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 return result_val;
1737 } else {
1738 return result;
1739 }
1740 }
1741 }
1742
1743 //------------------------------inline_exp-------------------------------------
1744 // Inline exp instructions, if possible. The Intel hardware only misses
1745 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1746 bool LibraryCallKit::inline_exp() {
1747 Node* arg = round_double_node(argument(0));
1748 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1749
1750 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1751 set_result(n);
1752
1753 C->set_has_split_ifs(true); // Has chance for split-if optimization
1754 return true;
1755 }
1756
1757 //------------------------------inline_pow-------------------------------------
1758 // Inline power instructions, if possible.
1759 bool LibraryCallKit::inline_pow() {
1760 // Pseudocode for pow
1761 // if (y == 2) {
1762 // return x * x;
1763 // } else {
1764 // if (x <= 0.0) {
1765 // long longy = (long)y;
1766 // if ((double)longy == y) { // if y is long
1767 // if (y + 1 == y) longy = 0; // huge number: even
1768 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1769 // } else {
1770 // result = NaN;
1771 // }
1772 // } else {
1773 // result = DPow(x,y);
1774 // }
1775 // if (result != result)? {
1776 // result = uncommon_trap() or runtime_call();
1777 // }
1778 // return result;
1779 // }
1780
1781 Node* x = round_double_node(argument(0));
1782 Node* y = round_double_node(argument(2));
1783
1784 Node* result = NULL;
1785
1786 Node* const_two_node = makecon(TypeD::make(2.0));
1787 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1788 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1789 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1790 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1791 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1792
1793 RegionNode* region_node = new (C) RegionNode(3);
1794 region_node->init_req(1, if_true);
1795
1796 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1797 // special case for x^y where y == 2, we can convert it to x * x
1798 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1799
1800 // set control to if_false since we will now process the false branch
1801 set_control(if_false);
1802
1803 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1804 // Short form: skip the fancy tests and just check for NaN result.
1805 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1806 } else {
1807 // If this inlining ever returned NaN in the past, include all
1808 // checks + call to the runtime.
1809
1810 // Set the merge point for If node with condition of (x <= 0.0)
1811 // There are four possible paths to region node and phi node
1812 RegionNode *r = new (C) RegionNode(4);
1813 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1814
1815 // Build the first if node: if (x <= 0.0)
1816 // Node for 0 constant
1817 Node *zeronode = makecon(TypeD::ZERO);
1818 // Check x:0
1819 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1820 // Check: If (x<=0) then go complex path
1821 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1822 // Branch either way
1823 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1824 // Fast path taken; set region slot 3
1825 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1826 r->init_req(3,fast_taken); // Capture fast-control
1827
1828 // Fast path not-taken, i.e. slow path
1829 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1830
1831 // Set fast path result
1832 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1833 phi->init_req(3, fast_result);
1834
1835 // Complex path
1836 // Build the second if node (if y is long)
1837 // Node for (long)y
1838 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1839 // Node for (double)((long) y)
1840 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1841 // Check (double)((long) y) : y
1842 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1843 // Check if (y isn't long) then go to slow path
1844
1845 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1846 // Branch either way
1847 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1848 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1849
1850 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1851
1852 // Calculate DPow(abs(x), y)*(1 & (long)y)
1853 // Node for constant 1
1854 Node *conone = longcon(1);
1855 // 1& (long)y
1856 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1857
1858 // A huge number is always even. Detect a huge number by checking
1859 // if y + 1 == y and set integer to be tested for parity to 0.
1860 // Required for corner case:
1861 // (long)9.223372036854776E18 = max_jlong
1862 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1863 // max_jlong is odd but 9.223372036854776E18 is even
1864 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1865 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1866 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1867 Node* correctedsign = NULL;
1868 if (ConditionalMoveLimit != 0) {
1869 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1870 } else {
1871 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1872 RegionNode *r = new (C) RegionNode(3);
1873 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1874 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1875 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1876 phi->init_req(1, signnode);
1877 phi->init_req(2, longcon(0));
1878 correctedsign = _gvn.transform(phi);
1879 ylong_path = _gvn.transform(r);
1880 record_for_igvn(r);
1881 }
1882
1883 // zero node
1884 Node *conzero = longcon(0);
1885 // Check (1&(long)y)==0?
1886 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1887 // Check if (1&(long)y)!=0?, if so the result is negative
1888 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1889 // abs(x)
1890 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1891 // abs(x)^y
1892 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1893 // -abs(x)^y
1894 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1895 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1896 Node *signresult = NULL;
1897 if (ConditionalMoveLimit != 0) {
1898 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1899 } else {
1900 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1901 RegionNode *r = new (C) RegionNode(3);
1902 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1903 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1904 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1905 phi->init_req(1, absxpowy);
1906 phi->init_req(2, negabsxpowy);
1907 signresult = _gvn.transform(phi);
1908 ylong_path = _gvn.transform(r);
1909 record_for_igvn(r);
1910 }
1911 // Set complex path fast result
1912 r->init_req(2, ylong_path);
1913 phi->init_req(2, signresult);
1914
1915 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1916 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1917 r->init_req(1,slow_path);
1918 phi->init_req(1,slow_result);
1919
1920 // Post merge
1921 set_control(_gvn.transform(r));
1922 record_for_igvn(r);
1923 result = _gvn.transform(phi);
1924 }
1925
1926 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1927
1928 // control from finish_pow_exp is now input to the region node
1929 region_node->set_req(2, control());
1930 // the result from finish_pow_exp is now input to the phi node
1931 phi_node->init_req(2, _gvn.transform(result));
1932 set_control(_gvn.transform(region_node));
1933 record_for_igvn(region_node);
1934 set_result(_gvn.transform(phi_node));
1935
1936 C->set_has_split_ifs(true); // Has chance for split-if optimization
1937 return true;
1938 }
1939
1940 //------------------------------runtime_math-----------------------------
1941 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1942 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1943 "must be (DD)D or (D)D type");
1944
1945 // Inputs
1946 Node* a = round_double_node(argument(0));
1947 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1948
1949 const TypePtr* no_memory_effects = NULL;
1950 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1951 no_memory_effects,
1952 a, top(), b, b ? top() : NULL);
1953 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1954 #ifdef ASSERT
1955 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1956 assert(value_top == top(), "second value must be top");
1957 #endif
1958
1959 set_result(value);
1960 return true;
1961 }
1962
1963 //------------------------------inline_math_native-----------------------------
1964 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1965 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1966 switch (id) {
1967 // These intrinsics are not properly supported on all hardware
1968 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1969 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1970 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1971 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1972 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1973 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1974
1975 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1976 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1977 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1978 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1979
1980 // These intrinsics are supported on all hardware
1981 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1982 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1983
1984 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1985 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1986 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1987 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1988 #undef FN_PTR
1989
1990 // These intrinsics are not yet correctly implemented
1991 case vmIntrinsics::_datan2:
1992 return false;
1993
1994 default:
1995 fatal_unexpected_iid(id);
1996 return false;
1997 }
1998 }
1999
2000 static bool is_simple_name(Node* n) {
2001 return (n->req() == 1 // constant
2002 || (n->is_Type() && n->as_Type()->type()->singleton())
2003 || n->is_Proj() // parameter or return value
2004 || n->is_Phi() // local of some sort
2005 );
2006 }
2007
2008 //----------------------------inline_min_max-----------------------------------
2009 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2010 set_result(generate_min_max(id, argument(0), argument(1)));
2011 return true;
2012 }
2013
2014 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2015 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2016 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2017 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2018 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2019
2020 {
2021 PreserveJVMState pjvms(this);
2022 PreserveReexecuteState preexecs(this);
2023 jvms()->set_should_reexecute(true);
2024
2025 set_control(slow_path);
2026 set_i_o(i_o());
2027
2028 uncommon_trap(Deoptimization::Reason_intrinsic,
2029 Deoptimization::Action_none);
2030 }
2031
2032 set_control(fast_path);
2033 set_result(math);
2034 }
2035
2036 template <typename OverflowOp>
2037 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2038 typedef typename OverflowOp::MathOp MathOp;
2039
2040 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2041 Node* operation = _gvn.transform( mathOp );
2042 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2043 inline_math_mathExact(operation, ofcheck);
2044 return true;
2045 }
2046
2047 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2048 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2049 }
2050
2051 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2052 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2053 }
2054
2055 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2056 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2057 }
2058
2059 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2060 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2061 }
2062
2063 bool LibraryCallKit::inline_math_negateExactI() {
2064 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2065 }
2066
2067 bool LibraryCallKit::inline_math_negateExactL() {
2068 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2069 }
2070
2071 bool LibraryCallKit::inline_math_multiplyExactI() {
2072 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2073 }
2074
2075 bool LibraryCallKit::inline_math_multiplyExactL() {
2076 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2077 }
2078
2079 Node*
2080 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2081 // These are the candidate return value:
2082 Node* xvalue = x0;
2083 Node* yvalue = y0;
2084
2085 if (xvalue == yvalue) {
2086 return xvalue;
2087 }
2088
2089 bool want_max = (id == vmIntrinsics::_max);
2090
2091 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2092 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2093 if (txvalue == NULL || tyvalue == NULL) return top();
2094 // This is not really necessary, but it is consistent with a
2095 // hypothetical MaxINode::Value method:
2096 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2097
2098 // %%% This folding logic should (ideally) be in a different place.
2099 // Some should be inside IfNode, and there to be a more reliable
2100 // transformation of ?: style patterns into cmoves. We also want
2101 // more powerful optimizations around cmove and min/max.
2102
2103 // Try to find a dominating comparison of these guys.
2104 // It can simplify the index computation for Arrays.copyOf
2105 // and similar uses of System.arraycopy.
2106 // First, compute the normalized version of CmpI(x, y).
2107 int cmp_op = Op_CmpI;
2108 Node* xkey = xvalue;
2109 Node* ykey = yvalue;
2110 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2111 if (ideal_cmpxy->is_Cmp()) {
2112 // E.g., if we have CmpI(length - offset, count),
2113 // it might idealize to CmpI(length, count + offset)
2114 cmp_op = ideal_cmpxy->Opcode();
2115 xkey = ideal_cmpxy->in(1);
2116 ykey = ideal_cmpxy->in(2);
2117 }
2118
2119 // Start by locating any relevant comparisons.
2120 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2121 Node* cmpxy = NULL;
2122 Node* cmpyx = NULL;
2123 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2124 Node* cmp = start_from->fast_out(k);
2125 if (cmp->outcnt() > 0 && // must have prior uses
2126 cmp->in(0) == NULL && // must be context-independent
2127 cmp->Opcode() == cmp_op) { // right kind of compare
2128 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2129 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2130 }
2131 }
2132
2133 const int NCMPS = 2;
2134 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2135 int cmpn;
2136 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2137 if (cmps[cmpn] != NULL) break; // find a result
2138 }
2139 if (cmpn < NCMPS) {
2140 // Look for a dominating test that tells us the min and max.
2141 int depth = 0; // Limit search depth for speed
2142 Node* dom = control();
2143 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2144 if (++depth >= 100) break;
2145 Node* ifproj = dom;
2146 if (!ifproj->is_Proj()) continue;
2147 Node* iff = ifproj->in(0);
2148 if (!iff->is_If()) continue;
2149 Node* bol = iff->in(1);
2150 if (!bol->is_Bool()) continue;
2151 Node* cmp = bol->in(1);
2152 if (cmp == NULL) continue;
2153 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2154 if (cmps[cmpn] == cmp) break;
2155 if (cmpn == NCMPS) continue;
2156 BoolTest::mask btest = bol->as_Bool()->_test._test;
2157 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2158 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2159 // At this point, we know that 'x btest y' is true.
2160 switch (btest) {
2161 case BoolTest::eq:
2162 // They are proven equal, so we can collapse the min/max.
2163 // Either value is the answer. Choose the simpler.
2164 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2165 return yvalue;
2166 return xvalue;
2167 case BoolTest::lt: // x < y
2168 case BoolTest::le: // x <= y
2169 return (want_max ? yvalue : xvalue);
2170 case BoolTest::gt: // x > y
2171 case BoolTest::ge: // x >= y
2172 return (want_max ? xvalue : yvalue);
2173 }
2174 }
2175 }
2176
2177 // We failed to find a dominating test.
2178 // Let's pick a test that might GVN with prior tests.
2179 Node* best_bol = NULL;
2180 BoolTest::mask best_btest = BoolTest::illegal;
2181 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2182 Node* cmp = cmps[cmpn];
2183 if (cmp == NULL) continue;
2184 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2185 Node* bol = cmp->fast_out(j);
2186 if (!bol->is_Bool()) continue;
2187 BoolTest::mask btest = bol->as_Bool()->_test._test;
2188 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2189 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2190 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2191 best_bol = bol->as_Bool();
2192 best_btest = btest;
2193 }
2194 }
2195 }
2196
2197 Node* answer_if_true = NULL;
2198 Node* answer_if_false = NULL;
2199 switch (best_btest) {
2200 default:
2201 if (cmpxy == NULL)
2202 cmpxy = ideal_cmpxy;
2203 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2204 // and fall through:
2205 case BoolTest::lt: // x < y
2206 case BoolTest::le: // x <= y
2207 answer_if_true = (want_max ? yvalue : xvalue);
2208 answer_if_false = (want_max ? xvalue : yvalue);
2209 break;
2210 case BoolTest::gt: // x > y
2211 case BoolTest::ge: // x >= y
2212 answer_if_true = (want_max ? xvalue : yvalue);
2213 answer_if_false = (want_max ? yvalue : xvalue);
2214 break;
2215 }
2216
2217 jint hi, lo;
2218 if (want_max) {
2219 // We can sharpen the minimum.
2220 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2221 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2222 } else {
2223 // We can sharpen the maximum.
2224 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2225 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2226 }
2227
2228 // Use a flow-free graph structure, to avoid creating excess control edges
2229 // which could hinder other optimizations.
2230 // Since Math.min/max is often used with arraycopy, we want
2231 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2232 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2233 answer_if_false, answer_if_true,
2234 TypeInt::make(lo, hi, widen));
2235
2236 return _gvn.transform(cmov);
2237
2238 /*
2239 // This is not as desirable as it may seem, since Min and Max
2240 // nodes do not have a full set of optimizations.
2241 // And they would interfere, anyway, with 'if' optimizations
2242 // and with CMoveI canonical forms.
2243 switch (id) {
2244 case vmIntrinsics::_min:
2245 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2246 case vmIntrinsics::_max:
2247 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2248 default:
2249 ShouldNotReachHere();
2250 }
2251 */
2252 }
2253
2254 inline int
2255 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2256 const TypePtr* base_type = TypePtr::NULL_PTR;
2257 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2258 if (base_type == NULL) {
2259 // Unknown type.
2260 return Type::AnyPtr;
2261 } else if (base_type == TypePtr::NULL_PTR) {
2262 // Since this is a NULL+long form, we have to switch to a rawptr.
2263 base = _gvn.transform(new (C) CastX2PNode(offset));
2264 offset = MakeConX(0);
2265 return Type::RawPtr;
2266 } else if (base_type->base() == Type::RawPtr) {
2267 return Type::RawPtr;
2268 } else if (base_type->isa_oopptr()) {
2269 // Base is never null => always a heap address.
2270 if (base_type->ptr() == TypePtr::NotNull) {
2271 return Type::OopPtr;
2272 }
2273 // Offset is small => always a heap address.
2274 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2275 if (offset_type != NULL &&
2276 base_type->offset() == 0 && // (should always be?)
2277 offset_type->_lo >= 0 &&
2278 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2279 return Type::OopPtr;
2280 }
2281 // Otherwise, it might either be oop+off or NULL+addr.
2282 return Type::AnyPtr;
2283 } else {
2284 // No information:
2285 return Type::AnyPtr;
2286 }
2287 }
2288
2289 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2290 int kind = classify_unsafe_addr(base, offset);
2291 if (kind == Type::RawPtr) {
2292 return basic_plus_adr(top(), base, offset);
2293 } else {
2294 return basic_plus_adr(base, offset);
2295 }
2296 }
2297
2298 //--------------------------inline_number_methods-----------------------------
2299 // inline int Integer.numberOfLeadingZeros(int)
2300 // inline int Long.numberOfLeadingZeros(long)
2301 //
2302 // inline int Integer.numberOfTrailingZeros(int)
2303 // inline int Long.numberOfTrailingZeros(long)
2304 //
2305 // inline int Integer.bitCount(int)
2306 // inline int Long.bitCount(long)
2307 //
2308 // inline char Character.reverseBytes(char)
2309 // inline short Short.reverseBytes(short)
2310 // inline int Integer.reverseBytes(int)
2311 // inline long Long.reverseBytes(long)
2312 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2313 Node* arg = argument(0);
2314 Node* n;
2315 switch (id) {
2316 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2317 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2318 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2319 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2320 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2321 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2322 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2323 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2324 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2325 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2326 default: fatal_unexpected_iid(id); break;
2327 }
2328 set_result(_gvn.transform(n));
2329 return true;
2330 }
2331
2332 //----------------------------inline_unsafe_access----------------------------
2333
2334 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2335
2336 // Helper that guards and inserts a pre-barrier.
2337 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2338 Node* pre_val, bool need_mem_bar) {
2339 // We could be accessing the referent field of a reference object. If so, when G1
2340 // is enabled, we need to log the value in the referent field in an SATB buffer.
2341 // This routine performs some compile time filters and generates suitable
2342 // runtime filters that guard the pre-barrier code.
2343 // Also add memory barrier for non volatile load from the referent field
2344 // to prevent commoning of loads across safepoint.
2345 if (!UseG1GC && !need_mem_bar)
2346 return;
2347
2348 // Some compile time checks.
2349
2350 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2351 const TypeX* otype = offset->find_intptr_t_type();
2352 if (otype != NULL && otype->is_con() &&
2353 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2354 // Constant offset but not the reference_offset so just return
2355 return;
2356 }
2357
2358 // We only need to generate the runtime guards for instances.
2359 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2360 if (btype != NULL) {
2361 if (btype->isa_aryptr()) {
2362 // Array type so nothing to do
2363 return;
2364 }
2365
2366 const TypeInstPtr* itype = btype->isa_instptr();
2367 if (itype != NULL) {
2368 // Can the klass of base_oop be statically determined to be
2369 // _not_ a sub-class of Reference and _not_ Object?
2370 ciKlass* klass = itype->klass();
2371 if ( klass->is_loaded() &&
2372 !klass->is_subtype_of(env()->Reference_klass()) &&
2373 !env()->Object_klass()->is_subtype_of(klass)) {
2374 return;
2375 }
2376 }
2377 }
2378
2379 // The compile time filters did not reject base_oop/offset so
2380 // we need to generate the following runtime filters
2381 //
2382 // if (offset == java_lang_ref_Reference::_reference_offset) {
2383 // if (instance_of(base, java.lang.ref.Reference)) {
2384 // pre_barrier(_, pre_val, ...);
2385 // }
2386 // }
2387
2388 float likely = PROB_LIKELY( 0.999);
2389 float unlikely = PROB_UNLIKELY(0.999);
2390
2391 IdealKit ideal(this);
2392 #define __ ideal.
2393
2394 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2395
2396 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2397 // Update graphKit memory and control from IdealKit.
2398 sync_kit(ideal);
2399
2400 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2401 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2402
2403 // Update IdealKit memory and control from graphKit.
2404 __ sync_kit(this);
2405
2406 Node* one = __ ConI(1);
2407 // is_instof == 0 if base_oop == NULL
2408 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2409
2410 // Update graphKit from IdeakKit.
2411 sync_kit(ideal);
2412
2413 // Use the pre-barrier to record the value in the referent field
2414 pre_barrier(false /* do_load */,
2415 __ ctrl(),
2416 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2417 pre_val /* pre_val */,
2418 T_OBJECT);
2419 if (need_mem_bar) {
2420 // Add memory barrier to prevent commoning reads from this field
2421 // across safepoint since GC can change its value.
2422 insert_mem_bar(Op_MemBarCPUOrder);
2423 }
2424 // Update IdealKit from graphKit.
2425 __ sync_kit(this);
2426
2427 } __ end_if(); // _ref_type != ref_none
2428 } __ end_if(); // offset == referent_offset
2429
2430 // Final sync IdealKit and GraphKit.
2431 final_sync(ideal);
2432 #undef __
2433 }
2434
2435
2436 // Interpret Unsafe.fieldOffset cookies correctly:
2437 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2438
2439 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2440 // Attempt to infer a sharper value type from the offset and base type.
2441 ciKlass* sharpened_klass = NULL;
2442
2443 // See if it is an instance field, with an object type.
2444 if (alias_type->field() != NULL) {
2445 assert(!is_native_ptr, "native pointer op cannot use a java address");
2446 if (alias_type->field()->type()->is_klass()) {
2447 sharpened_klass = alias_type->field()->type()->as_klass();
2448 }
2449 }
2450
2451 // See if it is a narrow oop array.
2452 if (adr_type->isa_aryptr()) {
2453 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2454 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2455 if (elem_type != NULL) {
2456 sharpened_klass = elem_type->klass();
2457 }
2458 }
2459 }
2460
2461 // The sharpened class might be unloaded if there is no class loader
2462 // contraint in place.
2463 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2464 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2465
2466 #ifndef PRODUCT
2467 if (C->print_intrinsics() || C->print_inlining()) {
2468 tty->print(" from base type: "); adr_type->dump();
2469 tty->print(" sharpened value: "); tjp->dump();
2470 }
2471 #endif
2472 // Sharpen the value type.
2473 return tjp;
2474 }
2475 return NULL;
2476 }
2477
2478 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2479 if (callee()->is_static()) return false; // caller must have the capability!
2480
2481 #ifndef PRODUCT
2482 {
2483 ResourceMark rm;
2484 // Check the signatures.
2485 ciSignature* sig = callee()->signature();
2486 #ifdef ASSERT
2487 if (!is_store) {
2488 // Object getObject(Object base, int/long offset), etc.
2489 BasicType rtype = sig->return_type()->basic_type();
2490 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2491 rtype = T_ADDRESS; // it is really a C void*
2492 assert(rtype == type, "getter must return the expected value");
2493 if (!is_native_ptr) {
2494 assert(sig->count() == 2, "oop getter has 2 arguments");
2495 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2496 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2497 } else {
2498 assert(sig->count() == 1, "native getter has 1 argument");
2499 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2500 }
2501 } else {
2502 // void putObject(Object base, int/long offset, Object x), etc.
2503 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2504 if (!is_native_ptr) {
2505 assert(sig->count() == 3, "oop putter has 3 arguments");
2506 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2507 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2508 } else {
2509 assert(sig->count() == 2, "native putter has 2 arguments");
2510 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2511 }
2512 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2513 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2514 vtype = T_ADDRESS; // it is really a C void*
2515 assert(vtype == type, "putter must accept the expected value");
2516 }
2517 #endif // ASSERT
2518 }
2519 #endif //PRODUCT
2520
2521 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2522
2523 Node* receiver = argument(0); // type: oop
2524
2525 // Build address expression. See the code in inline_unsafe_prefetch.
2526 Node* adr;
2527 Node* heap_base_oop = top();
2528 Node* offset = top();
2529 Node* val;
2530
2531 if (!is_native_ptr) {
2532 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2533 Node* base = argument(1); // type: oop
2534 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2535 offset = argument(2); // type: long
2536 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2537 // to be plain byte offsets, which are also the same as those accepted
2538 // by oopDesc::field_base.
2539 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2540 "fieldOffset must be byte-scaled");
2541 // 32-bit machines ignore the high half!
2542 offset = ConvL2X(offset);
2543 adr = make_unsafe_address(base, offset);
2544 heap_base_oop = base;
2545 val = is_store ? argument(4) : NULL;
2546 } else {
2547 Node* ptr = argument(1); // type: long
2548 ptr = ConvL2X(ptr); // adjust Java long to machine word
2549 adr = make_unsafe_address(NULL, ptr);
2550 val = is_store ? argument(3) : NULL;
2551 }
2552
2553 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2554
2555 // First guess at the value type.
2556 const Type *value_type = Type::get_const_basic_type(type);
2557
2558 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2559 // there was not enough information to nail it down.
2560 Compile::AliasType* alias_type = C->alias_type(adr_type);
2561 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2562
2563 // We will need memory barriers unless we can determine a unique
2564 // alias category for this reference. (Note: If for some reason
2565 // the barriers get omitted and the unsafe reference begins to "pollute"
2566 // the alias analysis of the rest of the graph, either Compile::can_alias
2567 // or Compile::must_alias will throw a diagnostic assert.)
2568 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2569
2570 // If we are reading the value of the referent field of a Reference
2571 // object (either by using Unsafe directly or through reflection)
2572 // then, if G1 is enabled, we need to record the referent in an
2573 // SATB log buffer using the pre-barrier mechanism.
2574 // Also we need to add memory barrier to prevent commoning reads
2575 // from this field across safepoint since GC can change its value.
2576 bool need_read_barrier = !is_native_ptr && !is_store &&
2577 offset != top() && heap_base_oop != top();
2578
2579 if (!is_store && type == T_OBJECT) {
2580 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2581 if (tjp != NULL) {
2582 value_type = tjp;
2583 }
2584 }
2585
2586 receiver = null_check(receiver);
2587 if (stopped()) {
2588 return true;
2589 }
2590 // Heap pointers get a null-check from the interpreter,
2591 // as a courtesy. However, this is not guaranteed by Unsafe,
2592 // and it is not possible to fully distinguish unintended nulls
2593 // from intended ones in this API.
2594
2595 if (is_volatile) {
2596 // We need to emit leading and trailing CPU membars (see below) in
2597 // addition to memory membars when is_volatile. This is a little
2598 // too strong, but avoids the need to insert per-alias-type
2599 // volatile membars (for stores; compare Parse::do_put_xxx), which
2600 // we cannot do effectively here because we probably only have a
2601 // rough approximation of type.
2602 need_mem_bar = true;
2603 // For Stores, place a memory ordering barrier now.
2604 if (is_store) {
2605 insert_mem_bar(Op_MemBarRelease);
2606 } else {
2607 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2608 insert_mem_bar(Op_MemBarVolatile);
2609 }
2610 }
2611 }
2612
2613 // Memory barrier to prevent normal and 'unsafe' accesses from
2614 // bypassing each other. Happens after null checks, so the
2615 // exception paths do not take memory state from the memory barrier,
2616 // so there's no problems making a strong assert about mixing users
2617 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2618 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2619 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2620
2621 if (!is_store) {
2622 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2623 // load value
2624 switch (type) {
2625 case T_BOOLEAN:
2626 case T_CHAR:
2627 case T_BYTE:
2628 case T_SHORT:
2629 case T_INT:
2630 case T_LONG:
2631 case T_FLOAT:
2632 case T_DOUBLE:
2633 break;
2634 case T_OBJECT:
2635 if (need_read_barrier) {
2636 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2637 }
2638 break;
2639 case T_ADDRESS:
2640 // Cast to an int type.
2641 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2642 p = ConvX2UL(p);
2643 break;
2644 default:
2645 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2646 break;
2647 }
2648 // The load node has the control of the preceding MemBarCPUOrder. All
2649 // following nodes will have the control of the MemBarCPUOrder inserted at
2650 // the end of this method. So, pushing the load onto the stack at a later
2651 // point is fine.
2652 set_result(p);
2653 } else {
2654 // place effect of store into memory
2655 switch (type) {
2656 case T_DOUBLE:
2657 val = dstore_rounding(val);
2658 break;
2659 case T_ADDRESS:
2660 // Repackage the long as a pointer.
2661 val = ConvL2X(val);
2662 val = _gvn.transform(new (C) CastX2PNode(val));
2663 break;
2664 }
2665
2666 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2667 if (type != T_OBJECT ) {
2668 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2669 } else {
2670 // Possibly an oop being stored to Java heap or native memory
2671 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2672 // oop to Java heap.
2673 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2674 } else {
2675 // We can't tell at compile time if we are storing in the Java heap or outside
2676 // of it. So we need to emit code to conditionally do the proper type of
2677 // store.
2678
2679 IdealKit ideal(this);
2680 #define __ ideal.
2681 // QQQ who knows what probability is here??
2682 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2683 // Sync IdealKit and graphKit.
2684 sync_kit(ideal);
2685 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2686 // Update IdealKit memory.
2687 __ sync_kit(this);
2688 } __ else_(); {
2689 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2690 } __ end_if();
2691 // Final sync IdealKit and GraphKit.
2692 final_sync(ideal);
2693 #undef __
2694 }
2695 }
2696 }
2697
2698 if (is_volatile) {
2699 if (!is_store) {
2700 insert_mem_bar(Op_MemBarAcquire);
2701 } else {
2702 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2703 insert_mem_bar(Op_MemBarVolatile);
2704 }
2705 }
2706 }
2707
2708 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2709
2710 return true;
2711 }
2712
2713 //----------------------------inline_unsafe_prefetch----------------------------
2714
2715 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2716 #ifndef PRODUCT
2717 {
2718 ResourceMark rm;
2719 // Check the signatures.
2720 ciSignature* sig = callee()->signature();
2721 #ifdef ASSERT
2722 // Object getObject(Object base, int/long offset), etc.
2723 BasicType rtype = sig->return_type()->basic_type();
2724 if (!is_native_ptr) {
2725 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2726 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2727 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2728 } else {
2729 assert(sig->count() == 1, "native prefetch has 1 argument");
2730 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2731 }
2732 #endif // ASSERT
2733 }
2734 #endif // !PRODUCT
2735
2736 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2737
2738 const int idx = is_static ? 0 : 1;
2739 if (!is_static) {
2740 null_check_receiver();
2741 if (stopped()) {
2742 return true;
2743 }
2744 }
2745
2746 // Build address expression. See the code in inline_unsafe_access.
2747 Node *adr;
2748 if (!is_native_ptr) {
2749 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2750 Node* base = argument(idx + 0); // type: oop
2751 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2752 Node* offset = argument(idx + 1); // type: long
2753 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2754 // to be plain byte offsets, which are also the same as those accepted
2755 // by oopDesc::field_base.
2756 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2757 "fieldOffset must be byte-scaled");
2758 // 32-bit machines ignore the high half!
2759 offset = ConvL2X(offset);
2760 adr = make_unsafe_address(base, offset);
2761 } else {
2762 Node* ptr = argument(idx + 0); // type: long
2763 ptr = ConvL2X(ptr); // adjust Java long to machine word
2764 adr = make_unsafe_address(NULL, ptr);
2765 }
2766
2767 // Generate the read or write prefetch
2768 Node *prefetch;
2769 if (is_store) {
2770 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2771 } else {
2772 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2773 }
2774 prefetch->init_req(0, control());
2775 set_i_o(_gvn.transform(prefetch));
2776
2777 return true;
2778 }
2779
2780 //----------------------------inline_unsafe_load_store----------------------------
2781 // This method serves a couple of different customers (depending on LoadStoreKind):
2782 //
2783 // LS_cmpxchg:
2784 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2785 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2786 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2787 //
2788 // LS_xadd:
2789 // public int getAndAddInt( Object o, long offset, int delta)
2790 // public long getAndAddLong(Object o, long offset, long delta)
2791 //
2792 // LS_xchg:
2793 // int getAndSet(Object o, long offset, int newValue)
2794 // long getAndSet(Object o, long offset, long newValue)
2795 // Object getAndSet(Object o, long offset, Object newValue)
2796 //
2797 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2798 // This basic scheme here is the same as inline_unsafe_access, but
2799 // differs in enough details that combining them would make the code
2800 // overly confusing. (This is a true fact! I originally combined
2801 // them, but even I was confused by it!) As much code/comments as
2802 // possible are retained from inline_unsafe_access though to make
2803 // the correspondences clearer. - dl
2804
2805 if (callee()->is_static()) return false; // caller must have the capability!
2806
2807 #ifndef PRODUCT
2808 BasicType rtype;
2809 {
2810 ResourceMark rm;
2811 // Check the signatures.
2812 ciSignature* sig = callee()->signature();
2813 rtype = sig->return_type()->basic_type();
2814 if (kind == LS_xadd || kind == LS_xchg) {
2815 // Check the signatures.
2816 #ifdef ASSERT
2817 assert(rtype == type, "get and set must return the expected type");
2818 assert(sig->count() == 3, "get and set has 3 arguments");
2819 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2820 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2821 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2822 #endif // ASSERT
2823 } else if (kind == LS_cmpxchg) {
2824 // Check the signatures.
2825 #ifdef ASSERT
2826 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2827 assert(sig->count() == 4, "CAS has 4 arguments");
2828 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2829 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2830 #endif // ASSERT
2831 } else {
2832 ShouldNotReachHere();
2833 }
2834 }
2835 #endif //PRODUCT
2836
2837 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2838
2839 // Get arguments:
2840 Node* receiver = NULL;
2841 Node* base = NULL;
2842 Node* offset = NULL;
2843 Node* oldval = NULL;
2844 Node* newval = NULL;
2845 if (kind == LS_cmpxchg) {
2846 const bool two_slot_type = type2size[type] == 2;
2847 receiver = argument(0); // type: oop
2848 base = argument(1); // type: oop
2849 offset = argument(2); // type: long
2850 oldval = argument(4); // type: oop, int, or long
2851 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2852 } else if (kind == LS_xadd || kind == LS_xchg){
2853 receiver = argument(0); // type: oop
2854 base = argument(1); // type: oop
2855 offset = argument(2); // type: long
2856 oldval = NULL;
2857 newval = argument(4); // type: oop, int, or long
2858 }
2859
2860 // Null check receiver.
2861 receiver = null_check(receiver);
2862 if (stopped()) {
2863 return true;
2864 }
2865
2866 // Build field offset expression.
2867 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2868 // to be plain byte offsets, which are also the same as those accepted
2869 // by oopDesc::field_base.
2870 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2871 // 32-bit machines ignore the high half of long offsets
2872 offset = ConvL2X(offset);
2873 Node* adr = make_unsafe_address(base, offset);
2874 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2875
2876 // For CAS, unlike inline_unsafe_access, there seems no point in
2877 // trying to refine types. Just use the coarse types here.
2878 const Type *value_type = Type::get_const_basic_type(type);
2879 Compile::AliasType* alias_type = C->alias_type(adr_type);
2880 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2881
2882 if (kind == LS_xchg && type == T_OBJECT) {
2883 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2884 if (tjp != NULL) {
2885 value_type = tjp;
2886 }
2887 }
2888
2889 int alias_idx = C->get_alias_index(adr_type);
2890
2891 // Memory-model-wise, a LoadStore acts like a little synchronized
2892 // block, so needs barriers on each side. These don't translate
2893 // into actual barriers on most machines, but we still need rest of
2894 // compiler to respect ordering.
2895
2896 insert_mem_bar(Op_MemBarRelease);
2897 insert_mem_bar(Op_MemBarCPUOrder);
2898
2899 // 4984716: MemBars must be inserted before this
2900 // memory node in order to avoid a false
2901 // dependency which will confuse the scheduler.
2902 Node *mem = memory(alias_idx);
2903
2904 // For now, we handle only those cases that actually exist: ints,
2905 // longs, and Object. Adding others should be straightforward.
2906 Node* load_store;
2907 switch(type) {
2908 case T_INT:
2909 if (kind == LS_xadd) {
2910 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2911 } else if (kind == LS_xchg) {
2912 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2913 } else if (kind == LS_cmpxchg) {
2914 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2915 } else {
2916 ShouldNotReachHere();
2917 }
2918 break;
2919 case T_LONG:
2920 if (kind == LS_xadd) {
2921 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2922 } else if (kind == LS_xchg) {
2923 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2924 } else if (kind == LS_cmpxchg) {
2925 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2926 } else {
2927 ShouldNotReachHere();
2928 }
2929 break;
2930 case T_OBJECT:
2931 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2932 // could be delayed during Parse (for example, in adjust_map_after_if()).
2933 // Execute transformation here to avoid barrier generation in such case.
2934 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2935 newval = _gvn.makecon(TypePtr::NULL_PTR);
2936
2937 // Reference stores need a store barrier.
2938 if (kind == LS_xchg) {
2939 // If pre-barrier must execute before the oop store, old value will require do_load here.
2940 if (!can_move_pre_barrier()) {
2941 pre_barrier(true /* do_load*/,
2942 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2943 NULL /* pre_val*/,
2944 T_OBJECT);
2945 } // Else move pre_barrier to use load_store value, see below.
2946 } else if (kind == LS_cmpxchg) {
2947 // Same as for newval above:
2948 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2949 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2950 }
2951 // The only known value which might get overwritten is oldval.
2952 pre_barrier(false /* do_load */,
2953 control(), NULL, NULL, max_juint, NULL, NULL,
2954 oldval /* pre_val */,
2955 T_OBJECT);
2956 } else {
2957 ShouldNotReachHere();
2958 }
2959
2960 #ifdef _LP64
2961 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2962 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2963 if (kind == LS_xchg) {
2964 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2965 newval_enc, adr_type, value_type->make_narrowoop()));
2966 } else {
2967 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2968 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2969 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2970 newval_enc, oldval_enc));
2971 }
2972 } else
2973 #endif
2974 {
2975 if (kind == LS_xchg) {
2976 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2977 } else {
2978 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2979 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2980 }
2981 }
2982 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2983 break;
2984 default:
2985 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2986 break;
2987 }
2988
2989 // SCMemProjNodes represent the memory state of a LoadStore. Their
2990 // main role is to prevent LoadStore nodes from being optimized away
2991 // when their results aren't used.
2992 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2993 set_memory(proj, alias_idx);
2994
2995 if (type == T_OBJECT && kind == LS_xchg) {
2996 #ifdef _LP64
2997 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2998 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2999 }
3000 #endif
3001 if (can_move_pre_barrier()) {
3002 // Don't need to load pre_val. The old value is returned by load_store.
3003 // The pre_barrier can execute after the xchg as long as no safepoint
3004 // gets inserted between them.
3005 pre_barrier(false /* do_load */,
3006 control(), NULL, NULL, max_juint, NULL, NULL,
3007 load_store /* pre_val */,
3008 T_OBJECT);
3009 }
3010 }
3011
3012 // Add the trailing membar surrounding the access
3013 insert_mem_bar(Op_MemBarCPUOrder);
3014 insert_mem_bar(Op_MemBarAcquire);
3015
3016 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3017 set_result(load_store);
3018 return true;
3019 }
3020
3021 //----------------------------inline_unsafe_ordered_store----------------------
3022 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3023 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3024 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3025 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3026 // This is another variant of inline_unsafe_access, differing in
3027 // that it always issues store-store ("release") barrier and ensures
3028 // store-atomicity (which only matters for "long").
3029
3030 if (callee()->is_static()) return false; // caller must have the capability!
3031
3032 #ifndef PRODUCT
3033 {
3034 ResourceMark rm;
3035 // Check the signatures.
3036 ciSignature* sig = callee()->signature();
3037 #ifdef ASSERT
3038 BasicType rtype = sig->return_type()->basic_type();
3039 assert(rtype == T_VOID, "must return void");
3040 assert(sig->count() == 3, "has 3 arguments");
3041 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3042 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3043 #endif // ASSERT
3044 }
3045 #endif //PRODUCT
3046
3047 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3048
3049 // Get arguments:
3050 Node* receiver = argument(0); // type: oop
3051 Node* base = argument(1); // type: oop
3052 Node* offset = argument(2); // type: long
3053 Node* val = argument(4); // type: oop, int, or long
3054
3055 // Null check receiver.
3056 receiver = null_check(receiver);
3057 if (stopped()) {
3058 return true;
3059 }
3060
3061 // Build field offset expression.
3062 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3063 // 32-bit machines ignore the high half of long offsets
3064 offset = ConvL2X(offset);
3065 Node* adr = make_unsafe_address(base, offset);
3066 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3067 const Type *value_type = Type::get_const_basic_type(type);
3068 Compile::AliasType* alias_type = C->alias_type(adr_type);
3069
3070 insert_mem_bar(Op_MemBarRelease);
3071 insert_mem_bar(Op_MemBarCPUOrder);
3072 // Ensure that the store is atomic for longs:
3073 const bool require_atomic_access = true;
3074 Node* store;
3075 if (type == T_OBJECT) // reference stores need a store barrier.
3076 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3077 else {
3078 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3079 }
3080 insert_mem_bar(Op_MemBarCPUOrder);
3081 return true;
3082 }
3083
3084 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3085 // Regardless of form, don't allow previous ld/st to move down,
3086 // then issue acquire, release, or volatile mem_bar.
3087 insert_mem_bar(Op_MemBarCPUOrder);
3088 switch(id) {
3089 case vmIntrinsics::_loadFence:
3090 insert_mem_bar(Op_LoadFence);
3091 return true;
3092 case vmIntrinsics::_storeFence:
3093 insert_mem_bar(Op_StoreFence);
3094 return true;
3095 case vmIntrinsics::_fullFence:
3096 insert_mem_bar(Op_MemBarVolatile);
3097 return true;
3098 default:
3099 fatal_unexpected_iid(id);
3100 return false;
3101 }
3102 }
3103
3104 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3105 if (!kls->is_Con()) {
3106 return true;
3107 }
3108 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3109 if (klsptr == NULL) {
3110 return true;
3111 }
3112 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3113 // don't need a guard for a klass that is already initialized
3114 return !ik->is_initialized();
3115 }
3116
3117 //----------------------------inline_unsafe_allocate---------------------------
3118 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3119 bool LibraryCallKit::inline_unsafe_allocate() {
3120 if (callee()->is_static()) return false; // caller must have the capability!
3121
3122 null_check_receiver(); // null-check, then ignore
3123 Node* cls = null_check(argument(1));
3124 if (stopped()) return true;
3125
3126 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3127 kls = null_check(kls);
3128 if (stopped()) return true; // argument was like int.class
3129
3130 Node* test = NULL;
3131 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3132 // Note: The argument might still be an illegal value like
3133 // Serializable.class or Object[].class. The runtime will handle it.
3134 // But we must make an explicit check for initialization.
3135 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3136 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3137 // can generate code to load it as unsigned byte.
3138 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3139 Node* bits = intcon(InstanceKlass::fully_initialized);
3140 test = _gvn.transform(new (C) SubINode(inst, bits));
3141 // The 'test' is non-zero if we need to take a slow path.
3142 }
3143
3144 Node* obj = new_instance(kls, test);
3145 set_result(obj);
3146 return true;
3147 }
3148
3149 #ifdef TRACE_HAVE_INTRINSICS
3150 /*
3151 * oop -> myklass
3152 * myklass->trace_id |= USED
3153 * return myklass->trace_id & ~0x3
3154 */
3155 bool LibraryCallKit::inline_native_classID() {
3156 null_check_receiver(); // null-check, then ignore
3157 Node* cls = null_check(argument(1), T_OBJECT);
3158 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3159 kls = null_check(kls, T_OBJECT);
3160 ByteSize offset = TRACE_ID_OFFSET;
3161 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3162 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3163 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3164 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3165 Node* clsused = longcon(0x01l); // set the class bit
3166 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3167
3168 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3169 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3170 set_result(andl);
3171 return true;
3172 }
3173
3174 bool LibraryCallKit::inline_native_threadID() {
3175 Node* tls_ptr = NULL;
3176 Node* cur_thr = generate_current_thread(tls_ptr);
3177 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3178 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3179 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3180
3181 Node* threadid = NULL;
3182 size_t thread_id_size = OSThread::thread_id_size();
3183 if (thread_id_size == (size_t) BytesPerLong) {
3184 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3185 } else if (thread_id_size == (size_t) BytesPerInt) {
3186 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3187 } else {
3188 ShouldNotReachHere();
3189 }
3190 set_result(threadid);
3191 return true;
3192 }
3193 #endif
3194
3195 //------------------------inline_native_time_funcs--------------
3196 // inline code for System.currentTimeMillis() and System.nanoTime()
3197 // these have the same type and signature
3198 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3199 const TypeFunc* tf = OptoRuntime::void_long_Type();
3200 const TypePtr* no_memory_effects = NULL;
3201 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3202 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3203 #ifdef ASSERT
3204 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3205 assert(value_top == top(), "second value must be top");
3206 #endif
3207 set_result(value);
3208 return true;
3209 }
3210
3211 //------------------------inline_native_currentThread------------------
3212 bool LibraryCallKit::inline_native_currentThread() {
3213 Node* junk = NULL;
3214 set_result(generate_current_thread(junk));
3215 return true;
3216 }
3217
3218 //------------------------inline_native_isInterrupted------------------
3219 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3220 bool LibraryCallKit::inline_native_isInterrupted() {
3221 // Add a fast path to t.isInterrupted(clear_int):
3222 // (t == Thread.current() &&
3223 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3224 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3225 // So, in the common case that the interrupt bit is false,
3226 // we avoid making a call into the VM. Even if the interrupt bit
3227 // is true, if the clear_int argument is false, we avoid the VM call.
3228 // However, if the receiver is not currentThread, we must call the VM,
3229 // because there must be some locking done around the operation.
3230
3231 // We only go to the fast case code if we pass two guards.
3232 // Paths which do not pass are accumulated in the slow_region.
3233
3234 enum {
3235 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3236 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3237 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3238 PATH_LIMIT
3239 };
3240
3241 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3242 // out of the function.
3243 insert_mem_bar(Op_MemBarCPUOrder);
3244
3245 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3246 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3247
3248 RegionNode* slow_region = new (C) RegionNode(1);
3249 record_for_igvn(slow_region);
3250
3251 // (a) Receiving thread must be the current thread.
3252 Node* rec_thr = argument(0);
3253 Node* tls_ptr = NULL;
3254 Node* cur_thr = generate_current_thread(tls_ptr);
3255 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3256 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3257
3258 generate_slow_guard(bol_thr, slow_region);
3259
3260 // (b) Interrupt bit on TLS must be false.
3261 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3262 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3263 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3264
3265 // Set the control input on the field _interrupted read to prevent it floating up.
3266 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3267 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3268 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3269
3270 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3271
3272 // First fast path: if (!TLS._interrupted) return false;
3273 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3274 result_rgn->init_req(no_int_result_path, false_bit);
3275 result_val->init_req(no_int_result_path, intcon(0));
3276
3277 // drop through to next case
3278 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3279
3280 #ifndef TARGET_OS_FAMILY_windows
3281 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3282 Node* clr_arg = argument(1);
3283 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3284 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3285 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3286
3287 // Second fast path: ... else if (!clear_int) return true;
3288 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3289 result_rgn->init_req(no_clear_result_path, false_arg);
3290 result_val->init_req(no_clear_result_path, intcon(1));
3291
3292 // drop through to next case
3293 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3294 #else
3295 // To return true on Windows you must read the _interrupted field
3296 // and check the the event state i.e. take the slow path.
3297 #endif // TARGET_OS_FAMILY_windows
3298
3299 // (d) Otherwise, go to the slow path.
3300 slow_region->add_req(control());
3301 set_control( _gvn.transform(slow_region));
3302
3303 if (stopped()) {
3304 // There is no slow path.
3305 result_rgn->init_req(slow_result_path, top());
3306 result_val->init_req(slow_result_path, top());
3307 } else {
3308 // non-virtual because it is a private non-static
3309 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3310
3311 Node* slow_val = set_results_for_java_call(slow_call);
3312 // this->control() comes from set_results_for_java_call
3313
3314 Node* fast_io = slow_call->in(TypeFunc::I_O);
3315 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3316
3317 // These two phis are pre-filled with copies of of the fast IO and Memory
3318 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3319 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3320
3321 result_rgn->init_req(slow_result_path, control());
3322 result_io ->init_req(slow_result_path, i_o());
3323 result_mem->init_req(slow_result_path, reset_memory());
3324 result_val->init_req(slow_result_path, slow_val);
3325
3326 set_all_memory(_gvn.transform(result_mem));
3327 set_i_o( _gvn.transform(result_io));
3328 }
3329
3330 C->set_has_split_ifs(true); // Has chance for split-if optimization
3331 set_result(result_rgn, result_val);
3332 return true;
3333 }
3334
3335 //---------------------------load_mirror_from_klass----------------------------
3336 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3337 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3338 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3339 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3340 }
3341
3342 //-----------------------load_klass_from_mirror_common-------------------------
3343 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3344 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3345 // and branch to the given path on the region.
3346 // If never_see_null, take an uncommon trap on null, so we can optimistically
3347 // compile for the non-null case.
3348 // If the region is NULL, force never_see_null = true.
3349 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3350 bool never_see_null,
3351 RegionNode* region,
3352 int null_path,
3353 int offset) {
3354 if (region == NULL) never_see_null = true;
3355 Node* p = basic_plus_adr(mirror, offset);
3356 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3357 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3358 Node* null_ctl = top();
3359 kls = null_check_oop(kls, &null_ctl, never_see_null);
3360 if (region != NULL) {
3361 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3362 region->init_req(null_path, null_ctl);
3363 } else {
3364 assert(null_ctl == top(), "no loose ends");
3365 }
3366 return kls;
3367 }
3368
3369 //--------------------(inline_native_Class_query helpers)---------------------
3370 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3371 // Fall through if (mods & mask) == bits, take the guard otherwise.
3372 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3373 // Branch around if the given klass has the given modifier bit set.
3374 // Like generate_guard, adds a new path onto the region.
3375 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3376 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3377 Node* mask = intcon(modifier_mask);
3378 Node* bits = intcon(modifier_bits);
3379 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3380 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3381 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3382 return generate_fair_guard(bol, region);
3383 }
3384 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3385 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3386 }
3387
3388 //-------------------------inline_native_Class_query-------------------
3389 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3390 const Type* return_type = TypeInt::BOOL;
3391 Node* prim_return_value = top(); // what happens if it's a primitive class?
3392 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3393 bool expect_prim = false; // most of these guys expect to work on refs
3394
3395 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3396
3397 Node* mirror = argument(0);
3398 Node* obj = top();
3399
3400 switch (id) {
3401 case vmIntrinsics::_isInstance:
3402 // nothing is an instance of a primitive type
3403 prim_return_value = intcon(0);
3404 obj = argument(1);
3405 break;
3406 case vmIntrinsics::_getModifiers:
3407 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3408 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3409 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3410 break;
3411 case vmIntrinsics::_isInterface:
3412 prim_return_value = intcon(0);
3413 break;
3414 case vmIntrinsics::_isArray:
3415 prim_return_value = intcon(0);
3416 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3417 break;
3418 case vmIntrinsics::_isPrimitive:
3419 prim_return_value = intcon(1);
3420 expect_prim = true; // obviously
3421 break;
3422 case vmIntrinsics::_getSuperclass:
3423 prim_return_value = null();
3424 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3425 break;
3426 case vmIntrinsics::_getComponentType:
3427 prim_return_value = null();
3428 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3429 break;
3430 case vmIntrinsics::_getClassAccessFlags:
3431 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3432 return_type = TypeInt::INT; // not bool! 6297094
3433 break;
3434 default:
3435 fatal_unexpected_iid(id);
3436 break;
3437 }
3438
3439 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3440 if (mirror_con == NULL) return false; // cannot happen?
3441
3442 #ifndef PRODUCT
3443 if (C->print_intrinsics() || C->print_inlining()) {
3444 ciType* k = mirror_con->java_mirror_type();
3445 if (k) {
3446 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3447 k->print_name();
3448 tty->cr();
3449 }
3450 }
3451 #endif
3452
3453 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3454 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3455 record_for_igvn(region);
3456 PhiNode* phi = new (C) PhiNode(region, return_type);
3457
3458 // The mirror will never be null of Reflection.getClassAccessFlags, however
3459 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3460 // if it is. See bug 4774291.
3461
3462 // For Reflection.getClassAccessFlags(), the null check occurs in
3463 // the wrong place; see inline_unsafe_access(), above, for a similar
3464 // situation.
3465 mirror = null_check(mirror);
3466 // If mirror or obj is dead, only null-path is taken.
3467 if (stopped()) return true;
3468
3469 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3470
3471 // Now load the mirror's klass metaobject, and null-check it.
3472 // Side-effects region with the control path if the klass is null.
3473 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3474 // If kls is null, we have a primitive mirror.
3475 phi->init_req(_prim_path, prim_return_value);
3476 if (stopped()) { set_result(region, phi); return true; }
3477 bool safe_for_replace = (region->in(_prim_path) == top());
3478
3479 Node* p; // handy temp
3480 Node* null_ctl;
3481
3482 // Now that we have the non-null klass, we can perform the real query.
3483 // For constant classes, the query will constant-fold in LoadNode::Value.
3484 Node* query_value = top();
3485 switch (id) {
3486 case vmIntrinsics::_isInstance:
3487 // nothing is an instance of a primitive type
3488 query_value = gen_instanceof(obj, kls, safe_for_replace);
3489 break;
3490
3491 case vmIntrinsics::_getModifiers:
3492 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3493 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3494 break;
3495
3496 case vmIntrinsics::_isInterface:
3497 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3498 if (generate_interface_guard(kls, region) != NULL)
3499 // A guard was added. If the guard is taken, it was an interface.
3500 phi->add_req(intcon(1));
3501 // If we fall through, it's a plain class.
3502 query_value = intcon(0);
3503 break;
3504
3505 case vmIntrinsics::_isArray:
3506 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3507 if (generate_array_guard(kls, region) != NULL)
3508 // A guard was added. If the guard is taken, it was an array.
3509 phi->add_req(intcon(1));
3510 // If we fall through, it's a plain class.
3511 query_value = intcon(0);
3512 break;
3513
3514 case vmIntrinsics::_isPrimitive:
3515 query_value = intcon(0); // "normal" path produces false
3516 break;
3517
3518 case vmIntrinsics::_getSuperclass:
3519 // The rules here are somewhat unfortunate, but we can still do better
3520 // with random logic than with a JNI call.
3521 // Interfaces store null or Object as _super, but must report null.
3522 // Arrays store an intermediate super as _super, but must report Object.
3523 // Other types can report the actual _super.
3524 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3525 if (generate_interface_guard(kls, region) != NULL)
3526 // A guard was added. If the guard is taken, it was an interface.
3527 phi->add_req(null());
3528 if (generate_array_guard(kls, region) != NULL)
3529 // A guard was added. If the guard is taken, it was an array.
3530 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3531 // If we fall through, it's a plain class. Get its _super.
3532 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3533 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3534 null_ctl = top();
3535 kls = null_check_oop(kls, &null_ctl);
3536 if (null_ctl != top()) {
3537 // If the guard is taken, Object.superClass is null (both klass and mirror).
3538 region->add_req(null_ctl);
3539 phi ->add_req(null());
3540 }
3541 if (!stopped()) {
3542 query_value = load_mirror_from_klass(kls);
3543 }
3544 break;
3545
3546 case vmIntrinsics::_getComponentType:
3547 if (generate_array_guard(kls, region) != NULL) {
3548 // Be sure to pin the oop load to the guard edge just created:
3549 Node* is_array_ctrl = region->in(region->req()-1);
3550 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3551 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3552 phi->add_req(cmo);
3553 }
3554 query_value = null(); // non-array case is null
3555 break;
3556
3557 case vmIntrinsics::_getClassAccessFlags:
3558 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3559 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3560 break;
3561
3562 default:
3563 fatal_unexpected_iid(id);
3564 break;
3565 }
3566
3567 // Fall-through is the normal case of a query to a real class.
3568 phi->init_req(1, query_value);
3569 region->init_req(1, control());
3570
3571 C->set_has_split_ifs(true); // Has chance for split-if optimization
3572 set_result(region, phi);
3573 return true;
3574 }
3575
3576 //--------------------------inline_native_subtype_check------------------------
3577 // This intrinsic takes the JNI calls out of the heart of
3578 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3579 bool LibraryCallKit::inline_native_subtype_check() {
3580 // Pull both arguments off the stack.
3581 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3582 args[0] = argument(0);
3583 args[1] = argument(1);
3584 Node* klasses[2]; // corresponding Klasses: superk, subk
3585 klasses[0] = klasses[1] = top();
3586
3587 enum {
3588 // A full decision tree on {superc is prim, subc is prim}:
3589 _prim_0_path = 1, // {P,N} => false
3590 // {P,P} & superc!=subc => false
3591 _prim_same_path, // {P,P} & superc==subc => true
3592 _prim_1_path, // {N,P} => false
3593 _ref_subtype_path, // {N,N} & subtype check wins => true
3594 _both_ref_path, // {N,N} & subtype check loses => false
3595 PATH_LIMIT
3596 };
3597
3598 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3599 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3600 record_for_igvn(region);
3601
3602 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3603 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3604 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3605
3606 // First null-check both mirrors and load each mirror's klass metaobject.
3607 int which_arg;
3608 for (which_arg = 0; which_arg <= 1; which_arg++) {
3609 Node* arg = args[which_arg];
3610 arg = null_check(arg);
3611 if (stopped()) break;
3612 args[which_arg] = arg;
3613
3614 Node* p = basic_plus_adr(arg, class_klass_offset);
3615 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3616 klasses[which_arg] = _gvn.transform(kls);
3617 }
3618
3619 // Having loaded both klasses, test each for null.
3620 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3621 for (which_arg = 0; which_arg <= 1; which_arg++) {
3622 Node* kls = klasses[which_arg];
3623 Node* null_ctl = top();
3624 kls = null_check_oop(kls, &null_ctl, never_see_null);
3625 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3626 region->init_req(prim_path, null_ctl);
3627 if (stopped()) break;
3628 klasses[which_arg] = kls;
3629 }
3630
3631 if (!stopped()) {
3632 // now we have two reference types, in klasses[0..1]
3633 Node* subk = klasses[1]; // the argument to isAssignableFrom
3634 Node* superk = klasses[0]; // the receiver
3635 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3636 // now we have a successful reference subtype check
3637 region->set_req(_ref_subtype_path, control());
3638 }
3639
3640 // If both operands are primitive (both klasses null), then
3641 // we must return true when they are identical primitives.
3642 // It is convenient to test this after the first null klass check.
3643 set_control(region->in(_prim_0_path)); // go back to first null check
3644 if (!stopped()) {
3645 // Since superc is primitive, make a guard for the superc==subc case.
3646 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3647 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3648 generate_guard(bol_eq, region, PROB_FAIR);
3649 if (region->req() == PATH_LIMIT+1) {
3650 // A guard was added. If the added guard is taken, superc==subc.
3651 region->swap_edges(PATH_LIMIT, _prim_same_path);
3652 region->del_req(PATH_LIMIT);
3653 }
3654 region->set_req(_prim_0_path, control()); // Not equal after all.
3655 }
3656
3657 // these are the only paths that produce 'true':
3658 phi->set_req(_prim_same_path, intcon(1));
3659 phi->set_req(_ref_subtype_path, intcon(1));
3660
3661 // pull together the cases:
3662 assert(region->req() == PATH_LIMIT, "sane region");
3663 for (uint i = 1; i < region->req(); i++) {
3664 Node* ctl = region->in(i);
3665 if (ctl == NULL || ctl == top()) {
3666 region->set_req(i, top());
3667 phi ->set_req(i, top());
3668 } else if (phi->in(i) == NULL) {
3669 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3670 }
3671 }
3672
3673 set_control(_gvn.transform(region));
3674 set_result(_gvn.transform(phi));
3675 return true;
3676 }
3677
3678 //---------------------generate_array_guard_common------------------------
3679 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3680 bool obj_array, bool not_array) {
3681 // If obj_array/non_array==false/false:
3682 // Branch around if the given klass is in fact an array (either obj or prim).
3683 // If obj_array/non_array==false/true:
3684 // Branch around if the given klass is not an array klass of any kind.
3685 // If obj_array/non_array==true/true:
3686 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3687 // If obj_array/non_array==true/false:
3688 // Branch around if the kls is an oop array (Object[] or subtype)
3689 //
3690 // Like generate_guard, adds a new path onto the region.
3691 jint layout_con = 0;
3692 Node* layout_val = get_layout_helper(kls, layout_con);
3693 if (layout_val == NULL) {
3694 bool query = (obj_array
3695 ? Klass::layout_helper_is_objArray(layout_con)
3696 : Klass::layout_helper_is_array(layout_con));
3697 if (query == not_array) {
3698 return NULL; // never a branch
3699 } else { // always a branch
3700 Node* always_branch = control();
3701 if (region != NULL)
3702 region->add_req(always_branch);
3703 set_control(top());
3704 return always_branch;
3705 }
3706 }
3707 // Now test the correct condition.
3708 jint nval = (obj_array
3709 ? ((jint)Klass::_lh_array_tag_type_value
3710 << Klass::_lh_array_tag_shift)
3711 : Klass::_lh_neutral_value);
3712 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3713 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3714 // invert the test if we are looking for a non-array
3715 if (not_array) btest = BoolTest(btest).negate();
3716 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3717 return generate_fair_guard(bol, region);
3718 }
3719
3720
3721 //-----------------------inline_native_newArray--------------------------
3722 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3723 bool LibraryCallKit::inline_native_newArray() {
3724 Node* mirror = argument(0);
3725 Node* count_val = argument(1);
3726
3727 mirror = null_check(mirror);
3728 // If mirror or obj is dead, only null-path is taken.
3729 if (stopped()) return true;
3730
3731 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3732 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3733 PhiNode* result_val = new(C) PhiNode(result_reg,
3734 TypeInstPtr::NOTNULL);
3735 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3736 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3737 TypePtr::BOTTOM);
3738
3739 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3740 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3741 result_reg, _slow_path);
3742 Node* normal_ctl = control();
3743 Node* no_array_ctl = result_reg->in(_slow_path);
3744
3745 // Generate code for the slow case. We make a call to newArray().
3746 set_control(no_array_ctl);
3747 if (!stopped()) {
3748 // Either the input type is void.class, or else the
3749 // array klass has not yet been cached. Either the
3750 // ensuing call will throw an exception, or else it
3751 // will cache the array klass for next time.
3752 PreserveJVMState pjvms(this);
3753 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3754 Node* slow_result = set_results_for_java_call(slow_call);
3755 // this->control() comes from set_results_for_java_call
3756 result_reg->set_req(_slow_path, control());
3757 result_val->set_req(_slow_path, slow_result);
3758 result_io ->set_req(_slow_path, i_o());
3759 result_mem->set_req(_slow_path, reset_memory());
3760 }
3761
3762 set_control(normal_ctl);
3763 if (!stopped()) {
3764 // Normal case: The array type has been cached in the java.lang.Class.
3765 // The following call works fine even if the array type is polymorphic.
3766 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3767 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3768 result_reg->init_req(_normal_path, control());
3769 result_val->init_req(_normal_path, obj);
3770 result_io ->init_req(_normal_path, i_o());
3771 result_mem->init_req(_normal_path, reset_memory());
3772 }
3773
3774 // Return the combined state.
3775 set_i_o( _gvn.transform(result_io) );
3776 set_all_memory( _gvn.transform(result_mem));
3777
3778 C->set_has_split_ifs(true); // Has chance for split-if optimization
3779 set_result(result_reg, result_val);
3780 return true;
3781 }
3782
3783 //----------------------inline_native_getLength--------------------------
3784 // public static native int java.lang.reflect.Array.getLength(Object array);
3785 bool LibraryCallKit::inline_native_getLength() {
3786 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3787
3788 Node* array = null_check(argument(0));
3789 // If array is dead, only null-path is taken.
3790 if (stopped()) return true;
3791
3792 // Deoptimize if it is a non-array.
3793 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3794
3795 if (non_array != NULL) {
3796 PreserveJVMState pjvms(this);
3797 set_control(non_array);
3798 uncommon_trap(Deoptimization::Reason_intrinsic,
3799 Deoptimization::Action_maybe_recompile);
3800 }
3801
3802 // If control is dead, only non-array-path is taken.
3803 if (stopped()) return true;
3804
3805 // The works fine even if the array type is polymorphic.
3806 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3807 Node* result = load_array_length(array);
3808
3809 C->set_has_split_ifs(true); // Has chance for split-if optimization
3810 set_result(result);
3811 return true;
3812 }
3813
3814 //------------------------inline_array_copyOf----------------------------
3815 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3816 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3817 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3818 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3819
3820 // Get the arguments.
3821 Node* original = argument(0);
3822 Node* start = is_copyOfRange? argument(1): intcon(0);
3823 Node* end = is_copyOfRange? argument(2): argument(1);
3824 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3825
3826 Node* newcopy;
3827
3828 // Set the original stack and the reexecute bit for the interpreter to reexecute
3829 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3830 { PreserveReexecuteState preexecs(this);
3831 jvms()->set_should_reexecute(true);
3832
3833 array_type_mirror = null_check(array_type_mirror);
3834 original = null_check(original);
3835
3836 // Check if a null path was taken unconditionally.
3837 if (stopped()) return true;
3838
3839 Node* orig_length = load_array_length(original);
3840
3841 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3842 klass_node = null_check(klass_node);
3843
3844 RegionNode* bailout = new (C) RegionNode(1);
3845 record_for_igvn(bailout);
3846
3847 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3848 // Bail out if that is so.
3849 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3850 if (not_objArray != NULL) {
3851 // Improve the klass node's type from the new optimistic assumption:
3852 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3853 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3854 Node* cast = new (C) CastPPNode(klass_node, akls);
3855 cast->init_req(0, control());
3856 klass_node = _gvn.transform(cast);
3857 }
3858
3859 // Bail out if either start or end is negative.
3860 generate_negative_guard(start, bailout, &start);
3861 generate_negative_guard(end, bailout, &end);
3862
3863 Node* length = end;
3864 if (_gvn.type(start) != TypeInt::ZERO) {
3865 length = _gvn.transform(new (C) SubINode(end, start));
3866 }
3867
3868 // Bail out if length is negative.
3869 // Without this the new_array would throw
3870 // NegativeArraySizeException but IllegalArgumentException is what
3871 // should be thrown
3872 generate_negative_guard(length, bailout, &length);
3873
3874 if (bailout->req() > 1) {
3875 PreserveJVMState pjvms(this);
3876 set_control(_gvn.transform(bailout));
3877 uncommon_trap(Deoptimization::Reason_intrinsic,
3878 Deoptimization::Action_maybe_recompile);
3879 }
3880
3881 if (!stopped()) {
3882 // How many elements will we copy from the original?
3883 // The answer is MinI(orig_length - start, length).
3884 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3885 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3886
3887 newcopy = new_array(klass_node, length, 0); // no argments to push
3888
3889 // Generate a direct call to the right arraycopy function(s).
3890 // We know the copy is disjoint but we might not know if the
3891 // oop stores need checking.
3892 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3893 // This will fail a store-check if x contains any non-nulls.
3894 bool disjoint_bases = true;
3895 // if start > orig_length then the length of the copy may be
3896 // negative.
3897 bool length_never_negative = !is_copyOfRange;
3898 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3899 original, start, newcopy, intcon(0), moved,
3900 disjoint_bases, length_never_negative);
3901 }
3902 } // original reexecute is set back here
3903
3904 C->set_has_split_ifs(true); // Has chance for split-if optimization
3905 if (!stopped()) {
3906 set_result(newcopy);
3907 }
3908 return true;
3909 }
3910
3911
3912 //----------------------generate_virtual_guard---------------------------
3913 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3914 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3915 RegionNode* slow_region) {
3916 ciMethod* method = callee();
3917 int vtable_index = method->vtable_index();
3918 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3919 err_msg_res("bad index %d", vtable_index));
3920 // Get the Method* out of the appropriate vtable entry.
3921 int entry_offset = (InstanceKlass::vtable_start_offset() +
3922 vtable_index*vtableEntry::size()) * wordSize +
3923 vtableEntry::method_offset_in_bytes();
3924 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3925 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3926
3927 // Compare the target method with the expected method (e.g., Object.hashCode).
3928 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3929
3930 Node* native_call = makecon(native_call_addr);
3931 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3932 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3933
3934 return generate_slow_guard(test_native, slow_region);
3935 }
3936
3937 //-----------------------generate_method_call----------------------------
3938 // Use generate_method_call to make a slow-call to the real
3939 // method if the fast path fails. An alternative would be to
3940 // use a stub like OptoRuntime::slow_arraycopy_Java.
3941 // This only works for expanding the current library call,
3942 // not another intrinsic. (E.g., don't use this for making an
3943 // arraycopy call inside of the copyOf intrinsic.)
3944 CallJavaNode*
3945 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3946 // When compiling the intrinsic method itself, do not use this technique.
3947 guarantee(callee() != C->method(), "cannot make slow-call to self");
3948
3949 ciMethod* method = callee();
3950 // ensure the JVMS we have will be correct for this call
3951 guarantee(method_id == method->intrinsic_id(), "must match");
3952
3953 const TypeFunc* tf = TypeFunc::make(method);
3954 CallJavaNode* slow_call;
3955 if (is_static) {
3956 assert(!is_virtual, "");
3957 slow_call = new(C) CallStaticJavaNode(C, tf,
3958 SharedRuntime::get_resolve_static_call_stub(),
3959 method, bci());
3960 } else if (is_virtual) {
3961 null_check_receiver();
3962 int vtable_index = Method::invalid_vtable_index;
3963 if (UseInlineCaches) {
3964 // Suppress the vtable call
3965 } else {
3966 // hashCode and clone are not a miranda methods,
3967 // so the vtable index is fixed.
3968 // No need to use the linkResolver to get it.
3969 vtable_index = method->vtable_index();
3970 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3971 err_msg_res("bad index %d", vtable_index));
3972 }
3973 slow_call = new(C) CallDynamicJavaNode(tf,
3974 SharedRuntime::get_resolve_virtual_call_stub(),
3975 method, vtable_index, bci());
3976 } else { // neither virtual nor static: opt_virtual
3977 null_check_receiver();
3978 slow_call = new(C) CallStaticJavaNode(C, tf,
3979 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3980 method, bci());
3981 slow_call->set_optimized_virtual(true);
3982 }
3983 set_arguments_for_java_call(slow_call);
3984 set_edges_for_java_call(slow_call);
3985 return slow_call;
3986 }
3987
3988
3989 //------------------------------inline_native_hashcode--------------------
3990 // Build special case code for calls to hashCode on an object.
3991 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3992 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3993 assert(!(is_virtual && is_static), "either virtual, special, or static");
3994
3995 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3996
3997 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3998 PhiNode* result_val = new(C) PhiNode(result_reg,
3999 TypeInt::INT);
4000 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4001 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4002 TypePtr::BOTTOM);
4003 Node* obj = NULL;
4004 if (!is_static) {
4005 // Check for hashing null object
4006 obj = null_check_receiver();
4007 if (stopped()) return true; // unconditionally null
4008 result_reg->init_req(_null_path, top());
4009 result_val->init_req(_null_path, top());
4010 } else {
4011 // Do a null check, and return zero if null.
4012 // System.identityHashCode(null) == 0
4013 obj = argument(0);
4014 Node* null_ctl = top();
4015 obj = null_check_oop(obj, &null_ctl);
4016 result_reg->init_req(_null_path, null_ctl);
4017 result_val->init_req(_null_path, _gvn.intcon(0));
4018 }
4019
4020 // Unconditionally null? Then return right away.
4021 if (stopped()) {
4022 set_control( result_reg->in(_null_path));
4023 if (!stopped())
4024 set_result(result_val->in(_null_path));
4025 return true;
4026 }
4027
4028 // After null check, get the object's klass.
4029 Node* obj_klass = load_object_klass(obj);
4030
4031 // This call may be virtual (invokevirtual) or bound (invokespecial).
4032 // For each case we generate slightly different code.
4033
4034 // We only go to the fast case code if we pass a number of guards. The
4035 // paths which do not pass are accumulated in the slow_region.
4036 RegionNode* slow_region = new (C) RegionNode(1);
4037 record_for_igvn(slow_region);
4038
4039 // If this is a virtual call, we generate a funny guard. We pull out
4040 // the vtable entry corresponding to hashCode() from the target object.
4041 // If the target method which we are calling happens to be the native
4042 // Object hashCode() method, we pass the guard. We do not need this
4043 // guard for non-virtual calls -- the caller is known to be the native
4044 // Object hashCode().
4045 if (is_virtual) {
4046 generate_virtual_guard(obj_klass, slow_region);
4047 }
4048
4049 // Get the header out of the object, use LoadMarkNode when available
4050 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4051 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4052
4053 // Test the header to see if it is unlocked.
4054 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4055 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4056 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4057 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4058 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4059
4060 generate_slow_guard(test_unlocked, slow_region);
4061
4062 // Get the hash value and check to see that it has been properly assigned.
4063 // We depend on hash_mask being at most 32 bits and avoid the use of
4064 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4065 // vm: see markOop.hpp.
4066 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4067 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4068 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4069 // This hack lets the hash bits live anywhere in the mark object now, as long
4070 // as the shift drops the relevant bits into the low 32 bits. Note that
4071 // Java spec says that HashCode is an int so there's no point in capturing
4072 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4073 hshifted_header = ConvX2I(hshifted_header);
4074 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4075
4076 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4077 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4078 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4079
4080 generate_slow_guard(test_assigned, slow_region);
4081
4082 Node* init_mem = reset_memory();
4083 // fill in the rest of the null path:
4084 result_io ->init_req(_null_path, i_o());
4085 result_mem->init_req(_null_path, init_mem);
4086
4087 result_val->init_req(_fast_path, hash_val);
4088 result_reg->init_req(_fast_path, control());
4089 result_io ->init_req(_fast_path, i_o());
4090 result_mem->init_req(_fast_path, init_mem);
4091
4092 // Generate code for the slow case. We make a call to hashCode().
4093 set_control(_gvn.transform(slow_region));
4094 if (!stopped()) {
4095 // No need for PreserveJVMState, because we're using up the present state.
4096 set_all_memory(init_mem);
4097 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4098 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4099 Node* slow_result = set_results_for_java_call(slow_call);
4100 // this->control() comes from set_results_for_java_call
4101 result_reg->init_req(_slow_path, control());
4102 result_val->init_req(_slow_path, slow_result);
4103 result_io ->set_req(_slow_path, i_o());
4104 result_mem ->set_req(_slow_path, reset_memory());
4105 }
4106
4107 // Return the combined state.
4108 set_i_o( _gvn.transform(result_io) );
4109 set_all_memory( _gvn.transform(result_mem));
4110
4111 set_result(result_reg, result_val);
4112 return true;
4113 }
4114
4115 //---------------------------inline_native_getClass----------------------------
4116 // public final native Class<?> java.lang.Object.getClass();
4117 //
4118 // Build special case code for calls to getClass on an object.
4119 bool LibraryCallKit::inline_native_getClass() {
4120 Node* obj = null_check_receiver();
4121 if (stopped()) return true;
4122 set_result(load_mirror_from_klass(load_object_klass(obj)));
4123 return true;
4124 }
4125
4126 //-----------------inline_native_Reflection_getCallerClass---------------------
4127 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4128 //
4129 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4130 //
4131 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4132 // in that it must skip particular security frames and checks for
4133 // caller sensitive methods.
4134 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4135 #ifndef PRODUCT
4136 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4137 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4138 }
4139 #endif
4140
4141 if (!jvms()->has_method()) {
4142 #ifndef PRODUCT
4143 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4144 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4145 }
4146 #endif
4147 return false;
4148 }
4149
4150 // Walk back up the JVM state to find the caller at the required
4151 // depth.
4152 JVMState* caller_jvms = jvms();
4153
4154 // Cf. JVM_GetCallerClass
4155 // NOTE: Start the loop at depth 1 because the current JVM state does
4156 // not include the Reflection.getCallerClass() frame.
4157 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4158 ciMethod* m = caller_jvms->method();
4159 switch (n) {
4160 case 0:
4161 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4162 break;
4163 case 1:
4164 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4165 if (!m->caller_sensitive()) {
4166 #ifndef PRODUCT
4167 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4168 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4169 }
4170 #endif
4171 return false; // bail-out; let JVM_GetCallerClass do the work
4172 }
4173 break;
4174 default:
4175 if (!m->is_ignored_by_security_stack_walk()) {
4176 // We have reached the desired frame; return the holder class.
4177 // Acquire method holder as java.lang.Class and push as constant.
4178 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4179 ciInstance* caller_mirror = caller_klass->java_mirror();
4180 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4181
4182 #ifndef PRODUCT
4183 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4184 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());
4185 tty->print_cr(" JVM state at this point:");
4186 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4187 ciMethod* m = jvms()->of_depth(i)->method();
4188 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4189 }
4190 }
4191 #endif
4192 return true;
4193 }
4194 break;
4195 }
4196 }
4197
4198 #ifndef PRODUCT
4199 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4200 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4201 tty->print_cr(" JVM state at this point:");
4202 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4203 ciMethod* m = jvms()->of_depth(i)->method();
4204 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4205 }
4206 }
4207 #endif
4208
4209 return false; // bail-out; let JVM_GetCallerClass do the work
4210 }
4211
4212 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4213 Node* arg = argument(0);
4214 Node* result;
4215
4216 switch (id) {
4217 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4218 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4219 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4220 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4221
4222 case vmIntrinsics::_doubleToLongBits: {
4223 // two paths (plus control) merge in a wood
4224 RegionNode *r = new (C) RegionNode(3);
4225 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4226
4227 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4228 // Build the boolean node
4229 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4230
4231 // Branch either way.
4232 // NaN case is less traveled, which makes all the difference.
4233 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4234 Node *opt_isnan = _gvn.transform(ifisnan);
4235 assert( opt_isnan->is_If(), "Expect an IfNode");
4236 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4237 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4238
4239 set_control(iftrue);
4240
4241 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4242 Node *slow_result = longcon(nan_bits); // return NaN
4243 phi->init_req(1, _gvn.transform( slow_result ));
4244 r->init_req(1, iftrue);
4245
4246 // Else fall through
4247 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4248 set_control(iffalse);
4249
4250 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4251 r->init_req(2, iffalse);
4252
4253 // Post merge
4254 set_control(_gvn.transform(r));
4255 record_for_igvn(r);
4256
4257 C->set_has_split_ifs(true); // Has chance for split-if optimization
4258 result = phi;
4259 assert(result->bottom_type()->isa_long(), "must be");
4260 break;
4261 }
4262
4263 case vmIntrinsics::_floatToIntBits: {
4264 // two paths (plus control) merge in a wood
4265 RegionNode *r = new (C) RegionNode(3);
4266 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4267
4268 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4269 // Build the boolean node
4270 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4271
4272 // Branch either way.
4273 // NaN case is less traveled, which makes all the difference.
4274 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4275 Node *opt_isnan = _gvn.transform(ifisnan);
4276 assert( opt_isnan->is_If(), "Expect an IfNode");
4277 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4278 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4279
4280 set_control(iftrue);
4281
4282 static const jint nan_bits = 0x7fc00000;
4283 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4284 phi->init_req(1, _gvn.transform( slow_result ));
4285 r->init_req(1, iftrue);
4286
4287 // Else fall through
4288 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4289 set_control(iffalse);
4290
4291 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4292 r->init_req(2, iffalse);
4293
4294 // Post merge
4295 set_control(_gvn.transform(r));
4296 record_for_igvn(r);
4297
4298 C->set_has_split_ifs(true); // Has chance for split-if optimization
4299 result = phi;
4300 assert(result->bottom_type()->isa_int(), "must be");
4301 break;
4302 }
4303
4304 default:
4305 fatal_unexpected_iid(id);
4306 break;
4307 }
4308 set_result(_gvn.transform(result));
4309 return true;
4310 }
4311
4312 #ifdef _LP64
4313 #define XTOP ,top() /*additional argument*/
4314 #else //_LP64
4315 #define XTOP /*no additional argument*/
4316 #endif //_LP64
4317
4318 //----------------------inline_unsafe_copyMemory-------------------------
4319 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4320 bool LibraryCallKit::inline_unsafe_copyMemory() {
4321 if (callee()->is_static()) return false; // caller must have the capability!
4322 null_check_receiver(); // null-check receiver
4323 if (stopped()) return true;
4324
4325 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4326
4327 Node* src_ptr = argument(1); // type: oop
4328 Node* src_off = ConvL2X(argument(2)); // type: long
4329 Node* dst_ptr = argument(4); // type: oop
4330 Node* dst_off = ConvL2X(argument(5)); // type: long
4331 Node* size = ConvL2X(argument(7)); // type: long
4332
4333 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4334 "fieldOffset must be byte-scaled");
4335
4336 Node* src = make_unsafe_address(src_ptr, src_off);
4337 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4338
4339 // Conservatively insert a memory barrier on all memory slices.
4340 // Do not let writes of the copy source or destination float below the copy.
4341 insert_mem_bar(Op_MemBarCPUOrder);
4342
4343 // Call it. Note that the length argument is not scaled.
4344 make_runtime_call(RC_LEAF|RC_NO_FP,
4345 OptoRuntime::fast_arraycopy_Type(),
4346 StubRoutines::unsafe_arraycopy(),
4347 "unsafe_arraycopy",
4348 TypeRawPtr::BOTTOM,
4349 src, dst, size XTOP);
4350
4351 // Do not let reads of the copy destination float above the copy.
4352 insert_mem_bar(Op_MemBarCPUOrder);
4353
4354 return true;
4355 }
4356
4357 //------------------------clone_coping-----------------------------------
4358 // Helper function for inline_native_clone.
4359 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4360 assert(obj_size != NULL, "");
4361 Node* raw_obj = alloc_obj->in(1);
4362 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4363
4364 AllocateNode* alloc = NULL;
4365 if (ReduceBulkZeroing) {
4366 // We will be completely responsible for initializing this object -
4367 // mark Initialize node as complete.
4368 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4369 // The object was just allocated - there should be no any stores!
4370 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4371 // Mark as complete_with_arraycopy so that on AllocateNode
4372 // expansion, we know this AllocateNode is initialized by an array
4373 // copy and a StoreStore barrier exists after the array copy.
4374 alloc->initialization()->set_complete_with_arraycopy();
4375 }
4376
4377 // Copy the fastest available way.
4378 // TODO: generate fields copies for small objects instead.
4379 Node* src = obj;
4380 Node* dest = alloc_obj;
4381 Node* size = _gvn.transform(obj_size);
4382
4383 // Exclude the header but include array length to copy by 8 bytes words.
4384 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4385 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4386 instanceOopDesc::base_offset_in_bytes();
4387 // base_off:
4388 // 8 - 32-bit VM
4389 // 12 - 64-bit VM, compressed klass
4390 // 16 - 64-bit VM, normal klass
4391 if (base_off % BytesPerLong != 0) {
4392 assert(UseCompressedClassPointers, "");
4393 if (is_array) {
4394 // Exclude length to copy by 8 bytes words.
4395 base_off += sizeof(int);
4396 } else {
4397 // Include klass to copy by 8 bytes words.
4398 base_off = instanceOopDesc::klass_offset_in_bytes();
4399 }
4400 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4401 }
4402 src = basic_plus_adr(src, base_off);
4403 dest = basic_plus_adr(dest, base_off);
4404
4405 // Compute the length also, if needed:
4406 Node* countx = size;
4407 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4408 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4409
4410 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4411 bool disjoint_bases = true;
4412 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4413 src, NULL, dest, NULL, countx,
4414 /*dest_uninitialized*/true);
4415
4416 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4417 if (card_mark) {
4418 assert(!is_array, "");
4419 // Put in store barrier for any and all oops we are sticking
4420 // into this object. (We could avoid this if we could prove
4421 // that the object type contains no oop fields at all.)
4422 Node* no_particular_value = NULL;
4423 Node* no_particular_field = NULL;
4424 int raw_adr_idx = Compile::AliasIdxRaw;
4425 post_barrier(control(),
4426 memory(raw_adr_type),
4427 alloc_obj,
4428 no_particular_field,
4429 raw_adr_idx,
4430 no_particular_value,
4431 T_OBJECT,
4432 false);
4433 }
4434
4435 // Do not let reads from the cloned object float above the arraycopy.
4436 if (alloc != NULL) {
4437 // Do not let stores that initialize this object be reordered with
4438 // a subsequent store that would make this object accessible by
4439 // other threads.
4440 // Record what AllocateNode this StoreStore protects so that
4441 // escape analysis can go from the MemBarStoreStoreNode to the
4442 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4443 // based on the escape status of the AllocateNode.
4444 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4445 } else {
4446 insert_mem_bar(Op_MemBarCPUOrder);
4447 }
4448 }
4449
4450 //------------------------inline_native_clone----------------------------
4451 // protected native Object java.lang.Object.clone();
4452 //
4453 // Here are the simple edge cases:
4454 // null receiver => normal trap
4455 // virtual and clone was overridden => slow path to out-of-line clone
4456 // not cloneable or finalizer => slow path to out-of-line Object.clone
4457 //
4458 // The general case has two steps, allocation and copying.
4459 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4460 //
4461 // Copying also has two cases, oop arrays and everything else.
4462 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4463 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4464 //
4465 // These steps fold up nicely if and when the cloned object's klass
4466 // can be sharply typed as an object array, a type array, or an instance.
4467 //
4468 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4469 PhiNode* result_val;
4470
4471 // Set the reexecute bit for the interpreter to reexecute
4472 // the bytecode that invokes Object.clone if deoptimization happens.
4473 { PreserveReexecuteState preexecs(this);
4474 jvms()->set_should_reexecute(true);
4475
4476 Node* obj = null_check_receiver();
4477 if (stopped()) return true;
4478
4479 Node* obj_klass = load_object_klass(obj);
4480 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4481 const TypeOopPtr* toop = ((tklass != NULL)
4482 ? tklass->as_instance_type()
4483 : TypeInstPtr::NOTNULL);
4484
4485 // Conservatively insert a memory barrier on all memory slices.
4486 // Do not let writes into the original float below the clone.
4487 insert_mem_bar(Op_MemBarCPUOrder);
4488
4489 // paths into result_reg:
4490 enum {
4491 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4492 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4493 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4494 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4495 PATH_LIMIT
4496 };
4497 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4498 result_val = new(C) PhiNode(result_reg,
4499 TypeInstPtr::NOTNULL);
4500 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4501 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4502 TypePtr::BOTTOM);
4503 record_for_igvn(result_reg);
4504
4505 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4506 int raw_adr_idx = Compile::AliasIdxRaw;
4507
4508 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4509 if (array_ctl != NULL) {
4510 // It's an array.
4511 PreserveJVMState pjvms(this);
4512 set_control(array_ctl);
4513 Node* obj_length = load_array_length(obj);
4514 Node* obj_size = NULL;
4515 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4516
4517 if (!use_ReduceInitialCardMarks()) {
4518 // If it is an oop array, it requires very special treatment,
4519 // because card marking is required on each card of the array.
4520 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4521 if (is_obja != NULL) {
4522 PreserveJVMState pjvms2(this);
4523 set_control(is_obja);
4524 // Generate a direct call to the right arraycopy function(s).
4525 bool disjoint_bases = true;
4526 bool length_never_negative = true;
4527 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4528 obj, intcon(0), alloc_obj, intcon(0),
4529 obj_length,
4530 disjoint_bases, length_never_negative);
4531 result_reg->init_req(_objArray_path, control());
4532 result_val->init_req(_objArray_path, alloc_obj);
4533 result_i_o ->set_req(_objArray_path, i_o());
4534 result_mem ->set_req(_objArray_path, reset_memory());
4535 }
4536 }
4537 // Otherwise, there are no card marks to worry about.
4538 // (We can dispense with card marks if we know the allocation
4539 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4540 // causes the non-eden paths to take compensating steps to
4541 // simulate a fresh allocation, so that no further
4542 // card marks are required in compiled code to initialize
4543 // the object.)
4544
4545 if (!stopped()) {
4546 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4547
4548 // Present the results of the copy.
4549 result_reg->init_req(_array_path, control());
4550 result_val->init_req(_array_path, alloc_obj);
4551 result_i_o ->set_req(_array_path, i_o());
4552 result_mem ->set_req(_array_path, reset_memory());
4553 }
4554 }
4555
4556 // We only go to the instance fast case code if we pass a number of guards.
4557 // The paths which do not pass are accumulated in the slow_region.
4558 RegionNode* slow_region = new (C) RegionNode(1);
4559 record_for_igvn(slow_region);
4560 if (!stopped()) {
4561 // It's an instance (we did array above). Make the slow-path tests.
4562 // If this is a virtual call, we generate a funny guard. We grab
4563 // the vtable entry corresponding to clone() from the target object.
4564 // If the target method which we are calling happens to be the
4565 // Object clone() method, we pass the guard. We do not need this
4566 // guard for non-virtual calls; the caller is known to be the native
4567 // Object clone().
4568 if (is_virtual) {
4569 generate_virtual_guard(obj_klass, slow_region);
4570 }
4571
4572 // The object must be cloneable and must not have a finalizer.
4573 // Both of these conditions may be checked in a single test.
4574 // We could optimize the cloneable test further, but we don't care.
4575 generate_access_flags_guard(obj_klass,
4576 // Test both conditions:
4577 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4578 // Must be cloneable but not finalizer:
4579 JVM_ACC_IS_CLONEABLE,
4580 slow_region);
4581 }
4582
4583 if (!stopped()) {
4584 // It's an instance, and it passed the slow-path tests.
4585 PreserveJVMState pjvms(this);
4586 Node* obj_size = NULL;
4587 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4588
4589 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4590
4591 // Present the results of the slow call.
4592 result_reg->init_req(_instance_path, control());
4593 result_val->init_req(_instance_path, alloc_obj);
4594 result_i_o ->set_req(_instance_path, i_o());
4595 result_mem ->set_req(_instance_path, reset_memory());
4596 }
4597
4598 // Generate code for the slow case. We make a call to clone().
4599 set_control(_gvn.transform(slow_region));
4600 if (!stopped()) {
4601 PreserveJVMState pjvms(this);
4602 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4603 Node* slow_result = set_results_for_java_call(slow_call);
4604 // this->control() comes from set_results_for_java_call
4605 result_reg->init_req(_slow_path, control());
4606 result_val->init_req(_slow_path, slow_result);
4607 result_i_o ->set_req(_slow_path, i_o());
4608 result_mem ->set_req(_slow_path, reset_memory());
4609 }
4610
4611 // Return the combined state.
4612 set_control( _gvn.transform(result_reg));
4613 set_i_o( _gvn.transform(result_i_o));
4614 set_all_memory( _gvn.transform(result_mem));
4615 } // original reexecute is set back here
4616
4617 set_result(_gvn.transform(result_val));
4618 return true;
4619 }
4620
4621 //------------------------------basictype2arraycopy----------------------------
4622 address LibraryCallKit::basictype2arraycopy(BasicType t,
4623 Node* src_offset,
4624 Node* dest_offset,
4625 bool disjoint_bases,
4626 const char* &name,
4627 bool dest_uninitialized) {
4628 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4629 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4630
4631 bool aligned = false;
4632 bool disjoint = disjoint_bases;
4633
4634 // if the offsets are the same, we can treat the memory regions as
4635 // disjoint, because either the memory regions are in different arrays,
4636 // or they are identical (which we can treat as disjoint.) We can also
4637 // treat a copy with a destination index less that the source index
4638 // as disjoint since a low->high copy will work correctly in this case.
4639 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4640 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4641 // both indices are constants
4642 int s_offs = src_offset_inttype->get_con();
4643 int d_offs = dest_offset_inttype->get_con();
4644 int element_size = type2aelembytes(t);
4645 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4646 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4647 if (s_offs >= d_offs) disjoint = true;
4648 } else if (src_offset == dest_offset && src_offset != NULL) {
4649 // This can occur if the offsets are identical non-constants.
4650 disjoint = true;
4651 }
4652
4653 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4654 }
4655
4656
4657 //------------------------------inline_arraycopy-----------------------
4658 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4659 // Object dest, int destPos,
4660 // int length);
4661 bool LibraryCallKit::inline_arraycopy() {
4662 // Get the arguments.
4663 Node* src = argument(0); // type: oop
4664 Node* src_offset = argument(1); // type: int
4665 Node* dest = argument(2); // type: oop
4666 Node* dest_offset = argument(3); // type: int
4667 Node* length = argument(4); // type: int
4668
4669 // Compile time checks. If any of these checks cannot be verified at compile time,
4670 // we do not make a fast path for this call. Instead, we let the call remain as it
4671 // is. The checks we choose to mandate at compile time are:
4672 //
4673 // (1) src and dest are arrays.
4674 const Type* src_type = src->Value(&_gvn);
4675 const Type* dest_type = dest->Value(&_gvn);
4676 const TypeAryPtr* top_src = src_type->isa_aryptr();
4677 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4678
4679 // Do we have the type of src?
4680 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4681 // Do we have the type of dest?
4682 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4683 // Is the type for src from speculation?
4684 bool src_spec = false;
4685 // Is the type for dest from speculation?
4686 bool dest_spec = false;
4687
4688 if (!has_src || !has_dest) {
4689 // We don't have sufficient type information, let's see if
4690 // speculative types can help. We need to have types for both src
4691 // and dest so that it pays off.
4692
4693 // Do we already have or could we have type information for src
4694 bool could_have_src = has_src;
4695 // Do we already have or could we have type information for dest
4696 bool could_have_dest = has_dest;
4697
4698 ciKlass* src_k = NULL;
4699 if (!has_src) {
4700 src_k = src_type->speculative_type_not_null();
4701 if (src_k != NULL && src_k->is_array_klass()) {
4702 could_have_src = true;
4703 }
4704 }
4705
4706 ciKlass* dest_k = NULL;
4707 if (!has_dest) {
4708 dest_k = dest_type->speculative_type_not_null();
4709 if (dest_k != NULL && dest_k->is_array_klass()) {
4710 could_have_dest = true;
4711 }
4712 }
4713
4714 if (could_have_src && could_have_dest) {
4715 // This is going to pay off so emit the required guards
4716 if (!has_src) {
4717 src = maybe_cast_profiled_obj(src, src_k);
4718 src_type = _gvn.type(src);
4719 top_src = src_type->isa_aryptr();
4720 has_src = (top_src != NULL && top_src->klass() != NULL);
4721 src_spec = true;
4722 }
4723 if (!has_dest) {
4724 dest = maybe_cast_profiled_obj(dest, dest_k);
4725 dest_type = _gvn.type(dest);
4726 top_dest = dest_type->isa_aryptr();
4727 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4728 dest_spec = true;
4729 }
4730 }
4731 }
4732
4733 if (!has_src || !has_dest) {
4734 // Conservatively insert a memory barrier on all memory slices.
4735 // Do not let writes into the source float below the arraycopy.
4736 insert_mem_bar(Op_MemBarCPUOrder);
4737
4738 // Call StubRoutines::generic_arraycopy stub.
4739 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4740 src, src_offset, dest, dest_offset, length);
4741
4742 // Do not let reads from the destination float above the arraycopy.
4743 // Since we cannot type the arrays, we don't know which slices
4744 // might be affected. We could restrict this barrier only to those
4745 // memory slices which pertain to array elements--but don't bother.
4746 if (!InsertMemBarAfterArraycopy)
4747 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4748 insert_mem_bar(Op_MemBarCPUOrder);
4749 return true;
4750 }
4751
4752 // (2) src and dest arrays must have elements of the same BasicType
4753 // Figure out the size and type of the elements we will be copying.
4754 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4755 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4756 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4757 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4758
4759 if (src_elem != dest_elem || dest_elem == T_VOID) {
4760 // The component types are not the same or are not recognized. Punt.
4761 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4762 generate_slow_arraycopy(TypePtr::BOTTOM,
4763 src, src_offset, dest, dest_offset, length,
4764 /*dest_uninitialized*/false);
4765 return true;
4766 }
4767
4768 if (src_elem == T_OBJECT) {
4769 // If both arrays are object arrays then having the exact types
4770 // for both will remove the need for a subtype check at runtime
4771 // before the call and may make it possible to pick a faster copy
4772 // routine (without a subtype check on every element)
4773 // Do we have the exact type of src?
4774 bool could_have_src = src_spec;
4775 // Do we have the exact type of dest?
4776 bool could_have_dest = dest_spec;
4777 ciKlass* src_k = top_src->klass();
4778 ciKlass* dest_k = top_dest->klass();
4779 if (!src_spec) {
4780 src_k = src_type->speculative_type_not_null();
4781 if (src_k != NULL && src_k->is_array_klass()) {
4782 could_have_src = true;
4783 }
4784 }
4785 if (!dest_spec) {
4786 dest_k = dest_type->speculative_type_not_null();
4787 if (dest_k != NULL && dest_k->is_array_klass()) {
4788 could_have_dest = true;
4789 }
4790 }
4791 if (could_have_src && could_have_dest) {
4792 // If we can have both exact types, emit the missing guards
4793 if (could_have_src && !src_spec) {
4794 src = maybe_cast_profiled_obj(src, src_k);
4795 }
4796 if (could_have_dest && !dest_spec) {
4797 dest = maybe_cast_profiled_obj(dest, dest_k);
4798 }
4799 }
4800 }
4801
4802 //---------------------------------------------------------------------------
4803 // We will make a fast path for this call to arraycopy.
4804
4805 // We have the following tests left to perform:
4806 //
4807 // (3) src and dest must not be null.
4808 // (4) src_offset must not be negative.
4809 // (5) dest_offset must not be negative.
4810 // (6) length must not be negative.
4811 // (7) src_offset + length must not exceed length of src.
4812 // (8) dest_offset + length must not exceed length of dest.
4813 // (9) each element of an oop array must be assignable
4814
4815 RegionNode* slow_region = new (C) RegionNode(1);
4816 record_for_igvn(slow_region);
4817
4818 // (3) operands must not be null
4819 // We currently perform our null checks with the null_check routine.
4820 // This means that the null exceptions will be reported in the caller
4821 // rather than (correctly) reported inside of the native arraycopy call.
4822 // This should be corrected, given time. We do our null check with the
4823 // stack pointer restored.
4824 src = null_check(src, T_ARRAY);
4825 dest = null_check(dest, T_ARRAY);
4826
4827 // (4) src_offset must not be negative.
4828 generate_negative_guard(src_offset, slow_region);
4829
4830 // (5) dest_offset must not be negative.
4831 generate_negative_guard(dest_offset, slow_region);
4832
4833 // (6) length must not be negative (moved to generate_arraycopy()).
4834 // generate_negative_guard(length, slow_region);
4835
4836 // (7) src_offset + length must not exceed length of src.
4837 generate_limit_guard(src_offset, length,
4838 load_array_length(src),
4839 slow_region);
4840
4841 // (8) dest_offset + length must not exceed length of dest.
4842 generate_limit_guard(dest_offset, length,
4843 load_array_length(dest),
4844 slow_region);
4845
4846 // (9) each element of an oop array must be assignable
4847 // The generate_arraycopy subroutine checks this.
4848
4849 // This is where the memory effects are placed:
4850 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4851 generate_arraycopy(adr_type, dest_elem,
4852 src, src_offset, dest, dest_offset, length,
4853 false, false, slow_region);
4854
4855 return true;
4856 }
4857
4858 //-----------------------------generate_arraycopy----------------------
4859 // Generate an optimized call to arraycopy.
4860 // Caller must guard against non-arrays.
4861 // Caller must determine a common array basic-type for both arrays.
4862 // Caller must validate offsets against array bounds.
4863 // The slow_region has already collected guard failure paths
4864 // (such as out of bounds length or non-conformable array types).
4865 // The generated code has this shape, in general:
4866 //
4867 // if (length == 0) return // via zero_path
4868 // slowval = -1
4869 // if (types unknown) {
4870 // slowval = call generic copy loop
4871 // if (slowval == 0) return // via checked_path
4872 // } else if (indexes in bounds) {
4873 // if ((is object array) && !(array type check)) {
4874 // slowval = call checked copy loop
4875 // if (slowval == 0) return // via checked_path
4876 // } else {
4877 // call bulk copy loop
4878 // return // via fast_path
4879 // }
4880 // }
4881 // // adjust params for remaining work:
4882 // if (slowval != -1) {
4883 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4884 // }
4885 // slow_region:
4886 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4887 // return // via slow_call_path
4888 //
4889 // This routine is used from several intrinsics: System.arraycopy,
4890 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4891 //
4892 void
4893 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4894 BasicType basic_elem_type,
4895 Node* src, Node* src_offset,
4896 Node* dest, Node* dest_offset,
4897 Node* copy_length,
4898 bool disjoint_bases,
4899 bool length_never_negative,
4900 RegionNode* slow_region) {
4901
4902 if (slow_region == NULL) {
4903 slow_region = new(C) RegionNode(1);
4904 record_for_igvn(slow_region);
4905 }
4906
4907 Node* original_dest = dest;
4908 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4909 bool dest_uninitialized = false;
4910
4911 // See if this is the initialization of a newly-allocated array.
4912 // If so, we will take responsibility here for initializing it to zero.
4913 // (Note: Because tightly_coupled_allocation performs checks on the
4914 // out-edges of the dest, we need to avoid making derived pointers
4915 // from it until we have checked its uses.)
4916 if (ReduceBulkZeroing
4917 && !ZeroTLAB // pointless if already zeroed
4918 && basic_elem_type != T_CONFLICT // avoid corner case
4919 && !src->eqv_uncast(dest)
4920 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4921 != NULL)
4922 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4923 && alloc->maybe_set_complete(&_gvn)) {
4924 // "You break it, you buy it."
4925 InitializeNode* init = alloc->initialization();
4926 assert(init->is_complete(), "we just did this");
4927 init->set_complete_with_arraycopy();
4928 assert(dest->is_CheckCastPP(), "sanity");
4929 assert(dest->in(0)->in(0) == init, "dest pinned");
4930 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4931 // From this point on, every exit path is responsible for
4932 // initializing any non-copied parts of the object to zero.
4933 // Also, if this flag is set we make sure that arraycopy interacts properly
4934 // with G1, eliding pre-barriers. See CR 6627983.
4935 dest_uninitialized = true;
4936 } else {
4937 // No zeroing elimination here.
4938 alloc = NULL;
4939 //original_dest = dest;
4940 //dest_uninitialized = false;
4941 }
4942
4943 // Results are placed here:
4944 enum { fast_path = 1, // normal void-returning assembly stub
4945 checked_path = 2, // special assembly stub with cleanup
4946 slow_call_path = 3, // something went wrong; call the VM
4947 zero_path = 4, // bypass when length of copy is zero
4948 bcopy_path = 5, // copy primitive array by 64-bit blocks
4949 PATH_LIMIT = 6
4950 };
4951 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4952 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4953 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4954 record_for_igvn(result_region);
4955 _gvn.set_type_bottom(result_i_o);
4956 _gvn.set_type_bottom(result_memory);
4957 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4958
4959 // The slow_control path:
4960 Node* slow_control;
4961 Node* slow_i_o = i_o();
4962 Node* slow_mem = memory(adr_type);
4963 debug_only(slow_control = (Node*) badAddress);
4964
4965 // Checked control path:
4966 Node* checked_control = top();
4967 Node* checked_mem = NULL;
4968 Node* checked_i_o = NULL;
4969 Node* checked_value = NULL;
4970
4971 if (basic_elem_type == T_CONFLICT) {
4972 assert(!dest_uninitialized, "");
4973 Node* cv = generate_generic_arraycopy(adr_type,
4974 src, src_offset, dest, dest_offset,
4975 copy_length, dest_uninitialized);
4976 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4977 checked_control = control();
4978 checked_i_o = i_o();
4979 checked_mem = memory(adr_type);
4980 checked_value = cv;
4981 set_control(top()); // no fast path
4982 }
4983
4984 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4985 if (not_pos != NULL) {
4986 PreserveJVMState pjvms(this);
4987 set_control(not_pos);
4988
4989 // (6) length must not be negative.
4990 if (!length_never_negative) {
4991 generate_negative_guard(copy_length, slow_region);
4992 }
4993
4994 // copy_length is 0.
4995 if (!stopped() && dest_uninitialized) {
4996 Node* dest_length = alloc->in(AllocateNode::ALength);
4997 if (copy_length->eqv_uncast(dest_length)
4998 || _gvn.find_int_con(dest_length, 1) <= 0) {
4999 // There is no zeroing to do. No need for a secondary raw memory barrier.
5000 } else {
5001 // Clear the whole thing since there are no source elements to copy.
5002 generate_clear_array(adr_type, dest, basic_elem_type,
5003 intcon(0), NULL,
5004 alloc->in(AllocateNode::AllocSize));
5005 // Use a secondary InitializeNode as raw memory barrier.
5006 // Currently it is needed only on this path since other
5007 // paths have stub or runtime calls as raw memory barriers.
5008 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5009 Compile::AliasIdxRaw,
5010 top())->as_Initialize();
5011 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5012 }
5013 }
5014
5015 // Present the results of the fast call.
5016 result_region->init_req(zero_path, control());
5017 result_i_o ->init_req(zero_path, i_o());
5018 result_memory->init_req(zero_path, memory(adr_type));
5019 }
5020
5021 if (!stopped() && dest_uninitialized) {
5022 // We have to initialize the *uncopied* part of the array to zero.
5023 // The copy destination is the slice dest[off..off+len]. The other slices
5024 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5025 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5026 Node* dest_length = alloc->in(AllocateNode::ALength);
5027 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5028 copy_length));
5029
5030 // If there is a head section that needs zeroing, do it now.
5031 if (find_int_con(dest_offset, -1) != 0) {
5032 generate_clear_array(adr_type, dest, basic_elem_type,
5033 intcon(0), dest_offset,
5034 NULL);
5035 }
5036
5037 // Next, perform a dynamic check on the tail length.
5038 // It is often zero, and we can win big if we prove this.
5039 // There are two wins: Avoid generating the ClearArray
5040 // with its attendant messy index arithmetic, and upgrade
5041 // the copy to a more hardware-friendly word size of 64 bits.
5042 Node* tail_ctl = NULL;
5043 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5044 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5045 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5046 tail_ctl = generate_slow_guard(bol_lt, NULL);
5047 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5048 }
5049
5050 // At this point, let's assume there is no tail.
5051 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5052 // There is no tail. Try an upgrade to a 64-bit copy.
5053 bool didit = false;
5054 { PreserveJVMState pjvms(this);
5055 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5056 src, src_offset, dest, dest_offset,
5057 dest_size, dest_uninitialized);
5058 if (didit) {
5059 // Present the results of the block-copying fast call.
5060 result_region->init_req(bcopy_path, control());
5061 result_i_o ->init_req(bcopy_path, i_o());
5062 result_memory->init_req(bcopy_path, memory(adr_type));
5063 }
5064 }
5065 if (didit)
5066 set_control(top()); // no regular fast path
5067 }
5068
5069 // Clear the tail, if any.
5070 if (tail_ctl != NULL) {
5071 Node* notail_ctl = stopped() ? NULL : control();
5072 set_control(tail_ctl);
5073 if (notail_ctl == NULL) {
5074 generate_clear_array(adr_type, dest, basic_elem_type,
5075 dest_tail, NULL,
5076 dest_size);
5077 } else {
5078 // Make a local merge.
5079 Node* done_ctl = new(C) RegionNode(3);
5080 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5081 done_ctl->init_req(1, notail_ctl);
5082 done_mem->init_req(1, memory(adr_type));
5083 generate_clear_array(adr_type, dest, basic_elem_type,
5084 dest_tail, NULL,
5085 dest_size);
5086 done_ctl->init_req(2, control());
5087 done_mem->init_req(2, memory(adr_type));
5088 set_control( _gvn.transform(done_ctl));
5089 set_memory( _gvn.transform(done_mem), adr_type );
5090 }
5091 }
5092 }
5093
5094 BasicType copy_type = basic_elem_type;
5095 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5096 if (!stopped() && copy_type == T_OBJECT) {
5097 // If src and dest have compatible element types, we can copy bits.
5098 // Types S[] and D[] are compatible if D is a supertype of S.
5099 //
5100 // If they are not, we will use checked_oop_disjoint_arraycopy,
5101 // which performs a fast optimistic per-oop check, and backs off
5102 // further to JVM_ArrayCopy on the first per-oop check that fails.
5103 // (Actually, we don't move raw bits only; the GC requires card marks.)
5104
5105 // Get the Klass* for both src and dest
5106 Node* src_klass = load_object_klass(src);
5107 Node* dest_klass = load_object_klass(dest);
5108
5109 // Generate the subtype check.
5110 // This might fold up statically, or then again it might not.
5111 //
5112 // Non-static example: Copying List<String>.elements to a new String[].
5113 // The backing store for a List<String> is always an Object[],
5114 // but its elements are always type String, if the generic types
5115 // are correct at the source level.
5116 //
5117 // Test S[] against D[], not S against D, because (probably)
5118 // the secondary supertype cache is less busy for S[] than S.
5119 // This usually only matters when D is an interface.
5120 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5121 // Plug failing path into checked_oop_disjoint_arraycopy
5122 if (not_subtype_ctrl != top()) {
5123 PreserveJVMState pjvms(this);
5124 set_control(not_subtype_ctrl);
5125 // (At this point we can assume disjoint_bases, since types differ.)
5126 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5127 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5128 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5129 Node* dest_elem_klass = _gvn.transform(n1);
5130 Node* cv = generate_checkcast_arraycopy(adr_type,
5131 dest_elem_klass,
5132 src, src_offset, dest, dest_offset,
5133 ConvI2X(copy_length), dest_uninitialized);
5134 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5135 checked_control = control();
5136 checked_i_o = i_o();
5137 checked_mem = memory(adr_type);
5138 checked_value = cv;
5139 }
5140 // At this point we know we do not need type checks on oop stores.
5141
5142 // Let's see if we need card marks:
5143 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5144 // If we do not need card marks, copy using the jint or jlong stub.
5145 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5146 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5147 "sizes agree");
5148 }
5149 }
5150
5151 if (!stopped()) {
5152 // Generate the fast path, if possible.
5153 PreserveJVMState pjvms(this);
5154 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5155 src, src_offset, dest, dest_offset,
5156 ConvI2X(copy_length), dest_uninitialized);
5157
5158 // Present the results of the fast call.
5159 result_region->init_req(fast_path, control());
5160 result_i_o ->init_req(fast_path, i_o());
5161 result_memory->init_req(fast_path, memory(adr_type));
5162 }
5163
5164 // Here are all the slow paths up to this point, in one bundle:
5165 slow_control = top();
5166 if (slow_region != NULL)
5167 slow_control = _gvn.transform(slow_region);
5168 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5169
5170 set_control(checked_control);
5171 if (!stopped()) {
5172 // Clean up after the checked call.
5173 // The returned value is either 0 or -1^K,
5174 // where K = number of partially transferred array elements.
5175 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5176 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5177 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5178
5179 // If it is 0, we are done, so transfer to the end.
5180 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5181 result_region->init_req(checked_path, checks_done);
5182 result_i_o ->init_req(checked_path, checked_i_o);
5183 result_memory->init_req(checked_path, checked_mem);
5184
5185 // If it is not zero, merge into the slow call.
5186 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5187 RegionNode* slow_reg2 = new(C) RegionNode(3);
5188 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5189 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5190 record_for_igvn(slow_reg2);
5191 slow_reg2 ->init_req(1, slow_control);
5192 slow_i_o2 ->init_req(1, slow_i_o);
5193 slow_mem2 ->init_req(1, slow_mem);
5194 slow_reg2 ->init_req(2, control());
5195 slow_i_o2 ->init_req(2, checked_i_o);
5196 slow_mem2 ->init_req(2, checked_mem);
5197
5198 slow_control = _gvn.transform(slow_reg2);
5199 slow_i_o = _gvn.transform(slow_i_o2);
5200 slow_mem = _gvn.transform(slow_mem2);
5201
5202 if (alloc != NULL) {
5203 // We'll restart from the very beginning, after zeroing the whole thing.
5204 // This can cause double writes, but that's OK since dest is brand new.
5205 // So we ignore the low 31 bits of the value returned from the stub.
5206 } else {
5207 // We must continue the copy exactly where it failed, or else
5208 // another thread might see the wrong number of writes to dest.
5209 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5210 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5211 slow_offset->init_req(1, intcon(0));
5212 slow_offset->init_req(2, checked_offset);
5213 slow_offset = _gvn.transform(slow_offset);
5214
5215 // Adjust the arguments by the conditionally incoming offset.
5216 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5217 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5218 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5219
5220 // Tweak the node variables to adjust the code produced below:
5221 src_offset = src_off_plus;
5222 dest_offset = dest_off_plus;
5223 copy_length = length_minus;
5224 }
5225 }
5226
5227 set_control(slow_control);
5228 if (!stopped()) {
5229 // Generate the slow path, if needed.
5230 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5231
5232 set_memory(slow_mem, adr_type);
5233 set_i_o(slow_i_o);
5234
5235 if (dest_uninitialized) {
5236 generate_clear_array(adr_type, dest, basic_elem_type,
5237 intcon(0), NULL,
5238 alloc->in(AllocateNode::AllocSize));
5239 }
5240
5241 generate_slow_arraycopy(adr_type,
5242 src, src_offset, dest, dest_offset,
5243 copy_length, /*dest_uninitialized*/false);
5244
5245 result_region->init_req(slow_call_path, control());
5246 result_i_o ->init_req(slow_call_path, i_o());
5247 result_memory->init_req(slow_call_path, memory(adr_type));
5248 }
5249
5250 // Remove unused edges.
5251 for (uint i = 1; i < result_region->req(); i++) {
5252 if (result_region->in(i) == NULL)
5253 result_region->init_req(i, top());
5254 }
5255
5256 // Finished; return the combined state.
5257 set_control( _gvn.transform(result_region));
5258 set_i_o( _gvn.transform(result_i_o) );
5259 set_memory( _gvn.transform(result_memory), adr_type );
5260
5261 // The memory edges above are precise in order to model effects around
5262 // array copies accurately to allow value numbering of field loads around
5263 // arraycopy. Such field loads, both before and after, are common in Java
5264 // collections and similar classes involving header/array data structures.
5265 //
5266 // But with low number of register or when some registers are used or killed
5267 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5268 // The next memory barrier is added to avoid it. If the arraycopy can be
5269 // optimized away (which it can, sometimes) then we can manually remove
5270 // the membar also.
5271 //
5272 // Do not let reads from the cloned object float above the arraycopy.
5273 if (alloc != NULL) {
5274 // Do not let stores that initialize this object be reordered with
5275 // a subsequent store that would make this object accessible by
5276 // other threads.
5277 // Record what AllocateNode this StoreStore protects so that
5278 // escape analysis can go from the MemBarStoreStoreNode to the
5279 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5280 // based on the escape status of the AllocateNode.
5281 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5282 } else if (InsertMemBarAfterArraycopy)
5283 insert_mem_bar(Op_MemBarCPUOrder);
5284 }
5285
5286
5287 // Helper function which determines if an arraycopy immediately follows
5288 // an allocation, with no intervening tests or other escapes for the object.
5289 AllocateArrayNode*
5290 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5291 RegionNode* slow_region) {
5292 if (stopped()) return NULL; // no fast path
5293 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5294
5295 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5296 if (alloc == NULL) return NULL;
5297
5298 Node* rawmem = memory(Compile::AliasIdxRaw);
5299 // Is the allocation's memory state untouched?
5300 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5301 // Bail out if there have been raw-memory effects since the allocation.
5302 // (Example: There might have been a call or safepoint.)
5303 return NULL;
5304 }
5305 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5306 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5307 return NULL;
5308 }
5309
5310 // There must be no unexpected observers of this allocation.
5311 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5312 Node* obs = ptr->fast_out(i);
5313 if (obs != this->map()) {
5314 return NULL;
5315 }
5316 }
5317
5318 // This arraycopy must unconditionally follow the allocation of the ptr.
5319 Node* alloc_ctl = ptr->in(0);
5320 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5321
5322 Node* ctl = control();
5323 while (ctl != alloc_ctl) {
5324 // There may be guards which feed into the slow_region.
5325 // Any other control flow means that we might not get a chance
5326 // to finish initializing the allocated object.
5327 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5328 IfNode* iff = ctl->in(0)->as_If();
5329 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5330 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5331 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5332 ctl = iff->in(0); // This test feeds the known slow_region.
5333 continue;
5334 }
5335 // One more try: Various low-level checks bottom out in
5336 // uncommon traps. If the debug-info of the trap omits
5337 // any reference to the allocation, as we've already
5338 // observed, then there can be no objection to the trap.
5339 bool found_trap = false;
5340 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5341 Node* obs = not_ctl->fast_out(j);
5342 if (obs->in(0) == not_ctl && obs->is_Call() &&
5343 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5344 found_trap = true; break;
5345 }
5346 }
5347 if (found_trap) {
5348 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5349 continue;
5350 }
5351 }
5352 return NULL;
5353 }
5354
5355 // If we get this far, we have an allocation which immediately
5356 // precedes the arraycopy, and we can take over zeroing the new object.
5357 // The arraycopy will finish the initialization, and provide
5358 // a new control state to which we will anchor the destination pointer.
5359
5360 return alloc;
5361 }
5362
5363 // Helper for initialization of arrays, creating a ClearArray.
5364 // It writes zero bits in [start..end), within the body of an array object.
5365 // The memory effects are all chained onto the 'adr_type' alias category.
5366 //
5367 // Since the object is otherwise uninitialized, we are free
5368 // to put a little "slop" around the edges of the cleared area,
5369 // as long as it does not go back into the array's header,
5370 // or beyond the array end within the heap.
5371 //
5372 // The lower edge can be rounded down to the nearest jint and the
5373 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5374 //
5375 // Arguments:
5376 // adr_type memory slice where writes are generated
5377 // dest oop of the destination array
5378 // basic_elem_type element type of the destination
5379 // slice_idx array index of first element to store
5380 // slice_len number of elements to store (or NULL)
5381 // dest_size total size in bytes of the array object
5382 //
5383 // Exactly one of slice_len or dest_size must be non-NULL.
5384 // If dest_size is non-NULL, zeroing extends to the end of the object.
5385 // If slice_len is non-NULL, the slice_idx value must be a constant.
5386 void
5387 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5388 Node* dest,
5389 BasicType basic_elem_type,
5390 Node* slice_idx,
5391 Node* slice_len,
5392 Node* dest_size) {
5393 // one or the other but not both of slice_len and dest_size:
5394 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5395 if (slice_len == NULL) slice_len = top();
5396 if (dest_size == NULL) dest_size = top();
5397
5398 // operate on this memory slice:
5399 Node* mem = memory(adr_type); // memory slice to operate on
5400
5401 // scaling and rounding of indexes:
5402 int scale = exact_log2(type2aelembytes(basic_elem_type));
5403 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5404 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5405 int bump_bit = (-1 << scale) & BytesPerInt;
5406
5407 // determine constant starts and ends
5408 const intptr_t BIG_NEG = -128;
5409 assert(BIG_NEG + 2*abase < 0, "neg enough");
5410 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5411 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5412 if (slice_len_con == 0) {
5413 return; // nothing to do here
5414 }
5415 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5416 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5417 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5418 assert(end_con < 0, "not two cons");
5419 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5420 BytesPerLong);
5421 }
5422
5423 if (start_con >= 0 && end_con >= 0) {
5424 // Constant start and end. Simple.
5425 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5426 start_con, end_con, &_gvn);
5427 } else if (start_con >= 0 && dest_size != top()) {
5428 // Constant start, pre-rounded end after the tail of the array.
5429 Node* end = dest_size;
5430 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5431 start_con, end, &_gvn);
5432 } else if (start_con >= 0 && slice_len != top()) {
5433 // Constant start, non-constant end. End needs rounding up.
5434 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5435 intptr_t end_base = abase + (slice_idx_con << scale);
5436 int end_round = (-1 << scale) & (BytesPerLong - 1);
5437 Node* end = ConvI2X(slice_len);
5438 if (scale != 0)
5439 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5440 end_base += end_round;
5441 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5442 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5443 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5444 start_con, end, &_gvn);
5445 } else if (start_con < 0 && dest_size != top()) {
5446 // Non-constant start, pre-rounded end after the tail of the array.
5447 // This is almost certainly a "round-to-end" operation.
5448 Node* start = slice_idx;
5449 start = ConvI2X(start);
5450 if (scale != 0)
5451 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5452 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5453 if ((bump_bit | clear_low) != 0) {
5454 int to_clear = (bump_bit | clear_low);
5455 // Align up mod 8, then store a jint zero unconditionally
5456 // just before the mod-8 boundary.
5457 if (((abase + bump_bit) & ~to_clear) - bump_bit
5458 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5459 bump_bit = 0;
5460 assert((abase & to_clear) == 0, "array base must be long-aligned");
5461 } else {
5462 // Bump 'start' up to (or past) the next jint boundary:
5463 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5464 assert((abase & clear_low) == 0, "array base must be int-aligned");
5465 }
5466 // Round bumped 'start' down to jlong boundary in body of array.
5467 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5468 if (bump_bit != 0) {
5469 // Store a zero to the immediately preceding jint:
5470 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5471 Node* p1 = basic_plus_adr(dest, x1);
5472 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5473 mem = _gvn.transform(mem);
5474 }
5475 }
5476 Node* end = dest_size; // pre-rounded
5477 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5478 start, end, &_gvn);
5479 } else {
5480 // Non-constant start, unrounded non-constant end.
5481 // (Nobody zeroes a random midsection of an array using this routine.)
5482 ShouldNotReachHere(); // fix caller
5483 }
5484
5485 // Done.
5486 set_memory(mem, adr_type);
5487 }
5488
5489
5490 bool
5491 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5492 BasicType basic_elem_type,
5493 AllocateNode* alloc,
5494 Node* src, Node* src_offset,
5495 Node* dest, Node* dest_offset,
5496 Node* dest_size, bool dest_uninitialized) {
5497 // See if there is an advantage from block transfer.
5498 int scale = exact_log2(type2aelembytes(basic_elem_type));
5499 if (scale >= LogBytesPerLong)
5500 return false; // it is already a block transfer
5501
5502 // Look at the alignment of the starting offsets.
5503 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5504
5505 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5506 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5507 if (src_off_con < 0 || dest_off_con < 0)
5508 // At present, we can only understand constants.
5509 return false;
5510
5511 intptr_t src_off = abase + (src_off_con << scale);
5512 intptr_t dest_off = abase + (dest_off_con << scale);
5513
5514 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5515 // Non-aligned; too bad.
5516 // One more chance: Pick off an initial 32-bit word.
5517 // This is a common case, since abase can be odd mod 8.
5518 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5519 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5520 Node* sptr = basic_plus_adr(src, src_off);
5521 Node* dptr = basic_plus_adr(dest, dest_off);
5522 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5523 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5524 src_off += BytesPerInt;
5525 dest_off += BytesPerInt;
5526 } else {
5527 return false;
5528 }
5529 }
5530 assert(src_off % BytesPerLong == 0, "");
5531 assert(dest_off % BytesPerLong == 0, "");
5532
5533 // Do this copy by giant steps.
5534 Node* sptr = basic_plus_adr(src, src_off);
5535 Node* dptr = basic_plus_adr(dest, dest_off);
5536 Node* countx = dest_size;
5537 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5538 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5539
5540 bool disjoint_bases = true; // since alloc != NULL
5541 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5542 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5543
5544 return true;
5545 }
5546
5547
5548 // Helper function; generates code for the slow case.
5549 // We make a call to a runtime method which emulates the native method,
5550 // but without the native wrapper overhead.
5551 void
5552 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5553 Node* src, Node* src_offset,
5554 Node* dest, Node* dest_offset,
5555 Node* copy_length, bool dest_uninitialized) {
5556 assert(!dest_uninitialized, "Invariant");
5557 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5558 OptoRuntime::slow_arraycopy_Type(),
5559 OptoRuntime::slow_arraycopy_Java(),
5560 "slow_arraycopy", adr_type,
5561 src, src_offset, dest, dest_offset,
5562 copy_length);
5563
5564 // Handle exceptions thrown by this fellow:
5565 make_slow_call_ex(call, env()->Throwable_klass(), false);
5566 }
5567
5568 // Helper function; generates code for cases requiring runtime checks.
5569 Node*
5570 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5571 Node* dest_elem_klass,
5572 Node* src, Node* src_offset,
5573 Node* dest, Node* dest_offset,
5574 Node* copy_length, bool dest_uninitialized) {
5575 if (stopped()) return NULL;
5576
5577 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5578 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5579 return NULL;
5580 }
5581
5582 // Pick out the parameters required to perform a store-check
5583 // for the target array. This is an optimistic check. It will
5584 // look in each non-null element's class, at the desired klass's
5585 // super_check_offset, for the desired klass.
5586 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5587 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5588 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5589 Node* check_offset = ConvI2X(_gvn.transform(n3));
5590 Node* check_value = dest_elem_klass;
5591
5592 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5593 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5594
5595 // (We know the arrays are never conjoint, because their types differ.)
5596 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5597 OptoRuntime::checkcast_arraycopy_Type(),
5598 copyfunc_addr, "checkcast_arraycopy", adr_type,
5599 // five arguments, of which two are
5600 // intptr_t (jlong in LP64)
5601 src_start, dest_start,
5602 copy_length XTOP,
5603 check_offset XTOP,
5604 check_value);
5605
5606 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5607 }
5608
5609
5610 // Helper function; generates code for cases requiring runtime checks.
5611 Node*
5612 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5613 Node* src, Node* src_offset,
5614 Node* dest, Node* dest_offset,
5615 Node* copy_length, bool dest_uninitialized) {
5616 assert(!dest_uninitialized, "Invariant");
5617 if (stopped()) return NULL;
5618 address copyfunc_addr = StubRoutines::generic_arraycopy();
5619 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5620 return NULL;
5621 }
5622
5623 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5624 OptoRuntime::generic_arraycopy_Type(),
5625 copyfunc_addr, "generic_arraycopy", adr_type,
5626 src, src_offset, dest, dest_offset, copy_length);
5627
5628 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5629 }
5630
5631 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5632 void
5633 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5634 BasicType basic_elem_type,
5635 bool disjoint_bases,
5636 Node* src, Node* src_offset,
5637 Node* dest, Node* dest_offset,
5638 Node* copy_length, bool dest_uninitialized) {
5639 if (stopped()) return; // nothing to do
5640
5641 Node* src_start = src;
5642 Node* dest_start = dest;
5643 if (src_offset != NULL || dest_offset != NULL) {
5644 assert(src_offset != NULL && dest_offset != NULL, "");
5645 src_start = array_element_address(src, src_offset, basic_elem_type);
5646 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5647 }
5648
5649 // Figure out which arraycopy runtime method to call.
5650 const char* copyfunc_name = "arraycopy";
5651 address copyfunc_addr =
5652 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5653 disjoint_bases, copyfunc_name, dest_uninitialized);
5654
5655 // Call it. Note that the count_ix value is not scaled to a byte-size.
5656 make_runtime_call(RC_LEAF|RC_NO_FP,
5657 OptoRuntime::fast_arraycopy_Type(),
5658 copyfunc_addr, copyfunc_name, adr_type,
5659 src_start, dest_start, copy_length XTOP);
5660 }
5661
5662 //-------------inline_encodeISOArray-----------------------------------
5663 // encode char[] to byte[] in ISO_8859_1
5664 bool LibraryCallKit::inline_encodeISOArray() {
5665 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5666 // no receiver since it is static method
5667 Node *src = argument(0);
5668 Node *src_offset = argument(1);
5669 Node *dst = argument(2);
5670 Node *dst_offset = argument(3);
5671 Node *length = argument(4);
5672
5673 const Type* src_type = src->Value(&_gvn);
5674 const Type* dst_type = dst->Value(&_gvn);
5675 const TypeAryPtr* top_src = src_type->isa_aryptr();
5676 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5677 if (top_src == NULL || top_src->klass() == NULL ||
5678 top_dest == NULL || top_dest->klass() == NULL) {
5679 // failed array check
5680 return false;
5681 }
5682
5683 // Figure out the size and type of the elements we will be copying.
5684 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5685 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5686 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5687 return false;
5688 }
5689 Node* src_start = array_element_address(src, src_offset, src_elem);
5690 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5691 // 'src_start' points to src array + scaled offset
5692 // 'dst_start' points to dst array + scaled offset
5693
5694 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5695 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5696 enc = _gvn.transform(enc);
5697 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5698 set_memory(res_mem, mtype);
5699 set_result(enc);
5700 return true;
5701 }
5702
5703 /**
5704 * Calculate CRC32 for byte.
5705 * int java.util.zip.CRC32.update(int crc, int b)
5706 */
5707 bool LibraryCallKit::inline_updateCRC32() {
5708 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5709 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5710 // no receiver since it is static method
5711 Node* crc = argument(0); // type: int
5712 Node* b = argument(1); // type: int
5713
5714 /*
5715 * int c = ~ crc;
5716 * b = timesXtoThe32[(b ^ c) & 0xFF];
5717 * b = b ^ (c >>> 8);
5718 * crc = ~b;
5719 */
5720
5721 Node* M1 = intcon(-1);
5722 crc = _gvn.transform(new (C) XorINode(crc, M1));
5723 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5724 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5725
5726 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5727 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5728 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5729 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5730
5731 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5732 result = _gvn.transform(new (C) XorINode(crc, result));
5733 result = _gvn.transform(new (C) XorINode(result, M1));
5734 set_result(result);
5735 return true;
5736 }
5737
5738 /**
5739 * Calculate CRC32 for byte[] array.
5740 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5741 */
5742 bool LibraryCallKit::inline_updateBytesCRC32() {
5743 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5744 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5745 // no receiver since it is static method
5746 Node* crc = argument(0); // type: int
5747 Node* src = argument(1); // type: oop
5748 Node* offset = argument(2); // type: int
5749 Node* length = argument(3); // type: int
5750
5751 const Type* src_type = src->Value(&_gvn);
5752 const TypeAryPtr* top_src = src_type->isa_aryptr();
5753 if (top_src == NULL || top_src->klass() == NULL) {
5754 // failed array check
5755 return false;
5756 }
5757
5758 // Figure out the size and type of the elements we will be copying.
5759 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5760 if (src_elem != T_BYTE) {
5761 return false;
5762 }
5763
5764 // 'src_start' points to src array + scaled offset
5765 Node* src_start = array_element_address(src, offset, src_elem);
5766
5767 // We assume that range check is done by caller.
5768 // TODO: generate range check (offset+length < src.length) in debug VM.
5769
5770 // Call the stub.
5771 address stubAddr = StubRoutines::updateBytesCRC32();
5772 const char *stubName = "updateBytesCRC32";
5773
5774 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5775 stubAddr, stubName, TypePtr::BOTTOM,
5776 crc, src_start, length);
5777 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5778 set_result(result);
5779 return true;
5780 }
5781
5782 /**
5783 * Calculate CRC32 for ByteBuffer.
5784 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5785 */
5786 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5787 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5788 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5789 // no receiver since it is static method
5790 Node* crc = argument(0); // type: int
5791 Node* src = argument(1); // type: long
5792 Node* offset = argument(3); // type: int
5793 Node* length = argument(4); // type: int
5794
5795 src = ConvL2X(src); // adjust Java long to machine word
5796 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5797 offset = ConvI2X(offset);
5798
5799 // 'src_start' points to src array + scaled offset
5800 Node* src_start = basic_plus_adr(top(), base, offset);
5801
5802 // Call the stub.
5803 address stubAddr = StubRoutines::updateBytesCRC32();
5804 const char *stubName = "updateBytesCRC32";
5805
5806 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5807 stubAddr, stubName, TypePtr::BOTTOM,
5808 crc, src_start, length);
5809 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5810 set_result(result);
5811 return true;
5812 }
5813
5814 //----------------------------inline_reference_get----------------------------
5815 // public T java.lang.ref.Reference.get();
5816 bool LibraryCallKit::inline_reference_get() {
5817 const int referent_offset = java_lang_ref_Reference::referent_offset;
5818 guarantee(referent_offset > 0, "should have already been set");
5819
5820 // Get the argument:
5821 Node* reference_obj = null_check_receiver();
5822 if (stopped()) return true;
5823
5824 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5825
5826 ciInstanceKlass* klass = env()->Object_klass();
5827 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5828
5829 Node* no_ctrl = NULL;
5830 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5831
5832 // Use the pre-barrier to record the value in the referent field
5833 pre_barrier(false /* do_load */,
5834 control(),
5835 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5836 result /* pre_val */,
5837 T_OBJECT);
5838
5839 // Add memory barrier to prevent commoning reads from this field
5840 // across safepoint since GC can change its value.
5841 insert_mem_bar(Op_MemBarCPUOrder);
5842
5843 set_result(result);
5844 return true;
5845 }
5846
5847
5848 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5849 bool is_exact=true, bool is_static=false) {
5850
5851 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5852 assert(tinst != NULL, "obj is null");
5853 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5854 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5855
5856 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5857 ciSymbol::make(fieldTypeString),
5858 is_static);
5859 if (field == NULL) return (Node *) NULL;
5860 assert (field != NULL, "undefined field");
5861
5862 // Next code copied from Parse::do_get_xxx():
5863
5864 // Compute address and memory type.
5865 int offset = field->offset_in_bytes();
5866 bool is_vol = field->is_volatile();
5867 ciType* field_klass = field->type();
5868 assert(field_klass->is_loaded(), "should be loaded");
5869 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5870 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5871 BasicType bt = field->layout_type();
5872
5873 // Build the resultant type of the load
5874 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5875
5876 // Build the load.
5877 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5878 return loadedField;
5879 }
5880
5881
5882 //------------------------------inline_aescrypt_Block-----------------------
5883 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5884 address stubAddr;
5885 const char *stubName;
5886 assert(UseAES, "need AES instruction support");
5887
5888 switch(id) {
5889 case vmIntrinsics::_aescrypt_encryptBlock:
5890 stubAddr = StubRoutines::aescrypt_encryptBlock();
5891 stubName = "aescrypt_encryptBlock";
5892 break;
5893 case vmIntrinsics::_aescrypt_decryptBlock:
5894 stubAddr = StubRoutines::aescrypt_decryptBlock();
5895 stubName = "aescrypt_decryptBlock";
5896 break;
5897 }
5898 if (stubAddr == NULL) return false;
5899
5900 Node* aescrypt_object = argument(0);
5901 Node* src = argument(1);
5902 Node* src_offset = argument(2);
5903 Node* dest = argument(3);
5904 Node* dest_offset = argument(4);
5905
5906 // (1) src and dest are arrays.
5907 const Type* src_type = src->Value(&_gvn);
5908 const Type* dest_type = dest->Value(&_gvn);
5909 const TypeAryPtr* top_src = src_type->isa_aryptr();
5910 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5911 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5912
5913 // for the quick and dirty code we will skip all the checks.
5914 // we are just trying to get the call to be generated.
5915 Node* src_start = src;
5916 Node* dest_start = dest;
5917 if (src_offset != NULL || dest_offset != NULL) {
5918 assert(src_offset != NULL && dest_offset != NULL, "");
5919 src_start = array_element_address(src, src_offset, T_BYTE);
5920 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5921 }
5922
5923 // now need to get the start of its expanded key array
5924 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5925 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5926 if (k_start == NULL) return false;
5927
5928 if (Matcher::pass_original_key_for_aes()) {
5929 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5930 // compatibility issues between Java key expansion and SPARC crypto instructions
5931 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5932 if (original_k_start == NULL) return false;
5933
5934 // Call the stub.
5935 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5936 stubAddr, stubName, TypePtr::BOTTOM,
5937 src_start, dest_start, k_start, original_k_start);
5938 } else {
5939 // Call the stub.
5940 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5941 stubAddr, stubName, TypePtr::BOTTOM,
5942 src_start, dest_start, k_start);
5943 }
5944
5945 return true;
5946 }
5947
5948 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5949 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5950 address stubAddr;
5951 const char *stubName;
5952
5953 assert(UseAES, "need AES instruction support");
5954
5955 switch(id) {
5956 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5957 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5958 stubName = "cipherBlockChaining_encryptAESCrypt";
5959 break;
5960 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5961 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5962 stubName = "cipherBlockChaining_decryptAESCrypt";
5963 break;
5964 }
5965 if (stubAddr == NULL) return false;
5966
5967 Node* cipherBlockChaining_object = argument(0);
5968 Node* src = argument(1);
5969 Node* src_offset = argument(2);
5970 Node* len = argument(3);
5971 Node* dest = argument(4);
5972 Node* dest_offset = argument(5);
5973
5974 // (1) src and dest are arrays.
5975 const Type* src_type = src->Value(&_gvn);
5976 const Type* dest_type = dest->Value(&_gvn);
5977 const TypeAryPtr* top_src = src_type->isa_aryptr();
5978 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5979 assert (top_src != NULL && top_src->klass() != NULL
5980 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5981
5982 // checks are the responsibility of the caller
5983 Node* src_start = src;
5984 Node* dest_start = dest;
5985 if (src_offset != NULL || dest_offset != NULL) {
5986 assert(src_offset != NULL && dest_offset != NULL, "");
5987 src_start = array_element_address(src, src_offset, T_BYTE);
5988 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5989 }
5990
5991 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5992 // (because of the predicated logic executed earlier).
5993 // so we cast it here safely.
5994 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5995
5996 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5997 if (embeddedCipherObj == NULL) return false;
5998
5999 // cast it to what we know it will be at runtime
6000 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6001 assert(tinst != NULL, "CBC obj is null");
6002 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6003 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6004 if (!klass_AESCrypt->is_loaded()) return false;
6005
6006 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6007 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6008 const TypeOopPtr* xtype = aklass->as_instance_type();
6009 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6010 aescrypt_object = _gvn.transform(aescrypt_object);
6011
6012 // we need to get the start of the aescrypt_object's expanded key array
6013 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6014 if (k_start == NULL) return false;
6015
6016 // similarly, get the start address of the r vector
6017 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6018 if (objRvec == NULL) return false;
6019 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6020
6021 Node* cbcCrypt;
6022 if (Matcher::pass_original_key_for_aes()) {
6023 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6024 // compatibility issues between Java key expansion and SPARC crypto instructions
6025 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6026 if (original_k_start == NULL) return false;
6027
6028 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6029 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6030 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6031 stubAddr, stubName, TypePtr::BOTTOM,
6032 src_start, dest_start, k_start, r_start, len, original_k_start);
6033 } else {
6034 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6035 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6036 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6037 stubAddr, stubName, TypePtr::BOTTOM,
6038 src_start, dest_start, k_start, r_start, len);
6039 }
6040
6041 // return cipher length (int)
6042 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6043 set_result(retvalue);
6044 return true;
6045 }
6046
6047 //------------------------------get_key_start_from_aescrypt_object-----------------------
6048 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6049 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6050 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6051 if (objAESCryptKey == NULL) return (Node *) NULL;
6052
6053 // now have the array, need to get the start address of the K array
6054 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6055 return k_start;
6056 }
6057
6058 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6059 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6060 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6061 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6062 if (objAESCryptKey == NULL) return (Node *) NULL;
6063
6064 // now have the array, need to get the start address of the lastKey array
6065 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6066 return original_k_start;
6067 }
6068
6069 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6070 // Return node representing slow path of predicate check.
6071 // the pseudo code we want to emulate with this predicate is:
6072 // for encryption:
6073 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6074 // for decryption:
6075 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6076 // note cipher==plain is more conservative than the original java code but that's OK
6077 //
6078 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6079 // First, check receiver for NULL since it is virtual method.
6080 Node* objCBC = argument(0);
6081 objCBC = null_check(objCBC);
6082
6083 if (stopped()) return NULL; // Always NULL
6084
6085 // Load embeddedCipher field of CipherBlockChaining object.
6086 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6087
6088 // get AESCrypt klass for instanceOf check
6089 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6090 // will have same classloader as CipherBlockChaining object
6091 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6092 assert(tinst != NULL, "CBCobj is null");
6093 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6094
6095 // we want to do an instanceof comparison against the AESCrypt class
6096 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6097 if (!klass_AESCrypt->is_loaded()) {
6098 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6099 Node* ctrl = control();
6100 set_control(top()); // no regular fast path
6101 return ctrl;
6102 }
6103 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6104
6105 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6106 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6107 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6108
6109 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6110
6111 // for encryption, we are done
6112 if (!decrypting)
6113 return instof_false; // even if it is NULL
6114
6115 // for decryption, we need to add a further check to avoid
6116 // taking the intrinsic path when cipher and plain are the same
6117 // see the original java code for why.
6118 RegionNode* region = new(C) RegionNode(3);
6119 region->init_req(1, instof_false);
6120 Node* src = argument(1);
6121 Node* dest = argument(4);
6122 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6123 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6124 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6125 region->init_req(2, src_dest_conjoint);
6126
6127 record_for_igvn(region);
6128 return _gvn.transform(region);
6129 }