085 // more powerful optimizations around cmove and min/max.
2086
2087 // Try to find a dominating comparison of these guys.
2088 // It can simplify the index computation for Arrays.copyOf
2089 // and similar uses of System.arraycopy.
2090 // First, compute the normalized version of CmpI(x, y).
2091 int cmp_op = Op_CmpI;
2092 Node* xkey = xvalue;
2093 Node* ykey = yvalue;
2094 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2095 if (ideal_cmpxy->is_Cmp()) {
2096 // E.g., if we have CmpI(length - offset, count),
2097 // it might idealize to CmpI(length, count + offset)
2098 cmp_op = ideal_cmpxy->Opcode();
2099 xkey = ideal_cmpxy->in(1);
2100 ykey = ideal_cmpxy->in(2);
2101 }
2102
2103 // Start by locating any relevant comparisons.
2104 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2105 Node* cmpxy = NULL;
2106 Node* cmpyx = NULL;
2107 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2108 Node* cmp = start_from->fast_out(k);
2109 if (cmp->outcnt() > 0 && // must have prior uses
2110 cmp->in(0) == NULL && // must be context-independent
2111 cmp->Opcode() == cmp_op) { // right kind of compare
2112 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2113 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2114 }
2115 }
2116
2117 const int NCMPS = 2;
2118 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2119 int cmpn;
2120 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2121 if (cmps[cmpn] != NULL) break; // find a result
2122 }
2123 if (cmpn < NCMPS) {
2124 // Look for a dominating test that tells us the min and max.
2125 int depth = 0; // Limit search depth for speed
2126 Node* dom = control();
2127 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2128 if (++depth >= 100) break;
2129 Node* ifproj = dom;
2130 if (!ifproj->is_Proj()) continue;
2131 Node* iff = ifproj->in(0);
2132 if (!iff->is_If()) continue;
2133 Node* bol = iff->in(1);
2134 if (!bol->is_Bool()) continue;
2135 Node* cmp = bol->in(1);
2136 if (cmp == NULL) continue;
2137 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2138 if (cmps[cmpn] == cmp) break;
2139 if (cmpn == NCMPS) continue;
2140 BoolTest::mask btest = bol->as_Bool()->_test._test;
2141 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2142 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2143 // At this point, we know that 'x btest y' is true.
2144 switch (btest) {
2145 case BoolTest::eq:
2146 // They are proven equal, so we can collapse the min/max.
2147 // Either value is the answer. Choose the simpler.
2148 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2149 return yvalue;
2150 return xvalue;
2151 case BoolTest::lt: // x < y
2152 case BoolTest::le: // x <= y
2153 return (want_max ? yvalue : xvalue);
2154 case BoolTest::gt: // x > y
2155 case BoolTest::ge: // x >= y
2156 return (want_max ? xvalue : yvalue);
2157 }
2158 }
2159 }
2160
2161 // We failed to find a dominating test.
2162 // Let's pick a test that might GVN with prior tests.
2163 Node* best_bol = NULL;
2164 BoolTest::mask best_btest = BoolTest::illegal;
2165 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2166 Node* cmp = cmps[cmpn];
2167 if (cmp == NULL) continue;
2168 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2169 Node* bol = cmp->fast_out(j);
2170 if (!bol->is_Bool()) continue;
2171 BoolTest::mask btest = bol->as_Bool()->_test._test;
2172 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2173 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2174 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2175 best_bol = bol->as_Bool();
2176 best_btest = btest;
2177 }
2178 }
2179 }
2180
2181 Node* answer_if_true = NULL;
2182 Node* answer_if_false = NULL;
2183 switch (best_btest) {
2184 default:
2185 if (cmpxy == NULL)
2186 cmpxy = ideal_cmpxy;
2187 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2188 // and fall through:
2189 case BoolTest::lt: // x < y
2190 case BoolTest::le: // x <= y
2191 answer_if_true = (want_max ? yvalue : xvalue);
2192 answer_if_false = (want_max ? xvalue : yvalue);
2193 break;
2194 case BoolTest::gt: // x > y
2195 case BoolTest::ge: // x >= y
2196 answer_if_true = (want_max ? xvalue : yvalue);
2197 answer_if_false = (want_max ? yvalue : xvalue);
2198 break;
2199 }
2200
2201 jint hi, lo;
2202 if (want_max) {
2203 // We can sharpen the minimum.
2204 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2205 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2206 } else {
2207 // We can sharpen the maximum.
2208 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2209 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2210 }
2211
2212 // Use a flow-free graph structure, to avoid creating excess control edges
2213 // which could hinder other optimizations.
2214 // Since Math.min/max is often used with arraycopy, we want
2215 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2216 Node* cmov = CMoveNode::make(NULL, best_bol,
2217 answer_if_false, answer_if_true,
2218 TypeInt::make(lo, hi, widen));
2219
2220 return _gvn.transform(cmov);
2221
2222 /*
2223 // This is not as desirable as it may seem, since Min and Max
2224 // nodes do not have a full set of optimizations.
2225 // And they would interfere, anyway, with 'if' optimizations
2226 // and with CMoveI canonical forms.
2227 switch (id) {
2228 case vmIntrinsics::_min:
2229 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2230 case vmIntrinsics::_max:
2231 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2232 default:
2233 ShouldNotReachHere();
2234 }
2235 */
2236 }
2237
2238 inline int
2239 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2240 const TypePtr* base_type = TypePtr::NULL_PTR;
2241 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2242 if (base_type == NULL) {
2243 // Unknown type.
2244 return Type::AnyPtr;
2245 } else if (base_type == TypePtr::NULL_PTR) {
2246 // Since this is a NULL+long form, we have to switch to a rawptr.
2247 base = _gvn.transform(new CastX2PNode(offset));
2248 offset = MakeConX(0);
2249 return Type::RawPtr;
2250 } else if (base_type->base() == Type::RawPtr) {
2251 return Type::RawPtr;
2252 } else if (base_type->isa_oopptr()) {
2253 // Base is never null => always a heap address.
2254 if (base_type->ptr() == TypePtr::NotNull) {
2255 return Type::OopPtr;
2256 }
2257 // Offset is small => always a heap address.
2258 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2259 if (offset_type != NULL &&
2260 base_type->offset() == 0 && // (should always be?)
2261 offset_type->_lo >= 0 &&
2262 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2263 return Type::OopPtr;
2264 }
2265 // Otherwise, it might either be oop+off or NULL+addr.
2266 return Type::AnyPtr;
2267 } else {
2268 // No information:
2269 return Type::AnyPtr;
2270 }
2271 }
2272
2273 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2274 int kind = classify_unsafe_addr(base, offset);
2275 if (kind == Type::RawPtr) {
2276 return basic_plus_adr(top(), base, offset);
2277 } else {
2278 return basic_plus_adr(base, offset);
2279 }
2280 }
2281
2282 //--------------------------inline_number_methods-----------------------------
2283 // inline int Integer.numberOfLeadingZeros(int)
2284 // inline int Long.numberOfLeadingZeros(long)
2285 //
2286 // inline int Integer.numberOfTrailingZeros(int)
2287 // inline int Long.numberOfTrailingZeros(long)
2288 //
2289 // inline int Integer.bitCount(int)
2290 // inline int Long.bitCount(long)
2291 //
2292 // inline char Character.reverseBytes(char)
2293 // inline short Short.reverseBytes(short)
2294 // inline int Integer.reverseBytes(int)
2295 // inline long Long.reverseBytes(long)
2296 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2297 Node* arg = argument(0);
2298 Node* n;
2299 switch (id) {
2300 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2301 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2302 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2303 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2304 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2305 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2306 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2307 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2308 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2309 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2310 default: fatal_unexpected_iid(id); break;
2311 }
2312 set_result(_gvn.transform(n));
2313 return true;
2314 }
2315
2316 //----------------------------inline_unsafe_access----------------------------
2317
2318 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2319
2320 // Helper that guards and inserts a pre-barrier.
2321 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2322 Node* pre_val, bool need_mem_bar) {
2323 // We could be accessing the referent field of a reference object. If so, when G1
2324 // is enabled, we need to log the value in the referent field in an SATB buffer.
2325 // This routine performs some compile time filters and generates suitable
2326 // runtime filters that guard the pre-barrier code.
2327 // Also add memory barrier for non volatile load from the referent field
2328 // to prevent commoning of loads across safepoint.
2329 if (!UseG1GC && !need_mem_bar)
2330 return;
2331
2332 // Some compile time checks.
2333
2334 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2335 const TypeX* otype = offset->find_intptr_t_type();
2336 if (otype != NULL && otype->is_con() &&
2337 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2338 // Constant offset but not the reference_offset so just return
2339 return;
2340 }
2341
2342 // We only need to generate the runtime guards for instances.
2343 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2344 if (btype != NULL) {
2345 if (btype->isa_aryptr()) {
2346 // Array type so nothing to do
2347 return;
2348 }
2349
2350 const TypeInstPtr* itype = btype->isa_instptr();
2351 if (itype != NULL) {
2352 // Can the klass of base_oop be statically determined to be
2353 // _not_ a sub-class of Reference and _not_ Object?
2354 ciKlass* klass = itype->klass();
2355 if ( klass->is_loaded() &&
2356 !klass->is_subtype_of(env()->Reference_klass()) &&
2357 !env()->Object_klass()->is_subtype_of(klass)) {
2358 return;
2359 }
2360 }
2361 }
2362
2363 // The compile time filters did not reject base_oop/offset so
2364 // we need to generate the following runtime filters
2365 //
2366 // if (offset == java_lang_ref_Reference::_reference_offset) {
2367 // if (instance_of(base, java.lang.ref.Reference)) {
2368 // pre_barrier(_, pre_val, ...);
2369 // }
2370 // }
2371
2372 float likely = PROB_LIKELY( 0.999);
2373 float unlikely = PROB_UNLIKELY(0.999);
2374
2375 IdealKit ideal(this);
2376 #define __ ideal.
2377
2378 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2379
2380 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2381 // Update graphKit memory and control from IdealKit.
2382 sync_kit(ideal);
2383
2384 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2385 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2386
2387 // Update IdealKit memory and control from graphKit.
2388 __ sync_kit(this);
2389
2390 Node* one = __ ConI(1);
2391 // is_instof == 0 if base_oop == NULL
2392 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2393
2394 // Update graphKit from IdeakKit.
2395 sync_kit(ideal);
2396
2397 // Use the pre-barrier to record the value in the referent field
2398 pre_barrier(false /* do_load */,
2399 __ ctrl(),
2400 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2401 pre_val /* pre_val */,
2402 T_OBJECT);
2403 if (need_mem_bar) {
2404 // Add memory barrier to prevent commoning reads from this field
2405 // across safepoint since GC can change its value.
2406 insert_mem_bar(Op_MemBarCPUOrder);
2407 }
2408 // Update IdealKit from graphKit.
2409 __ sync_kit(this);
2410
2411 } __ end_if(); // _ref_type != ref_none
2412 } __ end_if(); // offset == referent_offset
2413
2414 // Final sync IdealKit and GraphKit.
2415 final_sync(ideal);
2416 #undef __
2417 }
2418
2419
2420 // Interpret Unsafe.fieldOffset cookies correctly:
2421 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2422
2423 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2424 // Attempt to infer a sharper value type from the offset and base type.
2425 ciKlass* sharpened_klass = NULL;
2426
2427 // See if it is an instance field, with an object type.
2428 if (alias_type->field() != NULL) {
2429 assert(!is_native_ptr, "native pointer op cannot use a java address");
2430 if (alias_type->field()->type()->is_klass()) {
2431 sharpened_klass = alias_type->field()->type()->as_klass();
2432 }
2433 }
2434
2435 // See if it is a narrow oop array.
2436 if (adr_type->isa_aryptr()) {
2437 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2438 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2439 if (elem_type != NULL) {
2440 sharpened_klass = elem_type->klass();
2441 }
2442 }
2443 }
2444
2445 // The sharpened class might be unloaded if there is no class loader
2446 // contraint in place.
2447 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2448 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2449
2450 #ifndef PRODUCT
2451 if (C->print_intrinsics() || C->print_inlining()) {
2452 tty->print(" from base type: "); adr_type->dump();
2453 tty->print(" sharpened value: "); tjp->dump();
2454 }
2455 #endif
2456 // Sharpen the value type.
2457 return tjp;
2458 }
2459 return NULL;
2460 }
2461
2462 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2463 if (callee()->is_static()) return false; // caller must have the capability!
2464
2465 #ifndef PRODUCT
2466 {
2467 ResourceMark rm;
2468 // Check the signatures.
2469 ciSignature* sig = callee()->signature();
2470 #ifdef ASSERT
2471 if (!is_store) {
2472 // Object getObject(Object base, int/long offset), etc.
2473 BasicType rtype = sig->return_type()->basic_type();
2474 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2475 rtype = T_ADDRESS; // it is really a C void*
2476 assert(rtype == type, "getter must return the expected value");
2477 if (!is_native_ptr) {
2478 assert(sig->count() == 2, "oop getter has 2 arguments");
2479 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2480 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2481 } else {
2482 assert(sig->count() == 1, "native getter has 1 argument");
2483 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2484 }
2485 } else {
2486 // void putObject(Object base, int/long offset, Object x), etc.
2487 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2488 if (!is_native_ptr) {
2489 assert(sig->count() == 3, "oop putter has 3 arguments");
2490 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2491 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2492 } else {
2493 assert(sig->count() == 2, "native putter has 2 arguments");
2494 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2495 }
2496 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2497 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2498 vtype = T_ADDRESS; // it is really a C void*
2499 assert(vtype == type, "putter must accept the expected value");
2500 }
2501 #endif // ASSERT
2502 }
2503 #endif //PRODUCT
2504
2505 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2506
2507 Node* receiver = argument(0); // type: oop
2508
2509 // Build address expression.
2510 Node* adr;
2511 Node* heap_base_oop = top();
2512 Node* offset = top();
2513 Node* val;
2514
2515 if (!is_native_ptr) {
2516 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2517 Node* base = argument(1); // type: oop
2518 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2519 offset = argument(2); // type: long
2520 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2521 // to be plain byte offsets, which are also the same as those accepted
2522 // by oopDesc::field_base.
2523 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2524 "fieldOffset must be byte-scaled");
2525 // 32-bit machines ignore the high half!
2526 offset = ConvL2X(offset);
2527 adr = make_unsafe_address(base, offset);
2528 heap_base_oop = base;
2529 val = is_store ? argument(4) : NULL;
2530 } else {
2531 Node* ptr = argument(1); // type: long
2532 ptr = ConvL2X(ptr); // adjust Java long to machine word
2533 adr = make_unsafe_address(NULL, ptr);
2534 val = is_store ? argument(3) : NULL;
2535 }
2536
2537 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2538
2539 // First guess at the value type.
2540 const Type *value_type = Type::get_const_basic_type(type);
2541
2542 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2543 // there was not enough information to nail it down.
2544 Compile::AliasType* alias_type = C->alias_type(adr_type);
2545 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2546
2547 // We will need memory barriers unless we can determine a unique
2548 // alias category for this reference. (Note: If for some reason
2549 // the barriers get omitted and the unsafe reference begins to "pollute"
2550 // the alias analysis of the rest of the graph, either Compile::can_alias
2551 // or Compile::must_alias will throw a diagnostic assert.)
2552 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2553
2554 // If we are reading the value of the referent field of a Reference
2555 // object (either by using Unsafe directly or through reflection)
2556 // then, if G1 is enabled, we need to record the referent in an
2557 // SATB log buffer using the pre-barrier mechanism.
2558 // Also we need to add memory barrier to prevent commoning reads
2559 // from this field across safepoint since GC can change its value.
2560 bool need_read_barrier = !is_native_ptr && !is_store &&
2561 offset != top() && heap_base_oop != top();
2562
2563 if (!is_store && type == T_OBJECT) {
2564 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2565 if (tjp != NULL) {
2566 value_type = tjp;
2567 }
2568 }
2569
2570 receiver = null_check(receiver);
2571 if (stopped()) {
2572 return true;
2573 }
2574 // Heap pointers get a null-check from the interpreter,
2575 // as a courtesy. However, this is not guaranteed by Unsafe,
2576 // and it is not possible to fully distinguish unintended nulls
2577 // from intended ones in this API.
2578
2579 if (is_volatile) {
2580 // We need to emit leading and trailing CPU membars (see below) in
2581 // addition to memory membars when is_volatile. This is a little
2582 // too strong, but avoids the need to insert per-alias-type
2583 // volatile membars (for stores; compare Parse::do_put_xxx), which
2584 // we cannot do effectively here because we probably only have a
2585 // rough approximation of type.
2586 need_mem_bar = true;
2587 // For Stores, place a memory ordering barrier now.
2588 if (is_store) {
2589 insert_mem_bar(Op_MemBarRelease);
2590 } else {
2591 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2592 insert_mem_bar(Op_MemBarVolatile);
2593 }
2594 }
2595 }
2596
2597 // Memory barrier to prevent normal and 'unsafe' accesses from
2598 // bypassing each other. Happens after null checks, so the
2599 // exception paths do not take memory state from the memory barrier,
2600 // so there's no problems making a strong assert about mixing users
2601 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2602 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2603 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2604
2605 if (!is_store) {
2606 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2607 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2608 // load value
2609 switch (type) {
2610 case T_BOOLEAN:
2611 case T_CHAR:
2612 case T_BYTE:
2613 case T_SHORT:
2614 case T_INT:
2615 case T_LONG:
2616 case T_FLOAT:
2617 case T_DOUBLE:
2618 break;
2619 case T_OBJECT:
2620 if (need_read_barrier) {
2621 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2622 }
2623 break;
2624 case T_ADDRESS:
2625 // Cast to an int type.
2626 p = _gvn.transform(new CastP2XNode(NULL, p));
2627 p = ConvX2UL(p);
2628 break;
2629 default:
2630 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2631 break;
2632 }
2633 // The load node has the control of the preceding MemBarCPUOrder. All
2634 // following nodes will have the control of the MemBarCPUOrder inserted at
2635 // the end of this method. So, pushing the load onto the stack at a later
2636 // point is fine.
2637 set_result(p);
2638 } else {
2639 // place effect of store into memory
2640 switch (type) {
2641 case T_DOUBLE:
2642 val = dstore_rounding(val);
2643 break;
2644 case T_ADDRESS:
2645 // Repackage the long as a pointer.
2646 val = ConvL2X(val);
2647 val = _gvn.transform(new CastX2PNode(val));
2648 break;
2649 }
2650
2651 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2652 if (type != T_OBJECT ) {
2653 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2654 } else {
2655 // Possibly an oop being stored to Java heap or native memory
2656 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2657 // oop to Java heap.
2658 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2659 } else {
2660 // We can't tell at compile time if we are storing in the Java heap or outside
2661 // of it. So we need to emit code to conditionally do the proper type of
2662 // store.
2663
2664 IdealKit ideal(this);
2665 #define __ ideal.
2666 // QQQ who knows what probability is here??
2667 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2668 // Sync IdealKit and graphKit.
2669 sync_kit(ideal);
2670 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2671 // Update IdealKit memory.
2672 __ sync_kit(this);
2673 } __ else_(); {
2674 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2675 } __ end_if();
2676 // Final sync IdealKit and GraphKit.
2677 final_sync(ideal);
2678 #undef __
2679 }
2680 }
2681 }
2682
2683 if (is_volatile) {
2684 if (!is_store) {
2685 insert_mem_bar(Op_MemBarAcquire);
2686 } else {
2687 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2688 insert_mem_bar(Op_MemBarVolatile);
2689 }
2690 }
2691 }
2692
2693 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2694
2695 return true;
2696 }
2697
2698 //----------------------------inline_unsafe_load_store----------------------------
2699 // This method serves a couple of different customers (depending on LoadStoreKind):
2700 //
2701 // LS_cmpxchg:
2702 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2703 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2704 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2705 //
2706 // LS_xadd:
2707 // public int getAndAddInt( Object o, long offset, int delta)
2708 // public long getAndAddLong(Object o, long offset, long delta)
2709 //
2710 // LS_xchg:
2711 // int getAndSet(Object o, long offset, int newValue)
2712 // long getAndSet(Object o, long offset, long newValue)
2713 // Object getAndSet(Object o, long offset, Object newValue)
2714 //
2715 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2716 // This basic scheme here is the same as inline_unsafe_access, but
2717 // differs in enough details that combining them would make the code
2718 // overly confusing. (This is a true fact! I originally combined
2719 // them, but even I was confused by it!) As much code/comments as
2720 // possible are retained from inline_unsafe_access though to make
2721 // the correspondences clearer. - dl
2722
2723 if (callee()->is_static()) return false; // caller must have the capability!
2724
2725 #ifndef PRODUCT
2726 BasicType rtype;
2727 {
2728 ResourceMark rm;
2729 // Check the signatures.
2730 ciSignature* sig = callee()->signature();
2731 rtype = sig->return_type()->basic_type();
2732 if (kind == LS_xadd || kind == LS_xchg) {
2733 // Check the signatures.
2734 #ifdef ASSERT
2735 assert(rtype == type, "get and set must return the expected type");
2736 assert(sig->count() == 3, "get and set has 3 arguments");
2737 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2738 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2739 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2740 #endif // ASSERT
2741 } else if (kind == LS_cmpxchg) {
2742 // Check the signatures.
2743 #ifdef ASSERT
2744 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2745 assert(sig->count() == 4, "CAS has 4 arguments");
2746 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2747 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2748 #endif // ASSERT
2749 } else {
2750 ShouldNotReachHere();
2751 }
2752 }
2753 #endif //PRODUCT
2754
2755 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2756
2757 // Get arguments:
2758 Node* receiver = NULL;
2759 Node* base = NULL;
2760 Node* offset = NULL;
2761 Node* oldval = NULL;
2762 Node* newval = NULL;
2763 if (kind == LS_cmpxchg) {
2764 const bool two_slot_type = type2size[type] == 2;
2765 receiver = argument(0); // type: oop
2766 base = argument(1); // type: oop
2767 offset = argument(2); // type: long
2768 oldval = argument(4); // type: oop, int, or long
2769 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2770 } else if (kind == LS_xadd || kind == LS_xchg){
2771 receiver = argument(0); // type: oop
2772 base = argument(1); // type: oop
2773 offset = argument(2); // type: long
2774 oldval = NULL;
2775 newval = argument(4); // type: oop, int, or long
2776 }
2777
2778 // Null check receiver.
2779 receiver = null_check(receiver);
2780 if (stopped()) {
2781 return true;
2782 }
2783
2784 // Build field offset expression.
2785 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2786 // to be plain byte offsets, which are also the same as those accepted
2787 // by oopDesc::field_base.
2788 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2789 // 32-bit machines ignore the high half of long offsets
2790 offset = ConvL2X(offset);
2791 Node* adr = make_unsafe_address(base, offset);
2792 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2793
2794 // For CAS, unlike inline_unsafe_access, there seems no point in
2795 // trying to refine types. Just use the coarse types here.
2796 const Type *value_type = Type::get_const_basic_type(type);
2797 Compile::AliasType* alias_type = C->alias_type(adr_type);
2798 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2799
2800 if (kind == LS_xchg && type == T_OBJECT) {
2801 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2802 if (tjp != NULL) {
2803 value_type = tjp;
2804 }
2805 }
2806
2807 int alias_idx = C->get_alias_index(adr_type);
2808
2809 // Memory-model-wise, a LoadStore acts like a little synchronized
2810 // block, so needs barriers on each side. These don't translate
2811 // into actual barriers on most machines, but we still need rest of
2812 // compiler to respect ordering.
2813
2814 insert_mem_bar(Op_MemBarRelease);
2815 insert_mem_bar(Op_MemBarCPUOrder);
2816
2817 // 4984716: MemBars must be inserted before this
2818 // memory node in order to avoid a false
2819 // dependency which will confuse the scheduler.
2820 Node *mem = memory(alias_idx);
2821
2822 // For now, we handle only those cases that actually exist: ints,
2823 // longs, and Object. Adding others should be straightforward.
2824 Node* load_store;
2825 switch(type) {
2826 case T_INT:
2827 if (kind == LS_xadd) {
2828 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2829 } else if (kind == LS_xchg) {
2830 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2831 } else if (kind == LS_cmpxchg) {
2832 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2833 } else {
2834 ShouldNotReachHere();
2835 }
2836 break;
2837 case T_LONG:
2838 if (kind == LS_xadd) {
2839 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2840 } else if (kind == LS_xchg) {
2841 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2842 } else if (kind == LS_cmpxchg) {
2843 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2844 } else {
2845 ShouldNotReachHere();
2846 }
2847 break;
2848 case T_OBJECT:
2849 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2850 // could be delayed during Parse (for example, in adjust_map_after_if()).
2851 // Execute transformation here to avoid barrier generation in such case.
2852 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2853 newval = _gvn.makecon(TypePtr::NULL_PTR);
2854
2855 // Reference stores need a store barrier.
2856 if (kind == LS_xchg) {
2857 // If pre-barrier must execute before the oop store, old value will require do_load here.
2858 if (!can_move_pre_barrier()) {
2859 pre_barrier(true /* do_load*/,
2860 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2861 NULL /* pre_val*/,
2862 T_OBJECT);
2863 } // Else move pre_barrier to use load_store value, see below.
2864 } else if (kind == LS_cmpxchg) {
2865 // Same as for newval above:
2866 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2867 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2868 }
2869 // The only known value which might get overwritten is oldval.
2870 pre_barrier(false /* do_load */,
2871 control(), NULL, NULL, max_juint, NULL, NULL,
2872 oldval /* pre_val */,
2873 T_OBJECT);
2874 } else {
2875 ShouldNotReachHere();
2876 }
2877
2878 #ifdef _LP64
2879 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2880 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2881 if (kind == LS_xchg) {
2882 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2883 newval_enc, adr_type, value_type->make_narrowoop()));
2884 } else {
2885 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2886 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2887 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2888 newval_enc, oldval_enc));
2889 }
2890 } else
2891 #endif
2892 {
2893 if (kind == LS_xchg) {
2894 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2895 } else {
2896 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2897 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2898 }
2899 }
2900 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2901 break;
2902 default:
2903 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2904 break;
2905 }
2906
2907 // SCMemProjNodes represent the memory state of a LoadStore. Their
2908 // main role is to prevent LoadStore nodes from being optimized away
2909 // when their results aren't used.
2910 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2911 set_memory(proj, alias_idx);
2912
2913 if (type == T_OBJECT && kind == LS_xchg) {
2914 #ifdef _LP64
2915 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2916 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2917 }
2918 #endif
2919 if (can_move_pre_barrier()) {
2920 // Don't need to load pre_val. The old value is returned by load_store.
2921 // The pre_barrier can execute after the xchg as long as no safepoint
2922 // gets inserted between them.
2923 pre_barrier(false /* do_load */,
2924 control(), NULL, NULL, max_juint, NULL, NULL,
2925 load_store /* pre_val */,
2926 T_OBJECT);
2927 }
2928 }
2929
2930 // Add the trailing membar surrounding the access
2931 insert_mem_bar(Op_MemBarCPUOrder);
2932 insert_mem_bar(Op_MemBarAcquire);
2933
2934 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2935 set_result(load_store);
2936 return true;
2937 }
2938
2939 //----------------------------inline_unsafe_ordered_store----------------------
2940 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2941 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2942 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2943 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2944 // This is another variant of inline_unsafe_access, differing in
2945 // that it always issues store-store ("release") barrier and ensures
2946 // store-atomicity (which only matters for "long").
2947
2948 if (callee()->is_static()) return false; // caller must have the capability!
2949
2950 #ifndef PRODUCT
2951 {
2952 ResourceMark rm;
2953 // Check the signatures.
2954 ciSignature* sig = callee()->signature();
2955 #ifdef ASSERT
2956 BasicType rtype = sig->return_type()->basic_type();
2957 assert(rtype == T_VOID, "must return void");
2958 assert(sig->count() == 3, "has 3 arguments");
2959 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2960 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2961 #endif // ASSERT
2962 }
2963 #endif //PRODUCT
2964
2965 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2966
2967 // Get arguments:
2968 Node* receiver = argument(0); // type: oop
2969 Node* base = argument(1); // type: oop
2970 Node* offset = argument(2); // type: long
2971 Node* val = argument(4); // type: oop, int, or long
2972
2973 // Null check receiver.
2974 receiver = null_check(receiver);
2975 if (stopped()) {
2976 return true;
2977 }
2978
2979 // Build field offset expression.
2980 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2981 // 32-bit machines ignore the high half of long offsets
2982 offset = ConvL2X(offset);
2983 Node* adr = make_unsafe_address(base, offset);
2984 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2985 const Type *value_type = Type::get_const_basic_type(type);
2986 Compile::AliasType* alias_type = C->alias_type(adr_type);
2987
2988 insert_mem_bar(Op_MemBarRelease);
2989 insert_mem_bar(Op_MemBarCPUOrder);
2990 // Ensure that the store is atomic for longs:
2991 const bool require_atomic_access = true;
2992 Node* store;
2993 if (type == T_OBJECT) // reference stores need a store barrier.
2994 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2995 else {
2996 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2997 }
2998 insert_mem_bar(Op_MemBarCPUOrder);
2999 return true;
3000 }
3001
3002 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3003 // Regardless of form, don't allow previous ld/st to move down,
3004 // then issue acquire, release, or volatile mem_bar.
3005 insert_mem_bar(Op_MemBarCPUOrder);
3006 switch(id) {
3007 case vmIntrinsics::_loadFence:
3008 insert_mem_bar(Op_LoadFence);
3009 return true;
3010 case vmIntrinsics::_storeFence:
3011 insert_mem_bar(Op_StoreFence);
3012 return true;
3013 case vmIntrinsics::_fullFence:
3014 insert_mem_bar(Op_MemBarVolatile);
3015 return true;
3016 default:
3017 fatal_unexpected_iid(id);
3018 return false;
3019 }
3020 }
3021
3022 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3023 if (!kls->is_Con()) {
3024 return true;
3025 }
3026 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3027 if (klsptr == NULL) {
3028 return true;
3029 }
3030 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3031 // don't need a guard for a klass that is already initialized
3032 return !ik->is_initialized();
3033 }
3034
3035 //----------------------------inline_unsafe_allocate---------------------------
3036 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3037 bool LibraryCallKit::inline_unsafe_allocate() {
3038 if (callee()->is_static()) return false; // caller must have the capability!
3039
3040 null_check_receiver(); // null-check, then ignore
3041 Node* cls = null_check(argument(1));
3042 if (stopped()) return true;
3043
3044 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3045 kls = null_check(kls);
3046 if (stopped()) return true; // argument was like int.class
3047
3048 Node* test = NULL;
3049 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3050 // Note: The argument might still be an illegal value like
3051 // Serializable.class or Object[].class. The runtime will handle it.
3052 // But we must make an explicit check for initialization.
3053 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3054 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3055 // can generate code to load it as unsigned byte.
3056 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3057 Node* bits = intcon(InstanceKlass::fully_initialized);
3058 test = _gvn.transform(new SubINode(inst, bits));
3059 // The 'test' is non-zero if we need to take a slow path.
3060 }
3061
3062 Node* obj = new_instance(kls, test);
3063 set_result(obj);
3064 return true;
3065 }
3066
3067 #ifdef TRACE_HAVE_INTRINSICS
3068 /*
3069 * oop -> myklass
3070 * myklass->trace_id |= USED
3071 * return myklass->trace_id & ~0x3
3072 */
3073 bool LibraryCallKit::inline_native_classID() {
3074 null_check_receiver(); // null-check, then ignore
3075 Node* cls = null_check(argument(1), T_OBJECT);
3076 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3077 kls = null_check(kls, T_OBJECT);
3078 ByteSize offset = TRACE_ID_OFFSET;
3079 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3080 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3081 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3082 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3083 Node* clsused = longcon(0x01l); // set the class bit
3084 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3085
3086 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3087 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3088 set_result(andl);
3089 return true;
3090 }
3091
3092 bool LibraryCallKit::inline_native_threadID() {
3093 Node* tls_ptr = NULL;
3094 Node* cur_thr = generate_current_thread(tls_ptr);
3095 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3096 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3097 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3098
3099 Node* threadid = NULL;
3100 size_t thread_id_size = OSThread::thread_id_size();
3101 if (thread_id_size == (size_t) BytesPerLong) {
3102 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3103 } else if (thread_id_size == (size_t) BytesPerInt) {
3104 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3105 } else {
3106 ShouldNotReachHere();
3107 }
3108 set_result(threadid);
3109 return true;
3110 }
3111 #endif
3112
3113 //------------------------inline_native_time_funcs--------------
3114 // inline code for System.currentTimeMillis() and System.nanoTime()
3115 // these have the same type and signature
3116 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3117 const TypeFunc* tf = OptoRuntime::void_long_Type();
3118 const TypePtr* no_memory_effects = NULL;
3119 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3120 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3121 #ifdef ASSERT
3122 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3123 assert(value_top == top(), "second value must be top");
3124 #endif
3125 set_result(value);
3126 return true;
3127 }
3128
3129 //------------------------inline_native_currentThread------------------
3130 bool LibraryCallKit::inline_native_currentThread() {
3131 Node* junk = NULL;
3132 set_result(generate_current_thread(junk));
3133 return true;
3134 }
3135
3136 //------------------------inline_native_isInterrupted------------------
3137 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3138 bool LibraryCallKit::inline_native_isInterrupted() {
3139 // Add a fast path to t.isInterrupted(clear_int):
3140 // (t == Thread.current() &&
3141 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3142 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3143 // So, in the common case that the interrupt bit is false,
3144 // we avoid making a call into the VM. Even if the interrupt bit
3145 // is true, if the clear_int argument is false, we avoid the VM call.
3146 // However, if the receiver is not currentThread, we must call the VM,
3147 // because there must be some locking done around the operation.
3148
3149 // We only go to the fast case code if we pass two guards.
3150 // Paths which do not pass are accumulated in the slow_region.
3151
3152 enum {
3153 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3154 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3155 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3156 PATH_LIMIT
3157 };
3158
3159 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3160 // out of the function.
3161 insert_mem_bar(Op_MemBarCPUOrder);
3162
3163 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3164 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3165
3166 RegionNode* slow_region = new RegionNode(1);
3167 record_for_igvn(slow_region);
3168
3169 // (a) Receiving thread must be the current thread.
3170 Node* rec_thr = argument(0);
3171 Node* tls_ptr = NULL;
3172 Node* cur_thr = generate_current_thread(tls_ptr);
3173 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3174 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3175
3176 generate_slow_guard(bol_thr, slow_region);
3177
3178 // (b) Interrupt bit on TLS must be false.
3179 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3180 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3181 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3182
3183 // Set the control input on the field _interrupted read to prevent it floating up.
3184 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3185 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3186 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3187
3188 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3189
3190 // First fast path: if (!TLS._interrupted) return false;
3191 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3192 result_rgn->init_req(no_int_result_path, false_bit);
3193 result_val->init_req(no_int_result_path, intcon(0));
3194
3195 // drop through to next case
3196 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3197
3198 #ifndef TARGET_OS_FAMILY_windows
3199 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3200 Node* clr_arg = argument(1);
3201 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3202 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3203 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3204
3205 // Second fast path: ... else if (!clear_int) return true;
3206 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3207 result_rgn->init_req(no_clear_result_path, false_arg);
3208 result_val->init_req(no_clear_result_path, intcon(1));
3209
3210 // drop through to next case
3211 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3212 #else
3213 // To return true on Windows you must read the _interrupted field
3214 // and check the the event state i.e. take the slow path.
3215 #endif // TARGET_OS_FAMILY_windows
3216
3217 // (d) Otherwise, go to the slow path.
3218 slow_region->add_req(control());
3219 set_control( _gvn.transform(slow_region));
3220
3221 if (stopped()) {
3222 // There is no slow path.
3223 result_rgn->init_req(slow_result_path, top());
3224 result_val->init_req(slow_result_path, top());
3225 } else {
3226 // non-virtual because it is a private non-static
3227 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3228
3229 Node* slow_val = set_results_for_java_call(slow_call);
3230 // this->control() comes from set_results_for_java_call
3231
3232 Node* fast_io = slow_call->in(TypeFunc::I_O);
3233 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3234
3235 // These two phis are pre-filled with copies of of the fast IO and Memory
3236 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3237 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3238
3239 result_rgn->init_req(slow_result_path, control());
3240 result_io ->init_req(slow_result_path, i_o());
3241 result_mem->init_req(slow_result_path, reset_memory());
3242 result_val->init_req(slow_result_path, slow_val);
3243
3244 set_all_memory(_gvn.transform(result_mem));
3245 set_i_o( _gvn.transform(result_io));
3246 }
3247
3248 C->set_has_split_ifs(true); // Has chance for split-if optimization
3249 set_result(result_rgn, result_val);
3250 return true;
3251 }
3252
3253 //---------------------------load_mirror_from_klass----------------------------
3254 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3255 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3256 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3257 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3258 }
3259
3260 //-----------------------load_klass_from_mirror_common-------------------------
3261 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3262 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3263 // and branch to the given path on the region.
3264 // If never_see_null, take an uncommon trap on null, so we can optimistically
3265 // compile for the non-null case.
3266 // If the region is NULL, force never_see_null = true.
3267 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3268 bool never_see_null,
3269 RegionNode* region,
3270 int null_path,
3271 int offset) {
3272 if (region == NULL) never_see_null = true;
3273 Node* p = basic_plus_adr(mirror, offset);
3274 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3275 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3276 Node* null_ctl = top();
3277 kls = null_check_oop(kls, &null_ctl, never_see_null);
3278 if (region != NULL) {
3279 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3280 region->init_req(null_path, null_ctl);
3281 } else {
3282 assert(null_ctl == top(), "no loose ends");
3283 }
3284 return kls;
3285 }
3286
3287 //--------------------(inline_native_Class_query helpers)---------------------
3288 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3289 // Fall through if (mods & mask) == bits, take the guard otherwise.
3290 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3291 // Branch around if the given klass has the given modifier bit set.
3292 // Like generate_guard, adds a new path onto the region.
3293 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3294 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3295 Node* mask = intcon(modifier_mask);
3296 Node* bits = intcon(modifier_bits);
3297 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3298 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3299 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3300 return generate_fair_guard(bol, region);
3301 }
3302 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3303 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3304 }
3305
3306 //-------------------------inline_native_Class_query-------------------
3307 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3308 const Type* return_type = TypeInt::BOOL;
3309 Node* prim_return_value = top(); // what happens if it's a primitive class?
3310 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3311 bool expect_prim = false; // most of these guys expect to work on refs
3312
3313 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3314
3315 Node* mirror = argument(0);
3316 Node* obj = top();
3317
3318 switch (id) {
3319 case vmIntrinsics::_isInstance:
3320 // nothing is an instance of a primitive type
3321 prim_return_value = intcon(0);
3322 obj = argument(1);
3323 break;
3324 case vmIntrinsics::_getModifiers:
3325 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3326 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3327 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3328 break;
3329 case vmIntrinsics::_isInterface:
3330 prim_return_value = intcon(0);
3331 break;
3332 case vmIntrinsics::_isArray:
3333 prim_return_value = intcon(0);
3334 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3335 break;
3336 case vmIntrinsics::_isPrimitive:
3337 prim_return_value = intcon(1);
3338 expect_prim = true; // obviously
3339 break;
3340 case vmIntrinsics::_getSuperclass:
3341 prim_return_value = null();
3342 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3343 break;
3344 case vmIntrinsics::_getClassAccessFlags:
3345 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3346 return_type = TypeInt::INT; // not bool! 6297094
3347 break;
3348 default:
3349 fatal_unexpected_iid(id);
3350 break;
3351 }
3352
3353 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3354 if (mirror_con == NULL) return false; // cannot happen?
3355
3356 #ifndef PRODUCT
3357 if (C->print_intrinsics() || C->print_inlining()) {
3358 ciType* k = mirror_con->java_mirror_type();
3359 if (k) {
3360 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3361 k->print_name();
3362 tty->cr();
3363 }
3364 }
3365 #endif
3366
3367 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3368 RegionNode* region = new RegionNode(PATH_LIMIT);
3369 record_for_igvn(region);
3370 PhiNode* phi = new PhiNode(region, return_type);
3371
3372 // The mirror will never be null of Reflection.getClassAccessFlags, however
3373 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3374 // if it is. See bug 4774291.
3375
3376 // For Reflection.getClassAccessFlags(), the null check occurs in
3377 // the wrong place; see inline_unsafe_access(), above, for a similar
3378 // situation.
3379 mirror = null_check(mirror);
3380 // If mirror or obj is dead, only null-path is taken.
3381 if (stopped()) return true;
3382
3383 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3384
3385 // Now load the mirror's klass metaobject, and null-check it.
3386 // Side-effects region with the control path if the klass is null.
3387 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3388 // If kls is null, we have a primitive mirror.
3389 phi->init_req(_prim_path, prim_return_value);
3390 if (stopped()) { set_result(region, phi); return true; }
3391 bool safe_for_replace = (region->in(_prim_path) == top());
3392
3393 Node* p; // handy temp
3394 Node* null_ctl;
3395
3396 // Now that we have the non-null klass, we can perform the real query.
3397 // For constant classes, the query will constant-fold in LoadNode::Value.
3398 Node* query_value = top();
3399 switch (id) {
3400 case vmIntrinsics::_isInstance:
3401 // nothing is an instance of a primitive type
3402 query_value = gen_instanceof(obj, kls, safe_for_replace);
3403 break;
3404
3405 case vmIntrinsics::_getModifiers:
3406 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3407 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3408 break;
3409
3410 case vmIntrinsics::_isInterface:
3411 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3412 if (generate_interface_guard(kls, region) != NULL)
3413 // A guard was added. If the guard is taken, it was an interface.
3414 phi->add_req(intcon(1));
3415 // If we fall through, it's a plain class.
3416 query_value = intcon(0);
3417 break;
3418
3419 case vmIntrinsics::_isArray:
3420 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3421 if (generate_array_guard(kls, region) != NULL)
3422 // A guard was added. If the guard is taken, it was an array.
3423 phi->add_req(intcon(1));
3424 // If we fall through, it's a plain class.
3425 query_value = intcon(0);
3426 break;
3427
3428 case vmIntrinsics::_isPrimitive:
3429 query_value = intcon(0); // "normal" path produces false
3430 break;
3431
3432 case vmIntrinsics::_getSuperclass:
3433 // The rules here are somewhat unfortunate, but we can still do better
3434 // with random logic than with a JNI call.
3435 // Interfaces store null or Object as _super, but must report null.
3436 // Arrays store an intermediate super as _super, but must report Object.
3437 // Other types can report the actual _super.
3438 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3439 if (generate_interface_guard(kls, region) != NULL)
3440 // A guard was added. If the guard is taken, it was an interface.
3441 phi->add_req(null());
3442 if (generate_array_guard(kls, region) != NULL)
3443 // A guard was added. If the guard is taken, it was an array.
3444 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3445 // If we fall through, it's a plain class. Get its _super.
3446 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3447 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3448 null_ctl = top();
3449 kls = null_check_oop(kls, &null_ctl);
3450 if (null_ctl != top()) {
3451 // If the guard is taken, Object.superClass is null (both klass and mirror).
3452 region->add_req(null_ctl);
3453 phi ->add_req(null());
3454 }
3455 if (!stopped()) {
3456 query_value = load_mirror_from_klass(kls);
3457 }
3458 break;
3459
3460 case vmIntrinsics::_getClassAccessFlags:
3461 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3462 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3463 break;
3464
3465 default:
3466 fatal_unexpected_iid(id);
3467 break;
3468 }
3469
3470 // Fall-through is the normal case of a query to a real class.
3471 phi->init_req(1, query_value);
3472 region->init_req(1, control());
3473
3474 C->set_has_split_ifs(true); // Has chance for split-if optimization
3475 set_result(region, phi);
3476 return true;
3477 }
3478
3479 //-------------------------inline_Class_cast-------------------
3480 bool LibraryCallKit::inline_Class_cast() {
3481 Node* mirror = argument(0); // Class
3482 Node* obj = argument(1);
3483 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3484 if (mirror_con == NULL) {
3485 return false; // dead path (mirror->is_top()).
3486 }
3487 if (obj == NULL || obj->is_top()) {
3488 return false; // dead path
3489 }
3490 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3491
3492 // First, see if Class.cast() can be folded statically.
3493 // java_mirror_type() returns non-null for compile-time Class constants.
3494 ciType* tm = mirror_con->java_mirror_type();
3495 if (tm != NULL && tm->is_klass() &&
3496 tp != NULL && tp->klass() != NULL) {
3497 if (!tp->klass()->is_loaded()) {
3498 // Don't use intrinsic when class is not loaded.
3499 return false;
3500 } else {
3501 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3502 if (static_res == Compile::SSC_always_true) {
3503 // isInstance() is true - fold the code.
3504 set_result(obj);
3505 return true;
3506 } else if (static_res == Compile::SSC_always_false) {
3507 // Don't use intrinsic, have to throw ClassCastException.
3508 // If the reference is null, the non-intrinsic bytecode will
3509 // be optimized appropriately.
3510 return false;
3511 }
3512 }
3513 }
3514
3515 // Bailout intrinsic and do normal inlining if exception path is frequent.
3516 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3517 return false;
3518 }
3519
3520 // Generate dynamic checks.
3521 // Class.cast() is java implementation of _checkcast bytecode.
3522 // Do checkcast (Parse::do_checkcast()) optimizations here.
3523
3524 mirror = null_check(mirror);
3525 // If mirror is dead, only null-path is taken.
3526 if (stopped()) {
3527 return true;
3528 }
3529
3530 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3531 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3532 RegionNode* region = new RegionNode(PATH_LIMIT);
3533 record_for_igvn(region);
3534
3535 // Now load the mirror's klass metaobject, and null-check it.
3536 // If kls is null, we have a primitive mirror and
3537 // nothing is an instance of a primitive type.
3538 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3539
3540 Node* res = top();
3541 if (!stopped()) {
3542 Node* bad_type_ctrl = top();
3543 // Do checkcast optimizations.
3544 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3545 region->init_req(_bad_type_path, bad_type_ctrl);
3546 }
3547 if (region->in(_prim_path) != top() ||
3548 region->in(_bad_type_path) != top()) {
3549 // Let Interpreter throw ClassCastException.
3550 PreserveJVMState pjvms(this);
3551 set_control(_gvn.transform(region));
3552 uncommon_trap(Deoptimization::Reason_intrinsic,
3553 Deoptimization::Action_maybe_recompile);
3554 }
3555 if (!stopped()) {
3556 set_result(res);
3557 }
3558 return true;
3559 }
3560
3561
3562 //--------------------------inline_native_subtype_check------------------------
3563 // This intrinsic takes the JNI calls out of the heart of
3564 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3565 bool LibraryCallKit::inline_native_subtype_check() {
3566 // Pull both arguments off the stack.
3567 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3568 args[0] = argument(0);
3569 args[1] = argument(1);
3570 Node* klasses[2]; // corresponding Klasses: superk, subk
3571 klasses[0] = klasses[1] = top();
3572
3573 enum {
3574 // A full decision tree on {superc is prim, subc is prim}:
3575 _prim_0_path = 1, // {P,N} => false
3576 // {P,P} & superc!=subc => false
3577 _prim_same_path, // {P,P} & superc==subc => true
3578 _prim_1_path, // {N,P} => false
3579 _ref_subtype_path, // {N,N} & subtype check wins => true
3580 _both_ref_path, // {N,N} & subtype check loses => false
3581 PATH_LIMIT
3582 };
3583
3584 RegionNode* region = new RegionNode(PATH_LIMIT);
3585 Node* phi = new PhiNode(region, TypeInt::BOOL);
3586 record_for_igvn(region);
3587
3588 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3589 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3590 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3591
3592 // First null-check both mirrors and load each mirror's klass metaobject.
3593 int which_arg;
3594 for (which_arg = 0; which_arg <= 1; which_arg++) {
3595 Node* arg = args[which_arg];
3596 arg = null_check(arg);
3597 if (stopped()) break;
3598 args[which_arg] = arg;
3599
3600 Node* p = basic_plus_adr(arg, class_klass_offset);
3601 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3602 klasses[which_arg] = _gvn.transform(kls);
3603 }
3604
3605 // Having loaded both klasses, test each for null.
3606 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3607 for (which_arg = 0; which_arg <= 1; which_arg++) {
3608 Node* kls = klasses[which_arg];
3609 Node* null_ctl = top();
3610 kls = null_check_oop(kls, &null_ctl, never_see_null);
3611 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3612 region->init_req(prim_path, null_ctl);
3613 if (stopped()) break;
3614 klasses[which_arg] = kls;
3615 }
3616
3617 if (!stopped()) {
3618 // now we have two reference types, in klasses[0..1]
3619 Node* subk = klasses[1]; // the argument to isAssignableFrom
3620 Node* superk = klasses[0]; // the receiver
3621 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3622 // now we have a successful reference subtype check
3623 region->set_req(_ref_subtype_path, control());
3624 }
3625
3626 // If both operands are primitive (both klasses null), then
3627 // we must return true when they are identical primitives.
3628 // It is convenient to test this after the first null klass check.
3629 set_control(region->in(_prim_0_path)); // go back to first null check
3630 if (!stopped()) {
3631 // Since superc is primitive, make a guard for the superc==subc case.
3632 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3633 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3634 generate_guard(bol_eq, region, PROB_FAIR);
3635 if (region->req() == PATH_LIMIT+1) {
3636 // A guard was added. If the added guard is taken, superc==subc.
3637 region->swap_edges(PATH_LIMIT, _prim_same_path);
3638 region->del_req(PATH_LIMIT);
3639 }
3640 region->set_req(_prim_0_path, control()); // Not equal after all.
3641 }
3642
3643 // these are the only paths that produce 'true':
3644 phi->set_req(_prim_same_path, intcon(1));
3645 phi->set_req(_ref_subtype_path, intcon(1));
3646
3647 // pull together the cases:
3648 assert(region->req() == PATH_LIMIT, "sane region");
3649 for (uint i = 1; i < region->req(); i++) {
3650 Node* ctl = region->in(i);
3651 if (ctl == NULL || ctl == top()) {
3652 region->set_req(i, top());
3653 phi ->set_req(i, top());
3654 } else if (phi->in(i) == NULL) {
3655 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3656 }
3657 }
3658
3659 set_control(_gvn.transform(region));
3660 set_result(_gvn.transform(phi));
3661 return true;
3662 }
3663
3664 //---------------------generate_array_guard_common------------------------
3665 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3666 bool obj_array, bool not_array) {
3667 // If obj_array/non_array==false/false:
3668 // Branch around if the given klass is in fact an array (either obj or prim).
3669 // If obj_array/non_array==false/true:
3670 // Branch around if the given klass is not an array klass of any kind.
3671 // If obj_array/non_array==true/true:
3672 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3673 // If obj_array/non_array==true/false:
3674 // Branch around if the kls is an oop array (Object[] or subtype)
3675 //
3676 // Like generate_guard, adds a new path onto the region.
3677 jint layout_con = 0;
3678 Node* layout_val = get_layout_helper(kls, layout_con);
3679 if (layout_val == NULL) {
3680 bool query = (obj_array
3681 ? Klass::layout_helper_is_objArray(layout_con)
3682 : Klass::layout_helper_is_array(layout_con));
3683 if (query == not_array) {
3684 return NULL; // never a branch
3685 } else { // always a branch
3686 Node* always_branch = control();
3687 if (region != NULL)
3688 region->add_req(always_branch);
3689 set_control(top());
3690 return always_branch;
3691 }
3692 }
3693 // Now test the correct condition.
3694 jint nval = (obj_array
3695 ? ((jint)Klass::_lh_array_tag_type_value
3696 << Klass::_lh_array_tag_shift)
3697 : Klass::_lh_neutral_value);
3698 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3699 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3700 // invert the test if we are looking for a non-array
3701 if (not_array) btest = BoolTest(btest).negate();
3702 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3703 return generate_fair_guard(bol, region);
3704 }
3705
3706
3707 //-----------------------inline_native_newArray--------------------------
3708 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3709 bool LibraryCallKit::inline_native_newArray() {
3710 Node* mirror = argument(0);
3711 Node* count_val = argument(1);
3712
3713 mirror = null_check(mirror);
3714 // If mirror or obj is dead, only null-path is taken.
3715 if (stopped()) return true;
3716
3717 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3718 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3719 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3720 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3721 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3722
3723 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3724 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3725 result_reg, _slow_path);
3726 Node* normal_ctl = control();
3727 Node* no_array_ctl = result_reg->in(_slow_path);
3728
3729 // Generate code for the slow case. We make a call to newArray().
3730 set_control(no_array_ctl);
3731 if (!stopped()) {
3732 // Either the input type is void.class, or else the
3733 // array klass has not yet been cached. Either the
3734 // ensuing call will throw an exception, or else it
3735 // will cache the array klass for next time.
3736 PreserveJVMState pjvms(this);
3737 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3738 Node* slow_result = set_results_for_java_call(slow_call);
3739 // this->control() comes from set_results_for_java_call
3740 result_reg->set_req(_slow_path, control());
3741 result_val->set_req(_slow_path, slow_result);
3742 result_io ->set_req(_slow_path, i_o());
3743 result_mem->set_req(_slow_path, reset_memory());
3744 }
3745
3746 set_control(normal_ctl);
3747 if (!stopped()) {
3748 // Normal case: The array type has been cached in the java.lang.Class.
3749 // The following call works fine even if the array type is polymorphic.
3750 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3751 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3752 result_reg->init_req(_normal_path, control());
3753 result_val->init_req(_normal_path, obj);
3754 result_io ->init_req(_normal_path, i_o());
3755 result_mem->init_req(_normal_path, reset_memory());
3756 }
3757
3758 // Return the combined state.
3759 set_i_o( _gvn.transform(result_io) );
3760 set_all_memory( _gvn.transform(result_mem));
3761
3762 C->set_has_split_ifs(true); // Has chance for split-if optimization
3763 set_result(result_reg, result_val);
3764 return true;
3765 }
3766
3767 //----------------------inline_native_getLength--------------------------
3768 // public static native int java.lang.reflect.Array.getLength(Object array);
3769 bool LibraryCallKit::inline_native_getLength() {
3770 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3771
3772 Node* array = null_check(argument(0));
3773 // If array is dead, only null-path is taken.
3774 if (stopped()) return true;
3775
3776 // Deoptimize if it is a non-array.
3777 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3778
3779 if (non_array != NULL) {
3780 PreserveJVMState pjvms(this);
3781 set_control(non_array);
3782 uncommon_trap(Deoptimization::Reason_intrinsic,
3783 Deoptimization::Action_maybe_recompile);
3784 }
3785
3786 // If control is dead, only non-array-path is taken.
3787 if (stopped()) return true;
3788
3789 // The works fine even if the array type is polymorphic.
3790 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3791 Node* result = load_array_length(array);
3792
3793 C->set_has_split_ifs(true); // Has chance for split-if optimization
3794 set_result(result);
3795 return true;
3796 }
3797
3798 //------------------------inline_array_copyOf----------------------------
3799 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3800 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3801 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3802 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3803
3804 // Get the arguments.
3805 Node* original = argument(0);
3806 Node* start = is_copyOfRange? argument(1): intcon(0);
3807 Node* end = is_copyOfRange? argument(2): argument(1);
3808 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3809
3810 Node* newcopy;
3811
3812 // Set the original stack and the reexecute bit for the interpreter to reexecute
3813 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3814 { PreserveReexecuteState preexecs(this);
3815 jvms()->set_should_reexecute(true);
3816
3817 array_type_mirror = null_check(array_type_mirror);
3818 original = null_check(original);
3819
3820 // Check if a null path was taken unconditionally.
3821 if (stopped()) return true;
3822
3823 Node* orig_length = load_array_length(original);
3824
3825 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3826 klass_node = null_check(klass_node);
3827
3828 RegionNode* bailout = new RegionNode(1);
3829 record_for_igvn(bailout);
3830
3831 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3832 // Bail out if that is so.
3833 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3834 if (not_objArray != NULL) {
3835 // Improve the klass node's type from the new optimistic assumption:
3836 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3837 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3838 Node* cast = new CastPPNode(klass_node, akls);
3839 cast->init_req(0, control());
3840 klass_node = _gvn.transform(cast);
3841 }
3842
3843 // Bail out if either start or end is negative.
3844 generate_negative_guard(start, bailout, &start);
3845 generate_negative_guard(end, bailout, &end);
3846
3847 Node* length = end;
3848 if (_gvn.type(start) != TypeInt::ZERO) {
3849 length = _gvn.transform(new SubINode(end, start));
3850 }
3851
3852 // Bail out if length is negative.
3853 // Without this the new_array would throw
3854 // NegativeArraySizeException but IllegalArgumentException is what
3855 // should be thrown
3856 generate_negative_guard(length, bailout, &length);
3857
3858 if (bailout->req() > 1) {
3859 PreserveJVMState pjvms(this);
3860 set_control(_gvn.transform(bailout));
3861 uncommon_trap(Deoptimization::Reason_intrinsic,
3862 Deoptimization::Action_maybe_recompile);
3863 }
3864
3865 if (!stopped()) {
3866 // How many elements will we copy from the original?
3867 // The answer is MinI(orig_length - start, length).
3868 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3869 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3870
3871 newcopy = new_array(klass_node, length, 0); // no arguments to push
3872
3873 // Generate a direct call to the right arraycopy function(s).
3874 // We know the copy is disjoint but we might not know if the
3875 // oop stores need checking.
3876 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3877 // This will fail a store-check if x contains any non-nulls.
3878
3879 Node* alloc = tightly_coupled_allocation(newcopy, NULL);
3880
3881 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL,
3882 load_object_klass(original), klass_node);
3883 if (!is_copyOfRange) {
3884 ac->set_copyof();
3885 } else {
3886 ac->set_copyofrange();
3887 }
3888 Node* n = _gvn.transform(ac);
3889 assert(n == ac, "cannot disappear");
3890 ac->connect_outputs(this);
3891 }
3892 } // original reexecute is set back here
3893
3894 C->set_has_split_ifs(true); // Has chance for split-if optimization
3895 if (!stopped()) {
3896 set_result(newcopy);
3897 }
3898 return true;
3899 }
3900
3901
3902 //----------------------generate_virtual_guard---------------------------
3903 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3904 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3905 RegionNode* slow_region) {
3906 ciMethod* method = callee();
3907 int vtable_index = method->vtable_index();
3908 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3909 err_msg_res("bad index %d", vtable_index));
3910 // Get the Method* out of the appropriate vtable entry.
3911 int entry_offset = (InstanceKlass::vtable_start_offset() +
3912 vtable_index*vtableEntry::size()) * wordSize +
3913 vtableEntry::method_offset_in_bytes();
3914 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3915 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3916
3917 // Compare the target method with the expected method (e.g., Object.hashCode).
3918 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3919
3920 Node* native_call = makecon(native_call_addr);
3921 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3922 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3923
3924 return generate_slow_guard(test_native, slow_region);
3925 }
3926
3927 //-----------------------generate_method_call----------------------------
3928 // Use generate_method_call to make a slow-call to the real
3929 // method if the fast path fails. An alternative would be to
3930 // use a stub like OptoRuntime::slow_arraycopy_Java.
3931 // This only works for expanding the current library call,
3932 // not another intrinsic. (E.g., don't use this for making an
3933 // arraycopy call inside of the copyOf intrinsic.)
3934 CallJavaNode*
3935 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3936 // When compiling the intrinsic method itself, do not use this technique.
3937 guarantee(callee() != C->method(), "cannot make slow-call to self");
3938
3939 ciMethod* method = callee();
3940 // ensure the JVMS we have will be correct for this call
3941 guarantee(method_id == method->intrinsic_id(), "must match");
3942
3943 const TypeFunc* tf = TypeFunc::make(method);
3944 CallJavaNode* slow_call;
3945 if (is_static) {
3946 assert(!is_virtual, "");
3947 slow_call = new CallStaticJavaNode(C, tf,
3948 SharedRuntime::get_resolve_static_call_stub(),
3949 method, bci());
3950 } else if (is_virtual) {
3951 null_check_receiver();
3952 int vtable_index = Method::invalid_vtable_index;
3953 if (UseInlineCaches) {
3954 // Suppress the vtable call
3955 } else {
3956 // hashCode and clone are not a miranda methods,
3957 // so the vtable index is fixed.
3958 // No need to use the linkResolver to get it.
3959 vtable_index = method->vtable_index();
3960 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3961 err_msg_res("bad index %d", vtable_index));
3962 }
3963 slow_call = new CallDynamicJavaNode(tf,
3964 SharedRuntime::get_resolve_virtual_call_stub(),
3965 method, vtable_index, bci());
3966 } else { // neither virtual nor static: opt_virtual
3967 null_check_receiver();
3968 slow_call = new CallStaticJavaNode(C, tf,
3969 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3970 method, bci());
3971 slow_call->set_optimized_virtual(true);
3972 }
3973 set_arguments_for_java_call(slow_call);
3974 set_edges_for_java_call(slow_call);
3975 return slow_call;
3976 }
3977
3978
3979 /**
3980 * Build special case code for calls to hashCode on an object. This call may
3981 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3982 * slightly different code.
3983 */
3984 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3985 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3986 assert(!(is_virtual && is_static), "either virtual, special, or static");
3987
3988 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3989
3990 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3991 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3992 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3993 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3994 Node* obj = NULL;
3995 if (!is_static) {
3996 // Check for hashing null object
3997 obj = null_check_receiver();
3998 if (stopped()) return true; // unconditionally null
3999 result_reg->init_req(_null_path, top());
4000 result_val->init_req(_null_path, top());
4001 } else {
4002 // Do a null check, and return zero if null.
4003 // System.identityHashCode(null) == 0
4004 obj = argument(0);
4005 Node* null_ctl = top();
4006 obj = null_check_oop(obj, &null_ctl);
4007 result_reg->init_req(_null_path, null_ctl);
4008 result_val->init_req(_null_path, _gvn.intcon(0));
4009 }
4010
4011 // Unconditionally null? Then return right away.
4012 if (stopped()) {
4013 set_control( result_reg->in(_null_path));
4014 if (!stopped())
4015 set_result(result_val->in(_null_path));
4016 return true;
4017 }
4018
4019 // We only go to the fast case code if we pass a number of guards. The
4020 // paths which do not pass are accumulated in the slow_region.
4021 RegionNode* slow_region = new RegionNode(1);
4022 record_for_igvn(slow_region);
4023
4024 // If this is a virtual call, we generate a funny guard. We pull out
4025 // the vtable entry corresponding to hashCode() from the target object.
4026 // If the target method which we are calling happens to be the native
4027 // Object hashCode() method, we pass the guard. We do not need this
4028 // guard for non-virtual calls -- the caller is known to be the native
4029 // Object hashCode().
4030 if (is_virtual) {
4031 // After null check, get the object's klass.
4032 Node* obj_klass = load_object_klass(obj);
4033 generate_virtual_guard(obj_klass, slow_region);
4034 }
4035
4036 // Get the header out of the object, use LoadMarkNode when available
4037 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4038 // The control of the load must be NULL. Otherwise, the load can move before
4039 // the null check after castPP removal.
4040 Node* no_ctrl = NULL;
4041 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4042
4043 // Test the header to see if it is unlocked.
4044 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4045 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4046 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4047 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4048 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4049
4050 generate_slow_guard(test_unlocked, slow_region);
4051
4052 // Get the hash value and check to see that it has been properly assigned.
4053 // We depend on hash_mask being at most 32 bits and avoid the use of
4054 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4055 // vm: see markOop.hpp.
4056 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4057 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4058 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4059 // This hack lets the hash bits live anywhere in the mark object now, as long
4060 // as the shift drops the relevant bits into the low 32 bits. Note that
4061 // Java spec says that HashCode is an int so there's no point in capturing
4062 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4063 hshifted_header = ConvX2I(hshifted_header);
4064 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4065
4066 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4067 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4068 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4069
4070 generate_slow_guard(test_assigned, slow_region);
4071
4072 Node* init_mem = reset_memory();
4073 // fill in the rest of the null path:
4074 result_io ->init_req(_null_path, i_o());
4075 result_mem->init_req(_null_path, init_mem);
4076
4077 result_val->init_req(_fast_path, hash_val);
4078 result_reg->init_req(_fast_path, control());
4079 result_io ->init_req(_fast_path, i_o());
4080 result_mem->init_req(_fast_path, init_mem);
4081
4082 // Generate code for the slow case. We make a call to hashCode().
4083 set_control(_gvn.transform(slow_region));
4084 if (!stopped()) {
4085 // No need for PreserveJVMState, because we're using up the present state.
4086 set_all_memory(init_mem);
4087 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4088 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4089 Node* slow_result = set_results_for_java_call(slow_call);
4090 // this->control() comes from set_results_for_java_call
4091 result_reg->init_req(_slow_path, control());
4092 result_val->init_req(_slow_path, slow_result);
4093 result_io ->set_req(_slow_path, i_o());
4094 result_mem ->set_req(_slow_path, reset_memory());
4095 }
4096
4097 // Return the combined state.
4098 set_i_o( _gvn.transform(result_io) );
4099 set_all_memory( _gvn.transform(result_mem));
4100
4101 set_result(result_reg, result_val);
4102 return true;
4103 }
4104
4105 //---------------------------inline_native_getClass----------------------------
4106 // public final native Class<?> java.lang.Object.getClass();
4107 //
4108 // Build special case code for calls to getClass on an object.
4109 bool LibraryCallKit::inline_native_getClass() {
4110 Node* obj = null_check_receiver();
4111 if (stopped()) return true;
4112 set_result(load_mirror_from_klass(load_object_klass(obj)));
4113 return true;
4114 }
4115
4116 //-----------------inline_native_Reflection_getCallerClass---------------------
4117 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4118 //
4119 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4120 //
4121 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4122 // in that it must skip particular security frames and checks for
4123 // caller sensitive methods.
4124 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4125 #ifndef PRODUCT
4126 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4127 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4128 }
4129 #endif
4130
4131 if (!jvms()->has_method()) {
4132 #ifndef PRODUCT
4133 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4134 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4135 }
4136 #endif
4137 return false;
4138 }
4139
4140 // Walk back up the JVM state to find the caller at the required
4141 // depth.
4142 JVMState* caller_jvms = jvms();
4143
4144 // Cf. JVM_GetCallerClass
4145 // NOTE: Start the loop at depth 1 because the current JVM state does
4146 // not include the Reflection.getCallerClass() frame.
4147 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4148 ciMethod* m = caller_jvms->method();
4149 switch (n) {
4150 case 0:
4151 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4152 break;
4153 case 1:
4154 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4155 if (!m->caller_sensitive()) {
4156 #ifndef PRODUCT
4157 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4158 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4159 }
4160 #endif
4161 return false; // bail-out; let JVM_GetCallerClass do the work
4162 }
4163 break;
4164 default:
4165 if (!m->is_ignored_by_security_stack_walk()) {
4166 // We have reached the desired frame; return the holder class.
4167 // Acquire method holder as java.lang.Class and push as constant.
4168 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4169 ciInstance* caller_mirror = caller_klass->java_mirror();
4170 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4171
4172 #ifndef PRODUCT
4173 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4174 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());
4175 tty->print_cr(" JVM state at this point:");
4176 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4177 ciMethod* m = jvms()->of_depth(i)->method();
4178 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4179 }
4180 }
4181 #endif
4182 return true;
4183 }
4184 break;
4185 }
4186 }
4187
4188 #ifndef PRODUCT
4189 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4190 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4191 tty->print_cr(" JVM state at this point:");
4192 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4193 ciMethod* m = jvms()->of_depth(i)->method();
4194 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4195 }
4196 }
4197 #endif
4198
4199 return false; // bail-out; let JVM_GetCallerClass do the work
4200 }
4201
4202 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4203 Node* arg = argument(0);
4204 Node* result;
4205
4206 switch (id) {
4207 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4208 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4209 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4210 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4211
4212 case vmIntrinsics::_doubleToLongBits: {
4213 // two paths (plus control) merge in a wood
4214 RegionNode *r = new RegionNode(3);
4215 Node *phi = new PhiNode(r, TypeLong::LONG);
4216
4217 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4218 // Build the boolean node
4219 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4220
4221 // Branch either way.
4222 // NaN case is less traveled, which makes all the difference.
4223 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4224 Node *opt_isnan = _gvn.transform(ifisnan);
4225 assert( opt_isnan->is_If(), "Expect an IfNode");
4226 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4227 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4228
4229 set_control(iftrue);
4230
4231 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4232 Node *slow_result = longcon(nan_bits); // return NaN
4233 phi->init_req(1, _gvn.transform( slow_result ));
4234 r->init_req(1, iftrue);
4235
4236 // Else fall through
4237 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4238 set_control(iffalse);
4239
4240 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4241 r->init_req(2, iffalse);
4242
4243 // Post merge
4244 set_control(_gvn.transform(r));
4245 record_for_igvn(r);
4246
4247 C->set_has_split_ifs(true); // Has chance for split-if optimization
4248 result = phi;
4249 assert(result->bottom_type()->isa_long(), "must be");
4250 break;
4251 }
4252
4253 case vmIntrinsics::_floatToIntBits: {
4254 // two paths (plus control) merge in a wood
4255 RegionNode *r = new RegionNode(3);
4256 Node *phi = new PhiNode(r, TypeInt::INT);
4257
4258 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4259 // Build the boolean node
4260 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4261
4262 // Branch either way.
4263 // NaN case is less traveled, which makes all the difference.
4264 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4265 Node *opt_isnan = _gvn.transform(ifisnan);
4266 assert( opt_isnan->is_If(), "Expect an IfNode");
4267 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4268 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4269
4270 set_control(iftrue);
4271
4272 static const jint nan_bits = 0x7fc00000;
4273 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4274 phi->init_req(1, _gvn.transform( slow_result ));
4275 r->init_req(1, iftrue);
4276
4277 // Else fall through
4278 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4279 set_control(iffalse);
4280
4281 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4282 r->init_req(2, iffalse);
4283
4284 // Post merge
4285 set_control(_gvn.transform(r));
4286 record_for_igvn(r);
4287
4288 C->set_has_split_ifs(true); // Has chance for split-if optimization
4289 result = phi;
4290 assert(result->bottom_type()->isa_int(), "must be");
4291 break;
4292 }
4293
4294 default:
4295 fatal_unexpected_iid(id);
4296 break;
4297 }
4298 set_result(_gvn.transform(result));
4299 return true;
4300 }
4301
4302 #ifdef _LP64
4303 #define XTOP ,top() /*additional argument*/
4304 #else //_LP64
4305 #define XTOP /*no additional argument*/
4306 #endif //_LP64
4307
4308 //----------------------inline_unsafe_copyMemory-------------------------
4309 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4310 bool LibraryCallKit::inline_unsafe_copyMemory() {
4311 if (callee()->is_static()) return false; // caller must have the capability!
4312 null_check_receiver(); // null-check receiver
4313 if (stopped()) return true;
4314
4315 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4316
4317 Node* src_ptr = argument(1); // type: oop
4318 Node* src_off = ConvL2X(argument(2)); // type: long
4319 Node* dst_ptr = argument(4); // type: oop
4320 Node* dst_off = ConvL2X(argument(5)); // type: long
4321 Node* size = ConvL2X(argument(7)); // type: long
4322
4323 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4324 "fieldOffset must be byte-scaled");
4325
4326 Node* src = make_unsafe_address(src_ptr, src_off);
4327 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4328
4329 // Conservatively insert a memory barrier on all memory slices.
4330 // Do not let writes of the copy source or destination float below the copy.
4331 insert_mem_bar(Op_MemBarCPUOrder);
4332
4333 // Call it. Note that the length argument is not scaled.
4334 make_runtime_call(RC_LEAF|RC_NO_FP,
4335 OptoRuntime::fast_arraycopy_Type(),
4336 StubRoutines::unsafe_arraycopy(),
4337 "unsafe_arraycopy",
4338 TypeRawPtr::BOTTOM,
4339 src, dst, size XTOP);
4340
4341 // Do not let reads of the copy destination float above the copy.
4342 insert_mem_bar(Op_MemBarCPUOrder);
4343
4344 return true;
4345 }
4346
4347 //------------------------clone_coping-----------------------------------
4348 // Helper function for inline_native_clone.
4349 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4350 assert(obj_size != NULL, "");
4351 Node* raw_obj = alloc_obj->in(1);
4352 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4353
4354 AllocateNode* alloc = NULL;
4355 if (ReduceBulkZeroing) {
4356 // We will be completely responsible for initializing this object -
4357 // mark Initialize node as complete.
4358 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4359 // The object was just allocated - there should be no any stores!
4360 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4361 // Mark as complete_with_arraycopy so that on AllocateNode
4362 // expansion, we know this AllocateNode is initialized by an array
4363 // copy and a StoreStore barrier exists after the array copy.
4364 alloc->initialization()->set_complete_with_arraycopy();
4365 }
4366
4367 // Copy the fastest available way.
4368 // TODO: generate fields copies for small objects instead.
4369 Node* src = obj;
4370 Node* dest = alloc_obj;
4371 Node* size = _gvn.transform(obj_size);
4372
4373 // Exclude the header but include array length to copy by 8 bytes words.
4374 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4375 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4376 instanceOopDesc::base_offset_in_bytes();
4377 // base_off:
4378 // 8 - 32-bit VM
4379 // 12 - 64-bit VM, compressed klass
4380 // 16 - 64-bit VM, normal klass
4381 if (base_off % BytesPerLong != 0) {
4382 assert(UseCompressedClassPointers, "");
4383 if (is_array) {
4384 // Exclude length to copy by 8 bytes words.
4385 base_off += sizeof(int);
4386 } else {
4387 // Include klass to copy by 8 bytes words.
4388 base_off = instanceOopDesc::klass_offset_in_bytes();
4389 }
4390 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4391 }
4392 src = basic_plus_adr(src, base_off);
4393 dest = basic_plus_adr(dest, base_off);
4394
4395 // Compute the length also, if needed:
4396 Node* countx = size;
4397 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4398 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4399
4400 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4401
4402 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4403 ac->set_clonebasic();
4404 Node* n = _gvn.transform(ac);
4405 if (n == ac) {
4406 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4407 } else {
4408 set_all_memory(n);
4409 }
4410
4411 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4412 if (card_mark) {
4413 assert(!is_array, "");
4414 // Put in store barrier for any and all oops we are sticking
4415 // into this object. (We could avoid this if we could prove
4416 // that the object type contains no oop fields at all.)
4417 Node* no_particular_value = NULL;
4418 Node* no_particular_field = NULL;
4419 int raw_adr_idx = Compile::AliasIdxRaw;
4420 post_barrier(control(),
4421 memory(raw_adr_type),
4422 alloc_obj,
4423 no_particular_field,
4424 raw_adr_idx,
4425 no_particular_value,
4426 T_OBJECT,
4427 false);
4428 }
4429
4430 // Do not let reads from the cloned object float above the arraycopy.
4431 if (alloc != NULL) {
4432 // Do not let stores that initialize this object be reordered with
4433 // a subsequent store that would make this object accessible by
4434 // other threads.
4435 // Record what AllocateNode this StoreStore protects so that
4436 // escape analysis can go from the MemBarStoreStoreNode to the
4437 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4438 // based on the escape status of the AllocateNode.
4439 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4440 } else {
4441 insert_mem_bar(Op_MemBarCPUOrder);
4442 }
4443 }
4444
4445 //------------------------inline_native_clone----------------------------
4446 // protected native Object java.lang.Object.clone();
4447 //
4448 // Here are the simple edge cases:
4449 // null receiver => normal trap
4450 // virtual and clone was overridden => slow path to out-of-line clone
4451 // not cloneable or finalizer => slow path to out-of-line Object.clone
4452 //
4453 // The general case has two steps, allocation and copying.
4454 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4455 //
4456 // Copying also has two cases, oop arrays and everything else.
4457 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4458 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4459 //
4460 // These steps fold up nicely if and when the cloned object's klass
4461 // can be sharply typed as an object array, a type array, or an instance.
4462 //
4463 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4464 PhiNode* result_val;
4465
4466 // Set the reexecute bit for the interpreter to reexecute
4467 // the bytecode that invokes Object.clone if deoptimization happens.
4468 { PreserveReexecuteState preexecs(this);
4469 jvms()->set_should_reexecute(true);
4470
4471 Node* obj = null_check_receiver();
4472 if (stopped()) return true;
4473
4474 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4475
4476 // If we are going to clone an instance, we need its exact type to
4477 // know the number and types of fields to convert the clone to
4478 // loads/stores. Maybe a speculative type can help us.
4479 if (!obj_type->klass_is_exact() &&
4480 obj_type->speculative_type() != NULL &&
4481 obj_type->speculative_type()->is_instance_klass()) {
4482 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4483 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4484 !spec_ik->has_injected_fields()) {
4485 ciKlass* k = obj_type->klass();
4486 if (!k->is_instance_klass() ||
4487 k->as_instance_klass()->is_interface() ||
4488 k->as_instance_klass()->has_subklass()) {
4489 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4490 }
4491 }
4492 }
4493
4494 Node* obj_klass = load_object_klass(obj);
4495 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4496 const TypeOopPtr* toop = ((tklass != NULL)
4497 ? tklass->as_instance_type()
4498 : TypeInstPtr::NOTNULL);
4499
4500 // Conservatively insert a memory barrier on all memory slices.
4501 // Do not let writes into the original float below the clone.
4502 insert_mem_bar(Op_MemBarCPUOrder);
4503
4504 // paths into result_reg:
4505 enum {
4506 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4507 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4508 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4509 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4510 PATH_LIMIT
4511 };
4512 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4513 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4514 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4515 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4516 record_for_igvn(result_reg);
4517
4518 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4519 int raw_adr_idx = Compile::AliasIdxRaw;
4520
4521 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4522 if (array_ctl != NULL) {
4523 // It's an array.
4524 PreserveJVMState pjvms(this);
4525 set_control(array_ctl);
4526 Node* obj_length = load_array_length(obj);
4527 Node* obj_size = NULL;
4528 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4529
4530 if (!use_ReduceInitialCardMarks()) {
4531 // If it is an oop array, it requires very special treatment,
4532 // because card marking is required on each card of the array.
4533 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4534 if (is_obja != NULL) {
4535 PreserveJVMState pjvms2(this);
4536 set_control(is_obja);
4537 // Generate a direct call to the right arraycopy function(s).
4538 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4539 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4540 ac->set_cloneoop();
4541 Node* n = _gvn.transform(ac);
4542 assert(n == ac, "cannot disappear");
4543 ac->connect_outputs(this);
4544
4545 result_reg->init_req(_objArray_path, control());
4546 result_val->init_req(_objArray_path, alloc_obj);
4547 result_i_o ->set_req(_objArray_path, i_o());
4548 result_mem ->set_req(_objArray_path, reset_memory());
4549 }
4550 }
4551 // Otherwise, there are no card marks to worry about.
4552 // (We can dispense with card marks if we know the allocation
4553 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4554 // causes the non-eden paths to take compensating steps to
4555 // simulate a fresh allocation, so that no further
4556 // card marks are required in compiled code to initialize
4557 // the object.)
4558
4559 if (!stopped()) {
4560 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4561
4562 // Present the results of the copy.
4563 result_reg->init_req(_array_path, control());
4564 result_val->init_req(_array_path, alloc_obj);
4565 result_i_o ->set_req(_array_path, i_o());
4566 result_mem ->set_req(_array_path, reset_memory());
4567 }
4568 }
4569
4570 // We only go to the instance fast case code if we pass a number of guards.
4571 // The paths which do not pass are accumulated in the slow_region.
4572 RegionNode* slow_region = new RegionNode(1);
4573 record_for_igvn(slow_region);
4574 if (!stopped()) {
4575 // It's an instance (we did array above). Make the slow-path tests.
4576 // If this is a virtual call, we generate a funny guard. We grab
4577 // the vtable entry corresponding to clone() from the target object.
4578 // If the target method which we are calling happens to be the
4579 // Object clone() method, we pass the guard. We do not need this
4580 // guard for non-virtual calls; the caller is known to be the native
4581 // Object clone().
4582 if (is_virtual) {
4583 generate_virtual_guard(obj_klass, slow_region);
4584 }
4585
4586 // The object must be cloneable and must not have a finalizer.
4587 // Both of these conditions may be checked in a single test.
4588 // We could optimize the cloneable test further, but we don't care.
4589 generate_access_flags_guard(obj_klass,
4590 // Test both conditions:
4591 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4592 // Must be cloneable but not finalizer:
4593 JVM_ACC_IS_CLONEABLE,
4594 slow_region);
4595 }
4596
4597 if (!stopped()) {
4598 // It's an instance, and it passed the slow-path tests.
4599 PreserveJVMState pjvms(this);
4600 Node* obj_size = NULL;
4601 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4602 // is reexecuted if deoptimization occurs and there could be problems when merging
4603 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4604 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4605
4606 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4607
4608 // Present the results of the slow call.
4609 result_reg->init_req(_instance_path, control());
4610 result_val->init_req(_instance_path, alloc_obj);
4611 result_i_o ->set_req(_instance_path, i_o());
4612 result_mem ->set_req(_instance_path, reset_memory());
4613 }
4614
4615 // Generate code for the slow case. We make a call to clone().
4616 set_control(_gvn.transform(slow_region));
4617 if (!stopped()) {
4618 PreserveJVMState pjvms(this);
4619 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4620 Node* slow_result = set_results_for_java_call(slow_call);
4621 // this->control() comes from set_results_for_java_call
4622 result_reg->init_req(_slow_path, control());
4623 result_val->init_req(_slow_path, slow_result);
4624 result_i_o ->set_req(_slow_path, i_o());
4625 result_mem ->set_req(_slow_path, reset_memory());
4626 }
4627
4628 // Return the combined state.
4629 set_control( _gvn.transform(result_reg));
4630 set_i_o( _gvn.transform(result_i_o));
4631 set_all_memory( _gvn.transform(result_mem));
4632 } // original reexecute is set back here
4633
4634 set_result(_gvn.transform(result_val));
4635 return true;
4636 }
4637
4638 //------------------------------inline_arraycopy-----------------------
4639 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4640 // Object dest, int destPos,
4641 // int length);
4642 bool LibraryCallKit::inline_arraycopy() {
4643 // Get the arguments.
4644 Node* src = argument(0); // type: oop
4645 Node* src_offset = argument(1); // type: int
4646 Node* dest = argument(2); // type: oop
4647 Node* dest_offset = argument(3); // type: int
4648 Node* length = argument(4); // type: int
4649
4650 // The following tests must be performed
4651 // (1) src and dest are arrays.
4652 // (2) src and dest arrays must have elements of the same BasicType
4653 // (3) src and dest must not be null.
4654 // (4) src_offset must not be negative.
4655 // (5) dest_offset must not be negative.
4656 // (6) length must not be negative.
4657 // (7) src_offset + length must not exceed length of src.
4658 // (8) dest_offset + length must not exceed length of dest.
4659 // (9) each element of an oop array must be assignable
4660
4661 // (3) src and dest must not be null.
4662 // always do this here because we need the JVM state for uncommon traps
4663 src = null_check(src, T_ARRAY);
4664 dest = null_check(dest, T_ARRAY);
4665
4666 // Check for allocation before we add nodes that would confuse
4667 // tightly_coupled_allocation()
4668 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4669
4670 ciMethod* trap_method = method();
4671 int trap_bci = bci();
4672 SafePointNode* sfpt = NULL;
4673 if (alloc != NULL) {
4674 // The JVM state for uncommon traps between the allocation and
4675 // arraycopy is set to the state before the allocation: if the
4676 // initialization is performed by the array copy, we don't want to
4677 // go back to the interpreter with an unitialized array.
4678 JVMState* old_jvms = alloc->jvms();
4679 JVMState* jvms = old_jvms->clone_shallow(C);
4680 uint size = alloc->req();
4681 sfpt = new SafePointNode(size, jvms);
4682 jvms->set_map(sfpt);
4683 for (uint i = 0; i < size; i++) {
4684 sfpt->init_req(i, alloc->in(i));
4685 }
4686 // re-push array length for deoptimization
4687 sfpt->ins_req(jvms->stkoff() + jvms->sp(), alloc->in(AllocateNode::ALength));
4688 jvms->set_sp(jvms->sp()+1);
4689 jvms->set_monoff(jvms->monoff()+1);
4690 jvms->set_scloff(jvms->scloff()+1);
4691 jvms->set_endoff(jvms->endoff()+1);
4692 jvms->set_should_reexecute(true);
4693
4694 sfpt->set_i_o(map()->i_o());
4695 sfpt->set_memory(map()->memory());
4696
4697 trap_method = jvms->method();
4698 trap_bci = jvms->bci();
4699 }
4700
4701 bool validated = false;
4702
4703 const Type* src_type = _gvn.type(src);
4704 const Type* dest_type = _gvn.type(dest);
4705 const TypeAryPtr* top_src = src_type->isa_aryptr();
4706 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4707
4708 // Do we have the type of src?
4709 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4710 // Do we have the type of dest?
4711 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4712 // Is the type for src from speculation?
4713 bool src_spec = false;
4714 // Is the type for dest from speculation?
4715 bool dest_spec = false;
4716
4717 if (!has_src || !has_dest) {
4718 // We don't have sufficient type information, let's see if
4719 // speculative types can help. We need to have types for both src
4720 // and dest so that it pays off.
4721
4722 // Do we already have or could we have type information for src
4723 bool could_have_src = has_src;
4724 // Do we already have or could we have type information for dest
4725 bool could_have_dest = has_dest;
4726
4727 ciKlass* src_k = NULL;
4728 if (!has_src) {
4729 src_k = src_type->speculative_type_not_null();
4730 if (src_k != NULL && src_k->is_array_klass()) {
4731 could_have_src = true;
4732 }
4733 }
4734
4735 ciKlass* dest_k = NULL;
4736 if (!has_dest) {
4737 dest_k = dest_type->speculative_type_not_null();
4738 if (dest_k != NULL && dest_k->is_array_klass()) {
4739 could_have_dest = true;
4740 }
4741 }
4742
4743 if (could_have_src && could_have_dest) {
4744 // This is going to pay off so emit the required guards
4745 if (!has_src) {
4746 src = maybe_cast_profiled_obj(src, src_k, true, sfpt);
4747 src_type = _gvn.type(src);
4748 top_src = src_type->isa_aryptr();
4749 has_src = (top_src != NULL && top_src->klass() != NULL);
4750 src_spec = true;
4751 }
4752 if (!has_dest) {
4753 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4754 dest_type = _gvn.type(dest);
4755 top_dest = dest_type->isa_aryptr();
4756 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4757 dest_spec = true;
4758 }
4759 }
4760 }
4761
4762 if (has_src && has_dest) {
4763 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4764 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4765 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4766 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4767
4768 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4769 // If both arrays are object arrays then having the exact types
4770 // for both will remove the need for a subtype check at runtime
4771 // before the call and may make it possible to pick a faster copy
4772 // routine (without a subtype check on every element)
4773 // Do we have the exact type of src?
4774 bool could_have_src = src_spec;
4775 // Do we have the exact type of dest?
4776 bool could_have_dest = dest_spec;
4777 ciKlass* src_k = top_src->klass();
4778 ciKlass* dest_k = top_dest->klass();
4779 if (!src_spec) {
4780 src_k = src_type->speculative_type_not_null();
4781 if (src_k != NULL && src_k->is_array_klass()) {
4782 could_have_src = true;
4783 }
4784 }
4785 if (!dest_spec) {
4786 dest_k = dest_type->speculative_type_not_null();
4787 if (dest_k != NULL && dest_k->is_array_klass()) {
4788 could_have_dest = true;
4789 }
4790 }
4791 if (could_have_src && could_have_dest) {
4792 // If we can have both exact types, emit the missing guards
4793 if (could_have_src && !src_spec) {
4794 src = maybe_cast_profiled_obj(src, src_k, true, sfpt);
4795 }
4796 if (could_have_dest && !dest_spec) {
4797 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4798 }
4799 }
4800 }
4801 }
4802
4803 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) {
4804 // validate arguments: enables transformation the ArrayCopyNode
4805 validated = true;
4806
4807 RegionNode* slow_region = new RegionNode(1);
4808 record_for_igvn(slow_region);
4809
4810 // (1) src and dest are arrays.
4811 generate_non_array_guard(load_object_klass(src), slow_region);
4812 generate_non_array_guard(load_object_klass(dest), slow_region);
4813
4814 // (2) src and dest arrays must have elements of the same BasicType
4815 // done at macro expansion or at Ideal transformation time
4816
4817 // (4) src_offset must not be negative.
4818 generate_negative_guard(src_offset, slow_region);
4819
4820 // (5) dest_offset must not be negative.
4821 generate_negative_guard(dest_offset, slow_region);
4822
4823 // (7) src_offset + length must not exceed length of src.
4824 generate_limit_guard(src_offset, length,
4825 load_array_length(src),
4826 slow_region);
4827
4828 // (8) dest_offset + length must not exceed length of dest.
4829 generate_limit_guard(dest_offset, length,
4830 load_array_length(dest),
4831 slow_region);
4832
4833 // (9) each element of an oop array must be assignable
4834 Node* src_klass = load_object_klass(src);
4835 Node* dest_klass = load_object_klass(dest);
4836 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4837
4838 if (not_subtype_ctrl != top()) {
4839 if (sfpt != NULL) {
4840 GraphKit kit(sfpt->jvms());
4841 PreserveJVMState pjvms(&kit);
4842 kit.set_control(not_subtype_ctrl);
4843 kit.uncommon_trap(Deoptimization::Reason_intrinsic,
4844 Deoptimization::Action_make_not_entrant);
4845 assert(kit.stopped(), "Should be stopped");
4846 } else {
4847 PreserveJVMState pjvms(this);
4848 set_control(not_subtype_ctrl);
4849 uncommon_trap(Deoptimization::Reason_intrinsic,
4850 Deoptimization::Action_make_not_entrant);
4851 assert(stopped(), "Should be stopped");
4852 }
4853 }
4854 if (sfpt != NULL) {
4855 GraphKit kit(sfpt->jvms());
4856 kit.set_control(_gvn.transform(slow_region));
4857 kit.uncommon_trap(Deoptimization::Reason_intrinsic,
4858 Deoptimization::Action_make_not_entrant);
4859 assert(kit.stopped(), "Should be stopped");
4860 } else {
4861 PreserveJVMState pjvms(this);
4862 set_control(_gvn.transform(slow_region));
4863 uncommon_trap(Deoptimization::Reason_intrinsic,
4864 Deoptimization::Action_make_not_entrant);
4865 assert(stopped(), "Should be stopped");
4866 }
4867 }
4868
4869 if (stopped()) {
4870 return true;
4871 }
4872
4873 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4874 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4875 // so the compiler has a chance to eliminate them: during macro expansion,
4876 // we have to set their control (CastPP nodes are eliminated).
4877 load_object_klass(src), load_object_klass(dest),
4878 load_array_length(src), load_array_length(dest));
4879
4880 ac->set_arraycopy(validated);
4881
4882 Node* n = _gvn.transform(ac);
4883 if (n == ac) {
4884 ac->connect_outputs(this);
4885 } else {
4886 assert(validated, "shouldn't transform if all arguments not validated");
4887 set_all_memory(n);
4888 }
4889
4890 return true;
4891 }
4892
4893
4894 // Helper function which determines if an arraycopy immediately follows
4895 // an allocation, with no intervening tests or other escapes for the object.
4896 AllocateArrayNode*
4897 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4898 RegionNode* slow_region) {
4899 if (stopped()) return NULL; // no fast path
4900 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4901
4902 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4903 if (alloc == NULL) return NULL;
4904
4905 Node* rawmem = memory(Compile::AliasIdxRaw);
4906 // Is the allocation's memory state untouched?
4907 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4908 // Bail out if there have been raw-memory effects since the allocation.
4909 // (Example: There might have been a call or safepoint.)
4910 return NULL;
4911 }
4912 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4913 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4914 return NULL;
4915 }
4916
4917 // There must be no unexpected observers of this allocation.
4918 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4919 Node* obs = ptr->fast_out(i);
4920 if (obs != this->map()) {
4921 return NULL;
4922 }
4923 }
4924
4925 // This arraycopy must unconditionally follow the allocation of the ptr.
4926 Node* alloc_ctl = ptr->in(0);
4927 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4928
4929 Node* ctl = control();
4930 while (ctl != alloc_ctl) {
4931 // There may be guards which feed into the slow_region.
4932 // Any other control flow means that we might not get a chance
4933 // to finish initializing the allocated object.
4934 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4935 IfNode* iff = ctl->in(0)->as_If();
4936 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4937 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4938 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4939 ctl = iff->in(0); // This test feeds the known slow_region.
4940 continue;
4941 }
4942 // One more try: Various low-level checks bottom out in
4943 // uncommon traps. If the debug-info of the trap omits
4944 // any reference to the allocation, as we've already
4945 // observed, then there can be no objection to the trap.
4946 bool found_trap = false;
4947 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4948 Node* obs = not_ctl->fast_out(j);
4949 if (obs->in(0) == not_ctl && obs->is_Call() &&
4950 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4951 found_trap = true; break;
4952 }
4953 }
4954 if (found_trap) {
4955 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4956 continue;
4957 }
4958 }
4959 return NULL;
4960 }
4961
4962 // If we get this far, we have an allocation which immediately
4963 // precedes the arraycopy, and we can take over zeroing the new object.
4964 // The arraycopy will finish the initialization, and provide
4965 // a new control state to which we will anchor the destination pointer.
4966
4967 return alloc;
4968 }
4969
4970 //-------------inline_encodeISOArray-----------------------------------
4971 // encode char[] to byte[] in ISO_8859_1
4972 bool LibraryCallKit::inline_encodeISOArray() {
4973 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4974 // no receiver since it is static method
4975 Node *src = argument(0);
4976 Node *src_offset = argument(1);
4977 Node *dst = argument(2);
4978 Node *dst_offset = argument(3);
4979 Node *length = argument(4);
4980
4981 const Type* src_type = src->Value(&_gvn);
4982 const Type* dst_type = dst->Value(&_gvn);
4983 const TypeAryPtr* top_src = src_type->isa_aryptr();
4984 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4985 if (top_src == NULL || top_src->klass() == NULL ||
4986 top_dest == NULL || top_dest->klass() == NULL) {
4987 // failed array check
4988 return false;
4989 }
4990
4991 // Figure out the size and type of the elements we will be copying.
4992 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4993 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4994 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
4995 return false;
4996 }
4997 Node* src_start = array_element_address(src, src_offset, src_elem);
4998 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
4999 // 'src_start' points to src array + scaled offset
5000 // 'dst_start' points to dst array + scaled offset
5001
5002 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5003 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5004 enc = _gvn.transform(enc);
5005 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5006 set_memory(res_mem, mtype);
5007 set_result(enc);
5008 return true;
5009 }
5010
5011 //-------------inline_multiplyToLen-----------------------------------
5012 bool LibraryCallKit::inline_multiplyToLen() {
5013 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5014
5015 address stubAddr = StubRoutines::multiplyToLen();
5016 if (stubAddr == NULL) {
5017 return false; // Intrinsic's stub is not implemented on this platform
5018 }
5019 const char* stubName = "multiplyToLen";
5020
5021 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5022
5023 Node* x = argument(1);
5024 Node* xlen = argument(2);
5025 Node* y = argument(3);
5026 Node* ylen = argument(4);
5027 Node* z = argument(5);
5028
5029 const Type* x_type = x->Value(&_gvn);
5030 const Type* y_type = y->Value(&_gvn);
5031 const TypeAryPtr* top_x = x_type->isa_aryptr();
5032 const TypeAryPtr* top_y = y_type->isa_aryptr();
5033 if (top_x == NULL || top_x->klass() == NULL ||
5034 top_y == NULL || top_y->klass() == NULL) {
5035 // failed array check
5036 return false;
5037 }
5038
5039 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5040 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5041 if (x_elem != T_INT || y_elem != T_INT) {
5042 return false;
5043 }
5044
5045 // Set the original stack and the reexecute bit for the interpreter to reexecute
5046 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5047 // on the return from z array allocation in runtime.
5048 { PreserveReexecuteState preexecs(this);
5049 jvms()->set_should_reexecute(true);
5050
5051 Node* x_start = array_element_address(x, intcon(0), x_elem);
5052 Node* y_start = array_element_address(y, intcon(0), y_elem);
5053 // 'x_start' points to x array + scaled xlen
5054 // 'y_start' points to y array + scaled ylen
5055
5056 // Allocate the result array
5057 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5058 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5059 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5060
5061 IdealKit ideal(this);
5062
5063 #define __ ideal.
5064 Node* one = __ ConI(1);
5065 Node* zero = __ ConI(0);
5066 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5067 __ set(need_alloc, zero);
5068 __ set(z_alloc, z);
5069 __ if_then(z, BoolTest::eq, null()); {
5070 __ increment (need_alloc, one);
5071 } __ else_(); {
5072 // Update graphKit memory and control from IdealKit.
5073 sync_kit(ideal);
5074 Node* zlen_arg = load_array_length(z);
5075 // Update IdealKit memory and control from graphKit.
5076 __ sync_kit(this);
5077 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5078 __ increment (need_alloc, one);
5079 } __ end_if();
5080 } __ end_if();
5081
5082 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5083 // Update graphKit memory and control from IdealKit.
5084 sync_kit(ideal);
5085 Node * narr = new_array(klass_node, zlen, 1);
5086 // Update IdealKit memory and control from graphKit.
5087 __ sync_kit(this);
5088 __ set(z_alloc, narr);
5089 } __ end_if();
5090
5091 sync_kit(ideal);
5092 z = __ value(z_alloc);
5093 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5094 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5095 // Final sync IdealKit and GraphKit.
5096 final_sync(ideal);
5097 #undef __
5098
5099 Node* z_start = array_element_address(z, intcon(0), T_INT);
5100
5101 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5102 OptoRuntime::multiplyToLen_Type(),
5103 stubAddr, stubName, TypePtr::BOTTOM,
5104 x_start, xlen, y_start, ylen, z_start, zlen);
5105 } // original reexecute is set back here
5106
5107 C->set_has_split_ifs(true); // Has chance for split-if optimization
5108 set_result(z);
5109 return true;
5110 }
5111
5112
5113 /**
5114 * Calculate CRC32 for byte.
5115 * int java.util.zip.CRC32.update(int crc, int b)
5116 */
5117 bool LibraryCallKit::inline_updateCRC32() {
5118 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5119 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5120 // no receiver since it is static method
5121 Node* crc = argument(0); // type: int
5122 Node* b = argument(1); // type: int
5123
5124 /*
5125 * int c = ~ crc;
5126 * b = timesXtoThe32[(b ^ c) & 0xFF];
5127 * b = b ^ (c >>> 8);
5128 * crc = ~b;
5129 */
5130
5131 Node* M1 = intcon(-1);
5132 crc = _gvn.transform(new XorINode(crc, M1));
5133 Node* result = _gvn.transform(new XorINode(crc, b));
5134 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5135
5136 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5137 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5138 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5139 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5140
5141 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5142 result = _gvn.transform(new XorINode(crc, result));
5143 result = _gvn.transform(new XorINode(result, M1));
5144 set_result(result);
5145 return true;
5146 }
5147
5148 /**
5149 * Calculate CRC32 for byte[] array.
5150 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5151 */
5152 bool LibraryCallKit::inline_updateBytesCRC32() {
5153 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5154 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5155 // no receiver since it is static method
5156 Node* crc = argument(0); // type: int
5157 Node* src = argument(1); // type: oop
5158 Node* offset = argument(2); // type: int
5159 Node* length = argument(3); // type: int
5160
5161 const Type* src_type = src->Value(&_gvn);
5162 const TypeAryPtr* top_src = src_type->isa_aryptr();
5163 if (top_src == NULL || top_src->klass() == NULL) {
5164 // failed array check
5165 return false;
5166 }
5167
5168 // Figure out the size and type of the elements we will be copying.
5169 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5170 if (src_elem != T_BYTE) {
5171 return false;
5172 }
5173
5174 // 'src_start' points to src array + scaled offset
5175 Node* src_start = array_element_address(src, offset, src_elem);
5176
5177 // We assume that range check is done by caller.
5178 // TODO: generate range check (offset+length < src.length) in debug VM.
5179
5180 // Call the stub.
5181 address stubAddr = StubRoutines::updateBytesCRC32();
5182 const char *stubName = "updateBytesCRC32";
5183
5184 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5185 stubAddr, stubName, TypePtr::BOTTOM,
5186 crc, src_start, length);
5187 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5188 set_result(result);
5189 return true;
5190 }
5191
5192 /**
5193 * Calculate CRC32 for ByteBuffer.
5194 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5195 */
5196 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5197 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5198 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5199 // no receiver since it is static method
5200 Node* crc = argument(0); // type: int
5201 Node* src = argument(1); // type: long
5202 Node* offset = argument(3); // type: int
5203 Node* length = argument(4); // type: int
5204
5205 src = ConvL2X(src); // adjust Java long to machine word
5206 Node* base = _gvn.transform(new CastX2PNode(src));
5207 offset = ConvI2X(offset);
5208
5209 // 'src_start' points to src array + scaled offset
5210 Node* src_start = basic_plus_adr(top(), base, offset);
5211
5212 // Call the stub.
5213 address stubAddr = StubRoutines::updateBytesCRC32();
5214 const char *stubName = "updateBytesCRC32";
5215
5216 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5217 stubAddr, stubName, TypePtr::BOTTOM,
5218 crc, src_start, length);
5219 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5220 set_result(result);
5221 return true;
5222 }
5223
5224 //----------------------------inline_reference_get----------------------------
5225 // public T java.lang.ref.Reference.get();
5226 bool LibraryCallKit::inline_reference_get() {
5227 const int referent_offset = java_lang_ref_Reference::referent_offset;
5228 guarantee(referent_offset > 0, "should have already been set");
5229
5230 // Get the argument:
5231 Node* reference_obj = null_check_receiver();
5232 if (stopped()) return true;
5233
5234 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5235
5236 ciInstanceKlass* klass = env()->Object_klass();
5237 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5238
5239 Node* no_ctrl = NULL;
5240 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5241
5242 // Use the pre-barrier to record the value in the referent field
5243 pre_barrier(false /* do_load */,
5244 control(),
5245 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5246 result /* pre_val */,
5247 T_OBJECT);
5248
5249 // Add memory barrier to prevent commoning reads from this field
5250 // across safepoint since GC can change its value.
5251 insert_mem_bar(Op_MemBarCPUOrder);
5252
5253 set_result(result);
5254 return true;
5255 }
5256
5257
5258 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5259 bool is_exact=true, bool is_static=false) {
5260
5261 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5262 assert(tinst != NULL, "obj is null");
5263 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5264 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5265
5266 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5267 ciSymbol::make(fieldTypeString),
5268 is_static);
5269 if (field == NULL) return (Node *) NULL;
5270 assert (field != NULL, "undefined field");
5271
5272 // Next code copied from Parse::do_get_xxx():
5273
5274 // Compute address and memory type.
5275 int offset = field->offset_in_bytes();
5276 bool is_vol = field->is_volatile();
5277 ciType* field_klass = field->type();
5278 assert(field_klass->is_loaded(), "should be loaded");
5279 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5280 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5281 BasicType bt = field->layout_type();
5282
5283 // Build the resultant type of the load
5284 const Type *type;
5285 if (bt == T_OBJECT) {
5286 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5287 } else {
5288 type = Type::get_const_basic_type(bt);
5289 }
5290
5291 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5292 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5293 }
5294 // Build the load.
5295 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5296 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
5297 // If reference is volatile, prevent following memory ops from
5298 // floating up past the volatile read. Also prevents commoning
5299 // another volatile read.
5300 if (is_vol) {
5301 // Memory barrier includes bogus read of value to force load BEFORE membar
5302 insert_mem_bar(Op_MemBarAcquire, loadedField);
5303 }
5304 return loadedField;
5305 }
5306
5307
5308 //------------------------------inline_aescrypt_Block-----------------------
5309 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5310 address stubAddr;
5311 const char *stubName;
5312 assert(UseAES, "need AES instruction support");
5313
5314 switch(id) {
5315 case vmIntrinsics::_aescrypt_encryptBlock:
5316 stubAddr = StubRoutines::aescrypt_encryptBlock();
5317 stubName = "aescrypt_encryptBlock";
5318 break;
5319 case vmIntrinsics::_aescrypt_decryptBlock:
5320 stubAddr = StubRoutines::aescrypt_decryptBlock();
5321 stubName = "aescrypt_decryptBlock";
5322 break;
5323 }
5324 if (stubAddr == NULL) return false;
5325
5326 Node* aescrypt_object = argument(0);
5327 Node* src = argument(1);
5328 Node* src_offset = argument(2);
5329 Node* dest = argument(3);
5330 Node* dest_offset = argument(4);
5331
5332 // (1) src and dest are arrays.
5333 const Type* src_type = src->Value(&_gvn);
5334 const Type* dest_type = dest->Value(&_gvn);
5335 const TypeAryPtr* top_src = src_type->isa_aryptr();
5336 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5337 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5338
5339 // for the quick and dirty code we will skip all the checks.
5340 // we are just trying to get the call to be generated.
5341 Node* src_start = src;
5342 Node* dest_start = dest;
5343 if (src_offset != NULL || dest_offset != NULL) {
5344 assert(src_offset != NULL && dest_offset != NULL, "");
5345 src_start = array_element_address(src, src_offset, T_BYTE);
5346 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5347 }
5348
5349 // now need to get the start of its expanded key array
5350 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5351 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5352 if (k_start == NULL) return false;
5353
5354 if (Matcher::pass_original_key_for_aes()) {
5355 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5356 // compatibility issues between Java key expansion and SPARC crypto instructions
5357 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5358 if (original_k_start == NULL) return false;
5359
5360 // Call the stub.
5361 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5362 stubAddr, stubName, TypePtr::BOTTOM,
5363 src_start, dest_start, k_start, original_k_start);
5364 } else {
5365 // Call the stub.
5366 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5367 stubAddr, stubName, TypePtr::BOTTOM,
5368 src_start, dest_start, k_start);
5369 }
5370
5371 return true;
5372 }
5373
5374 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5375 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5376 address stubAddr;
5377 const char *stubName;
5378
5379 assert(UseAES, "need AES instruction support");
5380
5381 switch(id) {
5382 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5383 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5384 stubName = "cipherBlockChaining_encryptAESCrypt";
5385 break;
5386 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5387 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5388 stubName = "cipherBlockChaining_decryptAESCrypt";
5389 break;
5390 }
5391 if (stubAddr == NULL) return false;
5392
5393 Node* cipherBlockChaining_object = argument(0);
5394 Node* src = argument(1);
5395 Node* src_offset = argument(2);
5396 Node* len = argument(3);
5397 Node* dest = argument(4);
5398 Node* dest_offset = argument(5);
5399
5400 // (1) src and dest are arrays.
5401 const Type* src_type = src->Value(&_gvn);
5402 const Type* dest_type = dest->Value(&_gvn);
5403 const TypeAryPtr* top_src = src_type->isa_aryptr();
5404 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5405 assert (top_src != NULL && top_src->klass() != NULL
5406 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5407
5408 // checks are the responsibility of the caller
5409 Node* src_start = src;
5410 Node* dest_start = dest;
5411 if (src_offset != NULL || dest_offset != NULL) {
5412 assert(src_offset != NULL && dest_offset != NULL, "");
5413 src_start = array_element_address(src, src_offset, T_BYTE);
5414 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5415 }
5416
5417 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5418 // (because of the predicated logic executed earlier).
5419 // so we cast it here safely.
5420 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5421
5422 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5423 if (embeddedCipherObj == NULL) return false;
5424
5425 // cast it to what we know it will be at runtime
5426 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5427 assert(tinst != NULL, "CBC obj is null");
5428 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5429 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5430 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5431
5432 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5433 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5434 const TypeOopPtr* xtype = aklass->as_instance_type();
5435 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5436 aescrypt_object = _gvn.transform(aescrypt_object);
5437
5438 // we need to get the start of the aescrypt_object's expanded key array
5439 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5440 if (k_start == NULL) return false;
5441
5442 // similarly, get the start address of the r vector
5443 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5444 if (objRvec == NULL) return false;
5445 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5446
5447 Node* cbcCrypt;
5448 if (Matcher::pass_original_key_for_aes()) {
5449 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5450 // compatibility issues between Java key expansion and SPARC crypto instructions
5451 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5452 if (original_k_start == NULL) return false;
5453
5454 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5455 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5456 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5457 stubAddr, stubName, TypePtr::BOTTOM,
5458 src_start, dest_start, k_start, r_start, len, original_k_start);
5459 } else {
5460 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5461 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5462 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5463 stubAddr, stubName, TypePtr::BOTTOM,
5464 src_start, dest_start, k_start, r_start, len);
5465 }
5466
5467 // return cipher length (int)
5468 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5469 set_result(retvalue);
5470 return true;
5471 }
5472
5473 //------------------------------get_key_start_from_aescrypt_object-----------------------
5474 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5475 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5476 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5477 if (objAESCryptKey == NULL) return (Node *) NULL;
5478
5479 // now have the array, need to get the start address of the K array
5480 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5481 return k_start;
5482 }
5483
5484 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5485 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5486 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5487 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5488 if (objAESCryptKey == NULL) return (Node *) NULL;
5489
5490 // now have the array, need to get the start address of the lastKey array
5491 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5492 return original_k_start;
5493 }
5494
5495 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5496 // Return node representing slow path of predicate check.
5497 // the pseudo code we want to emulate with this predicate is:
5498 // for encryption:
5499 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5500 // for decryption:
5501 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5502 // note cipher==plain is more conservative than the original java code but that's OK
5503 //
5504 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5505 // The receiver was checked for NULL already.
5506 Node* objCBC = argument(0);
5507
5508 // Load embeddedCipher field of CipherBlockChaining object.
5509 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5510
5511 // get AESCrypt klass for instanceOf check
5512 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5513 // will have same classloader as CipherBlockChaining object
5514 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5515 assert(tinst != NULL, "CBCobj is null");
5516 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5517
5518 // we want to do an instanceof comparison against the AESCrypt class
5519 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5520 if (!klass_AESCrypt->is_loaded()) {
5521 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5522 Node* ctrl = control();
5523 set_control(top()); // no regular fast path
5524 return ctrl;
5525 }
5526 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5527
5528 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5529 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5530 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5531
5532 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5533
5534 // for encryption, we are done
5535 if (!decrypting)
5536 return instof_false; // even if it is NULL
5537
5538 // for decryption, we need to add a further check to avoid
5539 // taking the intrinsic path when cipher and plain are the same
5540 // see the original java code for why.
5541 RegionNode* region = new RegionNode(3);
5542 region->init_req(1, instof_false);
5543 Node* src = argument(1);
5544 Node* dest = argument(4);
5545 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5546 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5547 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5548 region->init_req(2, src_dest_conjoint);
5549
5550 record_for_igvn(region);
5551 return _gvn.transform(region);
5552 }
5553
5554 //------------------------------inline_sha_implCompress-----------------------
5555 //
5556 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5557 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5558 //
5559 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
5560 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
5561 //
5562 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
5563 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
5564 //
5565 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
5566 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
5567
5568 Node* sha_obj = argument(0);
5569 Node* src = argument(1); // type oop
5570 Node* ofs = argument(2); // type int
5571
5572 const Type* src_type = src->Value(&_gvn);
5573 const TypeAryPtr* top_src = src_type->isa_aryptr();
5574 if (top_src == NULL || top_src->klass() == NULL) {
5575 // failed array check
5576 return false;
5577 }
5578 // Figure out the size and type of the elements we will be copying.
5579 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5580 if (src_elem != T_BYTE) {
5581 return false;
5582 }
5583 // 'src_start' points to src array + offset
5584 Node* src_start = array_element_address(src, ofs, src_elem);
5585 Node* state = NULL;
5586 address stubAddr;
5587 const char *stubName;
5588
5589 switch(id) {
5590 case vmIntrinsics::_sha_implCompress:
5591 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
5592 state = get_state_from_sha_object(sha_obj);
5593 stubAddr = StubRoutines::sha1_implCompress();
5594 stubName = "sha1_implCompress";
5595 break;
5596 case vmIntrinsics::_sha2_implCompress:
5597 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
5598 state = get_state_from_sha_object(sha_obj);
5599 stubAddr = StubRoutines::sha256_implCompress();
5600 stubName = "sha256_implCompress";
5601 break;
5602 case vmIntrinsics::_sha5_implCompress:
5603 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
5604 state = get_state_from_sha5_object(sha_obj);
5605 stubAddr = StubRoutines::sha512_implCompress();
5606 stubName = "sha512_implCompress";
5607 break;
5608 default:
5609 fatal_unexpected_iid(id);
5610 return false;
5611 }
5612 if (state == NULL) return false;
5613
5614 // Call the stub.
5615 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
5616 stubAddr, stubName, TypePtr::BOTTOM,
5617 src_start, state);
5618
5619 return true;
5620 }
5621
5622 //------------------------------inline_digestBase_implCompressMB-----------------------
5623 //
5624 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
5625 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
5626 //
5627 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
5628 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5629 "need SHA1/SHA256/SHA512 instruction support");
5630 assert((uint)predicate < 3, "sanity");
5631 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
5632
5633 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
5634 Node* src = argument(1); // byte[] array
5635 Node* ofs = argument(2); // type int
5636 Node* limit = argument(3); // type int
5637
5638 const Type* src_type = src->Value(&_gvn);
5639 const TypeAryPtr* top_src = src_type->isa_aryptr();
5640 if (top_src == NULL || top_src->klass() == NULL) {
5641 // failed array check
5642 return false;
5643 }
5644 // Figure out the size and type of the elements we will be copying.
5645 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5646 if (src_elem != T_BYTE) {
5647 return false;
5648 }
5649 // 'src_start' points to src array + offset
5650 Node* src_start = array_element_address(src, ofs, src_elem);
5651
5652 const char* klass_SHA_name = NULL;
5653 const char* stub_name = NULL;
5654 address stub_addr = NULL;
5655 bool long_state = false;
5656
5657 switch (predicate) {
5658 case 0:
5659 if (UseSHA1Intrinsics) {
5660 klass_SHA_name = "sun/security/provider/SHA";
5661 stub_name = "sha1_implCompressMB";
5662 stub_addr = StubRoutines::sha1_implCompressMB();
5663 }
5664 break;
5665 case 1:
5666 if (UseSHA256Intrinsics) {
5667 klass_SHA_name = "sun/security/provider/SHA2";
5668 stub_name = "sha256_implCompressMB";
5669 stub_addr = StubRoutines::sha256_implCompressMB();
5670 }
5671 break;
5672 case 2:
5673 if (UseSHA512Intrinsics) {
5674 klass_SHA_name = "sun/security/provider/SHA5";
5675 stub_name = "sha512_implCompressMB";
5676 stub_addr = StubRoutines::sha512_implCompressMB();
5677 long_state = true;
5678 }
5679 break;
5680 default:
5681 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5682 }
5683 if (klass_SHA_name != NULL) {
5684 // get DigestBase klass to lookup for SHA klass
5685 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
5686 assert(tinst != NULL, "digestBase_obj is not instance???");
5687 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5688
5689 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5690 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
5691 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5692 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
5693 }
5694 return false;
5695 }
5696 //------------------------------inline_sha_implCompressMB-----------------------
5697 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
5698 bool long_state, address stubAddr, const char *stubName,
5699 Node* src_start, Node* ofs, Node* limit) {
5700 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
5701 const TypeOopPtr* xtype = aklass->as_instance_type();
5702 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
5703 sha_obj = _gvn.transform(sha_obj);
5704
5705 Node* state;
5706 if (long_state) {
5707 state = get_state_from_sha5_object(sha_obj);
5708 } else {
5709 state = get_state_from_sha_object(sha_obj);
5710 }
5711 if (state == NULL) return false;
5712
5713 // Call the stub.
5714 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5715 OptoRuntime::digestBase_implCompressMB_Type(),
5716 stubAddr, stubName, TypePtr::BOTTOM,
5717 src_start, state, ofs, limit);
5718 // return ofs (int)
5719 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5720 set_result(result);
5721
5722 return true;
5723 }
5724
5725 //------------------------------get_state_from_sha_object-----------------------
5726 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
5727 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
5728 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
5729 if (sha_state == NULL) return (Node *) NULL;
5730
5731 // now have the array, need to get the start address of the state array
5732 Node* state = array_element_address(sha_state, intcon(0), T_INT);
5733 return state;
5734 }
5735
5736 //------------------------------get_state_from_sha5_object-----------------------
5737 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
5738 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
5739 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
5740 if (sha_state == NULL) return (Node *) NULL;
5741
5742 // now have the array, need to get the start address of the state array
5743 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
5744 return state;
5745 }
5746
5747 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
5748 // Return node representing slow path of predicate check.
5749 // the pseudo code we want to emulate with this predicate is:
5750 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
5751 //
5752 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
5753 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5754 "need SHA1/SHA256/SHA512 instruction support");
5755 assert((uint)predicate < 3, "sanity");
5756
5757 // The receiver was checked for NULL already.
5758 Node* digestBaseObj = argument(0);
5759
5760 // get DigestBase klass for instanceOf check
5761 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
5762 assert(tinst != NULL, "digestBaseObj is null");
5763 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5764
5765 const char* klass_SHA_name = NULL;
5766 switch (predicate) {
5767 case 0:
5768 if (UseSHA1Intrinsics) {
5769 // we want to do an instanceof comparison against the SHA class
5770 klass_SHA_name = "sun/security/provider/SHA";
5771 }
5772 break;
5773 case 1:
5774 if (UseSHA256Intrinsics) {
5775 // we want to do an instanceof comparison against the SHA2 class
5776 klass_SHA_name = "sun/security/provider/SHA2";
5777 }
5778 break;
5779 case 2:
5780 if (UseSHA512Intrinsics) {
5781 // we want to do an instanceof comparison against the SHA5 class
5782 klass_SHA_name = "sun/security/provider/SHA5";
5783 }
5784 break;
5785 default:
5786 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5787 }
5788
5789 ciKlass* klass_SHA = NULL;
5790 if (klass_SHA_name != NULL) {
5791 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5792 }
5793 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
5794 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
5795 Node* ctrl = control();
5796 set_control(top()); // no intrinsic path
5797 return ctrl;
5798 }
5799 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5800
5801 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
5802 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
5803 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5804 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5805
5806 return instof_false; // even if it is NULL
5807 }
5808
5809 bool LibraryCallKit::inline_profileBoolean() {
5810 Node* counts = argument(1);
5811 const TypeAryPtr* ary = NULL;
5812 ciArray* aobj = NULL;
5813 if (counts->is_Con()
5814 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
5815 && (aobj = ary->const_oop()->as_array()) != NULL
5816 && (aobj->length() == 2)) {
5817 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
5818 jint false_cnt = aobj->element_value(0).as_int();
5819 jint true_cnt = aobj->element_value(1).as_int();
5820
5821 method()->set_injected_profile(true);
5822
5823 if (C->log() != NULL) {
5824 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
5825 false_cnt, true_cnt);
5826 }
5827
5828 if (false_cnt + true_cnt == 0) {
5829 // According to profile, never executed.
5830 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
5831 Deoptimization::Action_reinterpret);
5832 return true;
5833 }
5834 // Stop profiling.
5835 // MethodHandleImpl::profileBoolean() has profiling logic in it's bytecode.
5836 // By replacing method's body with profile data (represented as ProfileBooleanNode
5837 // on IR level) we effectively disable profiling.
5838 // It enables full speed execution once optimized code is generated.
5839 Node* profile = _gvn.transform(new ProfileBooleanNode(argument(0), false_cnt, true_cnt));
5840 C->record_for_igvn(profile);
5841 set_result(profile);
5842 return true;
5843 } else {
5844 // Continue profiling.
5845 // Profile data isn't available at the moment. So, execute method's bytecode version.
5846 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
5847 // is compiled and counters aren't available since corresponding MethodHandle
5848 // isn't a compile-time constant.
5849 return false;
5850 }
5851 }