1 /*
2 * Copyright (c) 1999, 2013, 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 "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/countbitsnode.hpp"
37 #include "opto/intrinsicnode.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/mathexactnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "prims/nativeLookup.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "trace/traceMacros.hpp"
49
50 class LibraryIntrinsic : public InlineCallGenerator {
51 // Extend the set of intrinsics known to the runtime:
52 public:
53 private:
54 bool _is_virtual;
55 bool _is_predicted;
56 bool _does_virtual_dispatch;
57 vmIntrinsics::ID _intrinsic_id;
58
59 public:
60 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
61 : InlineCallGenerator(m),
62 _is_virtual(is_virtual),
63 _is_predicted(is_predicted),
64 _does_virtual_dispatch(does_virtual_dispatch),
65 _intrinsic_id(id)
66 {
67 }
68 virtual bool is_intrinsic() const { return true; }
69 virtual bool is_virtual() const { return _is_virtual; }
70 virtual bool is_predicted() const { return _is_predicted; }
71 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
72 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
73 virtual Node* generate_predicate(JVMState* jvms);
74 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
75 };
76
77
78 // Local helper class for LibraryIntrinsic:
79 class LibraryCallKit : public GraphKit {
80 private:
81 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
82 Node* _result; // the result node, if any
83 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
84
85 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
86
87 public:
88 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
89 : GraphKit(jvms),
90 _intrinsic(intrinsic),
91 _result(NULL)
92 {
93 // Check if this is a root compile. In that case we don't have a caller.
94 if (!jvms->has_method()) {
95 _reexecute_sp = sp();
96 } else {
97 // Find out how many arguments the interpreter needs when deoptimizing
98 // and save the stack pointer value so it can used by uncommon_trap.
99 // We find the argument count by looking at the declared signature.
100 bool ignored_will_link;
101 ciSignature* declared_signature = NULL;
102 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
103 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
104 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
105 }
106 }
107
108 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
109
110 ciMethod* caller() const { return jvms()->method(); }
111 int bci() const { return jvms()->bci(); }
112 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
113 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
114 ciMethod* callee() const { return _intrinsic->method(); }
115
116 bool try_to_inline();
117 Node* try_to_predicate();
118
119 void push_result() {
120 // Push the result onto the stack.
121 if (!stopped() && result() != NULL) {
122 BasicType bt = result()->bottom_type()->basic_type();
123 push_node(bt, result());
124 }
125 }
126
127 private:
128 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
129 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
130 }
131
132 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
133 void set_result(RegionNode* region, PhiNode* value);
134 Node* result() { return _result; }
135
136 virtual int reexecute_sp() { return _reexecute_sp; }
137
138 // Helper functions to inline natives
139 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
140 Node* generate_slow_guard(Node* test, RegionNode* region);
141 Node* generate_fair_guard(Node* test, RegionNode* region);
142 Node* generate_negative_guard(Node* index, RegionNode* region,
143 // resulting CastII of index:
144 Node* *pos_index = NULL);
145 Node* generate_nonpositive_guard(Node* index, bool never_negative,
146 // resulting CastII of index:
147 Node* *pos_index = NULL);
148 Node* generate_limit_guard(Node* offset, Node* subseq_length,
149 Node* array_length,
150 RegionNode* region);
151 Node* generate_current_thread(Node* &tls_output);
152 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
153 bool disjoint_bases, const char* &name, bool dest_uninitialized);
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 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
237 static bool klass_needs_init_guard(Node* kls);
238 bool inline_unsafe_allocate();
239 bool inline_unsafe_copyMemory();
240 bool inline_native_currentThread();
241 #ifdef TRACE_HAVE_INTRINSICS
242 bool inline_native_classID();
243 bool inline_native_threadID();
244 #endif
245 bool inline_native_time_funcs(address method, const char* funcName);
246 bool inline_native_isInterrupted();
247 bool inline_native_Class_query(vmIntrinsics::ID id);
248 bool inline_native_subtype_check();
249
250 bool inline_native_newArray();
251 bool inline_native_getLength();
252 bool inline_array_copyOf(bool is_copyOfRange);
253 bool inline_array_equals();
254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
255 bool inline_native_clone(bool is_virtual);
256 bool inline_native_Reflection_getCallerClass();
257 // Helper function for inlining native object hash method
258 bool inline_native_hashcode(bool is_virtual, bool is_static);
259 bool inline_native_getClass();
260
261 // Helper functions for inlining arraycopy
262 bool inline_arraycopy();
263 void generate_arraycopy(const TypePtr* adr_type,
264 BasicType basic_elem_type,
265 Node* src, Node* src_offset,
266 Node* dest, Node* dest_offset,
267 Node* copy_length,
268 bool disjoint_bases = false,
269 bool length_never_negative = false,
270 RegionNode* slow_region = NULL);
271 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
272 RegionNode* slow_region);
273 void generate_clear_array(const TypePtr* adr_type,
274 Node* dest,
275 BasicType basic_elem_type,
276 Node* slice_off,
277 Node* slice_len,
278 Node* slice_end);
279 bool generate_block_arraycopy(const TypePtr* adr_type,
280 BasicType basic_elem_type,
281 AllocateNode* alloc,
282 Node* src, Node* src_offset,
283 Node* dest, Node* dest_offset,
284 Node* dest_size, bool dest_uninitialized);
285 void generate_slow_arraycopy(const TypePtr* adr_type,
286 Node* src, Node* src_offset,
287 Node* dest, Node* dest_offset,
288 Node* copy_length, bool dest_uninitialized);
289 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
290 Node* dest_elem_klass,
291 Node* src, Node* src_offset,
292 Node* dest, Node* dest_offset,
293 Node* copy_length, bool dest_uninitialized);
294 Node* generate_generic_arraycopy(const TypePtr* adr_type,
295 Node* src, Node* src_offset,
296 Node* dest, Node* dest_offset,
297 Node* copy_length, bool dest_uninitialized);
298 void generate_unchecked_arraycopy(const TypePtr* adr_type,
299 BasicType basic_elem_type,
300 bool disjoint_bases,
301 Node* src, Node* src_offset,
302 Node* dest, Node* dest_offset,
303 Node* copy_length, bool dest_uninitialized);
304 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
305 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
306 bool inline_unsafe_ordered_store(BasicType type);
307 bool inline_unsafe_fence(vmIntrinsics::ID id);
308 bool inline_fp_conversions(vmIntrinsics::ID id);
309 bool inline_number_methods(vmIntrinsics::ID id);
310 bool inline_reference_get();
311 bool inline_aescrypt_Block(vmIntrinsics::ID id);
312 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
313 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
314 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
315 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
316 bool inline_encodeISOArray();
317 bool inline_updateCRC32();
318 bool inline_updateBytesCRC32();
319 bool inline_updateByteBufferCRC32();
320 };
321
322
323 //---------------------------make_vm_intrinsic----------------------------
324 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
325 vmIntrinsics::ID id = m->intrinsic_id();
326 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
327
328 if (DisableIntrinsic[0] != '\0'
329 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
330 // disabled by a user request on the command line:
331 // example: -XX:DisableIntrinsic=_hashCode,_getClass
332 return NULL;
333 }
334
335 if (!m->is_loaded()) {
336 // do not attempt to inline unloaded methods
337 return NULL;
338 }
339
340 // Only a few intrinsics implement a virtual dispatch.
341 // They are expensive calls which are also frequently overridden.
342 if (is_virtual) {
343 switch (id) {
344 case vmIntrinsics::_hashCode:
345 case vmIntrinsics::_clone:
346 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
347 break;
348 default:
349 return NULL;
350 }
351 }
352
353 // -XX:-InlineNatives disables nearly all intrinsics:
354 if (!InlineNatives) {
355 switch (id) {
356 case vmIntrinsics::_indexOf:
357 case vmIntrinsics::_compareTo:
358 case vmIntrinsics::_equals:
359 case vmIntrinsics::_equalsC:
360 case vmIntrinsics::_getAndAddInt:
361 case vmIntrinsics::_getAndAddLong:
362 case vmIntrinsics::_getAndSetInt:
363 case vmIntrinsics::_getAndSetLong:
364 case vmIntrinsics::_getAndSetObject:
365 case vmIntrinsics::_loadFence:
366 case vmIntrinsics::_storeFence:
367 case vmIntrinsics::_fullFence:
368 break; // InlineNatives does not control String.compareTo
369 case vmIntrinsics::_Reference_get:
370 break; // InlineNatives does not control Reference.get
371 default:
372 return NULL;
373 }
374 }
375
376 bool is_predicted = false;
377 bool does_virtual_dispatch = false;
378
379 switch (id) {
380 case vmIntrinsics::_compareTo:
381 if (!SpecialStringCompareTo) return NULL;
382 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
383 break;
384 case vmIntrinsics::_indexOf:
385 if (!SpecialStringIndexOf) return NULL;
386 break;
387 case vmIntrinsics::_equals:
388 if (!SpecialStringEquals) return NULL;
389 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
390 break;
391 case vmIntrinsics::_equalsC:
392 if (!SpecialArraysEquals) return NULL;
393 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
394 break;
395 case vmIntrinsics::_arraycopy:
396 if (!InlineArrayCopy) return NULL;
397 break;
398 case vmIntrinsics::_copyMemory:
399 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
400 if (!InlineArrayCopy) return NULL;
401 break;
402 case vmIntrinsics::_hashCode:
403 if (!InlineObjectHash) return NULL;
404 does_virtual_dispatch = true;
405 break;
406 case vmIntrinsics::_clone:
407 does_virtual_dispatch = true;
408 case vmIntrinsics::_copyOf:
409 case vmIntrinsics::_copyOfRange:
410 if (!InlineObjectCopy) return NULL;
411 // These also use the arraycopy intrinsic mechanism:
412 if (!InlineArrayCopy) return NULL;
413 break;
414 case vmIntrinsics::_encodeISOArray:
415 if (!SpecialEncodeISOArray) return NULL;
416 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
417 break;
418 case vmIntrinsics::_checkIndex:
419 // We do not intrinsify this. The optimizer does fine with it.
420 return NULL;
421
422 case vmIntrinsics::_getCallerClass:
423 if (!UseNewReflection) return NULL;
424 if (!InlineReflectionGetCallerClass) return NULL;
425 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
426 break;
427
428 case vmIntrinsics::_bitCount_i:
429 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
430 break;
431
432 case vmIntrinsics::_bitCount_l:
433 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
434 break;
435
436 case vmIntrinsics::_numberOfLeadingZeros_i:
437 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
438 break;
439
440 case vmIntrinsics::_numberOfLeadingZeros_l:
441 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
442 break;
443
444 case vmIntrinsics::_numberOfTrailingZeros_i:
445 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
446 break;
447
448 case vmIntrinsics::_numberOfTrailingZeros_l:
449 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
450 break;
451
452 case vmIntrinsics::_reverseBytes_c:
453 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
454 break;
455 case vmIntrinsics::_reverseBytes_s:
456 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
457 break;
458 case vmIntrinsics::_reverseBytes_i:
459 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
460 break;
461 case vmIntrinsics::_reverseBytes_l:
462 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
463 break;
464
465 case vmIntrinsics::_Reference_get:
466 // Use the intrinsic version of Reference.get() so that the value in
467 // the referent field can be registered by the G1 pre-barrier code.
468 // Also add memory barrier to prevent commoning reads from this field
469 // across safepoint since GC can change it value.
470 break;
471
472 case vmIntrinsics::_compareAndSwapObject:
473 #ifdef _LP64
474 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
475 #endif
476 break;
477
478 case vmIntrinsics::_compareAndSwapLong:
479 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndAddInt:
483 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
484 break;
485
486 case vmIntrinsics::_getAndAddLong:
487 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
488 break;
489
490 case vmIntrinsics::_getAndSetInt:
491 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
492 break;
493
494 case vmIntrinsics::_getAndSetLong:
495 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
496 break;
497
498 case vmIntrinsics::_getAndSetObject:
499 #ifdef _LP64
500 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
501 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
502 break;
503 #else
504 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
505 break;
506 #endif
507
508 case vmIntrinsics::_aescrypt_encryptBlock:
509 case vmIntrinsics::_aescrypt_decryptBlock:
510 if (!UseAESIntrinsics) return NULL;
511 break;
512
513 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
514 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
515 if (!UseAESIntrinsics) return NULL;
516 // these two require the predicated logic
517 is_predicted = true;
518 break;
519
520 case vmIntrinsics::_updateCRC32:
521 case vmIntrinsics::_updateBytesCRC32:
522 case vmIntrinsics::_updateByteBufferCRC32:
523 if (!UseCRC32Intrinsics) return NULL;
524 break;
525
526 case vmIntrinsics::_incrementExactI:
527 case vmIntrinsics::_addExactI:
528 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
529 break;
530 case vmIntrinsics::_incrementExactL:
531 case vmIntrinsics::_addExactL:
532 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
533 break;
534 case vmIntrinsics::_decrementExactI:
535 case vmIntrinsics::_subtractExactI:
536 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
537 break;
538 case vmIntrinsics::_decrementExactL:
539 case vmIntrinsics::_subtractExactL:
540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
541 break;
542 case vmIntrinsics::_negateExactI:
543 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
544 break;
545 case vmIntrinsics::_negateExactL:
546 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
547 break;
548 case vmIntrinsics::_multiplyExactI:
549 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
550 break;
551 case vmIntrinsics::_multiplyExactL:
552 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
553 break;
554
555 default:
556 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
557 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
558 break;
559 }
560
561 // -XX:-InlineClassNatives disables natives from the Class class.
562 // The flag applies to all reflective calls, notably Array.newArray
563 // (visible to Java programmers as Array.newInstance).
564 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
565 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
566 if (!InlineClassNatives) return NULL;
567 }
568
569 // -XX:-InlineThreadNatives disables natives from the Thread class.
570 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
571 if (!InlineThreadNatives) return NULL;
572 }
573
574 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
575 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
576 m->holder()->name() == ciSymbol::java_lang_Float() ||
577 m->holder()->name() == ciSymbol::java_lang_Double()) {
578 if (!InlineMathNatives) return NULL;
579 }
580
581 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
582 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
583 if (!InlineUnsafeOps) return NULL;
584 }
585
586 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
587 }
588
589 //----------------------register_library_intrinsics-----------------------
590 // Initialize this file's data structures, for each Compile instance.
591 void Compile::register_library_intrinsics() {
592 // Nothing to do here.
593 }
594
595 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
596 LibraryCallKit kit(jvms, this);
597 Compile* C = kit.C;
598 int nodes = C->unique();
599 #ifndef PRODUCT
600 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
601 char buf[1000];
602 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
603 tty->print_cr("Intrinsic %s", str);
604 }
605 #endif
606 ciMethod* callee = kit.callee();
607 const int bci = kit.bci();
608
609 // Try to inline the intrinsic.
610 if (kit.try_to_inline()) {
611 if (C->print_intrinsics() || C->print_inlining()) {
612 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
613 }
614 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
615 if (C->log()) {
616 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
617 vmIntrinsics::name_at(intrinsic_id()),
618 (is_virtual() ? " virtual='1'" : ""),
619 C->unique() - nodes);
620 }
621 // Push the result from the inlined method onto the stack.
622 kit.push_result();
623 C->print_inlining_update(this);
624 return kit.transfer_exceptions_into_jvms();
625 }
626
627 // The intrinsic bailed out
628 if (C->print_intrinsics() || C->print_inlining()) {
629 if (jvms->has_method()) {
630 // Not a root compile.
631 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
632 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
633 } else {
634 // Root compile
635 tty->print("Did not generate intrinsic %s%s at bci:%d in",
636 vmIntrinsics::name_at(intrinsic_id()),
637 (is_virtual() ? " (virtual)" : ""), bci);
638 }
639 }
640 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
641 C->print_inlining_update(this);
642 return NULL;
643 }
644
645 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
646 LibraryCallKit kit(jvms, this);
647 Compile* C = kit.C;
648 int nodes = C->unique();
649 #ifndef PRODUCT
650 assert(is_predicted(), "sanity");
651 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
652 char buf[1000];
653 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
654 tty->print_cr("Predicate for intrinsic %s", str);
655 }
656 #endif
657 ciMethod* callee = kit.callee();
658 const int bci = kit.bci();
659
660 Node* slow_ctl = kit.try_to_predicate();
661 if (!kit.failing()) {
662 if (C->print_intrinsics() || C->print_inlining()) {
663 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
664 }
665 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
666 if (C->log()) {
667 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
668 vmIntrinsics::name_at(intrinsic_id()),
669 (is_virtual() ? " virtual='1'" : ""),
670 C->unique() - nodes);
671 }
672 return slow_ctl; // Could be NULL if the check folds.
673 }
674
675 // The intrinsic bailed out
676 if (C->print_intrinsics() || C->print_inlining()) {
677 if (jvms->has_method()) {
678 // Not a root compile.
679 const char* msg = "failed to generate predicate for intrinsic";
680 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
681 } else {
682 // Root compile
683 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
684 vmIntrinsics::name_at(intrinsic_id()),
685 (is_virtual() ? " (virtual)" : ""), bci);
686 }
687 }
688 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
689 return NULL;
690 }
691
692 bool LibraryCallKit::try_to_inline() {
693 // Handle symbolic names for otherwise undistinguished boolean switches:
694 const bool is_store = true;
695 const bool is_native_ptr = true;
696 const bool is_static = true;
697 const bool is_volatile = true;
698
699 if (!jvms()->has_method()) {
700 // Root JVMState has a null method.
701 assert(map()->memory()->Opcode() == Op_Parm, "");
702 // Insert the memory aliasing node
703 set_all_memory(reset_memory());
704 }
705 assert(merged_memory(), "");
706
707
708 switch (intrinsic_id()) {
709 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
710 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
711 case vmIntrinsics::_getClass: return inline_native_getClass();
712
713 case vmIntrinsics::_dsin:
714 case vmIntrinsics::_dcos:
715 case vmIntrinsics::_dtan:
716 case vmIntrinsics::_dabs:
717 case vmIntrinsics::_datan2:
718 case vmIntrinsics::_dsqrt:
719 case vmIntrinsics::_dexp:
720 case vmIntrinsics::_dlog:
721 case vmIntrinsics::_dlog10:
722 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
723
724 case vmIntrinsics::_min:
725 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
726
727 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
728 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
729 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
730 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
731 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
732 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
733 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
734 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
735 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
736 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
737 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
738 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
739
740 case vmIntrinsics::_arraycopy: return inline_arraycopy();
741
742 case vmIntrinsics::_compareTo: return inline_string_compareTo();
743 case vmIntrinsics::_indexOf: return inline_string_indexOf();
744 case vmIntrinsics::_equals: return inline_string_equals();
745
746 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
747 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
748 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
749 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
750 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
751 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
752 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
753 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
754 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
755
756 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
757 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
758 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
759 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
760 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
761 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
762 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
763 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
764 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
765
766 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
767 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
768 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
769 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
770 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
771 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
772 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
773 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
774
775 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
776 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
777 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
778 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
779 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
780 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
781 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
782 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
783
784 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
785 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
786 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
787 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
788 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
789 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
790 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
791 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
792 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
793
794 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
795 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
796 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
797 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
798 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
799 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
800 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
801 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
802 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
803
804 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
805 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
806 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
807 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
808
809 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
810 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
811 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
812
813 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
814 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
815 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
816
817 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
818 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
819 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
820 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
821 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
822
823 case vmIntrinsics::_loadFence:
824 case vmIntrinsics::_storeFence:
825 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
826
827 case vmIntrinsics::_currentThread: return inline_native_currentThread();
828 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
829
830 #ifdef TRACE_HAVE_INTRINSICS
831 case vmIntrinsics::_classID: return inline_native_classID();
832 case vmIntrinsics::_threadID: return inline_native_threadID();
833 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
834 #endif
835 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
836 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
837 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
838 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
839 case vmIntrinsics::_newArray: return inline_native_newArray();
840 case vmIntrinsics::_getLength: return inline_native_getLength();
841 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
842 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
843 case vmIntrinsics::_equalsC: return inline_array_equals();
844 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
845
846 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
847
848 case vmIntrinsics::_isInstance:
849 case vmIntrinsics::_getModifiers:
850 case vmIntrinsics::_isInterface:
851 case vmIntrinsics::_isArray:
852 case vmIntrinsics::_isPrimitive:
853 case vmIntrinsics::_getSuperclass:
854 case vmIntrinsics::_getComponentType:
855 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
856
857 case vmIntrinsics::_floatToRawIntBits:
858 case vmIntrinsics::_floatToIntBits:
859 case vmIntrinsics::_intBitsToFloat:
860 case vmIntrinsics::_doubleToRawLongBits:
861 case vmIntrinsics::_doubleToLongBits:
862 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
863
864 case vmIntrinsics::_numberOfLeadingZeros_i:
865 case vmIntrinsics::_numberOfLeadingZeros_l:
866 case vmIntrinsics::_numberOfTrailingZeros_i:
867 case vmIntrinsics::_numberOfTrailingZeros_l:
868 case vmIntrinsics::_bitCount_i:
869 case vmIntrinsics::_bitCount_l:
870 case vmIntrinsics::_reverseBytes_i:
871 case vmIntrinsics::_reverseBytes_l:
872 case vmIntrinsics::_reverseBytes_s:
873 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
874
875 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
876
877 case vmIntrinsics::_Reference_get: return inline_reference_get();
878
879 case vmIntrinsics::_aescrypt_encryptBlock:
880 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
881
882 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
883 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
884 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
885
886 case vmIntrinsics::_encodeISOArray:
887 return inline_encodeISOArray();
888
889 case vmIntrinsics::_updateCRC32:
890 return inline_updateCRC32();
891 case vmIntrinsics::_updateBytesCRC32:
892 return inline_updateBytesCRC32();
893 case vmIntrinsics::_updateByteBufferCRC32:
894 return inline_updateByteBufferCRC32();
895
896 default:
897 // If you get here, it may be that someone has added a new intrinsic
898 // to the list in vmSymbols.hpp without implementing it here.
899 #ifndef PRODUCT
900 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
901 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
902 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
903 }
904 #endif
905 return false;
906 }
907 }
908
909 Node* LibraryCallKit::try_to_predicate() {
910 if (!jvms()->has_method()) {
911 // Root JVMState has a null method.
912 assert(map()->memory()->Opcode() == Op_Parm, "");
913 // Insert the memory aliasing node
914 set_all_memory(reset_memory());
915 }
916 assert(merged_memory(), "");
917
918 switch (intrinsic_id()) {
919 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
920 return inline_cipherBlockChaining_AESCrypt_predicate(false);
921 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
922 return inline_cipherBlockChaining_AESCrypt_predicate(true);
923
924 default:
925 // If you get here, it may be that someone has added a new intrinsic
926 // to the list in vmSymbols.hpp without implementing it here.
927 #ifndef PRODUCT
928 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
929 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
930 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
931 }
932 #endif
933 Node* slow_ctl = control();
934 set_control(top()); // No fast path instrinsic
935 return slow_ctl;
936 }
937 }
938
939 //------------------------------set_result-------------------------------
940 // Helper function for finishing intrinsics.
941 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
942 record_for_igvn(region);
943 set_control(_gvn.transform(region));
944 set_result( _gvn.transform(value));
945 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
946 }
947
948 //------------------------------generate_guard---------------------------
949 // Helper function for generating guarded fast-slow graph structures.
950 // The given 'test', if true, guards a slow path. If the test fails
951 // then a fast path can be taken. (We generally hope it fails.)
952 // In all cases, GraphKit::control() is updated to the fast path.
953 // The returned value represents the control for the slow path.
954 // The return value is never 'top'; it is either a valid control
955 // or NULL if it is obvious that the slow path can never be taken.
956 // Also, if region and the slow control are not NULL, the slow edge
957 // is appended to the region.
958 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
959 if (stopped()) {
960 // Already short circuited.
961 return NULL;
962 }
963
964 // Build an if node and its projections.
965 // If test is true we take the slow path, which we assume is uncommon.
966 if (_gvn.type(test) == TypeInt::ZERO) {
967 // The slow branch is never taken. No need to build this guard.
968 return NULL;
969 }
970
971 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
972
973 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
974 if (if_slow == top()) {
975 // The slow branch is never taken. No need to build this guard.
976 return NULL;
977 }
978
979 if (region != NULL)
980 region->add_req(if_slow);
981
982 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
983 set_control(if_fast);
984
985 return if_slow;
986 }
987
988 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
989 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
990 }
991 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
992 return generate_guard(test, region, PROB_FAIR);
993 }
994
995 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
996 Node* *pos_index) {
997 if (stopped())
998 return NULL; // already stopped
999 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1000 return NULL; // index is already adequately typed
1001 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1002 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1003 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1004 if (is_neg != NULL && pos_index != NULL) {
1005 // Emulate effect of Parse::adjust_map_after_if.
1006 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1007 ccast->set_req(0, control());
1008 (*pos_index) = _gvn.transform(ccast);
1009 }
1010 return is_neg;
1011 }
1012
1013 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1014 Node* *pos_index) {
1015 if (stopped())
1016 return NULL; // already stopped
1017 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1018 return NULL; // index is already adequately typed
1019 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1020 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1021 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1022 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1023 if (is_notp != NULL && pos_index != NULL) {
1024 // Emulate effect of Parse::adjust_map_after_if.
1025 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1026 ccast->set_req(0, control());
1027 (*pos_index) = _gvn.transform(ccast);
1028 }
1029 return is_notp;
1030 }
1031
1032 // Make sure that 'position' is a valid limit index, in [0..length].
1033 // There are two equivalent plans for checking this:
1034 // A. (offset + copyLength) unsigned<= arrayLength
1035 // B. offset <= (arrayLength - copyLength)
1036 // We require that all of the values above, except for the sum and
1037 // difference, are already known to be non-negative.
1038 // Plan A is robust in the face of overflow, if offset and copyLength
1039 // are both hugely positive.
1040 //
1041 // Plan B is less direct and intuitive, but it does not overflow at
1042 // all, since the difference of two non-negatives is always
1043 // representable. Whenever Java methods must perform the equivalent
1044 // check they generally use Plan B instead of Plan A.
1045 // For the moment we use Plan A.
1046 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1047 Node* subseq_length,
1048 Node* array_length,
1049 RegionNode* region) {
1050 if (stopped())
1051 return NULL; // already stopped
1052 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1053 if (zero_offset && subseq_length->eqv_uncast(array_length))
1054 return NULL; // common case of whole-array copy
1055 Node* last = subseq_length;
1056 if (!zero_offset) // last += offset
1057 last = _gvn.transform(new (C) AddINode(last, offset));
1058 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1059 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1060 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1061 return is_over;
1062 }
1063
1064
1065 //--------------------------generate_current_thread--------------------
1066 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1067 ciKlass* thread_klass = env()->Thread_klass();
1068 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1069 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1070 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1071 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1072 tls_output = thread;
1073 return threadObj;
1074 }
1075
1076
1077 //------------------------------make_string_method_node------------------------
1078 // Helper method for String intrinsic functions. This version is called
1079 // with str1 and str2 pointing to String object nodes.
1080 //
1081 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1082 Node* no_ctrl = NULL;
1083
1084 // Get start addr of string
1085 Node* str1_value = load_String_value(no_ctrl, str1);
1086 Node* str1_offset = load_String_offset(no_ctrl, str1);
1087 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1088
1089 // Get length of string 1
1090 Node* str1_len = load_String_length(no_ctrl, str1);
1091
1092 Node* str2_value = load_String_value(no_ctrl, str2);
1093 Node* str2_offset = load_String_offset(no_ctrl, str2);
1094 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1095
1096 Node* str2_len = NULL;
1097 Node* result = NULL;
1098
1099 switch (opcode) {
1100 case Op_StrIndexOf:
1101 // Get length of string 2
1102 str2_len = load_String_length(no_ctrl, str2);
1103
1104 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1105 str1_start, str1_len, str2_start, str2_len);
1106 break;
1107 case Op_StrComp:
1108 // Get length of string 2
1109 str2_len = load_String_length(no_ctrl, str2);
1110
1111 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1112 str1_start, str1_len, str2_start, str2_len);
1113 break;
1114 case Op_StrEquals:
1115 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1116 str1_start, str2_start, str1_len);
1117 break;
1118 default:
1119 ShouldNotReachHere();
1120 return NULL;
1121 }
1122
1123 // All these intrinsics have checks.
1124 C->set_has_split_ifs(true); // Has chance for split-if optimization
1125
1126 return _gvn.transform(result);
1127 }
1128
1129 // Helper method for String intrinsic functions. This version is called
1130 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1131 // to Int nodes containing the lenghts of str1 and str2.
1132 //
1133 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1134 Node* result = NULL;
1135 switch (opcode) {
1136 case Op_StrIndexOf:
1137 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1138 str1_start, cnt1, str2_start, cnt2);
1139 break;
1140 case Op_StrComp:
1141 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1142 str1_start, cnt1, str2_start, cnt2);
1143 break;
1144 case Op_StrEquals:
1145 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1146 str1_start, str2_start, cnt1);
1147 break;
1148 default:
1149 ShouldNotReachHere();
1150 return NULL;
1151 }
1152
1153 // All these intrinsics have checks.
1154 C->set_has_split_ifs(true); // Has chance for split-if optimization
1155
1156 return _gvn.transform(result);
1157 }
1158
1159 //------------------------------inline_string_compareTo------------------------
1160 // public int java.lang.String.compareTo(String anotherString);
1161 bool LibraryCallKit::inline_string_compareTo() {
1162 Node* receiver = null_check(argument(0));
1163 Node* arg = null_check(argument(1));
1164 if (stopped()) {
1165 return true;
1166 }
1167 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1168 return true;
1169 }
1170
1171 //------------------------------inline_string_equals------------------------
1172 bool LibraryCallKit::inline_string_equals() {
1173 Node* receiver = null_check_receiver();
1174 // NOTE: Do not null check argument for String.equals() because spec
1175 // allows to specify NULL as argument.
1176 Node* argument = this->argument(1);
1177 if (stopped()) {
1178 return true;
1179 }
1180
1181 // paths (plus control) merge
1182 RegionNode* region = new (C) RegionNode(5);
1183 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1184
1185 // does source == target string?
1186 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1187 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1188
1189 Node* if_eq = generate_slow_guard(bol, NULL);
1190 if (if_eq != NULL) {
1191 // receiver == argument
1192 phi->init_req(2, intcon(1));
1193 region->init_req(2, if_eq);
1194 }
1195
1196 // get String klass for instanceOf
1197 ciInstanceKlass* klass = env()->String_klass();
1198
1199 if (!stopped()) {
1200 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1201 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1202 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1203
1204 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1205 //instanceOf == true, fallthrough
1206
1207 if (inst_false != NULL) {
1208 phi->init_req(3, intcon(0));
1209 region->init_req(3, inst_false);
1210 }
1211 }
1212
1213 if (!stopped()) {
1214 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1215
1216 // Properly cast the argument to String
1217 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1218 // This path is taken only when argument's type is String:NotNull.
1219 argument = cast_not_null(argument, false);
1220
1221 Node* no_ctrl = NULL;
1222
1223 // Get start addr of receiver
1224 Node* receiver_val = load_String_value(no_ctrl, receiver);
1225 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1226 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1227
1228 // Get length of receiver
1229 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1230
1231 // Get start addr of argument
1232 Node* argument_val = load_String_value(no_ctrl, argument);
1233 Node* argument_offset = load_String_offset(no_ctrl, argument);
1234 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1235
1236 // Get length of argument
1237 Node* argument_cnt = load_String_length(no_ctrl, argument);
1238
1239 // Check for receiver count != argument count
1240 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1241 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1242 Node* if_ne = generate_slow_guard(bol, NULL);
1243 if (if_ne != NULL) {
1244 phi->init_req(4, intcon(0));
1245 region->init_req(4, if_ne);
1246 }
1247
1248 // Check for count == 0 is done by assembler code for StrEquals.
1249
1250 if (!stopped()) {
1251 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1252 phi->init_req(1, equals);
1253 region->init_req(1, control());
1254 }
1255 }
1256
1257 // post merge
1258 set_control(_gvn.transform(region));
1259 record_for_igvn(region);
1260
1261 set_result(_gvn.transform(phi));
1262 return true;
1263 }
1264
1265 //------------------------------inline_array_equals----------------------------
1266 bool LibraryCallKit::inline_array_equals() {
1267 Node* arg1 = argument(0);
1268 Node* arg2 = argument(1);
1269 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1270 return true;
1271 }
1272
1273 // Java version of String.indexOf(constant string)
1274 // class StringDecl {
1275 // StringDecl(char[] ca) {
1276 // offset = 0;
1277 // count = ca.length;
1278 // value = ca;
1279 // }
1280 // int offset;
1281 // int count;
1282 // char[] value;
1283 // }
1284 //
1285 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1286 // int targetOffset, int cache_i, int md2) {
1287 // int cache = cache_i;
1288 // int sourceOffset = string_object.offset;
1289 // int sourceCount = string_object.count;
1290 // int targetCount = target_object.length;
1291 //
1292 // int targetCountLess1 = targetCount - 1;
1293 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1294 //
1295 // char[] source = string_object.value;
1296 // char[] target = target_object;
1297 // int lastChar = target[targetCountLess1];
1298 //
1299 // outer_loop:
1300 // for (int i = sourceOffset; i < sourceEnd; ) {
1301 // int src = source[i + targetCountLess1];
1302 // if (src == lastChar) {
1303 // // With random strings and a 4-character alphabet,
1304 // // reverse matching at this point sets up 0.8% fewer
1305 // // frames, but (paradoxically) makes 0.3% more probes.
1306 // // Since those probes are nearer the lastChar probe,
1307 // // there is may be a net D$ win with reverse matching.
1308 // // But, reversing loop inhibits unroll of inner loop
1309 // // for unknown reason. So, does running outer loop from
1310 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1311 // for (int j = 0; j < targetCountLess1; j++) {
1312 // if (target[targetOffset + j] != source[i+j]) {
1313 // if ((cache & (1 << source[i+j])) == 0) {
1314 // if (md2 < j+1) {
1315 // i += j+1;
1316 // continue outer_loop;
1317 // }
1318 // }
1319 // i += md2;
1320 // continue outer_loop;
1321 // }
1322 // }
1323 // return i - sourceOffset;
1324 // }
1325 // if ((cache & (1 << src)) == 0) {
1326 // i += targetCountLess1;
1327 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1328 // i++;
1329 // }
1330 // return -1;
1331 // }
1332
1333 //------------------------------string_indexOf------------------------
1334 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1335 jint cache_i, jint md2_i) {
1336
1337 Node* no_ctrl = NULL;
1338 float likely = PROB_LIKELY(0.9);
1339 float unlikely = PROB_UNLIKELY(0.9);
1340
1341 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1342
1343 Node* source = load_String_value(no_ctrl, string_object);
1344 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1345 Node* sourceCount = load_String_length(no_ctrl, string_object);
1346
1347 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1348 jint target_length = target_array->length();
1349 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1350 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1351
1352 // String.value field is known to be @Stable.
1353 if (UseImplicitStableValues) {
1354 target = cast_array_to_stable(target, target_type);
1355 }
1356
1357 IdealKit kit(this, false, true);
1358 #define __ kit.
1359 Node* zero = __ ConI(0);
1360 Node* one = __ ConI(1);
1361 Node* cache = __ ConI(cache_i);
1362 Node* md2 = __ ConI(md2_i);
1363 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1364 Node* targetCount = __ ConI(target_length);
1365 Node* targetCountLess1 = __ ConI(target_length - 1);
1366 Node* targetOffset = __ ConI(targetOffset_i);
1367 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1368
1369 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1370 Node* outer_loop = __ make_label(2 /* goto */);
1371 Node* return_ = __ make_label(1);
1372
1373 __ set(rtn,__ ConI(-1));
1374 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1375 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1376 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1377 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1378 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1379 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1380 Node* tpj = __ AddI(targetOffset, __ value(j));
1381 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1382 Node* ipj = __ AddI(__ value(i), __ value(j));
1383 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1384 __ if_then(targ, BoolTest::ne, src2); {
1385 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1386 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1387 __ increment(i, __ AddI(__ value(j), one));
1388 __ goto_(outer_loop);
1389 } __ end_if(); __ dead(j);
1390 }__ end_if(); __ dead(j);
1391 __ increment(i, md2);
1392 __ goto_(outer_loop);
1393 }__ end_if();
1394 __ increment(j, one);
1395 }__ end_loop(); __ dead(j);
1396 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1397 __ goto_(return_);
1398 }__ end_if();
1399 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1400 __ increment(i, targetCountLess1);
1401 }__ end_if();
1402 __ increment(i, one);
1403 __ bind(outer_loop);
1404 }__ end_loop(); __ dead(i);
1405 __ bind(return_);
1406
1407 // Final sync IdealKit and GraphKit.
1408 final_sync(kit);
1409 Node* result = __ value(rtn);
1410 #undef __
1411 C->set_has_loops(true);
1412 return result;
1413 }
1414
1415 //------------------------------inline_string_indexOf------------------------
1416 bool LibraryCallKit::inline_string_indexOf() {
1417 Node* receiver = argument(0);
1418 Node* arg = argument(1);
1419
1420 Node* result;
1421 // Disable the use of pcmpestri until it can be guaranteed that
1422 // the load doesn't cross into the uncommited space.
1423 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1424 UseSSE42Intrinsics) {
1425 // Generate SSE4.2 version of indexOf
1426 // We currently only have match rules that use SSE4.2
1427
1428 receiver = null_check(receiver);
1429 arg = null_check(arg);
1430 if (stopped()) {
1431 return true;
1432 }
1433
1434 ciInstanceKlass* str_klass = env()->String_klass();
1435 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1436
1437 // Make the merge point
1438 RegionNode* result_rgn = new (C) RegionNode(4);
1439 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1440 Node* no_ctrl = NULL;
1441
1442 // Get start addr of source string
1443 Node* source = load_String_value(no_ctrl, receiver);
1444 Node* source_offset = load_String_offset(no_ctrl, receiver);
1445 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1446
1447 // Get length of source string
1448 Node* source_cnt = load_String_length(no_ctrl, receiver);
1449
1450 // Get start addr of substring
1451 Node* substr = load_String_value(no_ctrl, arg);
1452 Node* substr_offset = load_String_offset(no_ctrl, arg);
1453 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1454
1455 // Get length of source string
1456 Node* substr_cnt = load_String_length(no_ctrl, arg);
1457
1458 // Check for substr count > string count
1459 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1460 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1461 Node* if_gt = generate_slow_guard(bol, NULL);
1462 if (if_gt != NULL) {
1463 result_phi->init_req(2, intcon(-1));
1464 result_rgn->init_req(2, if_gt);
1465 }
1466
1467 if (!stopped()) {
1468 // Check for substr count == 0
1469 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1470 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1471 Node* if_zero = generate_slow_guard(bol, NULL);
1472 if (if_zero != NULL) {
1473 result_phi->init_req(3, intcon(0));
1474 result_rgn->init_req(3, if_zero);
1475 }
1476 }
1477
1478 if (!stopped()) {
1479 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1480 result_phi->init_req(1, result);
1481 result_rgn->init_req(1, control());
1482 }
1483 set_control(_gvn.transform(result_rgn));
1484 record_for_igvn(result_rgn);
1485 result = _gvn.transform(result_phi);
1486
1487 } else { // Use LibraryCallKit::string_indexOf
1488 // don't intrinsify if argument isn't a constant string.
1489 if (!arg->is_Con()) {
1490 return false;
1491 }
1492 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1493 if (str_type == NULL) {
1494 return false;
1495 }
1496 ciInstanceKlass* klass = env()->String_klass();
1497 ciObject* str_const = str_type->const_oop();
1498 if (str_const == NULL || str_const->klass() != klass) {
1499 return false;
1500 }
1501 ciInstance* str = str_const->as_instance();
1502 assert(str != NULL, "must be instance");
1503
1504 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1505 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1506
1507 int o;
1508 int c;
1509 if (java_lang_String::has_offset_field()) {
1510 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1511 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1512 } else {
1513 o = 0;
1514 c = pat->length();
1515 }
1516
1517 // constant strings have no offset and count == length which
1518 // simplifies the resulting code somewhat so lets optimize for that.
1519 if (o != 0 || c != pat->length()) {
1520 return false;
1521 }
1522
1523 receiver = null_check(receiver, T_OBJECT);
1524 // NOTE: No null check on the argument is needed since it's a constant String oop.
1525 if (stopped()) {
1526 return true;
1527 }
1528
1529 // The null string as a pattern always returns 0 (match at beginning of string)
1530 if (c == 0) {
1531 set_result(intcon(0));
1532 return true;
1533 }
1534
1535 // Generate default indexOf
1536 jchar lastChar = pat->char_at(o + (c - 1));
1537 int cache = 0;
1538 int i;
1539 for (i = 0; i < c - 1; i++) {
1540 assert(i < pat->length(), "out of range");
1541 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1542 }
1543
1544 int md2 = c;
1545 for (i = 0; i < c - 1; i++) {
1546 assert(i < pat->length(), "out of range");
1547 if (pat->char_at(o + i) == lastChar) {
1548 md2 = (c - 1) - i;
1549 }
1550 }
1551
1552 result = string_indexOf(receiver, pat, o, cache, md2);
1553 }
1554 set_result(result);
1555 return true;
1556 }
1557
1558 //--------------------------round_double_node--------------------------------
1559 // Round a double node if necessary.
1560 Node* LibraryCallKit::round_double_node(Node* n) {
1561 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1562 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1563 return n;
1564 }
1565
1566 //------------------------------inline_math-----------------------------------
1567 // public static double Math.abs(double)
1568 // public static double Math.sqrt(double)
1569 // public static double Math.log(double)
1570 // public static double Math.log10(double)
1571 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1572 Node* arg = round_double_node(argument(0));
1573 Node* n;
1574 switch (id) {
1575 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1576 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1577 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1578 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1579 default: fatal_unexpected_iid(id); break;
1580 }
1581 set_result(_gvn.transform(n));
1582 return true;
1583 }
1584
1585 //------------------------------inline_trig----------------------------------
1586 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1587 // argument reduction which will turn into a fast/slow diamond.
1588 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1589 Node* arg = round_double_node(argument(0));
1590 Node* n = NULL;
1591
1592 switch (id) {
1593 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1594 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1595 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1596 default: fatal_unexpected_iid(id); break;
1597 }
1598 n = _gvn.transform(n);
1599
1600 // Rounding required? Check for argument reduction!
1601 if (Matcher::strict_fp_requires_explicit_rounding) {
1602 static const double pi_4 = 0.7853981633974483;
1603 static const double neg_pi_4 = -0.7853981633974483;
1604 // pi/2 in 80-bit extended precision
1605 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1606 // -pi/2 in 80-bit extended precision
1607 // 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};
1608 // Cutoff value for using this argument reduction technique
1609 //static const double pi_2_minus_epsilon = 1.564660403643354;
1610 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1611
1612 // Pseudocode for sin:
1613 // if (x <= Math.PI / 4.0) {
1614 // if (x >= -Math.PI / 4.0) return fsin(x);
1615 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1616 // } else {
1617 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1618 // }
1619 // return StrictMath.sin(x);
1620
1621 // Pseudocode for cos:
1622 // if (x <= Math.PI / 4.0) {
1623 // if (x >= -Math.PI / 4.0) return fcos(x);
1624 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1625 // } else {
1626 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1627 // }
1628 // return StrictMath.cos(x);
1629
1630 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1631 // requires a special machine instruction to load it. Instead we'll try
1632 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1633 // probably do the math inside the SIN encoding.
1634
1635 // Make the merge point
1636 RegionNode* r = new (C) RegionNode(3);
1637 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1638
1639 // Flatten arg so we need only 1 test
1640 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1641 // Node for PI/4 constant
1642 Node *pi4 = makecon(TypeD::make(pi_4));
1643 // Check PI/4 : abs(arg)
1644 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1645 // Check: If PI/4 < abs(arg) then go slow
1646 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1647 // Branch either way
1648 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1649 set_control(opt_iff(r,iff));
1650
1651 // Set fast path result
1652 phi->init_req(2, n);
1653
1654 // Slow path - non-blocking leaf call
1655 Node* call = NULL;
1656 switch (id) {
1657 case vmIntrinsics::_dsin:
1658 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1659 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1660 "Sin", NULL, arg, top());
1661 break;
1662 case vmIntrinsics::_dcos:
1663 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1664 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1665 "Cos", NULL, arg, top());
1666 break;
1667 case vmIntrinsics::_dtan:
1668 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1669 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1670 "Tan", NULL, arg, top());
1671 break;
1672 }
1673 assert(control()->in(0) == call, "");
1674 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1675 r->init_req(1, control());
1676 phi->init_req(1, slow_result);
1677
1678 // Post-merge
1679 set_control(_gvn.transform(r));
1680 record_for_igvn(r);
1681 n = _gvn.transform(phi);
1682
1683 C->set_has_split_ifs(true); // Has chance for split-if optimization
1684 }
1685 set_result(n);
1686 return true;
1687 }
1688
1689 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1690 //-------------------
1691 //result=(result.isNaN())? funcAddr():result;
1692 // Check: If isNaN() by checking result!=result? then either trap
1693 // or go to runtime
1694 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1695 // Build the boolean node
1696 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1697
1698 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1699 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1700 // The pow or exp intrinsic returned a NaN, which requires a call
1701 // to the runtime. Recompile with the runtime call.
1702 uncommon_trap(Deoptimization::Reason_intrinsic,
1703 Deoptimization::Action_make_not_entrant);
1704 }
1705 return result;
1706 } else {
1707 // If this inlining ever returned NaN in the past, we compile a call
1708 // to the runtime to properly handle corner cases
1709
1710 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1711 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1712 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1713
1714 if (!if_slow->is_top()) {
1715 RegionNode* result_region = new (C) RegionNode(3);
1716 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1717
1718 result_region->init_req(1, if_fast);
1719 result_val->init_req(1, result);
1720
1721 set_control(if_slow);
1722
1723 const TypePtr* no_memory_effects = NULL;
1724 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1725 no_memory_effects,
1726 x, top(), y, y ? top() : NULL);
1727 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1728 #ifdef ASSERT
1729 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1730 assert(value_top == top(), "second value must be top");
1731 #endif
1732
1733 result_region->init_req(2, control());
1734 result_val->init_req(2, value);
1735 set_control(_gvn.transform(result_region));
1736 // Make sure type of PhiNode is recorded
1737 _gvn.transform(result_val);
1738 return result_val;
1739 } else {
1740 return result;
1741 }
1742 }
1743 }
1744
1745 //------------------------------inline_exp-------------------------------------
1746 // Inline exp instructions, if possible. The Intel hardware only misses
1747 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1748 bool LibraryCallKit::inline_exp() {
1749 Node* arg = round_double_node(argument(0));
1750 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1751
1752 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1753 set_result(n);
1754
1755 C->set_has_split_ifs(true); // Has chance for split-if optimization
1756 return true;
1757 }
1758
1759 //------------------------------inline_pow-------------------------------------
1760 // Inline power instructions, if possible.
1761 bool LibraryCallKit::inline_pow() {
1762 // Pseudocode for pow
1763 // if (y == 2) {
1764 // return x * x;
1765 // } else {
1766 // if (x <= 0.0) {
1767 // long longy = (long)y;
1768 // if ((double)longy == y) { // if y is long
1769 // if (y + 1 == y) longy = 0; // huge number: even
1770 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1771 // } else {
1772 // result = NaN;
1773 // }
1774 // } else {
1775 // result = DPow(x,y);
1776 // }
1777 // if (result != result)? {
1778 // result = uncommon_trap() or runtime_call();
1779 // }
1780 // return result;
1781 // }
1782
1783 Node* x = round_double_node(argument(0));
1784 Node* y = round_double_node(argument(2));
1785
1786 Node* result = NULL;
1787
1788 Node* const_two_node = makecon(TypeD::make(2.0));
1789 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1790 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1791 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1792 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1793 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1794
1795 RegionNode* region_node = new (C) RegionNode(3);
1796 region_node->init_req(1, if_true);
1797
1798 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1799 // special case for x^y where y == 2, we can convert it to x * x
1800 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1801
1802 // set control to if_false since we will now process the false branch
1803 set_control(if_false);
1804
1805 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1806 // Short form: skip the fancy tests and just check for NaN result.
1807 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1808 } else {
1809 // If this inlining ever returned NaN in the past, include all
1810 // checks + call to the runtime.
1811
1812 // Set the merge point for If node with condition of (x <= 0.0)
1813 // There are four possible paths to region node and phi node
1814 RegionNode *r = new (C) RegionNode(4);
1815 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1816
1817 // Build the first if node: if (x <= 0.0)
1818 // Node for 0 constant
1819 Node *zeronode = makecon(TypeD::ZERO);
1820 // Check x:0
1821 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1822 // Check: If (x<=0) then go complex path
1823 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1824 // Branch either way
1825 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1826 // Fast path taken; set region slot 3
1827 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1828 r->init_req(3,fast_taken); // Capture fast-control
1829
1830 // Fast path not-taken, i.e. slow path
1831 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1832
1833 // Set fast path result
1834 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1835 phi->init_req(3, fast_result);
1836
1837 // Complex path
1838 // Build the second if node (if y is long)
1839 // Node for (long)y
1840 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1841 // Node for (double)((long) y)
1842 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1843 // Check (double)((long) y) : y
1844 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1845 // Check if (y isn't long) then go to slow path
1846
1847 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1848 // Branch either way
1849 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1850 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1851
1852 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1853
1854 // Calculate DPow(abs(x), y)*(1 & (long)y)
1855 // Node for constant 1
1856 Node *conone = longcon(1);
1857 // 1& (long)y
1858 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1859
1860 // A huge number is always even. Detect a huge number by checking
1861 // if y + 1 == y and set integer to be tested for parity to 0.
1862 // Required for corner case:
1863 // (long)9.223372036854776E18 = max_jlong
1864 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1865 // max_jlong is odd but 9.223372036854776E18 is even
1866 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1867 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1868 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1869 Node* correctedsign = NULL;
1870 if (ConditionalMoveLimit != 0) {
1871 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1872 } else {
1873 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1874 RegionNode *r = new (C) RegionNode(3);
1875 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1876 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1877 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1878 phi->init_req(1, signnode);
1879 phi->init_req(2, longcon(0));
1880 correctedsign = _gvn.transform(phi);
1881 ylong_path = _gvn.transform(r);
1882 record_for_igvn(r);
1883 }
1884
1885 // zero node
1886 Node *conzero = longcon(0);
1887 // Check (1&(long)y)==0?
1888 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1889 // Check if (1&(long)y)!=0?, if so the result is negative
1890 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1891 // abs(x)
1892 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1893 // abs(x)^y
1894 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1895 // -abs(x)^y
1896 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1897 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1898 Node *signresult = NULL;
1899 if (ConditionalMoveLimit != 0) {
1900 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1901 } else {
1902 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1903 RegionNode *r = new (C) RegionNode(3);
1904 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1905 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1906 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1907 phi->init_req(1, absxpowy);
1908 phi->init_req(2, negabsxpowy);
1909 signresult = _gvn.transform(phi);
1910 ylong_path = _gvn.transform(r);
1911 record_for_igvn(r);
1912 }
1913 // Set complex path fast result
1914 r->init_req(2, ylong_path);
1915 phi->init_req(2, signresult);
1916
1917 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1918 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1919 r->init_req(1,slow_path);
1920 phi->init_req(1,slow_result);
1921
1922 // Post merge
1923 set_control(_gvn.transform(r));
1924 record_for_igvn(r);
1925 result = _gvn.transform(phi);
1926 }
1927
1928 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1929
1930 // control from finish_pow_exp is now input to the region node
1931 region_node->set_req(2, control());
1932 // the result from finish_pow_exp is now input to the phi node
1933 phi_node->init_req(2, _gvn.transform(result));
1934 set_control(_gvn.transform(region_node));
1935 record_for_igvn(region_node);
1936 set_result(_gvn.transform(phi_node));
1937
1938 C->set_has_split_ifs(true); // Has chance for split-if optimization
1939 return true;
1940 }
1941
1942 //------------------------------runtime_math-----------------------------
1943 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1944 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1945 "must be (DD)D or (D)D type");
1946
1947 // Inputs
1948 Node* a = round_double_node(argument(0));
1949 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1950
1951 const TypePtr* no_memory_effects = NULL;
1952 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1953 no_memory_effects,
1954 a, top(), b, b ? top() : NULL);
1955 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1956 #ifdef ASSERT
1957 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1958 assert(value_top == top(), "second value must be top");
1959 #endif
1960
1961 set_result(value);
1962 return true;
1963 }
1964
1965 //------------------------------inline_math_native-----------------------------
1966 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1967 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1968 switch (id) {
1969 // These intrinsics are not properly supported on all hardware
1970 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1971 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1972 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1973 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1974 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1975 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1976
1977 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1978 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1979 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1980 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1981
1982 // These intrinsics are supported on all hardware
1983 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1984 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1985
1986 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1987 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1988 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1989 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1990 #undef FN_PTR
1991
1992 // These intrinsics are not yet correctly implemented
1993 case vmIntrinsics::_datan2:
1994 return false;
1995
1996 default:
1997 fatal_unexpected_iid(id);
1998 return false;
1999 }
2000 }
2001
2002 static bool is_simple_name(Node* n) {
2003 return (n->req() == 1 // constant
2004 || (n->is_Type() && n->as_Type()->type()->singleton())
2005 || n->is_Proj() // parameter or return value
2006 || n->is_Phi() // local of some sort
2007 );
2008 }
2009
2010 //----------------------------inline_min_max-----------------------------------
2011 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2012 set_result(generate_min_max(id, argument(0), argument(1)));
2013 return true;
2014 }
2015
2016 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2017 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2018 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2019 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2020 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2021
2022 {
2023 PreserveJVMState pjvms(this);
2024 PreserveReexecuteState preexecs(this);
2025 jvms()->set_should_reexecute(true);
2026
2027 set_control(slow_path);
2028 set_i_o(i_o());
2029
2030 uncommon_trap(Deoptimization::Reason_intrinsic,
2031 Deoptimization::Action_none);
2032 }
2033
2034 set_control(fast_path);
2035 set_result(math);
2036 }
2037
2038 template <typename OverflowOp>
2039 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2040 typedef typename OverflowOp::MathOp MathOp;
2041
2042 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2043 Node* operation = _gvn.transform( mathOp );
2044 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2045 inline_math_mathExact(operation, ofcheck);
2046 return true;
2047 }
2048
2049 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2050 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2051 }
2052
2053 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2054 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2055 }
2056
2057 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2058 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2059 }
2060
2061 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2062 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2063 }
2064
2065 bool LibraryCallKit::inline_math_negateExactI() {
2066 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2067 }
2068
2069 bool LibraryCallKit::inline_math_negateExactL() {
2070 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2071 }
2072
2073 bool LibraryCallKit::inline_math_multiplyExactI() {
2074 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2075 }
2076
2077 bool LibraryCallKit::inline_math_multiplyExactL() {
2078 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2079 }
2080
2081 Node*
2082 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2083 // These are the candidate return value:
2084 Node* xvalue = x0;
2085 Node* yvalue = y0;
2086
2087 if (xvalue == yvalue) {
2088 return xvalue;
2089 }
2090
2091 bool want_max = (id == vmIntrinsics::_max);
2092
2093 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2094 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2095 if (txvalue == NULL || tyvalue == NULL) return top();
2096 // This is not really necessary, but it is consistent with a
2097 // hypothetical MaxINode::Value method:
2098 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2099
2100 // %%% This folding logic should (ideally) be in a different place.
2101 // Some should be inside IfNode, and there to be a more reliable
2102 // transformation of ?: style patterns into cmoves. We also want
2103 // more powerful optimizations around cmove and min/max.
2104
2105 // Try to find a dominating comparison of these guys.
2106 // It can simplify the index computation for Arrays.copyOf
2107 // and similar uses of System.arraycopy.
2108 // First, compute the normalized version of CmpI(x, y).
2109 int cmp_op = Op_CmpI;
2110 Node* xkey = xvalue;
2111 Node* ykey = yvalue;
2112 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2113 if (ideal_cmpxy->is_Cmp()) {
2114 // E.g., if we have CmpI(length - offset, count),
2115 // it might idealize to CmpI(length, count + offset)
2116 cmp_op = ideal_cmpxy->Opcode();
2117 xkey = ideal_cmpxy->in(1);
2118 ykey = ideal_cmpxy->in(2);
2119 }
2120
2121 // Start by locating any relevant comparisons.
2122 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2123 Node* cmpxy = NULL;
2124 Node* cmpyx = NULL;
2125 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2126 Node* cmp = start_from->fast_out(k);
2127 if (cmp->outcnt() > 0 && // must have prior uses
2128 cmp->in(0) == NULL && // must be context-independent
2129 cmp->Opcode() == cmp_op) { // right kind of compare
2130 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2131 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2132 }
2133 }
2134
2135 const int NCMPS = 2;
2136 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2137 int cmpn;
2138 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2139 if (cmps[cmpn] != NULL) break; // find a result
2140 }
2141 if (cmpn < NCMPS) {
2142 // Look for a dominating test that tells us the min and max.
2143 int depth = 0; // Limit search depth for speed
2144 Node* dom = control();
2145 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2146 if (++depth >= 100) break;
2147 Node* ifproj = dom;
2148 if (!ifproj->is_Proj()) continue;
2149 Node* iff = ifproj->in(0);
2150 if (!iff->is_If()) continue;
2151 Node* bol = iff->in(1);
2152 if (!bol->is_Bool()) continue;
2153 Node* cmp = bol->in(1);
2154 if (cmp == NULL) continue;
2155 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2156 if (cmps[cmpn] == cmp) break;
2157 if (cmpn == NCMPS) continue;
2158 BoolTest::mask btest = bol->as_Bool()->_test._test;
2159 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2160 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2161 // At this point, we know that 'x btest y' is true.
2162 switch (btest) {
2163 case BoolTest::eq:
2164 // They are proven equal, so we can collapse the min/max.
2165 // Either value is the answer. Choose the simpler.
2166 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2167 return yvalue;
2168 return xvalue;
2169 case BoolTest::lt: // x < y
2170 case BoolTest::le: // x <= y
2171 return (want_max ? yvalue : xvalue);
2172 case BoolTest::gt: // x > y
2173 case BoolTest::ge: // x >= y
2174 return (want_max ? xvalue : yvalue);
2175 }
2176 }
2177 }
2178
2179 // We failed to find a dominating test.
2180 // Let's pick a test that might GVN with prior tests.
2181 Node* best_bol = NULL;
2182 BoolTest::mask best_btest = BoolTest::illegal;
2183 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2184 Node* cmp = cmps[cmpn];
2185 if (cmp == NULL) continue;
2186 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2187 Node* bol = cmp->fast_out(j);
2188 if (!bol->is_Bool()) continue;
2189 BoolTest::mask btest = bol->as_Bool()->_test._test;
2190 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2191 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2192 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2193 best_bol = bol->as_Bool();
2194 best_btest = btest;
2195 }
2196 }
2197 }
2198
2199 Node* answer_if_true = NULL;
2200 Node* answer_if_false = NULL;
2201 switch (best_btest) {
2202 default:
2203 if (cmpxy == NULL)
2204 cmpxy = ideal_cmpxy;
2205 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2206 // and fall through:
2207 case BoolTest::lt: // x < y
2208 case BoolTest::le: // x <= y
2209 answer_if_true = (want_max ? yvalue : xvalue);
2210 answer_if_false = (want_max ? xvalue : yvalue);
2211 break;
2212 case BoolTest::gt: // x > y
2213 case BoolTest::ge: // x >= y
2214 answer_if_true = (want_max ? xvalue : yvalue);
2215 answer_if_false = (want_max ? yvalue : xvalue);
2216 break;
2217 }
2218
2219 jint hi, lo;
2220 if (want_max) {
2221 // We can sharpen the minimum.
2222 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2223 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2224 } else {
2225 // We can sharpen the maximum.
2226 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2227 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2228 }
2229
2230 // Use a flow-free graph structure, to avoid creating excess control edges
2231 // which could hinder other optimizations.
2232 // Since Math.min/max is often used with arraycopy, we want
2233 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2234 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2235 answer_if_false, answer_if_true,
2236 TypeInt::make(lo, hi, widen));
2237
2238 return _gvn.transform(cmov);
2239
2240 /*
2241 // This is not as desirable as it may seem, since Min and Max
2242 // nodes do not have a full set of optimizations.
2243 // And they would interfere, anyway, with 'if' optimizations
2244 // and with CMoveI canonical forms.
2245 switch (id) {
2246 case vmIntrinsics::_min:
2247 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2248 case vmIntrinsics::_max:
2249 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2250 default:
2251 ShouldNotReachHere();
2252 }
2253 */
2254 }
2255
2256 inline int
2257 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2258 const TypePtr* base_type = TypePtr::NULL_PTR;
2259 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2260 if (base_type == NULL) {
2261 // Unknown type.
2262 return Type::AnyPtr;
2263 } else if (base_type == TypePtr::NULL_PTR) {
2264 // Since this is a NULL+long form, we have to switch to a rawptr.
2265 base = _gvn.transform(new (C) CastX2PNode(offset));
2266 offset = MakeConX(0);
2267 return Type::RawPtr;
2268 } else if (base_type->base() == Type::RawPtr) {
2269 return Type::RawPtr;
2270 } else if (base_type->isa_oopptr()) {
2271 // Base is never null => always a heap address.
2272 if (base_type->ptr() == TypePtr::NotNull) {
2273 return Type::OopPtr;
2274 }
2275 // Offset is small => always a heap address.
2276 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2277 if (offset_type != NULL &&
2278 base_type->offset() == 0 && // (should always be?)
2279 offset_type->_lo >= 0 &&
2280 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2281 return Type::OopPtr;
2282 }
2283 // Otherwise, it might either be oop+off or NULL+addr.
2284 return Type::AnyPtr;
2285 } else {
2286 // No information:
2287 return Type::AnyPtr;
2288 }
2289 }
2290
2291 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2292 int kind = classify_unsafe_addr(base, offset);
2293 if (kind == Type::RawPtr) {
2294 return basic_plus_adr(top(), base, offset);
2295 } else {
2296 return basic_plus_adr(base, offset);
2297 }
2298 }
2299
2300 //--------------------------inline_number_methods-----------------------------
2301 // inline int Integer.numberOfLeadingZeros(int)
2302 // inline int Long.numberOfLeadingZeros(long)
2303 //
2304 // inline int Integer.numberOfTrailingZeros(int)
2305 // inline int Long.numberOfTrailingZeros(long)
2306 //
2307 // inline int Integer.bitCount(int)
2308 // inline int Long.bitCount(long)
2309 //
2310 // inline char Character.reverseBytes(char)
2311 // inline short Short.reverseBytes(short)
2312 // inline int Integer.reverseBytes(int)
2313 // inline long Long.reverseBytes(long)
2314 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2315 Node* arg = argument(0);
2316 Node* n;
2317 switch (id) {
2318 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2319 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2320 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2321 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2322 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2323 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2324 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2325 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2326 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2327 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2328 default: fatal_unexpected_iid(id); break;
2329 }
2330 set_result(_gvn.transform(n));
2331 return true;
2332 }
2333
2334 //----------------------------inline_unsafe_access----------------------------
2335
2336 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2337
2338 // Helper that guards and inserts a pre-barrier.
2339 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2340 Node* pre_val, bool need_mem_bar) {
2341 // We could be accessing the referent field of a reference object. If so, when G1
2342 // is enabled, we need to log the value in the referent field in an SATB buffer.
2343 // This routine performs some compile time filters and generates suitable
2344 // runtime filters that guard the pre-barrier code.
2345 // Also add memory barrier for non volatile load from the referent field
2346 // to prevent commoning of loads across safepoint.
2347 if (!UseG1GC && !need_mem_bar)
2348 return;
2349
2350 // Some compile time checks.
2351
2352 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2353 const TypeX* otype = offset->find_intptr_t_type();
2354 if (otype != NULL && otype->is_con() &&
2355 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2356 // Constant offset but not the reference_offset so just return
2357 return;
2358 }
2359
2360 // We only need to generate the runtime guards for instances.
2361 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2362 if (btype != NULL) {
2363 if (btype->isa_aryptr()) {
2364 // Array type so nothing to do
2365 return;
2366 }
2367
2368 const TypeInstPtr* itype = btype->isa_instptr();
2369 if (itype != NULL) {
2370 // Can the klass of base_oop be statically determined to be
2371 // _not_ a sub-class of Reference and _not_ Object?
2372 ciKlass* klass = itype->klass();
2373 if ( klass->is_loaded() &&
2374 !klass->is_subtype_of(env()->Reference_klass()) &&
2375 !env()->Object_klass()->is_subtype_of(klass)) {
2376 return;
2377 }
2378 }
2379 }
2380
2381 // The compile time filters did not reject base_oop/offset so
2382 // we need to generate the following runtime filters
2383 //
2384 // if (offset == java_lang_ref_Reference::_reference_offset) {
2385 // if (instance_of(base, java.lang.ref.Reference)) {
2386 // pre_barrier(_, pre_val, ...);
2387 // }
2388 // }
2389
2390 float likely = PROB_LIKELY( 0.999);
2391 float unlikely = PROB_UNLIKELY(0.999);
2392
2393 IdealKit ideal(this);
2394 #define __ ideal.
2395
2396 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2397
2398 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2399 // Update graphKit memory and control from IdealKit.
2400 sync_kit(ideal);
2401
2402 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2403 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2404
2405 // Update IdealKit memory and control from graphKit.
2406 __ sync_kit(this);
2407
2408 Node* one = __ ConI(1);
2409 // is_instof == 0 if base_oop == NULL
2410 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2411
2412 // Update graphKit from IdeakKit.
2413 sync_kit(ideal);
2414
2415 // Use the pre-barrier to record the value in the referent field
2416 pre_barrier(false /* do_load */,
2417 __ ctrl(),
2418 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2419 pre_val /* pre_val */,
2420 T_OBJECT);
2421 if (need_mem_bar) {
2422 // Add memory barrier to prevent commoning reads from this field
2423 // across safepoint since GC can change its value.
2424 insert_mem_bar(Op_MemBarCPUOrder);
2425 }
2426 // Update IdealKit from graphKit.
2427 __ sync_kit(this);
2428
2429 } __ end_if(); // _ref_type != ref_none
2430 } __ end_if(); // offset == referent_offset
2431
2432 // Final sync IdealKit and GraphKit.
2433 final_sync(ideal);
2434 #undef __
2435 }
2436
2437
2438 // Interpret Unsafe.fieldOffset cookies correctly:
2439 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2440
2441 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2442 // Attempt to infer a sharper value type from the offset and base type.
2443 ciKlass* sharpened_klass = NULL;
2444
2445 // See if it is an instance field, with an object type.
2446 if (alias_type->field() != NULL) {
2447 assert(!is_native_ptr, "native pointer op cannot use a java address");
2448 if (alias_type->field()->type()->is_klass()) {
2449 sharpened_klass = alias_type->field()->type()->as_klass();
2450 }
2451 }
2452
2453 // See if it is a narrow oop array.
2454 if (adr_type->isa_aryptr()) {
2455 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2456 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2457 if (elem_type != NULL) {
2458 sharpened_klass = elem_type->klass();
2459 }
2460 }
2461 }
2462
2463 // The sharpened class might be unloaded if there is no class loader
2464 // contraint in place.
2465 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2466 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2467
2468 #ifndef PRODUCT
2469 if (C->print_intrinsics() || C->print_inlining()) {
2470 tty->print(" from base type: "); adr_type->dump();
2471 tty->print(" sharpened value: "); tjp->dump();
2472 }
2473 #endif
2474 // Sharpen the value type.
2475 return tjp;
2476 }
2477 return NULL;
2478 }
2479
2480 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2481 if (callee()->is_static()) return false; // caller must have the capability!
2482
2483 #ifndef PRODUCT
2484 {
2485 ResourceMark rm;
2486 // Check the signatures.
2487 ciSignature* sig = callee()->signature();
2488 #ifdef ASSERT
2489 if (!is_store) {
2490 // Object getObject(Object base, int/long offset), etc.
2491 BasicType rtype = sig->return_type()->basic_type();
2492 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2493 rtype = T_ADDRESS; // it is really a C void*
2494 assert(rtype == type, "getter must return the expected value");
2495 if (!is_native_ptr) {
2496 assert(sig->count() == 2, "oop getter has 2 arguments");
2497 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2498 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2499 } else {
2500 assert(sig->count() == 1, "native getter has 1 argument");
2501 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2502 }
2503 } else {
2504 // void putObject(Object base, int/long offset, Object x), etc.
2505 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2506 if (!is_native_ptr) {
2507 assert(sig->count() == 3, "oop putter has 3 arguments");
2508 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2509 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2510 } else {
2511 assert(sig->count() == 2, "native putter has 2 arguments");
2512 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2513 }
2514 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2515 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2516 vtype = T_ADDRESS; // it is really a C void*
2517 assert(vtype == type, "putter must accept the expected value");
2518 }
2519 #endif // ASSERT
2520 }
2521 #endif //PRODUCT
2522
2523 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2524
2525 Node* receiver = argument(0); // type: oop
2526
2527 // Build address expression. See the code in inline_unsafe_prefetch.
2528 Node* adr;
2529 Node* heap_base_oop = top();
2530 Node* offset = top();
2531 Node* val;
2532
2533 if (!is_native_ptr) {
2534 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2535 Node* base = argument(1); // type: oop
2536 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2537 offset = argument(2); // type: long
2538 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2539 // to be plain byte offsets, which are also the same as those accepted
2540 // by oopDesc::field_base.
2541 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2542 "fieldOffset must be byte-scaled");
2543 // 32-bit machines ignore the high half!
2544 offset = ConvL2X(offset);
2545 adr = make_unsafe_address(base, offset);
2546 heap_base_oop = base;
2547 val = is_store ? argument(4) : NULL;
2548 } else {
2549 Node* ptr = argument(1); // type: long
2550 ptr = ConvL2X(ptr); // adjust Java long to machine word
2551 adr = make_unsafe_address(NULL, ptr);
2552 val = is_store ? argument(3) : NULL;
2553 }
2554
2555 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2556
2557 // First guess at the value type.
2558 const Type *value_type = Type::get_const_basic_type(type);
2559
2560 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2561 // there was not enough information to nail it down.
2562 Compile::AliasType* alias_type = C->alias_type(adr_type);
2563 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2564
2565 // We will need memory barriers unless we can determine a unique
2566 // alias category for this reference. (Note: If for some reason
2567 // the barriers get omitted and the unsafe reference begins to "pollute"
2568 // the alias analysis of the rest of the graph, either Compile::can_alias
2569 // or Compile::must_alias will throw a diagnostic assert.)
2570 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2571
2572 // If we are reading the value of the referent field of a Reference
2573 // object (either by using Unsafe directly or through reflection)
2574 // then, if G1 is enabled, we need to record the referent in an
2575 // SATB log buffer using the pre-barrier mechanism.
2576 // Also we need to add memory barrier to prevent commoning reads
2577 // from this field across safepoint since GC can change its value.
2578 bool need_read_barrier = !is_native_ptr && !is_store &&
2579 offset != top() && heap_base_oop != top();
2580
2581 if (!is_store && type == T_OBJECT) {
2582 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2583 if (tjp != NULL) {
2584 value_type = tjp;
2585 }
2586 }
2587
2588 receiver = null_check(receiver);
2589 if (stopped()) {
2590 return true;
2591 }
2592 // Heap pointers get a null-check from the interpreter,
2593 // as a courtesy. However, this is not guaranteed by Unsafe,
2594 // and it is not possible to fully distinguish unintended nulls
2595 // from intended ones in this API.
2596
2597 if (is_volatile) {
2598 // We need to emit leading and trailing CPU membars (see below) in
2599 // addition to memory membars when is_volatile. This is a little
2600 // too strong, but avoids the need to insert per-alias-type
2601 // volatile membars (for stores; compare Parse::do_put_xxx), which
2602 // we cannot do effectively here because we probably only have a
2603 // rough approximation of type.
2604 need_mem_bar = true;
2605 // For Stores, place a memory ordering barrier now.
2606 if (is_store) {
2607 insert_mem_bar(Op_MemBarRelease);
2608 } else {
2609 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2610 insert_mem_bar(Op_MemBarVolatile);
2611 }
2612 }
2613 }
2614
2615 // Memory barrier to prevent normal and 'unsafe' accesses from
2616 // bypassing each other. Happens after null checks, so the
2617 // exception paths do not take memory state from the memory barrier,
2618 // so there's no problems making a strong assert about mixing users
2619 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2620 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2621 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2622
2623 if (!is_store) {
2624 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2625 // load value
2626 switch (type) {
2627 case T_BOOLEAN:
2628 case T_CHAR:
2629 case T_BYTE:
2630 case T_SHORT:
2631 case T_INT:
2632 case T_LONG:
2633 case T_FLOAT:
2634 case T_DOUBLE:
2635 break;
2636 case T_OBJECT:
2637 if (need_read_barrier) {
2638 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2639 }
2640 break;
2641 case T_ADDRESS:
2642 // Cast to an int type.
2643 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2644 p = ConvX2UL(p);
2645 break;
2646 default:
2647 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2648 break;
2649 }
2650 // The load node has the control of the preceding MemBarCPUOrder. All
2651 // following nodes will have the control of the MemBarCPUOrder inserted at
2652 // the end of this method. So, pushing the load onto the stack at a later
2653 // point is fine.
2654 set_result(p);
2655 } else {
2656 // place effect of store into memory
2657 switch (type) {
2658 case T_DOUBLE:
2659 val = dstore_rounding(val);
2660 break;
2661 case T_ADDRESS:
2662 // Repackage the long as a pointer.
2663 val = ConvL2X(val);
2664 val = _gvn.transform(new (C) CastX2PNode(val));
2665 break;
2666 }
2667
2668 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2669 if (type != T_OBJECT ) {
2670 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2671 } else {
2672 // Possibly an oop being stored to Java heap or native memory
2673 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2674 // oop to Java heap.
2675 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2676 } else {
2677 // We can't tell at compile time if we are storing in the Java heap or outside
2678 // of it. So we need to emit code to conditionally do the proper type of
2679 // store.
2680
2681 IdealKit ideal(this);
2682 #define __ ideal.
2683 // QQQ who knows what probability is here??
2684 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2685 // Sync IdealKit and graphKit.
2686 sync_kit(ideal);
2687 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2688 // Update IdealKit memory.
2689 __ sync_kit(this);
2690 } __ else_(); {
2691 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2692 } __ end_if();
2693 // Final sync IdealKit and GraphKit.
2694 final_sync(ideal);
2695 #undef __
2696 }
2697 }
2698 }
2699
2700 if (is_volatile) {
2701 if (!is_store) {
2702 insert_mem_bar(Op_MemBarAcquire);
2703 } else {
2704 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2705 insert_mem_bar(Op_MemBarVolatile);
2706 }
2707 }
2708 }
2709
2710 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2711
2712 return true;
2713 }
2714
2715 //----------------------------inline_unsafe_prefetch----------------------------
2716
2717 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2718 #ifndef PRODUCT
2719 {
2720 ResourceMark rm;
2721 // Check the signatures.
2722 ciSignature* sig = callee()->signature();
2723 #ifdef ASSERT
2724 // Object getObject(Object base, int/long offset), etc.
2725 BasicType rtype = sig->return_type()->basic_type();
2726 if (!is_native_ptr) {
2727 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2728 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2729 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2730 } else {
2731 assert(sig->count() == 1, "native prefetch has 1 argument");
2732 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2733 }
2734 #endif // ASSERT
2735 }
2736 #endif // !PRODUCT
2737
2738 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2739
2740 const int idx = is_static ? 0 : 1;
2741 if (!is_static) {
2742 null_check_receiver();
2743 if (stopped()) {
2744 return true;
2745 }
2746 }
2747
2748 // Build address expression. See the code in inline_unsafe_access.
2749 Node *adr;
2750 if (!is_native_ptr) {
2751 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2752 Node* base = argument(idx + 0); // type: oop
2753 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2754 Node* offset = argument(idx + 1); // type: long
2755 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2756 // to be plain byte offsets, which are also the same as those accepted
2757 // by oopDesc::field_base.
2758 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2759 "fieldOffset must be byte-scaled");
2760 // 32-bit machines ignore the high half!
2761 offset = ConvL2X(offset);
2762 adr = make_unsafe_address(base, offset);
2763 } else {
2764 Node* ptr = argument(idx + 0); // type: long
2765 ptr = ConvL2X(ptr); // adjust Java long to machine word
2766 adr = make_unsafe_address(NULL, ptr);
2767 }
2768
2769 // Generate the read or write prefetch
2770 Node *prefetch;
2771 if (is_store) {
2772 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2773 } else {
2774 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2775 }
2776 prefetch->init_req(0, control());
2777 set_i_o(_gvn.transform(prefetch));
2778
2779 return true;
2780 }
2781
2782 //----------------------------inline_unsafe_load_store----------------------------
2783 // This method serves a couple of different customers (depending on LoadStoreKind):
2784 //
2785 // LS_cmpxchg:
2786 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2787 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2788 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2789 //
2790 // LS_xadd:
2791 // public int getAndAddInt( Object o, long offset, int delta)
2792 // public long getAndAddLong(Object o, long offset, long delta)
2793 //
2794 // LS_xchg:
2795 // int getAndSet(Object o, long offset, int newValue)
2796 // long getAndSet(Object o, long offset, long newValue)
2797 // Object getAndSet(Object o, long offset, Object newValue)
2798 //
2799 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2800 // This basic scheme here is the same as inline_unsafe_access, but
2801 // differs in enough details that combining them would make the code
2802 // overly confusing. (This is a true fact! I originally combined
2803 // them, but even I was confused by it!) As much code/comments as
2804 // possible are retained from inline_unsafe_access though to make
2805 // the correspondences clearer. - dl
2806
2807 if (callee()->is_static()) return false; // caller must have the capability!
2808
2809 #ifndef PRODUCT
2810 BasicType rtype;
2811 {
2812 ResourceMark rm;
2813 // Check the signatures.
2814 ciSignature* sig = callee()->signature();
2815 rtype = sig->return_type()->basic_type();
2816 if (kind == LS_xadd || kind == LS_xchg) {
2817 // Check the signatures.
2818 #ifdef ASSERT
2819 assert(rtype == type, "get and set must return the expected type");
2820 assert(sig->count() == 3, "get and set has 3 arguments");
2821 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2822 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2823 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2824 #endif // ASSERT
2825 } else if (kind == LS_cmpxchg) {
2826 // Check the signatures.
2827 #ifdef ASSERT
2828 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2829 assert(sig->count() == 4, "CAS has 4 arguments");
2830 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2831 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2832 #endif // ASSERT
2833 } else {
2834 ShouldNotReachHere();
2835 }
2836 }
2837 #endif //PRODUCT
2838
2839 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2840
2841 // Get arguments:
2842 Node* receiver = NULL;
2843 Node* base = NULL;
2844 Node* offset = NULL;
2845 Node* oldval = NULL;
2846 Node* newval = NULL;
2847 if (kind == LS_cmpxchg) {
2848 const bool two_slot_type = type2size[type] == 2;
2849 receiver = argument(0); // type: oop
2850 base = argument(1); // type: oop
2851 offset = argument(2); // type: long
2852 oldval = argument(4); // type: oop, int, or long
2853 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2854 } else if (kind == LS_xadd || kind == LS_xchg){
2855 receiver = argument(0); // type: oop
2856 base = argument(1); // type: oop
2857 offset = argument(2); // type: long
2858 oldval = NULL;
2859 newval = argument(4); // type: oop, int, or long
2860 }
2861
2862 // Null check receiver.
2863 receiver = null_check(receiver);
2864 if (stopped()) {
2865 return true;
2866 }
2867
2868 // Build field offset expression.
2869 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2870 // to be plain byte offsets, which are also the same as those accepted
2871 // by oopDesc::field_base.
2872 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2873 // 32-bit machines ignore the high half of long offsets
2874 offset = ConvL2X(offset);
2875 Node* adr = make_unsafe_address(base, offset);
2876 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2877
2878 // For CAS, unlike inline_unsafe_access, there seems no point in
2879 // trying to refine types. Just use the coarse types here.
2880 const Type *value_type = Type::get_const_basic_type(type);
2881 Compile::AliasType* alias_type = C->alias_type(adr_type);
2882 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2883
2884 if (kind == LS_xchg && type == T_OBJECT) {
2885 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2886 if (tjp != NULL) {
2887 value_type = tjp;
2888 }
2889 }
2890
2891 int alias_idx = C->get_alias_index(adr_type);
2892
2893 // Memory-model-wise, a LoadStore acts like a little synchronized
2894 // block, so needs barriers on each side. These don't translate
2895 // into actual barriers on most machines, but we still need rest of
2896 // compiler to respect ordering.
2897
2898 insert_mem_bar(Op_MemBarRelease);
2899 insert_mem_bar(Op_MemBarCPUOrder);
2900
2901 // 4984716: MemBars must be inserted before this
2902 // memory node in order to avoid a false
2903 // dependency which will confuse the scheduler.
2904 Node *mem = memory(alias_idx);
2905
2906 // For now, we handle only those cases that actually exist: ints,
2907 // longs, and Object. Adding others should be straightforward.
2908 Node* load_store;
2909 switch(type) {
2910 case T_INT:
2911 if (kind == LS_xadd) {
2912 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2913 } else if (kind == LS_xchg) {
2914 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2915 } else if (kind == LS_cmpxchg) {
2916 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2917 } else {
2918 ShouldNotReachHere();
2919 }
2920 break;
2921 case T_LONG:
2922 if (kind == LS_xadd) {
2923 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2924 } else if (kind == LS_xchg) {
2925 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2926 } else if (kind == LS_cmpxchg) {
2927 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2928 } else {
2929 ShouldNotReachHere();
2930 }
2931 break;
2932 case T_OBJECT:
2933 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2934 // could be delayed during Parse (for example, in adjust_map_after_if()).
2935 // Execute transformation here to avoid barrier generation in such case.
2936 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2937 newval = _gvn.makecon(TypePtr::NULL_PTR);
2938
2939 // Reference stores need a store barrier.
2940 if (kind == LS_xchg) {
2941 // If pre-barrier must execute before the oop store, old value will require do_load here.
2942 if (!can_move_pre_barrier()) {
2943 pre_barrier(true /* do_load*/,
2944 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2945 NULL /* pre_val*/,
2946 T_OBJECT);
2947 } // Else move pre_barrier to use load_store value, see below.
2948 } else if (kind == LS_cmpxchg) {
2949 // Same as for newval above:
2950 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2951 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2952 }
2953 // The only known value which might get overwritten is oldval.
2954 pre_barrier(false /* do_load */,
2955 control(), NULL, NULL, max_juint, NULL, NULL,
2956 oldval /* pre_val */,
2957 T_OBJECT);
2958 } else {
2959 ShouldNotReachHere();
2960 }
2961
2962 #ifdef _LP64
2963 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2964 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2965 if (kind == LS_xchg) {
2966 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2967 newval_enc, adr_type, value_type->make_narrowoop()));
2968 } else {
2969 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2970 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2971 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2972 newval_enc, oldval_enc));
2973 }
2974 } else
2975 #endif
2976 {
2977 if (kind == LS_xchg) {
2978 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2979 } else {
2980 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2981 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2982 }
2983 }
2984 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2985 break;
2986 default:
2987 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2988 break;
2989 }
2990
2991 // SCMemProjNodes represent the memory state of a LoadStore. Their
2992 // main role is to prevent LoadStore nodes from being optimized away
2993 // when their results aren't used.
2994 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2995 set_memory(proj, alias_idx);
2996
2997 if (type == T_OBJECT && kind == LS_xchg) {
2998 #ifdef _LP64
2999 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3000 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3001 }
3002 #endif
3003 if (can_move_pre_barrier()) {
3004 // Don't need to load pre_val. The old value is returned by load_store.
3005 // The pre_barrier can execute after the xchg as long as no safepoint
3006 // gets inserted between them.
3007 pre_barrier(false /* do_load */,
3008 control(), NULL, NULL, max_juint, NULL, NULL,
3009 load_store /* pre_val */,
3010 T_OBJECT);
3011 }
3012 }
3013
3014 // Add the trailing membar surrounding the access
3015 insert_mem_bar(Op_MemBarCPUOrder);
3016 insert_mem_bar(Op_MemBarAcquire);
3017
3018 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3019 set_result(load_store);
3020 return true;
3021 }
3022
3023 //----------------------------inline_unsafe_ordered_store----------------------
3024 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3025 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3026 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3027 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3028 // This is another variant of inline_unsafe_access, differing in
3029 // that it always issues store-store ("release") barrier and ensures
3030 // store-atomicity (which only matters for "long").
3031
3032 if (callee()->is_static()) return false; // caller must have the capability!
3033
3034 #ifndef PRODUCT
3035 {
3036 ResourceMark rm;
3037 // Check the signatures.
3038 ciSignature* sig = callee()->signature();
3039 #ifdef ASSERT
3040 BasicType rtype = sig->return_type()->basic_type();
3041 assert(rtype == T_VOID, "must return void");
3042 assert(sig->count() == 3, "has 3 arguments");
3043 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3044 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3045 #endif // ASSERT
3046 }
3047 #endif //PRODUCT
3048
3049 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3050
3051 // Get arguments:
3052 Node* receiver = argument(0); // type: oop
3053 Node* base = argument(1); // type: oop
3054 Node* offset = argument(2); // type: long
3055 Node* val = argument(4); // type: oop, int, or long
3056
3057 // Null check receiver.
3058 receiver = null_check(receiver);
3059 if (stopped()) {
3060 return true;
3061 }
3062
3063 // Build field offset expression.
3064 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3065 // 32-bit machines ignore the high half of long offsets
3066 offset = ConvL2X(offset);
3067 Node* adr = make_unsafe_address(base, offset);
3068 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3069 const Type *value_type = Type::get_const_basic_type(type);
3070 Compile::AliasType* alias_type = C->alias_type(adr_type);
3071
3072 insert_mem_bar(Op_MemBarRelease);
3073 insert_mem_bar(Op_MemBarCPUOrder);
3074 // Ensure that the store is atomic for longs:
3075 const bool require_atomic_access = true;
3076 Node* store;
3077 if (type == T_OBJECT) // reference stores need a store barrier.
3078 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3079 else {
3080 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3081 }
3082 insert_mem_bar(Op_MemBarCPUOrder);
3083 return true;
3084 }
3085
3086 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3087 // Regardless of form, don't allow previous ld/st to move down,
3088 // then issue acquire, release, or volatile mem_bar.
3089 insert_mem_bar(Op_MemBarCPUOrder);
3090 switch(id) {
3091 case vmIntrinsics::_loadFence:
3092 insert_mem_bar(Op_LoadFence);
3093 return true;
3094 case vmIntrinsics::_storeFence:
3095 insert_mem_bar(Op_StoreFence);
3096 return true;
3097 case vmIntrinsics::_fullFence:
3098 insert_mem_bar(Op_MemBarVolatile);
3099 return true;
3100 default:
3101 fatal_unexpected_iid(id);
3102 return false;
3103 }
3104 }
3105
3106 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3107 if (!kls->is_Con()) {
3108 return true;
3109 }
3110 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3111 if (klsptr == NULL) {
3112 return true;
3113 }
3114 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3115 // don't need a guard for a klass that is already initialized
3116 return !ik->is_initialized();
3117 }
3118
3119 //----------------------------inline_unsafe_allocate---------------------------
3120 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3121 bool LibraryCallKit::inline_unsafe_allocate() {
3122 if (callee()->is_static()) return false; // caller must have the capability!
3123
3124 null_check_receiver(); // null-check, then ignore
3125 Node* cls = null_check(argument(1));
3126 if (stopped()) return true;
3127
3128 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3129 kls = null_check(kls);
3130 if (stopped()) return true; // argument was like int.class
3131
3132 Node* test = NULL;
3133 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3134 // Note: The argument might still be an illegal value like
3135 // Serializable.class or Object[].class. The runtime will handle it.
3136 // But we must make an explicit check for initialization.
3137 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3138 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3139 // can generate code to load it as unsigned byte.
3140 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3141 Node* bits = intcon(InstanceKlass::fully_initialized);
3142 test = _gvn.transform(new (C) SubINode(inst, bits));
3143 // The 'test' is non-zero if we need to take a slow path.
3144 }
3145
3146 Node* obj = new_instance(kls, test);
3147 set_result(obj);
3148 return true;
3149 }
3150
3151 #ifdef TRACE_HAVE_INTRINSICS
3152 /*
3153 * oop -> myklass
3154 * myklass->trace_id |= USED
3155 * return myklass->trace_id & ~0x3
3156 */
3157 bool LibraryCallKit::inline_native_classID() {
3158 null_check_receiver(); // null-check, then ignore
3159 Node* cls = null_check(argument(1), T_OBJECT);
3160 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3161 kls = null_check(kls, T_OBJECT);
3162 ByteSize offset = TRACE_ID_OFFSET;
3163 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3164 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3165 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3166 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3167 Node* clsused = longcon(0x01l); // set the class bit
3168 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3169
3170 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3171 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3172 set_result(andl);
3173 return true;
3174 }
3175
3176 bool LibraryCallKit::inline_native_threadID() {
3177 Node* tls_ptr = NULL;
3178 Node* cur_thr = generate_current_thread(tls_ptr);
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::thread_id_offset()));
3182
3183 Node* threadid = NULL;
3184 size_t thread_id_size = OSThread::thread_id_size();
3185 if (thread_id_size == (size_t) BytesPerLong) {
3186 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3187 } else if (thread_id_size == (size_t) BytesPerInt) {
3188 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3189 } else {
3190 ShouldNotReachHere();
3191 }
3192 set_result(threadid);
3193 return true;
3194 }
3195 #endif
3196
3197 //------------------------inline_native_time_funcs--------------
3198 // inline code for System.currentTimeMillis() and System.nanoTime()
3199 // these have the same type and signature
3200 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3201 const TypeFunc* tf = OptoRuntime::void_long_Type();
3202 const TypePtr* no_memory_effects = NULL;
3203 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3204 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3205 #ifdef ASSERT
3206 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3207 assert(value_top == top(), "second value must be top");
3208 #endif
3209 set_result(value);
3210 return true;
3211 }
3212
3213 //------------------------inline_native_currentThread------------------
3214 bool LibraryCallKit::inline_native_currentThread() {
3215 Node* junk = NULL;
3216 set_result(generate_current_thread(junk));
3217 return true;
3218 }
3219
3220 //------------------------inline_native_isInterrupted------------------
3221 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3222 bool LibraryCallKit::inline_native_isInterrupted() {
3223 // Add a fast path to t.isInterrupted(clear_int):
3224 // (t == Thread.current() &&
3225 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3226 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3227 // So, in the common case that the interrupt bit is false,
3228 // we avoid making a call into the VM. Even if the interrupt bit
3229 // is true, if the clear_int argument is false, we avoid the VM call.
3230 // However, if the receiver is not currentThread, we must call the VM,
3231 // because there must be some locking done around the operation.
3232
3233 // We only go to the fast case code if we pass two guards.
3234 // Paths which do not pass are accumulated in the slow_region.
3235
3236 enum {
3237 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3238 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3239 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3240 PATH_LIMIT
3241 };
3242
3243 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3244 // out of the function.
3245 insert_mem_bar(Op_MemBarCPUOrder);
3246
3247 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3248 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3249
3250 RegionNode* slow_region = new (C) RegionNode(1);
3251 record_for_igvn(slow_region);
3252
3253 // (a) Receiving thread must be the current thread.
3254 Node* rec_thr = argument(0);
3255 Node* tls_ptr = NULL;
3256 Node* cur_thr = generate_current_thread(tls_ptr);
3257 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3258 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3259
3260 generate_slow_guard(bol_thr, slow_region);
3261
3262 // (b) Interrupt bit on TLS must be false.
3263 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3264 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3265 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3266
3267 // Set the control input on the field _interrupted read to prevent it floating up.
3268 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3269 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3270 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3271
3272 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3273
3274 // First fast path: if (!TLS._interrupted) return false;
3275 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3276 result_rgn->init_req(no_int_result_path, false_bit);
3277 result_val->init_req(no_int_result_path, intcon(0));
3278
3279 // drop through to next case
3280 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3281
3282 #ifndef TARGET_OS_FAMILY_windows
3283 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3284 Node* clr_arg = argument(1);
3285 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3286 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3287 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3288
3289 // Second fast path: ... else if (!clear_int) return true;
3290 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3291 result_rgn->init_req(no_clear_result_path, false_arg);
3292 result_val->init_req(no_clear_result_path, intcon(1));
3293
3294 // drop through to next case
3295 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3296 #else
3297 // To return true on Windows you must read the _interrupted field
3298 // and check the the event state i.e. take the slow path.
3299 #endif // TARGET_OS_FAMILY_windows
3300
3301 // (d) Otherwise, go to the slow path.
3302 slow_region->add_req(control());
3303 set_control( _gvn.transform(slow_region));
3304
3305 if (stopped()) {
3306 // There is no slow path.
3307 result_rgn->init_req(slow_result_path, top());
3308 result_val->init_req(slow_result_path, top());
3309 } else {
3310 // non-virtual because it is a private non-static
3311 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3312
3313 Node* slow_val = set_results_for_java_call(slow_call);
3314 // this->control() comes from set_results_for_java_call
3315
3316 Node* fast_io = slow_call->in(TypeFunc::I_O);
3317 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3318
3319 // These two phis are pre-filled with copies of of the fast IO and Memory
3320 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3321 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3322
3323 result_rgn->init_req(slow_result_path, control());
3324 result_io ->init_req(slow_result_path, i_o());
3325 result_mem->init_req(slow_result_path, reset_memory());
3326 result_val->init_req(slow_result_path, slow_val);
3327
3328 set_all_memory(_gvn.transform(result_mem));
3329 set_i_o( _gvn.transform(result_io));
3330 }
3331
3332 C->set_has_split_ifs(true); // Has chance for split-if optimization
3333 set_result(result_rgn, result_val);
3334 return true;
3335 }
3336
3337 //---------------------------load_mirror_from_klass----------------------------
3338 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3339 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3340 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3341 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3342 }
3343
3344 //-----------------------load_klass_from_mirror_common-------------------------
3345 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3346 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3347 // and branch to the given path on the region.
3348 // If never_see_null, take an uncommon trap on null, so we can optimistically
3349 // compile for the non-null case.
3350 // If the region is NULL, force never_see_null = true.
3351 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3352 bool never_see_null,
3353 RegionNode* region,
3354 int null_path,
3355 int offset) {
3356 if (region == NULL) never_see_null = true;
3357 Node* p = basic_plus_adr(mirror, offset);
3358 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3359 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3360 Node* null_ctl = top();
3361 kls = null_check_oop(kls, &null_ctl, never_see_null);
3362 if (region != NULL) {
3363 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3364 region->init_req(null_path, null_ctl);
3365 } else {
3366 assert(null_ctl == top(), "no loose ends");
3367 }
3368 return kls;
3369 }
3370
3371 //--------------------(inline_native_Class_query helpers)---------------------
3372 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3373 // Fall through if (mods & mask) == bits, take the guard otherwise.
3374 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3375 // Branch around if the given klass has the given modifier bit set.
3376 // Like generate_guard, adds a new path onto the region.
3377 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3378 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3379 Node* mask = intcon(modifier_mask);
3380 Node* bits = intcon(modifier_bits);
3381 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3382 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3383 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3384 return generate_fair_guard(bol, region);
3385 }
3386 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3387 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3388 }
3389
3390 //-------------------------inline_native_Class_query-------------------
3391 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3392 const Type* return_type = TypeInt::BOOL;
3393 Node* prim_return_value = top(); // what happens if it's a primitive class?
3394 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3395 bool expect_prim = false; // most of these guys expect to work on refs
3396
3397 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3398
3399 Node* mirror = argument(0);
3400 Node* obj = top();
3401
3402 switch (id) {
3403 case vmIntrinsics::_isInstance:
3404 // nothing is an instance of a primitive type
3405 prim_return_value = intcon(0);
3406 obj = argument(1);
3407 break;
3408 case vmIntrinsics::_getModifiers:
3409 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3410 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3411 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3412 break;
3413 case vmIntrinsics::_isInterface:
3414 prim_return_value = intcon(0);
3415 break;
3416 case vmIntrinsics::_isArray:
3417 prim_return_value = intcon(0);
3418 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3419 break;
3420 case vmIntrinsics::_isPrimitive:
3421 prim_return_value = intcon(1);
3422 expect_prim = true; // obviously
3423 break;
3424 case vmIntrinsics::_getSuperclass:
3425 prim_return_value = null();
3426 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3427 break;
3428 case vmIntrinsics::_getComponentType:
3429 prim_return_value = null();
3430 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3431 break;
3432 case vmIntrinsics::_getClassAccessFlags:
3433 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3434 return_type = TypeInt::INT; // not bool! 6297094
3435 break;
3436 default:
3437 fatal_unexpected_iid(id);
3438 break;
3439 }
3440
3441 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3442 if (mirror_con == NULL) return false; // cannot happen?
3443
3444 #ifndef PRODUCT
3445 if (C->print_intrinsics() || C->print_inlining()) {
3446 ciType* k = mirror_con->java_mirror_type();
3447 if (k) {
3448 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3449 k->print_name();
3450 tty->cr();
3451 }
3452 }
3453 #endif
3454
3455 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3456 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3457 record_for_igvn(region);
3458 PhiNode* phi = new (C) PhiNode(region, return_type);
3459
3460 // The mirror will never be null of Reflection.getClassAccessFlags, however
3461 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3462 // if it is. See bug 4774291.
3463
3464 // For Reflection.getClassAccessFlags(), the null check occurs in
3465 // the wrong place; see inline_unsafe_access(), above, for a similar
3466 // situation.
3467 mirror = null_check(mirror);
3468 // If mirror or obj is dead, only null-path is taken.
3469 if (stopped()) return true;
3470
3471 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3472
3473 // Now load the mirror's klass metaobject, and null-check it.
3474 // Side-effects region with the control path if the klass is null.
3475 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3476 // If kls is null, we have a primitive mirror.
3477 phi->init_req(_prim_path, prim_return_value);
3478 if (stopped()) { set_result(region, phi); return true; }
3479 bool safe_for_replace = (region->in(_prim_path) == top());
3480
3481 Node* p; // handy temp
3482 Node* null_ctl;
3483
3484 // Now that we have the non-null klass, we can perform the real query.
3485 // For constant classes, the query will constant-fold in LoadNode::Value.
3486 Node* query_value = top();
3487 switch (id) {
3488 case vmIntrinsics::_isInstance:
3489 // nothing is an instance of a primitive type
3490 query_value = gen_instanceof(obj, kls, safe_for_replace);
3491 break;
3492
3493 case vmIntrinsics::_getModifiers:
3494 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3495 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3496 break;
3497
3498 case vmIntrinsics::_isInterface:
3499 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3500 if (generate_interface_guard(kls, region) != NULL)
3501 // A guard was added. If the guard is taken, it was an interface.
3502 phi->add_req(intcon(1));
3503 // If we fall through, it's a plain class.
3504 query_value = intcon(0);
3505 break;
3506
3507 case vmIntrinsics::_isArray:
3508 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3509 if (generate_array_guard(kls, region) != NULL)
3510 // A guard was added. If the guard is taken, it was an array.
3511 phi->add_req(intcon(1));
3512 // If we fall through, it's a plain class.
3513 query_value = intcon(0);
3514 break;
3515
3516 case vmIntrinsics::_isPrimitive:
3517 query_value = intcon(0); // "normal" path produces false
3518 break;
3519
3520 case vmIntrinsics::_getSuperclass:
3521 // The rules here are somewhat unfortunate, but we can still do better
3522 // with random logic than with a JNI call.
3523 // Interfaces store null or Object as _super, but must report null.
3524 // Arrays store an intermediate super as _super, but must report Object.
3525 // Other types can report the actual _super.
3526 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3527 if (generate_interface_guard(kls, region) != NULL)
3528 // A guard was added. If the guard is taken, it was an interface.
3529 phi->add_req(null());
3530 if (generate_array_guard(kls, region) != NULL)
3531 // A guard was added. If the guard is taken, it was an array.
3532 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3533 // If we fall through, it's a plain class. Get its _super.
3534 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3535 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3536 null_ctl = top();
3537 kls = null_check_oop(kls, &null_ctl);
3538 if (null_ctl != top()) {
3539 // If the guard is taken, Object.superClass is null (both klass and mirror).
3540 region->add_req(null_ctl);
3541 phi ->add_req(null());
3542 }
3543 if (!stopped()) {
3544 query_value = load_mirror_from_klass(kls);
3545 }
3546 break;
3547
3548 case vmIntrinsics::_getComponentType:
3549 if (generate_array_guard(kls, region) != NULL) {
3550 // Be sure to pin the oop load to the guard edge just created:
3551 Node* is_array_ctrl = region->in(region->req()-1);
3552 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3553 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3554 phi->add_req(cmo);
3555 }
3556 query_value = null(); // non-array case is null
3557 break;
3558
3559 case vmIntrinsics::_getClassAccessFlags:
3560 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3561 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3562 break;
3563
3564 default:
3565 fatal_unexpected_iid(id);
3566 break;
3567 }
3568
3569 // Fall-through is the normal case of a query to a real class.
3570 phi->init_req(1, query_value);
3571 region->init_req(1, control());
3572
3573 C->set_has_split_ifs(true); // Has chance for split-if optimization
3574 set_result(region, phi);
3575 return true;
3576 }
3577
3578 //--------------------------inline_native_subtype_check------------------------
3579 // This intrinsic takes the JNI calls out of the heart of
3580 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3581 bool LibraryCallKit::inline_native_subtype_check() {
3582 // Pull both arguments off the stack.
3583 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3584 args[0] = argument(0);
3585 args[1] = argument(1);
3586 Node* klasses[2]; // corresponding Klasses: superk, subk
3587 klasses[0] = klasses[1] = top();
3588
3589 enum {
3590 // A full decision tree on {superc is prim, subc is prim}:
3591 _prim_0_path = 1, // {P,N} => false
3592 // {P,P} & superc!=subc => false
3593 _prim_same_path, // {P,P} & superc==subc => true
3594 _prim_1_path, // {N,P} => false
3595 _ref_subtype_path, // {N,N} & subtype check wins => true
3596 _both_ref_path, // {N,N} & subtype check loses => false
3597 PATH_LIMIT
3598 };
3599
3600 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3601 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3602 record_for_igvn(region);
3603
3604 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3605 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3606 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3607
3608 // First null-check both mirrors and load each mirror's klass metaobject.
3609 int which_arg;
3610 for (which_arg = 0; which_arg <= 1; which_arg++) {
3611 Node* arg = args[which_arg];
3612 arg = null_check(arg);
3613 if (stopped()) break;
3614 args[which_arg] = arg;
3615
3616 Node* p = basic_plus_adr(arg, class_klass_offset);
3617 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3618 klasses[which_arg] = _gvn.transform(kls);
3619 }
3620
3621 // Having loaded both klasses, test each for null.
3622 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3623 for (which_arg = 0; which_arg <= 1; which_arg++) {
3624 Node* kls = klasses[which_arg];
3625 Node* null_ctl = top();
3626 kls = null_check_oop(kls, &null_ctl, never_see_null);
3627 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3628 region->init_req(prim_path, null_ctl);
3629 if (stopped()) break;
3630 klasses[which_arg] = kls;
3631 }
3632
3633 if (!stopped()) {
3634 // now we have two reference types, in klasses[0..1]
3635 Node* subk = klasses[1]; // the argument to isAssignableFrom
3636 Node* superk = klasses[0]; // the receiver
3637 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3638 // now we have a successful reference subtype check
3639 region->set_req(_ref_subtype_path, control());
3640 }
3641
3642 // If both operands are primitive (both klasses null), then
3643 // we must return true when they are identical primitives.
3644 // It is convenient to test this after the first null klass check.
3645 set_control(region->in(_prim_0_path)); // go back to first null check
3646 if (!stopped()) {
3647 // Since superc is primitive, make a guard for the superc==subc case.
3648 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3649 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3650 generate_guard(bol_eq, region, PROB_FAIR);
3651 if (region->req() == PATH_LIMIT+1) {
3652 // A guard was added. If the added guard is taken, superc==subc.
3653 region->swap_edges(PATH_LIMIT, _prim_same_path);
3654 region->del_req(PATH_LIMIT);
3655 }
3656 region->set_req(_prim_0_path, control()); // Not equal after all.
3657 }
3658
3659 // these are the only paths that produce 'true':
3660 phi->set_req(_prim_same_path, intcon(1));
3661 phi->set_req(_ref_subtype_path, intcon(1));
3662
3663 // pull together the cases:
3664 assert(region->req() == PATH_LIMIT, "sane region");
3665 for (uint i = 1; i < region->req(); i++) {
3666 Node* ctl = region->in(i);
3667 if (ctl == NULL || ctl == top()) {
3668 region->set_req(i, top());
3669 phi ->set_req(i, top());
3670 } else if (phi->in(i) == NULL) {
3671 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3672 }
3673 }
3674
3675 set_control(_gvn.transform(region));
3676 set_result(_gvn.transform(phi));
3677 return true;
3678 }
3679
3680 //---------------------generate_array_guard_common------------------------
3681 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3682 bool obj_array, bool not_array) {
3683 // If obj_array/non_array==false/false:
3684 // Branch around if the given klass is in fact an array (either obj or prim).
3685 // If obj_array/non_array==false/true:
3686 // Branch around if the given klass is not an array klass of any kind.
3687 // If obj_array/non_array==true/true:
3688 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3689 // If obj_array/non_array==true/false:
3690 // Branch around if the kls is an oop array (Object[] or subtype)
3691 //
3692 // Like generate_guard, adds a new path onto the region.
3693 jint layout_con = 0;
3694 Node* layout_val = get_layout_helper(kls, layout_con);
3695 if (layout_val == NULL) {
3696 bool query = (obj_array
3697 ? Klass::layout_helper_is_objArray(layout_con)
3698 : Klass::layout_helper_is_array(layout_con));
3699 if (query == not_array) {
3700 return NULL; // never a branch
3701 } else { // always a branch
3702 Node* always_branch = control();
3703 if (region != NULL)
3704 region->add_req(always_branch);
3705 set_control(top());
3706 return always_branch;
3707 }
3708 }
3709 // Now test the correct condition.
3710 jint nval = (obj_array
3711 ? ((jint)Klass::_lh_array_tag_type_value
3712 << Klass::_lh_array_tag_shift)
3713 : Klass::_lh_neutral_value);
3714 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3715 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3716 // invert the test if we are looking for a non-array
3717 if (not_array) btest = BoolTest(btest).negate();
3718 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3719 return generate_fair_guard(bol, region);
3720 }
3721
3722
3723 //-----------------------inline_native_newArray--------------------------
3724 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3725 bool LibraryCallKit::inline_native_newArray() {
3726 Node* mirror = argument(0);
3727 Node* count_val = argument(1);
3728
3729 mirror = null_check(mirror);
3730 // If mirror or obj is dead, only null-path is taken.
3731 if (stopped()) return true;
3732
3733 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3734 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3735 PhiNode* result_val = new(C) PhiNode(result_reg,
3736 TypeInstPtr::NOTNULL);
3737 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3738 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3739 TypePtr::BOTTOM);
3740
3741 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3742 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3743 result_reg, _slow_path);
3744 Node* normal_ctl = control();
3745 Node* no_array_ctl = result_reg->in(_slow_path);
3746
3747 // Generate code for the slow case. We make a call to newArray().
3748 set_control(no_array_ctl);
3749 if (!stopped()) {
3750 // Either the input type is void.class, or else the
3751 // array klass has not yet been cached. Either the
3752 // ensuing call will throw an exception, or else it
3753 // will cache the array klass for next time.
3754 PreserveJVMState pjvms(this);
3755 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3756 Node* slow_result = set_results_for_java_call(slow_call);
3757 // this->control() comes from set_results_for_java_call
3758 result_reg->set_req(_slow_path, control());
3759 result_val->set_req(_slow_path, slow_result);
3760 result_io ->set_req(_slow_path, i_o());
3761 result_mem->set_req(_slow_path, reset_memory());
3762 }
3763
3764 set_control(normal_ctl);
3765 if (!stopped()) {
3766 // Normal case: The array type has been cached in the java.lang.Class.
3767 // The following call works fine even if the array type is polymorphic.
3768 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3769 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3770 result_reg->init_req(_normal_path, control());
3771 result_val->init_req(_normal_path, obj);
3772 result_io ->init_req(_normal_path, i_o());
3773 result_mem->init_req(_normal_path, reset_memory());
3774 }
3775
3776 // Return the combined state.
3777 set_i_o( _gvn.transform(result_io) );
3778 set_all_memory( _gvn.transform(result_mem));
3779
3780 C->set_has_split_ifs(true); // Has chance for split-if optimization
3781 set_result(result_reg, result_val);
3782 return true;
3783 }
3784
3785 //----------------------inline_native_getLength--------------------------
3786 // public static native int java.lang.reflect.Array.getLength(Object array);
3787 bool LibraryCallKit::inline_native_getLength() {
3788 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3789
3790 Node* array = null_check(argument(0));
3791 // If array is dead, only null-path is taken.
3792 if (stopped()) return true;
3793
3794 // Deoptimize if it is a non-array.
3795 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3796
3797 if (non_array != NULL) {
3798 PreserveJVMState pjvms(this);
3799 set_control(non_array);
3800 uncommon_trap(Deoptimization::Reason_intrinsic,
3801 Deoptimization::Action_maybe_recompile);
3802 }
3803
3804 // If control is dead, only non-array-path is taken.
3805 if (stopped()) return true;
3806
3807 // The works fine even if the array type is polymorphic.
3808 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3809 Node* result = load_array_length(array);
3810
3811 C->set_has_split_ifs(true); // Has chance for split-if optimization
3812 set_result(result);
3813 return true;
3814 }
3815
3816 //------------------------inline_array_copyOf----------------------------
3817 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3818 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3819 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3820 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3821
3822 // Get the arguments.
3823 Node* original = argument(0);
3824 Node* start = is_copyOfRange? argument(1): intcon(0);
3825 Node* end = is_copyOfRange? argument(2): argument(1);
3826 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3827
3828 Node* newcopy;
3829
3830 // Set the original stack and the reexecute bit for the interpreter to reexecute
3831 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3832 { PreserveReexecuteState preexecs(this);
3833 jvms()->set_should_reexecute(true);
3834
3835 array_type_mirror = null_check(array_type_mirror);
3836 original = null_check(original);
3837
3838 // Check if a null path was taken unconditionally.
3839 if (stopped()) return true;
3840
3841 Node* orig_length = load_array_length(original);
3842
3843 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3844 klass_node = null_check(klass_node);
3845
3846 RegionNode* bailout = new (C) RegionNode(1);
3847 record_for_igvn(bailout);
3848
3849 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3850 // Bail out if that is so.
3851 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3852 if (not_objArray != NULL) {
3853 // Improve the klass node's type from the new optimistic assumption:
3854 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3855 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3856 Node* cast = new (C) CastPPNode(klass_node, akls);
3857 cast->init_req(0, control());
3858 klass_node = _gvn.transform(cast);
3859 }
3860
3861 // Bail out if either start or end is negative.
3862 generate_negative_guard(start, bailout, &start);
3863 generate_negative_guard(end, bailout, &end);
3864
3865 Node* length = end;
3866 if (_gvn.type(start) != TypeInt::ZERO) {
3867 length = _gvn.transform(new (C) SubINode(end, start));
3868 }
3869
3870 // Bail out if length is negative.
3871 // Without this the new_array would throw
3872 // NegativeArraySizeException but IllegalArgumentException is what
3873 // should be thrown
3874 generate_negative_guard(length, bailout, &length);
3875
3876 if (bailout->req() > 1) {
3877 PreserveJVMState pjvms(this);
3878 set_control(_gvn.transform(bailout));
3879 uncommon_trap(Deoptimization::Reason_intrinsic,
3880 Deoptimization::Action_maybe_recompile);
3881 }
3882
3883 if (!stopped()) {
3884 // How many elements will we copy from the original?
3885 // The answer is MinI(orig_length - start, length).
3886 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3887 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3888
3889 newcopy = new_array(klass_node, length, 0); // no argments to push
3890
3891 // Generate a direct call to the right arraycopy function(s).
3892 // We know the copy is disjoint but we might not know if the
3893 // oop stores need checking.
3894 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3895 // This will fail a store-check if x contains any non-nulls.
3896 bool disjoint_bases = true;
3897 // if start > orig_length then the length of the copy may be
3898 // negative.
3899 bool length_never_negative = !is_copyOfRange;
3900 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3901 original, start, newcopy, intcon(0), moved,
3902 disjoint_bases, length_never_negative);
3903 }
3904 } // original reexecute is set back here
3905
3906 C->set_has_split_ifs(true); // Has chance for split-if optimization
3907 if (!stopped()) {
3908 set_result(newcopy);
3909 }
3910 return true;
3911 }
3912
3913
3914 //----------------------generate_virtual_guard---------------------------
3915 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3916 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3917 RegionNode* slow_region) {
3918 ciMethod* method = callee();
3919 int vtable_index = method->vtable_index();
3920 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3921 err_msg_res("bad index %d", vtable_index));
3922 // Get the Method* out of the appropriate vtable entry.
3923 int entry_offset = (InstanceKlass::vtable_start_offset() +
3924 vtable_index*vtableEntry::size()) * wordSize +
3925 vtableEntry::method_offset_in_bytes();
3926 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3927 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3928
3929 // Compare the target method with the expected method (e.g., Object.hashCode).
3930 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3931
3932 Node* native_call = makecon(native_call_addr);
3933 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3934 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3935
3936 return generate_slow_guard(test_native, slow_region);
3937 }
3938
3939 //-----------------------generate_method_call----------------------------
3940 // Use generate_method_call to make a slow-call to the real
3941 // method if the fast path fails. An alternative would be to
3942 // use a stub like OptoRuntime::slow_arraycopy_Java.
3943 // This only works for expanding the current library call,
3944 // not another intrinsic. (E.g., don't use this for making an
3945 // arraycopy call inside of the copyOf intrinsic.)
3946 CallJavaNode*
3947 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3948 // When compiling the intrinsic method itself, do not use this technique.
3949 guarantee(callee() != C->method(), "cannot make slow-call to self");
3950
3951 ciMethod* method = callee();
3952 // ensure the JVMS we have will be correct for this call
3953 guarantee(method_id == method->intrinsic_id(), "must match");
3954
3955 const TypeFunc* tf = TypeFunc::make(method);
3956 CallJavaNode* slow_call;
3957 if (is_static) {
3958 assert(!is_virtual, "");
3959 slow_call = new(C) CallStaticJavaNode(C, tf,
3960 SharedRuntime::get_resolve_static_call_stub(),
3961 method, bci());
3962 } else if (is_virtual) {
3963 null_check_receiver();
3964 int vtable_index = Method::invalid_vtable_index;
3965 if (UseInlineCaches) {
3966 // Suppress the vtable call
3967 } else {
3968 // hashCode and clone are not a miranda methods,
3969 // so the vtable index is fixed.
3970 // No need to use the linkResolver to get it.
3971 vtable_index = method->vtable_index();
3972 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3973 err_msg_res("bad index %d", vtable_index));
3974 }
3975 slow_call = new(C) CallDynamicJavaNode(tf,
3976 SharedRuntime::get_resolve_virtual_call_stub(),
3977 method, vtable_index, bci());
3978 } else { // neither virtual nor static: opt_virtual
3979 null_check_receiver();
3980 slow_call = new(C) CallStaticJavaNode(C, tf,
3981 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3982 method, bci());
3983 slow_call->set_optimized_virtual(true);
3984 }
3985 set_arguments_for_java_call(slow_call);
3986 set_edges_for_java_call(slow_call);
3987 return slow_call;
3988 }
3989
3990
3991 //------------------------------inline_native_hashcode--------------------
3992 // Build special case code for calls to hashCode on an object.
3993 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3994 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3995 assert(!(is_virtual && is_static), "either virtual, special, or static");
3996
3997 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3998
3999 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4000 PhiNode* result_val = new(C) PhiNode(result_reg,
4001 TypeInt::INT);
4002 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4003 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4004 TypePtr::BOTTOM);
4005 Node* obj = NULL;
4006 if (!is_static) {
4007 // Check for hashing null object
4008 obj = null_check_receiver();
4009 if (stopped()) return true; // unconditionally null
4010 result_reg->init_req(_null_path, top());
4011 result_val->init_req(_null_path, top());
4012 } else {
4013 // Do a null check, and return zero if null.
4014 // System.identityHashCode(null) == 0
4015 obj = argument(0);
4016 Node* null_ctl = top();
4017 obj = null_check_oop(obj, &null_ctl);
4018 result_reg->init_req(_null_path, null_ctl);
4019 result_val->init_req(_null_path, _gvn.intcon(0));
4020 }
4021
4022 // Unconditionally null? Then return right away.
4023 if (stopped()) {
4024 set_control( result_reg->in(_null_path));
4025 if (!stopped())
4026 set_result(result_val->in(_null_path));
4027 return true;
4028 }
4029
4030 // After null check, get the object's klass.
4031 Node* obj_klass = load_object_klass(obj);
4032
4033 // This call may be virtual (invokevirtual) or bound (invokespecial).
4034 // For each case we generate slightly different code.
4035
4036 // We only go to the fast case code if we pass a number of guards. The
4037 // paths which do not pass are accumulated in the slow_region.
4038 RegionNode* slow_region = new (C) RegionNode(1);
4039 record_for_igvn(slow_region);
4040
4041 // If this is a virtual call, we generate a funny guard. We pull out
4042 // the vtable entry corresponding to hashCode() from the target object.
4043 // If the target method which we are calling happens to be the native
4044 // Object hashCode() method, we pass the guard. We do not need this
4045 // guard for non-virtual calls -- the caller is known to be the native
4046 // Object hashCode().
4047 if (is_virtual) {
4048 generate_virtual_guard(obj_klass, slow_region);
4049 }
4050
4051 // Get the header out of the object, use LoadMarkNode when available
4052 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4053 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4054
4055 // Test the header to see if it is unlocked.
4056 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4057 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4058 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4059 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4060 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4061
4062 generate_slow_guard(test_unlocked, slow_region);
4063
4064 // Get the hash value and check to see that it has been properly assigned.
4065 // We depend on hash_mask being at most 32 bits and avoid the use of
4066 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4067 // vm: see markOop.hpp.
4068 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4069 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4070 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4071 // This hack lets the hash bits live anywhere in the mark object now, as long
4072 // as the shift drops the relevant bits into the low 32 bits. Note that
4073 // Java spec says that HashCode is an int so there's no point in capturing
4074 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4075 hshifted_header = ConvX2I(hshifted_header);
4076 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4077
4078 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4079 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4080 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4081
4082 generate_slow_guard(test_assigned, slow_region);
4083
4084 Node* init_mem = reset_memory();
4085 // fill in the rest of the null path:
4086 result_io ->init_req(_null_path, i_o());
4087 result_mem->init_req(_null_path, init_mem);
4088
4089 result_val->init_req(_fast_path, hash_val);
4090 result_reg->init_req(_fast_path, control());
4091 result_io ->init_req(_fast_path, i_o());
4092 result_mem->init_req(_fast_path, init_mem);
4093
4094 // Generate code for the slow case. We make a call to hashCode().
4095 set_control(_gvn.transform(slow_region));
4096 if (!stopped()) {
4097 // No need for PreserveJVMState, because we're using up the present state.
4098 set_all_memory(init_mem);
4099 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4100 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4101 Node* slow_result = set_results_for_java_call(slow_call);
4102 // this->control() comes from set_results_for_java_call
4103 result_reg->init_req(_slow_path, control());
4104 result_val->init_req(_slow_path, slow_result);
4105 result_io ->set_req(_slow_path, i_o());
4106 result_mem ->set_req(_slow_path, reset_memory());
4107 }
4108
4109 // Return the combined state.
4110 set_i_o( _gvn.transform(result_io) );
4111 set_all_memory( _gvn.transform(result_mem));
4112
4113 set_result(result_reg, result_val);
4114 return true;
4115 }
4116
4117 //---------------------------inline_native_getClass----------------------------
4118 // public final native Class<?> java.lang.Object.getClass();
4119 //
4120 // Build special case code for calls to getClass on an object.
4121 bool LibraryCallKit::inline_native_getClass() {
4122 Node* obj = null_check_receiver();
4123 if (stopped()) return true;
4124 set_result(load_mirror_from_klass(load_object_klass(obj)));
4125 return true;
4126 }
4127
4128 //-----------------inline_native_Reflection_getCallerClass---------------------
4129 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4130 //
4131 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4132 //
4133 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4134 // in that it must skip particular security frames and checks for
4135 // caller sensitive methods.
4136 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4137 #ifndef PRODUCT
4138 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4139 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4140 }
4141 #endif
4142
4143 if (!jvms()->has_method()) {
4144 #ifndef PRODUCT
4145 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4146 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4147 }
4148 #endif
4149 return false;
4150 }
4151
4152 // Walk back up the JVM state to find the caller at the required
4153 // depth.
4154 JVMState* caller_jvms = jvms();
4155
4156 // Cf. JVM_GetCallerClass
4157 // NOTE: Start the loop at depth 1 because the current JVM state does
4158 // not include the Reflection.getCallerClass() frame.
4159 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4160 ciMethod* m = caller_jvms->method();
4161 switch (n) {
4162 case 0:
4163 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4164 break;
4165 case 1:
4166 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4167 if (!m->caller_sensitive()) {
4168 #ifndef PRODUCT
4169 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4170 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4171 }
4172 #endif
4173 return false; // bail-out; let JVM_GetCallerClass do the work
4174 }
4175 break;
4176 default:
4177 if (!m->is_ignored_by_security_stack_walk()) {
4178 // We have reached the desired frame; return the holder class.
4179 // Acquire method holder as java.lang.Class and push as constant.
4180 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4181 ciInstance* caller_mirror = caller_klass->java_mirror();
4182 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4183
4184 #ifndef PRODUCT
4185 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4186 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());
4187 tty->print_cr(" JVM state at this point:");
4188 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4189 ciMethod* m = jvms()->of_depth(i)->method();
4190 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4191 }
4192 }
4193 #endif
4194 return true;
4195 }
4196 break;
4197 }
4198 }
4199
4200 #ifndef PRODUCT
4201 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4202 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4203 tty->print_cr(" JVM state at this point:");
4204 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4205 ciMethod* m = jvms()->of_depth(i)->method();
4206 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4207 }
4208 }
4209 #endif
4210
4211 return false; // bail-out; let JVM_GetCallerClass do the work
4212 }
4213
4214 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4215 Node* arg = argument(0);
4216 Node* result;
4217
4218 switch (id) {
4219 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4220 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4221 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4222 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4223
4224 case vmIntrinsics::_doubleToLongBits: {
4225 // two paths (plus control) merge in a wood
4226 RegionNode *r = new (C) RegionNode(3);
4227 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4228
4229 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4230 // Build the boolean node
4231 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4232
4233 // Branch either way.
4234 // NaN case is less traveled, which makes all the difference.
4235 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4236 Node *opt_isnan = _gvn.transform(ifisnan);
4237 assert( opt_isnan->is_If(), "Expect an IfNode");
4238 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4239 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4240
4241 set_control(iftrue);
4242
4243 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4244 Node *slow_result = longcon(nan_bits); // return NaN
4245 phi->init_req(1, _gvn.transform( slow_result ));
4246 r->init_req(1, iftrue);
4247
4248 // Else fall through
4249 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4250 set_control(iffalse);
4251
4252 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4253 r->init_req(2, iffalse);
4254
4255 // Post merge
4256 set_control(_gvn.transform(r));
4257 record_for_igvn(r);
4258
4259 C->set_has_split_ifs(true); // Has chance for split-if optimization
4260 result = phi;
4261 assert(result->bottom_type()->isa_long(), "must be");
4262 break;
4263 }
4264
4265 case vmIntrinsics::_floatToIntBits: {
4266 // two paths (plus control) merge in a wood
4267 RegionNode *r = new (C) RegionNode(3);
4268 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4269
4270 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4271 // Build the boolean node
4272 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4273
4274 // Branch either way.
4275 // NaN case is less traveled, which makes all the difference.
4276 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4277 Node *opt_isnan = _gvn.transform(ifisnan);
4278 assert( opt_isnan->is_If(), "Expect an IfNode");
4279 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4280 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4281
4282 set_control(iftrue);
4283
4284 static const jint nan_bits = 0x7fc00000;
4285 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4286 phi->init_req(1, _gvn.transform( slow_result ));
4287 r->init_req(1, iftrue);
4288
4289 // Else fall through
4290 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4291 set_control(iffalse);
4292
4293 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4294 r->init_req(2, iffalse);
4295
4296 // Post merge
4297 set_control(_gvn.transform(r));
4298 record_for_igvn(r);
4299
4300 C->set_has_split_ifs(true); // Has chance for split-if optimization
4301 result = phi;
4302 assert(result->bottom_type()->isa_int(), "must be");
4303 break;
4304 }
4305
4306 default:
4307 fatal_unexpected_iid(id);
4308 break;
4309 }
4310 set_result(_gvn.transform(result));
4311 return true;
4312 }
4313
4314 #ifdef _LP64
4315 #define XTOP ,top() /*additional argument*/
4316 #else //_LP64
4317 #define XTOP /*no additional argument*/
4318 #endif //_LP64
4319
4320 //----------------------inline_unsafe_copyMemory-------------------------
4321 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4322 bool LibraryCallKit::inline_unsafe_copyMemory() {
4323 if (callee()->is_static()) return false; // caller must have the capability!
4324 null_check_receiver(); // null-check receiver
4325 if (stopped()) return true;
4326
4327 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4328
4329 Node* src_ptr = argument(1); // type: oop
4330 Node* src_off = ConvL2X(argument(2)); // type: long
4331 Node* dst_ptr = argument(4); // type: oop
4332 Node* dst_off = ConvL2X(argument(5)); // type: long
4333 Node* size = ConvL2X(argument(7)); // type: long
4334
4335 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4336 "fieldOffset must be byte-scaled");
4337
4338 Node* src = make_unsafe_address(src_ptr, src_off);
4339 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4340
4341 // Conservatively insert a memory barrier on all memory slices.
4342 // Do not let writes of the copy source or destination float below the copy.
4343 insert_mem_bar(Op_MemBarCPUOrder);
4344
4345 // Call it. Note that the length argument is not scaled.
4346 make_runtime_call(RC_LEAF|RC_NO_FP,
4347 OptoRuntime::fast_arraycopy_Type(),
4348 StubRoutines::unsafe_arraycopy(),
4349 "unsafe_arraycopy",
4350 TypeRawPtr::BOTTOM,
4351 src, dst, size XTOP);
4352
4353 // Do not let reads of the copy destination float above the copy.
4354 insert_mem_bar(Op_MemBarCPUOrder);
4355
4356 return true;
4357 }
4358
4359 //------------------------clone_coping-----------------------------------
4360 // Helper function for inline_native_clone.
4361 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4362 assert(obj_size != NULL, "");
4363 Node* raw_obj = alloc_obj->in(1);
4364 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4365
4366 AllocateNode* alloc = NULL;
4367 if (ReduceBulkZeroing) {
4368 // We will be completely responsible for initializing this object -
4369 // mark Initialize node as complete.
4370 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4371 // The object was just allocated - there should be no any stores!
4372 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4373 // Mark as complete_with_arraycopy so that on AllocateNode
4374 // expansion, we know this AllocateNode is initialized by an array
4375 // copy and a StoreStore barrier exists after the array copy.
4376 alloc->initialization()->set_complete_with_arraycopy();
4377 }
4378
4379 // Copy the fastest available way.
4380 // TODO: generate fields copies for small objects instead.
4381 Node* src = obj;
4382 Node* dest = alloc_obj;
4383 Node* size = _gvn.transform(obj_size);
4384
4385 // Exclude the header but include array length to copy by 8 bytes words.
4386 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4387 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4388 instanceOopDesc::base_offset_in_bytes();
4389 // base_off:
4390 // 8 - 32-bit VM
4391 // 12 - 64-bit VM, compressed klass
4392 // 16 - 64-bit VM, normal klass
4393 if (base_off % BytesPerLong != 0) {
4394 assert(UseCompressedClassPointers, "");
4395 if (is_array) {
4396 // Exclude length to copy by 8 bytes words.
4397 base_off += sizeof(int);
4398 } else {
4399 // Include klass to copy by 8 bytes words.
4400 base_off = instanceOopDesc::klass_offset_in_bytes();
4401 }
4402 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4403 }
4404 src = basic_plus_adr(src, base_off);
4405 dest = basic_plus_adr(dest, base_off);
4406
4407 // Compute the length also, if needed:
4408 Node* countx = size;
4409 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4410 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4411
4412 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4413 bool disjoint_bases = true;
4414 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4415 src, NULL, dest, NULL, countx,
4416 /*dest_uninitialized*/true);
4417
4418 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4419 if (card_mark) {
4420 assert(!is_array, "");
4421 // Put in store barrier for any and all oops we are sticking
4422 // into this object. (We could avoid this if we could prove
4423 // that the object type contains no oop fields at all.)
4424 Node* no_particular_value = NULL;
4425 Node* no_particular_field = NULL;
4426 int raw_adr_idx = Compile::AliasIdxRaw;
4427 post_barrier(control(),
4428 memory(raw_adr_type),
4429 alloc_obj,
4430 no_particular_field,
4431 raw_adr_idx,
4432 no_particular_value,
4433 T_OBJECT,
4434 false);
4435 }
4436
4437 // Do not let reads from the cloned object float above the arraycopy.
4438 if (alloc != NULL) {
4439 // Do not let stores that initialize this object be reordered with
4440 // a subsequent store that would make this object accessible by
4441 // other threads.
4442 // Record what AllocateNode this StoreStore protects so that
4443 // escape analysis can go from the MemBarStoreStoreNode to the
4444 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4445 // based on the escape status of the AllocateNode.
4446 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4447 } else {
4448 insert_mem_bar(Op_MemBarCPUOrder);
4449 }
4450 }
4451
4452 //------------------------inline_native_clone----------------------------
4453 // protected native Object java.lang.Object.clone();
4454 //
4455 // Here are the simple edge cases:
4456 // null receiver => normal trap
4457 // virtual and clone was overridden => slow path to out-of-line clone
4458 // not cloneable or finalizer => slow path to out-of-line Object.clone
4459 //
4460 // The general case has two steps, allocation and copying.
4461 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4462 //
4463 // Copying also has two cases, oop arrays and everything else.
4464 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4465 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4466 //
4467 // These steps fold up nicely if and when the cloned object's klass
4468 // can be sharply typed as an object array, a type array, or an instance.
4469 //
4470 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4471 PhiNode* result_val;
4472
4473 // Set the reexecute bit for the interpreter to reexecute
4474 // the bytecode that invokes Object.clone if deoptimization happens.
4475 { PreserveReexecuteState preexecs(this);
4476 jvms()->set_should_reexecute(true);
4477
4478 Node* obj = null_check_receiver();
4479 if (stopped()) return true;
4480
4481 Node* obj_klass = load_object_klass(obj);
4482 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4483 const TypeOopPtr* toop = ((tklass != NULL)
4484 ? tklass->as_instance_type()
4485 : TypeInstPtr::NOTNULL);
4486
4487 // Conservatively insert a memory barrier on all memory slices.
4488 // Do not let writes into the original float below the clone.
4489 insert_mem_bar(Op_MemBarCPUOrder);
4490
4491 // paths into result_reg:
4492 enum {
4493 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4494 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4495 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4496 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4497 PATH_LIMIT
4498 };
4499 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4500 result_val = new(C) PhiNode(result_reg,
4501 TypeInstPtr::NOTNULL);
4502 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4503 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4504 TypePtr::BOTTOM);
4505 record_for_igvn(result_reg);
4506
4507 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4508 int raw_adr_idx = Compile::AliasIdxRaw;
4509
4510 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4511 if (array_ctl != NULL) {
4512 // It's an array.
4513 PreserveJVMState pjvms(this);
4514 set_control(array_ctl);
4515 Node* obj_length = load_array_length(obj);
4516 Node* obj_size = NULL;
4517 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4518
4519 if (!use_ReduceInitialCardMarks()) {
4520 // If it is an oop array, it requires very special treatment,
4521 // because card marking is required on each card of the array.
4522 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4523 if (is_obja != NULL) {
4524 PreserveJVMState pjvms2(this);
4525 set_control(is_obja);
4526 // Generate a direct call to the right arraycopy function(s).
4527 bool disjoint_bases = true;
4528 bool length_never_negative = true;
4529 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4530 obj, intcon(0), alloc_obj, intcon(0),
4531 obj_length,
4532 disjoint_bases, length_never_negative);
4533 result_reg->init_req(_objArray_path, control());
4534 result_val->init_req(_objArray_path, alloc_obj);
4535 result_i_o ->set_req(_objArray_path, i_o());
4536 result_mem ->set_req(_objArray_path, reset_memory());
4537 }
4538 }
4539 // Otherwise, there are no card marks to worry about.
4540 // (We can dispense with card marks if we know the allocation
4541 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4542 // causes the non-eden paths to take compensating steps to
4543 // simulate a fresh allocation, so that no further
4544 // card marks are required in compiled code to initialize
4545 // the object.)
4546
4547 if (!stopped()) {
4548 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4549
4550 // Present the results of the copy.
4551 result_reg->init_req(_array_path, control());
4552 result_val->init_req(_array_path, alloc_obj);
4553 result_i_o ->set_req(_array_path, i_o());
4554 result_mem ->set_req(_array_path, reset_memory());
4555 }
4556 }
4557
4558 // We only go to the instance fast case code if we pass a number of guards.
4559 // The paths which do not pass are accumulated in the slow_region.
4560 RegionNode* slow_region = new (C) RegionNode(1);
4561 record_for_igvn(slow_region);
4562 if (!stopped()) {
4563 // It's an instance (we did array above). Make the slow-path tests.
4564 // If this is a virtual call, we generate a funny guard. We grab
4565 // the vtable entry corresponding to clone() from the target object.
4566 // If the target method which we are calling happens to be the
4567 // Object clone() method, we pass the guard. We do not need this
4568 // guard for non-virtual calls; the caller is known to be the native
4569 // Object clone().
4570 if (is_virtual) {
4571 generate_virtual_guard(obj_klass, slow_region);
4572 }
4573
4574 // The object must be cloneable and must not have a finalizer.
4575 // Both of these conditions may be checked in a single test.
4576 // We could optimize the cloneable test further, but we don't care.
4577 generate_access_flags_guard(obj_klass,
4578 // Test both conditions:
4579 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4580 // Must be cloneable but not finalizer:
4581 JVM_ACC_IS_CLONEABLE,
4582 slow_region);
4583 }
4584
4585 if (!stopped()) {
4586 // It's an instance, and it passed the slow-path tests.
4587 PreserveJVMState pjvms(this);
4588 Node* obj_size = NULL;
4589 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4590
4591 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4592
4593 // Present the results of the slow call.
4594 result_reg->init_req(_instance_path, control());
4595 result_val->init_req(_instance_path, alloc_obj);
4596 result_i_o ->set_req(_instance_path, i_o());
4597 result_mem ->set_req(_instance_path, reset_memory());
4598 }
4599
4600 // Generate code for the slow case. We make a call to clone().
4601 set_control(_gvn.transform(slow_region));
4602 if (!stopped()) {
4603 PreserveJVMState pjvms(this);
4604 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4605 Node* slow_result = set_results_for_java_call(slow_call);
4606 // this->control() comes from set_results_for_java_call
4607 result_reg->init_req(_slow_path, control());
4608 result_val->init_req(_slow_path, slow_result);
4609 result_i_o ->set_req(_slow_path, i_o());
4610 result_mem ->set_req(_slow_path, reset_memory());
4611 }
4612
4613 // Return the combined state.
4614 set_control( _gvn.transform(result_reg));
4615 set_i_o( _gvn.transform(result_i_o));
4616 set_all_memory( _gvn.transform(result_mem));
4617 } // original reexecute is set back here
4618
4619 set_result(_gvn.transform(result_val));
4620 return true;
4621 }
4622
4623 //------------------------------basictype2arraycopy----------------------------
4624 address LibraryCallKit::basictype2arraycopy(BasicType t,
4625 Node* src_offset,
4626 Node* dest_offset,
4627 bool disjoint_bases,
4628 const char* &name,
4629 bool dest_uninitialized) {
4630 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4631 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4632
4633 bool aligned = false;
4634 bool disjoint = disjoint_bases;
4635
4636 // if the offsets are the same, we can treat the memory regions as
4637 // disjoint, because either the memory regions are in different arrays,
4638 // or they are identical (which we can treat as disjoint.) We can also
4639 // treat a copy with a destination index less that the source index
4640 // as disjoint since a low->high copy will work correctly in this case.
4641 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4642 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4643 // both indices are constants
4644 int s_offs = src_offset_inttype->get_con();
4645 int d_offs = dest_offset_inttype->get_con();
4646 int element_size = type2aelembytes(t);
4647 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4648 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4649 if (s_offs >= d_offs) disjoint = true;
4650 } else if (src_offset == dest_offset && src_offset != NULL) {
4651 // This can occur if the offsets are identical non-constants.
4652 disjoint = true;
4653 }
4654
4655 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4656 }
4657
4658
4659 //------------------------------inline_arraycopy-----------------------
4660 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4661 // Object dest, int destPos,
4662 // int length);
4663 bool LibraryCallKit::inline_arraycopy() {
4664 // Get the arguments.
4665 Node* src = argument(0); // type: oop
4666 Node* src_offset = argument(1); // type: int
4667 Node* dest = argument(2); // type: oop
4668 Node* dest_offset = argument(3); // type: int
4669 Node* length = argument(4); // type: int
4670
4671 // Compile time checks. If any of these checks cannot be verified at compile time,
4672 // we do not make a fast path for this call. Instead, we let the call remain as it
4673 // is. The checks we choose to mandate at compile time are:
4674 //
4675 // (1) src and dest are arrays.
4676 const Type* src_type = src->Value(&_gvn);
4677 const Type* dest_type = dest->Value(&_gvn);
4678 const TypeAryPtr* top_src = src_type->isa_aryptr();
4679 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4680
4681 // Do we have the type of src?
4682 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4683 // Do we have the type of dest?
4684 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4685 // Is the type for src from speculation?
4686 bool src_spec = false;
4687 // Is the type for dest from speculation?
4688 bool dest_spec = false;
4689
4690 if (!has_src || !has_dest) {
4691 // We don't have sufficient type information, let's see if
4692 // speculative types can help. We need to have types for both src
4693 // and dest so that it pays off.
4694
4695 // Do we already have or could we have type information for src
4696 bool could_have_src = has_src;
4697 // Do we already have or could we have type information for dest
4698 bool could_have_dest = has_dest;
4699
4700 ciKlass* src_k = NULL;
4701 if (!has_src) {
4702 src_k = src_type->speculative_type_not_null();
4703 if (src_k != NULL && src_k->is_array_klass()) {
4704 could_have_src = true;
4705 }
4706 }
4707
4708 ciKlass* dest_k = NULL;
4709 if (!has_dest) {
4710 dest_k = dest_type->speculative_type_not_null();
4711 if (dest_k != NULL && dest_k->is_array_klass()) {
4712 could_have_dest = true;
4713 }
4714 }
4715
4716 if (could_have_src && could_have_dest) {
4717 // This is going to pay off so emit the required guards
4718 if (!has_src) {
4719 src = maybe_cast_profiled_obj(src, src_k);
4720 src_type = _gvn.type(src);
4721 top_src = src_type->isa_aryptr();
4722 has_src = (top_src != NULL && top_src->klass() != NULL);
4723 src_spec = true;
4724 }
4725 if (!has_dest) {
4726 dest = maybe_cast_profiled_obj(dest, dest_k);
4727 dest_type = _gvn.type(dest);
4728 top_dest = dest_type->isa_aryptr();
4729 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4730 dest_spec = true;
4731 }
4732 }
4733 }
4734
4735 if (!has_src || !has_dest) {
4736 // Conservatively insert a memory barrier on all memory slices.
4737 // Do not let writes into the source float below the arraycopy.
4738 insert_mem_bar(Op_MemBarCPUOrder);
4739
4740 // Call StubRoutines::generic_arraycopy stub.
4741 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4742 src, src_offset, dest, dest_offset, length);
4743
4744 // Do not let reads from the destination float above the arraycopy.
4745 // Since we cannot type the arrays, we don't know which slices
4746 // might be affected. We could restrict this barrier only to those
4747 // memory slices which pertain to array elements--but don't bother.
4748 if (!InsertMemBarAfterArraycopy)
4749 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4750 insert_mem_bar(Op_MemBarCPUOrder);
4751 return true;
4752 }
4753
4754 // (2) src and dest arrays must have elements of the same BasicType
4755 // Figure out the size and type of the elements we will be copying.
4756 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4757 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4758 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4759 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4760
4761 if (src_elem != dest_elem || dest_elem == T_VOID) {
4762 // The component types are not the same or are not recognized. Punt.
4763 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4764 generate_slow_arraycopy(TypePtr::BOTTOM,
4765 src, src_offset, dest, dest_offset, length,
4766 /*dest_uninitialized*/false);
4767 return true;
4768 }
4769
4770 if (src_elem == T_OBJECT) {
4771 // If both arrays are object arrays then having the exact types
4772 // for both will remove the need for a subtype check at runtime
4773 // before the call and may make it possible to pick a faster copy
4774 // routine (without a subtype check on every element)
4775 // Do we have the exact type of src?
4776 bool could_have_src = src_spec;
4777 // Do we have the exact type of dest?
4778 bool could_have_dest = dest_spec;
4779 ciKlass* src_k = top_src->klass();
4780 ciKlass* dest_k = top_dest->klass();
4781 if (!src_spec) {
4782 src_k = src_type->speculative_type_not_null();
4783 if (src_k != NULL && src_k->is_array_klass()) {
4784 could_have_src = true;
4785 }
4786 }
4787 if (!dest_spec) {
4788 dest_k = dest_type->speculative_type_not_null();
4789 if (dest_k != NULL && dest_k->is_array_klass()) {
4790 could_have_dest = true;
4791 }
4792 }
4793 if (could_have_src && could_have_dest) {
4794 // If we can have both exact types, emit the missing guards
4795 if (could_have_src && !src_spec) {
4796 src = maybe_cast_profiled_obj(src, src_k);
4797 }
4798 if (could_have_dest && !dest_spec) {
4799 dest = maybe_cast_profiled_obj(dest, dest_k);
4800 }
4801 }
4802 }
4803
4804 //---------------------------------------------------------------------------
4805 // We will make a fast path for this call to arraycopy.
4806
4807 // We have the following tests left to perform:
4808 //
4809 // (3) src and dest must not be null.
4810 // (4) src_offset must not be negative.
4811 // (5) dest_offset must not be negative.
4812 // (6) length must not be negative.
4813 // (7) src_offset + length must not exceed length of src.
4814 // (8) dest_offset + length must not exceed length of dest.
4815 // (9) each element of an oop array must be assignable
4816
4817 RegionNode* slow_region = new (C) RegionNode(1);
4818 record_for_igvn(slow_region);
4819
4820 // (3) operands must not be null
4821 // We currently perform our null checks with the null_check routine.
4822 // This means that the null exceptions will be reported in the caller
4823 // rather than (correctly) reported inside of the native arraycopy call.
4824 // This should be corrected, given time. We do our null check with the
4825 // stack pointer restored.
4826 src = null_check(src, T_ARRAY);
4827 dest = null_check(dest, T_ARRAY);
4828
4829 // (4) src_offset must not be negative.
4830 generate_negative_guard(src_offset, slow_region);
4831
4832 // (5) dest_offset must not be negative.
4833 generate_negative_guard(dest_offset, slow_region);
4834
4835 // (6) length must not be negative (moved to generate_arraycopy()).
4836 // generate_negative_guard(length, slow_region);
4837
4838 // (7) src_offset + length must not exceed length of src.
4839 generate_limit_guard(src_offset, length,
4840 load_array_length(src),
4841 slow_region);
4842
4843 // (8) dest_offset + length must not exceed length of dest.
4844 generate_limit_guard(dest_offset, length,
4845 load_array_length(dest),
4846 slow_region);
4847
4848 // (9) each element of an oop array must be assignable
4849 // The generate_arraycopy subroutine checks this.
4850
4851 // This is where the memory effects are placed:
4852 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4853 generate_arraycopy(adr_type, dest_elem,
4854 src, src_offset, dest, dest_offset, length,
4855 false, false, slow_region);
4856
4857 return true;
4858 }
4859
4860 //-----------------------------generate_arraycopy----------------------
4861 // Generate an optimized call to arraycopy.
4862 // Caller must guard against non-arrays.
4863 // Caller must determine a common array basic-type for both arrays.
4864 // Caller must validate offsets against array bounds.
4865 // The slow_region has already collected guard failure paths
4866 // (such as out of bounds length or non-conformable array types).
4867 // The generated code has this shape, in general:
4868 //
4869 // if (length == 0) return // via zero_path
4870 // slowval = -1
4871 // if (types unknown) {
4872 // slowval = call generic copy loop
4873 // if (slowval == 0) return // via checked_path
4874 // } else if (indexes in bounds) {
4875 // if ((is object array) && !(array type check)) {
4876 // slowval = call checked copy loop
4877 // if (slowval == 0) return // via checked_path
4878 // } else {
4879 // call bulk copy loop
4880 // return // via fast_path
4881 // }
4882 // }
4883 // // adjust params for remaining work:
4884 // if (slowval != -1) {
4885 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4886 // }
4887 // slow_region:
4888 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4889 // return // via slow_call_path
4890 //
4891 // This routine is used from several intrinsics: System.arraycopy,
4892 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4893 //
4894 void
4895 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4896 BasicType basic_elem_type,
4897 Node* src, Node* src_offset,
4898 Node* dest, Node* dest_offset,
4899 Node* copy_length,
4900 bool disjoint_bases,
4901 bool length_never_negative,
4902 RegionNode* slow_region) {
4903
4904 if (slow_region == NULL) {
4905 slow_region = new(C) RegionNode(1);
4906 record_for_igvn(slow_region);
4907 }
4908
4909 Node* original_dest = dest;
4910 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4911 bool dest_uninitialized = false;
4912
4913 // See if this is the initialization of a newly-allocated array.
4914 // If so, we will take responsibility here for initializing it to zero.
4915 // (Note: Because tightly_coupled_allocation performs checks on the
4916 // out-edges of the dest, we need to avoid making derived pointers
4917 // from it until we have checked its uses.)
4918 if (ReduceBulkZeroing
4919 && !ZeroTLAB // pointless if already zeroed
4920 && basic_elem_type != T_CONFLICT // avoid corner case
4921 && !src->eqv_uncast(dest)
4922 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4923 != NULL)
4924 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4925 && alloc->maybe_set_complete(&_gvn)) {
4926 // "You break it, you buy it."
4927 InitializeNode* init = alloc->initialization();
4928 assert(init->is_complete(), "we just did this");
4929 init->set_complete_with_arraycopy();
4930 assert(dest->is_CheckCastPP(), "sanity");
4931 assert(dest->in(0)->in(0) == init, "dest pinned");
4932 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4933 // From this point on, every exit path is responsible for
4934 // initializing any non-copied parts of the object to zero.
4935 // Also, if this flag is set we make sure that arraycopy interacts properly
4936 // with G1, eliding pre-barriers. See CR 6627983.
4937 dest_uninitialized = true;
4938 } else {
4939 // No zeroing elimination here.
4940 alloc = NULL;
4941 //original_dest = dest;
4942 //dest_uninitialized = false;
4943 }
4944
4945 // Results are placed here:
4946 enum { fast_path = 1, // normal void-returning assembly stub
4947 checked_path = 2, // special assembly stub with cleanup
4948 slow_call_path = 3, // something went wrong; call the VM
4949 zero_path = 4, // bypass when length of copy is zero
4950 bcopy_path = 5, // copy primitive array by 64-bit blocks
4951 PATH_LIMIT = 6
4952 };
4953 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4954 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4955 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4956 record_for_igvn(result_region);
4957 _gvn.set_type_bottom(result_i_o);
4958 _gvn.set_type_bottom(result_memory);
4959 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4960
4961 // The slow_control path:
4962 Node* slow_control;
4963 Node* slow_i_o = i_o();
4964 Node* slow_mem = memory(adr_type);
4965 debug_only(slow_control = (Node*) badAddress);
4966
4967 // Checked control path:
4968 Node* checked_control = top();
4969 Node* checked_mem = NULL;
4970 Node* checked_i_o = NULL;
4971 Node* checked_value = NULL;
4972
4973 if (basic_elem_type == T_CONFLICT) {
4974 assert(!dest_uninitialized, "");
4975 Node* cv = generate_generic_arraycopy(adr_type,
4976 src, src_offset, dest, dest_offset,
4977 copy_length, dest_uninitialized);
4978 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4979 checked_control = control();
4980 checked_i_o = i_o();
4981 checked_mem = memory(adr_type);
4982 checked_value = cv;
4983 set_control(top()); // no fast path
4984 }
4985
4986 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4987 if (not_pos != NULL) {
4988 PreserveJVMState pjvms(this);
4989 set_control(not_pos);
4990
4991 // (6) length must not be negative.
4992 if (!length_never_negative) {
4993 generate_negative_guard(copy_length, slow_region);
4994 }
4995
4996 // copy_length is 0.
4997 if (!stopped() && dest_uninitialized) {
4998 Node* dest_length = alloc->in(AllocateNode::ALength);
4999 if (copy_length->eqv_uncast(dest_length)
5000 || _gvn.find_int_con(dest_length, 1) <= 0) {
5001 // There is no zeroing to do. No need for a secondary raw memory barrier.
5002 } else {
5003 // Clear the whole thing since there are no source elements to copy.
5004 generate_clear_array(adr_type, dest, basic_elem_type,
5005 intcon(0), NULL,
5006 alloc->in(AllocateNode::AllocSize));
5007 // Use a secondary InitializeNode as raw memory barrier.
5008 // Currently it is needed only on this path since other
5009 // paths have stub or runtime calls as raw memory barriers.
5010 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5011 Compile::AliasIdxRaw,
5012 top())->as_Initialize();
5013 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5014 }
5015 }
5016
5017 // Present the results of the fast call.
5018 result_region->init_req(zero_path, control());
5019 result_i_o ->init_req(zero_path, i_o());
5020 result_memory->init_req(zero_path, memory(adr_type));
5021 }
5022
5023 if (!stopped() && dest_uninitialized) {
5024 // We have to initialize the *uncopied* part of the array to zero.
5025 // The copy destination is the slice dest[off..off+len]. The other slices
5026 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5027 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5028 Node* dest_length = alloc->in(AllocateNode::ALength);
5029 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5030 copy_length));
5031
5032 // If there is a head section that needs zeroing, do it now.
5033 if (find_int_con(dest_offset, -1) != 0) {
5034 generate_clear_array(adr_type, dest, basic_elem_type,
5035 intcon(0), dest_offset,
5036 NULL);
5037 }
5038
5039 // Next, perform a dynamic check on the tail length.
5040 // It is often zero, and we can win big if we prove this.
5041 // There are two wins: Avoid generating the ClearArray
5042 // with its attendant messy index arithmetic, and upgrade
5043 // the copy to a more hardware-friendly word size of 64 bits.
5044 Node* tail_ctl = NULL;
5045 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5046 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5047 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5048 tail_ctl = generate_slow_guard(bol_lt, NULL);
5049 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5050 }
5051
5052 // At this point, let's assume there is no tail.
5053 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5054 // There is no tail. Try an upgrade to a 64-bit copy.
5055 bool didit = false;
5056 { PreserveJVMState pjvms(this);
5057 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5058 src, src_offset, dest, dest_offset,
5059 dest_size, dest_uninitialized);
5060 if (didit) {
5061 // Present the results of the block-copying fast call.
5062 result_region->init_req(bcopy_path, control());
5063 result_i_o ->init_req(bcopy_path, i_o());
5064 result_memory->init_req(bcopy_path, memory(adr_type));
5065 }
5066 }
5067 if (didit)
5068 set_control(top()); // no regular fast path
5069 }
5070
5071 // Clear the tail, if any.
5072 if (tail_ctl != NULL) {
5073 Node* notail_ctl = stopped() ? NULL : control();
5074 set_control(tail_ctl);
5075 if (notail_ctl == NULL) {
5076 generate_clear_array(adr_type, dest, basic_elem_type,
5077 dest_tail, NULL,
5078 dest_size);
5079 } else {
5080 // Make a local merge.
5081 Node* done_ctl = new(C) RegionNode(3);
5082 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5083 done_ctl->init_req(1, notail_ctl);
5084 done_mem->init_req(1, memory(adr_type));
5085 generate_clear_array(adr_type, dest, basic_elem_type,
5086 dest_tail, NULL,
5087 dest_size);
5088 done_ctl->init_req(2, control());
5089 done_mem->init_req(2, memory(adr_type));
5090 set_control( _gvn.transform(done_ctl));
5091 set_memory( _gvn.transform(done_mem), adr_type );
5092 }
5093 }
5094 }
5095
5096 BasicType copy_type = basic_elem_type;
5097 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5098 if (!stopped() && copy_type == T_OBJECT) {
5099 // If src and dest have compatible element types, we can copy bits.
5100 // Types S[] and D[] are compatible if D is a supertype of S.
5101 //
5102 // If they are not, we will use checked_oop_disjoint_arraycopy,
5103 // which performs a fast optimistic per-oop check, and backs off
5104 // further to JVM_ArrayCopy on the first per-oop check that fails.
5105 // (Actually, we don't move raw bits only; the GC requires card marks.)
5106
5107 // Get the Klass* for both src and dest
5108 Node* src_klass = load_object_klass(src);
5109 Node* dest_klass = load_object_klass(dest);
5110
5111 // Generate the subtype check.
5112 // This might fold up statically, or then again it might not.
5113 //
5114 // Non-static example: Copying List<String>.elements to a new String[].
5115 // The backing store for a List<String> is always an Object[],
5116 // but its elements are always type String, if the generic types
5117 // are correct at the source level.
5118 //
5119 // Test S[] against D[], not S against D, because (probably)
5120 // the secondary supertype cache is less busy for S[] than S.
5121 // This usually only matters when D is an interface.
5122 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5123 // Plug failing path into checked_oop_disjoint_arraycopy
5124 if (not_subtype_ctrl != top()) {
5125 PreserveJVMState pjvms(this);
5126 set_control(not_subtype_ctrl);
5127 // (At this point we can assume disjoint_bases, since types differ.)
5128 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5129 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5130 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5131 Node* dest_elem_klass = _gvn.transform(n1);
5132 Node* cv = generate_checkcast_arraycopy(adr_type,
5133 dest_elem_klass,
5134 src, src_offset, dest, dest_offset,
5135 ConvI2X(copy_length), dest_uninitialized);
5136 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5137 checked_control = control();
5138 checked_i_o = i_o();
5139 checked_mem = memory(adr_type);
5140 checked_value = cv;
5141 }
5142 // At this point we know we do not need type checks on oop stores.
5143
5144 // Let's see if we need card marks:
5145 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5146 // If we do not need card marks, copy using the jint or jlong stub.
5147 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5148 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5149 "sizes agree");
5150 }
5151 }
5152
5153 if (!stopped()) {
5154 // Generate the fast path, if possible.
5155 PreserveJVMState pjvms(this);
5156 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5157 src, src_offset, dest, dest_offset,
5158 ConvI2X(copy_length), dest_uninitialized);
5159
5160 // Present the results of the fast call.
5161 result_region->init_req(fast_path, control());
5162 result_i_o ->init_req(fast_path, i_o());
5163 result_memory->init_req(fast_path, memory(adr_type));
5164 }
5165
5166 // Here are all the slow paths up to this point, in one bundle:
5167 slow_control = top();
5168 if (slow_region != NULL)
5169 slow_control = _gvn.transform(slow_region);
5170 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5171
5172 set_control(checked_control);
5173 if (!stopped()) {
5174 // Clean up after the checked call.
5175 // The returned value is either 0 or -1^K,
5176 // where K = number of partially transferred array elements.
5177 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5178 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5179 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5180
5181 // If it is 0, we are done, so transfer to the end.
5182 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5183 result_region->init_req(checked_path, checks_done);
5184 result_i_o ->init_req(checked_path, checked_i_o);
5185 result_memory->init_req(checked_path, checked_mem);
5186
5187 // If it is not zero, merge into the slow call.
5188 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5189 RegionNode* slow_reg2 = new(C) RegionNode(3);
5190 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5191 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5192 record_for_igvn(slow_reg2);
5193 slow_reg2 ->init_req(1, slow_control);
5194 slow_i_o2 ->init_req(1, slow_i_o);
5195 slow_mem2 ->init_req(1, slow_mem);
5196 slow_reg2 ->init_req(2, control());
5197 slow_i_o2 ->init_req(2, checked_i_o);
5198 slow_mem2 ->init_req(2, checked_mem);
5199
5200 slow_control = _gvn.transform(slow_reg2);
5201 slow_i_o = _gvn.transform(slow_i_o2);
5202 slow_mem = _gvn.transform(slow_mem2);
5203
5204 if (alloc != NULL) {
5205 // We'll restart from the very beginning, after zeroing the whole thing.
5206 // This can cause double writes, but that's OK since dest is brand new.
5207 // So we ignore the low 31 bits of the value returned from the stub.
5208 } else {
5209 // We must continue the copy exactly where it failed, or else
5210 // another thread might see the wrong number of writes to dest.
5211 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5212 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5213 slow_offset->init_req(1, intcon(0));
5214 slow_offset->init_req(2, checked_offset);
5215 slow_offset = _gvn.transform(slow_offset);
5216
5217 // Adjust the arguments by the conditionally incoming offset.
5218 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5219 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5220 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5221
5222 // Tweak the node variables to adjust the code produced below:
5223 src_offset = src_off_plus;
5224 dest_offset = dest_off_plus;
5225 copy_length = length_minus;
5226 }
5227 }
5228
5229 set_control(slow_control);
5230 if (!stopped()) {
5231 // Generate the slow path, if needed.
5232 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5233
5234 set_memory(slow_mem, adr_type);
5235 set_i_o(slow_i_o);
5236
5237 if (dest_uninitialized) {
5238 generate_clear_array(adr_type, dest, basic_elem_type,
5239 intcon(0), NULL,
5240 alloc->in(AllocateNode::AllocSize));
5241 }
5242
5243 generate_slow_arraycopy(adr_type,
5244 src, src_offset, dest, dest_offset,
5245 copy_length, /*dest_uninitialized*/false);
5246
5247 result_region->init_req(slow_call_path, control());
5248 result_i_o ->init_req(slow_call_path, i_o());
5249 result_memory->init_req(slow_call_path, memory(adr_type));
5250 }
5251
5252 // Remove unused edges.
5253 for (uint i = 1; i < result_region->req(); i++) {
5254 if (result_region->in(i) == NULL)
5255 result_region->init_req(i, top());
5256 }
5257
5258 // Finished; return the combined state.
5259 set_control( _gvn.transform(result_region));
5260 set_i_o( _gvn.transform(result_i_o) );
5261 set_memory( _gvn.transform(result_memory), adr_type );
5262
5263 // The memory edges above are precise in order to model effects around
5264 // array copies accurately to allow value numbering of field loads around
5265 // arraycopy. Such field loads, both before and after, are common in Java
5266 // collections and similar classes involving header/array data structures.
5267 //
5268 // But with low number of register or when some registers are used or killed
5269 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5270 // The next memory barrier is added to avoid it. If the arraycopy can be
5271 // optimized away (which it can, sometimes) then we can manually remove
5272 // the membar also.
5273 //
5274 // Do not let reads from the cloned object float above the arraycopy.
5275 if (alloc != NULL) {
5276 // Do not let stores that initialize this object be reordered with
5277 // a subsequent store that would make this object accessible by
5278 // other threads.
5279 // Record what AllocateNode this StoreStore protects so that
5280 // escape analysis can go from the MemBarStoreStoreNode to the
5281 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5282 // based on the escape status of the AllocateNode.
5283 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5284 } else if (InsertMemBarAfterArraycopy)
5285 insert_mem_bar(Op_MemBarCPUOrder);
5286 }
5287
5288
5289 // Helper function which determines if an arraycopy immediately follows
5290 // an allocation, with no intervening tests or other escapes for the object.
5291 AllocateArrayNode*
5292 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5293 RegionNode* slow_region) {
5294 if (stopped()) return NULL; // no fast path
5295 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5296
5297 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5298 if (alloc == NULL) return NULL;
5299
5300 Node* rawmem = memory(Compile::AliasIdxRaw);
5301 // Is the allocation's memory state untouched?
5302 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5303 // Bail out if there have been raw-memory effects since the allocation.
5304 // (Example: There might have been a call or safepoint.)
5305 return NULL;
5306 }
5307 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5308 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5309 return NULL;
5310 }
5311
5312 // There must be no unexpected observers of this allocation.
5313 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5314 Node* obs = ptr->fast_out(i);
5315 if (obs != this->map()) {
5316 return NULL;
5317 }
5318 }
5319
5320 // This arraycopy must unconditionally follow the allocation of the ptr.
5321 Node* alloc_ctl = ptr->in(0);
5322 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5323
5324 Node* ctl = control();
5325 while (ctl != alloc_ctl) {
5326 // There may be guards which feed into the slow_region.
5327 // Any other control flow means that we might not get a chance
5328 // to finish initializing the allocated object.
5329 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5330 IfNode* iff = ctl->in(0)->as_If();
5331 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5332 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5333 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5334 ctl = iff->in(0); // This test feeds the known slow_region.
5335 continue;
5336 }
5337 // One more try: Various low-level checks bottom out in
5338 // uncommon traps. If the debug-info of the trap omits
5339 // any reference to the allocation, as we've already
5340 // observed, then there can be no objection to the trap.
5341 bool found_trap = false;
5342 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5343 Node* obs = not_ctl->fast_out(j);
5344 if (obs->in(0) == not_ctl && obs->is_Call() &&
5345 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5346 found_trap = true; break;
5347 }
5348 }
5349 if (found_trap) {
5350 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5351 continue;
5352 }
5353 }
5354 return NULL;
5355 }
5356
5357 // If we get this far, we have an allocation which immediately
5358 // precedes the arraycopy, and we can take over zeroing the new object.
5359 // The arraycopy will finish the initialization, and provide
5360 // a new control state to which we will anchor the destination pointer.
5361
5362 return alloc;
5363 }
5364
5365 // Helper for initialization of arrays, creating a ClearArray.
5366 // It writes zero bits in [start..end), within the body of an array object.
5367 // The memory effects are all chained onto the 'adr_type' alias category.
5368 //
5369 // Since the object is otherwise uninitialized, we are free
5370 // to put a little "slop" around the edges of the cleared area,
5371 // as long as it does not go back into the array's header,
5372 // or beyond the array end within the heap.
5373 //
5374 // The lower edge can be rounded down to the nearest jint and the
5375 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5376 //
5377 // Arguments:
5378 // adr_type memory slice where writes are generated
5379 // dest oop of the destination array
5380 // basic_elem_type element type of the destination
5381 // slice_idx array index of first element to store
5382 // slice_len number of elements to store (or NULL)
5383 // dest_size total size in bytes of the array object
5384 //
5385 // Exactly one of slice_len or dest_size must be non-NULL.
5386 // If dest_size is non-NULL, zeroing extends to the end of the object.
5387 // If slice_len is non-NULL, the slice_idx value must be a constant.
5388 void
5389 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5390 Node* dest,
5391 BasicType basic_elem_type,
5392 Node* slice_idx,
5393 Node* slice_len,
5394 Node* dest_size) {
5395 // one or the other but not both of slice_len and dest_size:
5396 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5397 if (slice_len == NULL) slice_len = top();
5398 if (dest_size == NULL) dest_size = top();
5399
5400 // operate on this memory slice:
5401 Node* mem = memory(adr_type); // memory slice to operate on
5402
5403 // scaling and rounding of indexes:
5404 int scale = exact_log2(type2aelembytes(basic_elem_type));
5405 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5406 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5407 int bump_bit = (-1 << scale) & BytesPerInt;
5408
5409 // determine constant starts and ends
5410 const intptr_t BIG_NEG = -128;
5411 assert(BIG_NEG + 2*abase < 0, "neg enough");
5412 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5413 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5414 if (slice_len_con == 0) {
5415 return; // nothing to do here
5416 }
5417 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5418 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5419 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5420 assert(end_con < 0, "not two cons");
5421 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5422 BytesPerLong);
5423 }
5424
5425 if (start_con >= 0 && end_con >= 0) {
5426 // Constant start and end. Simple.
5427 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5428 start_con, end_con, &_gvn);
5429 } else if (start_con >= 0 && dest_size != top()) {
5430 // Constant start, pre-rounded end after the tail of the array.
5431 Node* end = dest_size;
5432 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5433 start_con, end, &_gvn);
5434 } else if (start_con >= 0 && slice_len != top()) {
5435 // Constant start, non-constant end. End needs rounding up.
5436 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5437 intptr_t end_base = abase + (slice_idx_con << scale);
5438 int end_round = (-1 << scale) & (BytesPerLong - 1);
5439 Node* end = ConvI2X(slice_len);
5440 if (scale != 0)
5441 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5442 end_base += end_round;
5443 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5444 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5445 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5446 start_con, end, &_gvn);
5447 } else if (start_con < 0 && dest_size != top()) {
5448 // Non-constant start, pre-rounded end after the tail of the array.
5449 // This is almost certainly a "round-to-end" operation.
5450 Node* start = slice_idx;
5451 start = ConvI2X(start);
5452 if (scale != 0)
5453 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5454 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5455 if ((bump_bit | clear_low) != 0) {
5456 int to_clear = (bump_bit | clear_low);
5457 // Align up mod 8, then store a jint zero unconditionally
5458 // just before the mod-8 boundary.
5459 if (((abase + bump_bit) & ~to_clear) - bump_bit
5460 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5461 bump_bit = 0;
5462 assert((abase & to_clear) == 0, "array base must be long-aligned");
5463 } else {
5464 // Bump 'start' up to (or past) the next jint boundary:
5465 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5466 assert((abase & clear_low) == 0, "array base must be int-aligned");
5467 }
5468 // Round bumped 'start' down to jlong boundary in body of array.
5469 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5470 if (bump_bit != 0) {
5471 // Store a zero to the immediately preceding jint:
5472 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5473 Node* p1 = basic_plus_adr(dest, x1);
5474 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5475 mem = _gvn.transform(mem);
5476 }
5477 }
5478 Node* end = dest_size; // pre-rounded
5479 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5480 start, end, &_gvn);
5481 } else {
5482 // Non-constant start, unrounded non-constant end.
5483 // (Nobody zeroes a random midsection of an array using this routine.)
5484 ShouldNotReachHere(); // fix caller
5485 }
5486
5487 // Done.
5488 set_memory(mem, adr_type);
5489 }
5490
5491
5492 bool
5493 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5494 BasicType basic_elem_type,
5495 AllocateNode* alloc,
5496 Node* src, Node* src_offset,
5497 Node* dest, Node* dest_offset,
5498 Node* dest_size, bool dest_uninitialized) {
5499 // See if there is an advantage from block transfer.
5500 int scale = exact_log2(type2aelembytes(basic_elem_type));
5501 if (scale >= LogBytesPerLong)
5502 return false; // it is already a block transfer
5503
5504 // Look at the alignment of the starting offsets.
5505 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5506
5507 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5508 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5509 if (src_off_con < 0 || dest_off_con < 0)
5510 // At present, we can only understand constants.
5511 return false;
5512
5513 intptr_t src_off = abase + (src_off_con << scale);
5514 intptr_t dest_off = abase + (dest_off_con << scale);
5515
5516 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5517 // Non-aligned; too bad.
5518 // One more chance: Pick off an initial 32-bit word.
5519 // This is a common case, since abase can be odd mod 8.
5520 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5521 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5522 Node* sptr = basic_plus_adr(src, src_off);
5523 Node* dptr = basic_plus_adr(dest, dest_off);
5524 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5525 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5526 src_off += BytesPerInt;
5527 dest_off += BytesPerInt;
5528 } else {
5529 return false;
5530 }
5531 }
5532 assert(src_off % BytesPerLong == 0, "");
5533 assert(dest_off % BytesPerLong == 0, "");
5534
5535 // Do this copy by giant steps.
5536 Node* sptr = basic_plus_adr(src, src_off);
5537 Node* dptr = basic_plus_adr(dest, dest_off);
5538 Node* countx = dest_size;
5539 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5540 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5541
5542 bool disjoint_bases = true; // since alloc != NULL
5543 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5544 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5545
5546 return true;
5547 }
5548
5549
5550 // Helper function; generates code for the slow case.
5551 // We make a call to a runtime method which emulates the native method,
5552 // but without the native wrapper overhead.
5553 void
5554 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5555 Node* src, Node* src_offset,
5556 Node* dest, Node* dest_offset,
5557 Node* copy_length, bool dest_uninitialized) {
5558 assert(!dest_uninitialized, "Invariant");
5559 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5560 OptoRuntime::slow_arraycopy_Type(),
5561 OptoRuntime::slow_arraycopy_Java(),
5562 "slow_arraycopy", adr_type,
5563 src, src_offset, dest, dest_offset,
5564 copy_length);
5565
5566 // Handle exceptions thrown by this fellow:
5567 make_slow_call_ex(call, env()->Throwable_klass(), false);
5568 }
5569
5570 // Helper function; generates code for cases requiring runtime checks.
5571 Node*
5572 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5573 Node* dest_elem_klass,
5574 Node* src, Node* src_offset,
5575 Node* dest, Node* dest_offset,
5576 Node* copy_length, bool dest_uninitialized) {
5577 if (stopped()) return NULL;
5578
5579 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5580 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5581 return NULL;
5582 }
5583
5584 // Pick out the parameters required to perform a store-check
5585 // for the target array. This is an optimistic check. It will
5586 // look in each non-null element's class, at the desired klass's
5587 // super_check_offset, for the desired klass.
5588 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5589 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5590 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5591 Node* check_offset = ConvI2X(_gvn.transform(n3));
5592 Node* check_value = dest_elem_klass;
5593
5594 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5595 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5596
5597 // (We know the arrays are never conjoint, because their types differ.)
5598 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5599 OptoRuntime::checkcast_arraycopy_Type(),
5600 copyfunc_addr, "checkcast_arraycopy", adr_type,
5601 // five arguments, of which two are
5602 // intptr_t (jlong in LP64)
5603 src_start, dest_start,
5604 copy_length XTOP,
5605 check_offset XTOP,
5606 check_value);
5607
5608 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5609 }
5610
5611
5612 // Helper function; generates code for cases requiring runtime checks.
5613 Node*
5614 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5615 Node* src, Node* src_offset,
5616 Node* dest, Node* dest_offset,
5617 Node* copy_length, bool dest_uninitialized) {
5618 assert(!dest_uninitialized, "Invariant");
5619 if (stopped()) return NULL;
5620 address copyfunc_addr = StubRoutines::generic_arraycopy();
5621 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5622 return NULL;
5623 }
5624
5625 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5626 OptoRuntime::generic_arraycopy_Type(),
5627 copyfunc_addr, "generic_arraycopy", adr_type,
5628 src, src_offset, dest, dest_offset, copy_length);
5629
5630 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5631 }
5632
5633 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5634 void
5635 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5636 BasicType basic_elem_type,
5637 bool disjoint_bases,
5638 Node* src, Node* src_offset,
5639 Node* dest, Node* dest_offset,
5640 Node* copy_length, bool dest_uninitialized) {
5641 if (stopped()) return; // nothing to do
5642
5643 Node* src_start = src;
5644 Node* dest_start = dest;
5645 if (src_offset != NULL || dest_offset != NULL) {
5646 assert(src_offset != NULL && dest_offset != NULL, "");
5647 src_start = array_element_address(src, src_offset, basic_elem_type);
5648 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5649 }
5650
5651 // Figure out which arraycopy runtime method to call.
5652 const char* copyfunc_name = "arraycopy";
5653 address copyfunc_addr =
5654 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5655 disjoint_bases, copyfunc_name, dest_uninitialized);
5656
5657 // Call it. Note that the count_ix value is not scaled to a byte-size.
5658 make_runtime_call(RC_LEAF|RC_NO_FP,
5659 OptoRuntime::fast_arraycopy_Type(),
5660 copyfunc_addr, copyfunc_name, adr_type,
5661 src_start, dest_start, copy_length XTOP);
5662 }
5663
5664 //-------------inline_encodeISOArray-----------------------------------
5665 // encode char[] to byte[] in ISO_8859_1
5666 bool LibraryCallKit::inline_encodeISOArray() {
5667 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5668 // no receiver since it is static method
5669 Node *src = argument(0);
5670 Node *src_offset = argument(1);
5671 Node *dst = argument(2);
5672 Node *dst_offset = argument(3);
5673 Node *length = argument(4);
5674
5675 const Type* src_type = src->Value(&_gvn);
5676 const Type* dst_type = dst->Value(&_gvn);
5677 const TypeAryPtr* top_src = src_type->isa_aryptr();
5678 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5679 if (top_src == NULL || top_src->klass() == NULL ||
5680 top_dest == NULL || top_dest->klass() == NULL) {
5681 // failed array check
5682 return false;
5683 }
5684
5685 // Figure out the size and type of the elements we will be copying.
5686 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5687 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5688 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5689 return false;
5690 }
5691 Node* src_start = array_element_address(src, src_offset, src_elem);
5692 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5693 // 'src_start' points to src array + scaled offset
5694 // 'dst_start' points to dst array + scaled offset
5695
5696 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5697 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5698 enc = _gvn.transform(enc);
5699 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5700 set_memory(res_mem, mtype);
5701 set_result(enc);
5702 return true;
5703 }
5704
5705 /**
5706 * Calculate CRC32 for byte.
5707 * int java.util.zip.CRC32.update(int crc, int b)
5708 */
5709 bool LibraryCallKit::inline_updateCRC32() {
5710 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5711 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5712 // no receiver since it is static method
5713 Node* crc = argument(0); // type: int
5714 Node* b = argument(1); // type: int
5715
5716 /*
5717 * int c = ~ crc;
5718 * b = timesXtoThe32[(b ^ c) & 0xFF];
5719 * b = b ^ (c >>> 8);
5720 * crc = ~b;
5721 */
5722
5723 Node* M1 = intcon(-1);
5724 crc = _gvn.transform(new (C) XorINode(crc, M1));
5725 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5726 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5727
5728 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5729 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5730 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5731 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5732
5733 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5734 result = _gvn.transform(new (C) XorINode(crc, result));
5735 result = _gvn.transform(new (C) XorINode(result, M1));
5736 set_result(result);
5737 return true;
5738 }
5739
5740 /**
5741 * Calculate CRC32 for byte[] array.
5742 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5743 */
5744 bool LibraryCallKit::inline_updateBytesCRC32() {
5745 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5746 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5747 // no receiver since it is static method
5748 Node* crc = argument(0); // type: int
5749 Node* src = argument(1); // type: oop
5750 Node* offset = argument(2); // type: int
5751 Node* length = argument(3); // type: int
5752
5753 const Type* src_type = src->Value(&_gvn);
5754 const TypeAryPtr* top_src = src_type->isa_aryptr();
5755 if (top_src == NULL || top_src->klass() == NULL) {
5756 // failed array check
5757 return false;
5758 }
5759
5760 // Figure out the size and type of the elements we will be copying.
5761 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5762 if (src_elem != T_BYTE) {
5763 return false;
5764 }
5765
5766 // 'src_start' points to src array + scaled offset
5767 Node* src_start = array_element_address(src, offset, src_elem);
5768
5769 // We assume that range check is done by caller.
5770 // TODO: generate range check (offset+length < src.length) in debug VM.
5771
5772 // Call the stub.
5773 address stubAddr = StubRoutines::updateBytesCRC32();
5774 const char *stubName = "updateBytesCRC32";
5775
5776 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5777 stubAddr, stubName, TypePtr::BOTTOM,
5778 crc, src_start, length);
5779 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5780 set_result(result);
5781 return true;
5782 }
5783
5784 /**
5785 * Calculate CRC32 for ByteBuffer.
5786 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5787 */
5788 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5789 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5790 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5791 // no receiver since it is static method
5792 Node* crc = argument(0); // type: int
5793 Node* src = argument(1); // type: long
5794 Node* offset = argument(3); // type: int
5795 Node* length = argument(4); // type: int
5796
5797 src = ConvL2X(src); // adjust Java long to machine word
5798 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5799 offset = ConvI2X(offset);
5800
5801 // 'src_start' points to src array + scaled offset
5802 Node* src_start = basic_plus_adr(top(), base, offset);
5803
5804 // Call the stub.
5805 address stubAddr = StubRoutines::updateBytesCRC32();
5806 const char *stubName = "updateBytesCRC32";
5807
5808 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5809 stubAddr, stubName, TypePtr::BOTTOM,
5810 crc, src_start, length);
5811 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5812 set_result(result);
5813 return true;
5814 }
5815
5816 //----------------------------inline_reference_get----------------------------
5817 // public T java.lang.ref.Reference.get();
5818 bool LibraryCallKit::inline_reference_get() {
5819 const int referent_offset = java_lang_ref_Reference::referent_offset;
5820 guarantee(referent_offset > 0, "should have already been set");
5821
5822 // Get the argument:
5823 Node* reference_obj = null_check_receiver();
5824 if (stopped()) return true;
5825
5826 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5827
5828 ciInstanceKlass* klass = env()->Object_klass();
5829 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5830
5831 Node* no_ctrl = NULL;
5832 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5833
5834 // Use the pre-barrier to record the value in the referent field
5835 pre_barrier(false /* do_load */,
5836 control(),
5837 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5838 result /* pre_val */,
5839 T_OBJECT);
5840
5841 // Add memory barrier to prevent commoning reads from this field
5842 // across safepoint since GC can change its value.
5843 insert_mem_bar(Op_MemBarCPUOrder);
5844
5845 set_result(result);
5846 return true;
5847 }
5848
5849
5850 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5851 bool is_exact=true, bool is_static=false) {
5852
5853 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5854 assert(tinst != NULL, "obj is null");
5855 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5856 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5857
5858 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5859 ciSymbol::make(fieldTypeString),
5860 is_static);
5861 if (field == NULL) return (Node *) NULL;
5862 assert (field != NULL, "undefined field");
5863
5864 // Next code copied from Parse::do_get_xxx():
5865
5866 // Compute address and memory type.
5867 int offset = field->offset_in_bytes();
5868 bool is_vol = field->is_volatile();
5869 ciType* field_klass = field->type();
5870 assert(field_klass->is_loaded(), "should be loaded");
5871 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5872 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5873 BasicType bt = field->layout_type();
5874
5875 // Build the resultant type of the load
5876 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5877
5878 // Build the load.
5879 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5880 return loadedField;
5881 }
5882
5883
5884 //------------------------------inline_aescrypt_Block-----------------------
5885 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5886 address stubAddr;
5887 const char *stubName;
5888 assert(UseAES, "need AES instruction support");
5889
5890 switch(id) {
5891 case vmIntrinsics::_aescrypt_encryptBlock:
5892 stubAddr = StubRoutines::aescrypt_encryptBlock();
5893 stubName = "aescrypt_encryptBlock";
5894 break;
5895 case vmIntrinsics::_aescrypt_decryptBlock:
5896 stubAddr = StubRoutines::aescrypt_decryptBlock();
5897 stubName = "aescrypt_decryptBlock";
5898 break;
5899 }
5900 if (stubAddr == NULL) return false;
5901
5902 Node* aescrypt_object = argument(0);
5903 Node* src = argument(1);
5904 Node* src_offset = argument(2);
5905 Node* dest = argument(3);
5906 Node* dest_offset = argument(4);
5907
5908 // (1) src and dest are arrays.
5909 const Type* src_type = src->Value(&_gvn);
5910 const Type* dest_type = dest->Value(&_gvn);
5911 const TypeAryPtr* top_src = src_type->isa_aryptr();
5912 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5913 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5914
5915 // for the quick and dirty code we will skip all the checks.
5916 // we are just trying to get the call to be generated.
5917 Node* src_start = src;
5918 Node* dest_start = dest;
5919 if (src_offset != NULL || dest_offset != NULL) {
5920 assert(src_offset != NULL && dest_offset != NULL, "");
5921 src_start = array_element_address(src, src_offset, T_BYTE);
5922 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5923 }
5924
5925 // now need to get the start of its expanded key array
5926 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5927 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5928 if (k_start == NULL) return false;
5929
5930 if (Matcher::pass_original_key_for_aes()) {
5931 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5932 // compatibility issues between Java key expansion and SPARC crypto instructions
5933 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5934 if (original_k_start == NULL) return false;
5935
5936 // Call the stub.
5937 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5938 stubAddr, stubName, TypePtr::BOTTOM,
5939 src_start, dest_start, k_start, original_k_start);
5940 } else {
5941 // Call the stub.
5942 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5943 stubAddr, stubName, TypePtr::BOTTOM,
5944 src_start, dest_start, k_start);
5945 }
5946
5947 return true;
5948 }
5949
5950 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5951 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5952 address stubAddr;
5953 const char *stubName;
5954
5955 assert(UseAES, "need AES instruction support");
5956
5957 switch(id) {
5958 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5959 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5960 stubName = "cipherBlockChaining_encryptAESCrypt";
5961 break;
5962 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5963 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5964 stubName = "cipherBlockChaining_decryptAESCrypt";
5965 break;
5966 }
5967 if (stubAddr == NULL) return false;
5968
5969 Node* cipherBlockChaining_object = argument(0);
5970 Node* src = argument(1);
5971 Node* src_offset = argument(2);
5972 Node* len = argument(3);
5973 Node* dest = argument(4);
5974 Node* dest_offset = argument(5);
5975
5976 // (1) src and dest are arrays.
5977 const Type* src_type = src->Value(&_gvn);
5978 const Type* dest_type = dest->Value(&_gvn);
5979 const TypeAryPtr* top_src = src_type->isa_aryptr();
5980 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5981 assert (top_src != NULL && top_src->klass() != NULL
5982 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5983
5984 // checks are the responsibility of the caller
5985 Node* src_start = src;
5986 Node* dest_start = dest;
5987 if (src_offset != NULL || dest_offset != NULL) {
5988 assert(src_offset != NULL && dest_offset != NULL, "");
5989 src_start = array_element_address(src, src_offset, T_BYTE);
5990 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5991 }
5992
5993 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5994 // (because of the predicated logic executed earlier).
5995 // so we cast it here safely.
5996 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5997
5998 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5999 if (embeddedCipherObj == NULL) return false;
6000
6001 // cast it to what we know it will be at runtime
6002 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6003 assert(tinst != NULL, "CBC obj is null");
6004 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6005 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6006 if (!klass_AESCrypt->is_loaded()) return false;
6007
6008 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6009 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6010 const TypeOopPtr* xtype = aklass->as_instance_type();
6011 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6012 aescrypt_object = _gvn.transform(aescrypt_object);
6013
6014 // we need to get the start of the aescrypt_object's expanded key array
6015 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6016 if (k_start == NULL) return false;
6017
6018 // similarly, get the start address of the r vector
6019 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6020 if (objRvec == NULL) return false;
6021 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6022
6023 Node* cbcCrypt;
6024 if (Matcher::pass_original_key_for_aes()) {
6025 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6026 // compatibility issues between Java key expansion and SPARC crypto instructions
6027 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6028 if (original_k_start == NULL) return false;
6029
6030 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6031 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6032 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6033 stubAddr, stubName, TypePtr::BOTTOM,
6034 src_start, dest_start, k_start, r_start, len, original_k_start);
6035 } else {
6036 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6037 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6038 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6039 stubAddr, stubName, TypePtr::BOTTOM,
6040 src_start, dest_start, k_start, r_start, len);
6041 }
6042
6043 // return cipher length (int)
6044 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6045 set_result(retvalue);
6046 return true;
6047 }
6048
6049 //------------------------------get_key_start_from_aescrypt_object-----------------------
6050 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6051 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6052 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6053 if (objAESCryptKey == NULL) return (Node *) NULL;
6054
6055 // now have the array, need to get the start address of the K array
6056 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6057 return k_start;
6058 }
6059
6060 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6061 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6062 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6063 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6064 if (objAESCryptKey == NULL) return (Node *) NULL;
6065
6066 // now have the array, need to get the start address of the lastKey array
6067 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6068 return original_k_start;
6069 }
6070
6071 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6072 // Return node representing slow path of predicate check.
6073 // the pseudo code we want to emulate with this predicate is:
6074 // for encryption:
6075 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6076 // for decryption:
6077 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6078 // note cipher==plain is more conservative than the original java code but that's OK
6079 //
6080 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6081 // First, check receiver for NULL since it is virtual method.
6082 Node* objCBC = argument(0);
6083 objCBC = null_check(objCBC);
6084
6085 if (stopped()) return NULL; // Always NULL
6086
6087 // Load embeddedCipher field of CipherBlockChaining object.
6088 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6089
6090 // get AESCrypt klass for instanceOf check
6091 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6092 // will have same classloader as CipherBlockChaining object
6093 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6094 assert(tinst != NULL, "CBCobj is null");
6095 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6096
6097 // we want to do an instanceof comparison against the AESCrypt class
6098 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6099 if (!klass_AESCrypt->is_loaded()) {
6100 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6101 Node* ctrl = control();
6102 set_control(top()); // no regular fast path
6103 return ctrl;
6104 }
6105 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6106
6107 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6108 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6109 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6110
6111 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6112
6113 // for encryption, we are done
6114 if (!decrypting)
6115 return instof_false; // even if it is NULL
6116
6117 // for decryption, we need to add a further check to avoid
6118 // taking the intrinsic path when cipher and plain are the same
6119 // see the original java code for why.
6120 RegionNode* region = new(C) RegionNode(3);
6121 region->init_req(1, instof_false);
6122 Node* src = argument(1);
6123 Node* dest = argument(4);
6124 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6125 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6126 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6127 region->init_req(2, src_dest_conjoint);
6128
6129 record_for_igvn(region);
6130 return _gvn.transform(region);
6131 }