1 /*
2 * Copyright (c) 1999, 2014, 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 (!InlineReflectionGetCallerClass) return NULL;
424 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
425 break;
426
427 case vmIntrinsics::_bitCount_i:
428 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
429 break;
430
431 case vmIntrinsics::_bitCount_l:
432 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
433 break;
434
435 case vmIntrinsics::_numberOfLeadingZeros_i:
436 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
437 break;
438
439 case vmIntrinsics::_numberOfLeadingZeros_l:
440 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
441 break;
442
443 case vmIntrinsics::_numberOfTrailingZeros_i:
444 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
445 break;
446
447 case vmIntrinsics::_numberOfTrailingZeros_l:
448 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
449 break;
450
451 case vmIntrinsics::_reverseBytes_c:
452 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
453 break;
454 case vmIntrinsics::_reverseBytes_s:
455 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
456 break;
457 case vmIntrinsics::_reverseBytes_i:
458 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
459 break;
460 case vmIntrinsics::_reverseBytes_l:
461 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
462 break;
463
464 case vmIntrinsics::_Reference_get:
465 // Use the intrinsic version of Reference.get() so that the value in
466 // the referent field can be registered by the G1 pre-barrier code.
467 // Also add memory barrier to prevent commoning reads from this field
468 // across safepoint since GC can change it value.
469 break;
470
471 case vmIntrinsics::_compareAndSwapObject:
472 #ifdef _LP64
473 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
474 #endif
475 break;
476
477 case vmIntrinsics::_compareAndSwapLong:
478 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
479 break;
480
481 case vmIntrinsics::_getAndAddInt:
482 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
483 break;
484
485 case vmIntrinsics::_getAndAddLong:
486 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
487 break;
488
489 case vmIntrinsics::_getAndSetInt:
490 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
491 break;
492
493 case vmIntrinsics::_getAndSetLong:
494 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
495 break;
496
497 case vmIntrinsics::_getAndSetObject:
498 #ifdef _LP64
499 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
500 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
501 break;
502 #else
503 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
504 break;
505 #endif
506
507 case vmIntrinsics::_aescrypt_encryptBlock:
508 case vmIntrinsics::_aescrypt_decryptBlock:
509 if (!UseAESIntrinsics) return NULL;
510 break;
511
512 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
513 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
514 if (!UseAESIntrinsics) return NULL;
515 // these two require the predicated logic
516 is_predicted = true;
517 break;
518
519 case vmIntrinsics::_updateCRC32:
520 case vmIntrinsics::_updateBytesCRC32:
521 case vmIntrinsics::_updateByteBufferCRC32:
522 if (!UseCRC32Intrinsics) return NULL;
523 break;
524
525 case vmIntrinsics::_incrementExactI:
526 case vmIntrinsics::_addExactI:
527 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
528 break;
529 case vmIntrinsics::_incrementExactL:
530 case vmIntrinsics::_addExactL:
531 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
532 break;
533 case vmIntrinsics::_decrementExactI:
534 case vmIntrinsics::_subtractExactI:
535 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
536 break;
537 case vmIntrinsics::_decrementExactL:
538 case vmIntrinsics::_subtractExactL:
539 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
540 break;
541 case vmIntrinsics::_negateExactI:
542 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
543 break;
544 case vmIntrinsics::_negateExactL:
545 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
546 break;
547 case vmIntrinsics::_multiplyExactI:
548 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
549 break;
550 case vmIntrinsics::_multiplyExactL:
551 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
552 break;
553
554 default:
555 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
556 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
557 break;
558 }
559
560 // -XX:-InlineClassNatives disables natives from the Class class.
561 // The flag applies to all reflective calls, notably Array.newArray
562 // (visible to Java programmers as Array.newInstance).
563 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
564 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
565 if (!InlineClassNatives) return NULL;
566 }
567
568 // -XX:-InlineThreadNatives disables natives from the Thread class.
569 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
570 if (!InlineThreadNatives) return NULL;
571 }
572
573 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
574 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
575 m->holder()->name() == ciSymbol::java_lang_Float() ||
576 m->holder()->name() == ciSymbol::java_lang_Double()) {
577 if (!InlineMathNatives) return NULL;
578 }
579
580 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
581 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
582 if (!InlineUnsafeOps) return NULL;
583 }
584
585 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
586 }
587
588 //----------------------register_library_intrinsics-----------------------
589 // Initialize this file's data structures, for each Compile instance.
590 void Compile::register_library_intrinsics() {
591 // Nothing to do here.
592 }
593
594 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
595 LibraryCallKit kit(jvms, this);
596 Compile* C = kit.C;
597 int nodes = C->unique();
598 #ifndef PRODUCT
599 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
600 char buf[1000];
601 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
602 tty->print_cr("Intrinsic %s", str);
603 }
604 #endif
605 ciMethod* callee = kit.callee();
606 const int bci = kit.bci();
607
608 // Try to inline the intrinsic.
609 if (kit.try_to_inline()) {
610 if (C->print_intrinsics() || C->print_inlining()) {
611 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
612 }
613 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
614 if (C->log()) {
615 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
616 vmIntrinsics::name_at(intrinsic_id()),
617 (is_virtual() ? " virtual='1'" : ""),
618 C->unique() - nodes);
619 }
620 // Push the result from the inlined method onto the stack.
621 kit.push_result();
622 C->print_inlining_update(this);
623 return kit.transfer_exceptions_into_jvms();
624 }
625
626 // The intrinsic bailed out
627 if (C->print_intrinsics() || C->print_inlining()) {
628 if (jvms->has_method()) {
629 // Not a root compile.
630 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
631 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
632 } else {
633 // Root compile
634 tty->print("Did not generate intrinsic %s%s at bci:%d in",
635 vmIntrinsics::name_at(intrinsic_id()),
636 (is_virtual() ? " (virtual)" : ""), bci);
637 }
638 }
639 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
640 C->print_inlining_update(this);
641 return NULL;
642 }
643
644 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
645 LibraryCallKit kit(jvms, this);
646 Compile* C = kit.C;
647 int nodes = C->unique();
648 #ifndef PRODUCT
649 assert(is_predicted(), "sanity");
650 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
651 char buf[1000];
652 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
653 tty->print_cr("Predicate for intrinsic %s", str);
654 }
655 #endif
656 ciMethod* callee = kit.callee();
657 const int bci = kit.bci();
658
659 Node* slow_ctl = kit.try_to_predicate();
660 if (!kit.failing()) {
661 if (C->print_intrinsics() || C->print_inlining()) {
662 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
663 }
664 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
665 if (C->log()) {
666 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
667 vmIntrinsics::name_at(intrinsic_id()),
668 (is_virtual() ? " virtual='1'" : ""),
669 C->unique() - nodes);
670 }
671 return slow_ctl; // Could be NULL if the check folds.
672 }
673
674 // The intrinsic bailed out
675 if (C->print_intrinsics() || C->print_inlining()) {
676 if (jvms->has_method()) {
677 // Not a root compile.
678 const char* msg = "failed to generate predicate for intrinsic";
679 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
680 } else {
681 // Root compile
682 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
683 vmIntrinsics::name_at(intrinsic_id()),
684 (is_virtual() ? " (virtual)" : ""), bci);
685 }
686 }
687 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
688 return NULL;
689 }
690
691 bool LibraryCallKit::try_to_inline() {
692 // Handle symbolic names for otherwise undistinguished boolean switches:
693 const bool is_store = true;
694 const bool is_native_ptr = true;
695 const bool is_static = true;
696 const bool is_volatile = true;
697
698 if (!jvms()->has_method()) {
699 // Root JVMState has a null method.
700 assert(map()->memory()->Opcode() == Op_Parm, "");
701 // Insert the memory aliasing node
702 set_all_memory(reset_memory());
703 }
704 assert(merged_memory(), "");
705
706
707 switch (intrinsic_id()) {
708 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
709 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
710 case vmIntrinsics::_getClass: return inline_native_getClass();
711
712 case vmIntrinsics::_dsin:
713 case vmIntrinsics::_dcos:
714 case vmIntrinsics::_dtan:
715 case vmIntrinsics::_dabs:
716 case vmIntrinsics::_datan2:
717 case vmIntrinsics::_dsqrt:
718 case vmIntrinsics::_dexp:
719 case vmIntrinsics::_dlog:
720 case vmIntrinsics::_dlog10:
721 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
722
723 case vmIntrinsics::_min:
724 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
725
726 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
727 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
728 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
729 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
730 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
731 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
732 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
733 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
734 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
735 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
736 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
737 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
738
739 case vmIntrinsics::_arraycopy: return inline_arraycopy();
740
741 case vmIntrinsics::_compareTo: return inline_string_compareTo();
742 case vmIntrinsics::_indexOf: return inline_string_indexOf();
743 case vmIntrinsics::_equals: return inline_string_equals();
744
745 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
746 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
747 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
748 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
749 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
750 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
751 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
752 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
753 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
754
755 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
756 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
757 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
758 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
759 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
760 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
761 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
762 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
763 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
764
765 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
766 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
767 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
768 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
769 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
770 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
771 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
772 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
773
774 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
775 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
776 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
777 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
778 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
779 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
780 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
781 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
782
783 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
784 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
785 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
786 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
787 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
788 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
789 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
790 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
791 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
792
793 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
794 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
795 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
796 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
797 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
798 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
799 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
800 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
801 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
802
803 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
804 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
805 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
806 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
807
808 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
809 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
810 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
811
812 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
813 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
814 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
815
816 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
817 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
818 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
819 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
820 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
821
822 case vmIntrinsics::_loadFence:
823 case vmIntrinsics::_storeFence:
824 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
825
826 case vmIntrinsics::_currentThread: return inline_native_currentThread();
827 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
828
829 #ifdef TRACE_HAVE_INTRINSICS
830 case vmIntrinsics::_classID: return inline_native_classID();
831 case vmIntrinsics::_threadID: return inline_native_threadID();
832 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
833 #endif
834 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
835 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
836 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
837 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
838 case vmIntrinsics::_newArray: return inline_native_newArray();
839 case vmIntrinsics::_getLength: return inline_native_getLength();
840 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
841 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
842 case vmIntrinsics::_equalsC: return inline_array_equals();
843 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
844
845 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
846
847 case vmIntrinsics::_isInstance:
848 case vmIntrinsics::_getModifiers:
849 case vmIntrinsics::_isInterface:
850 case vmIntrinsics::_isArray:
851 case vmIntrinsics::_isPrimitive:
852 case vmIntrinsics::_getSuperclass:
853 case vmIntrinsics::_getComponentType:
854 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
855
856 case vmIntrinsics::_floatToRawIntBits:
857 case vmIntrinsics::_floatToIntBits:
858 case vmIntrinsics::_intBitsToFloat:
859 case vmIntrinsics::_doubleToRawLongBits:
860 case vmIntrinsics::_doubleToLongBits:
861 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
862
863 case vmIntrinsics::_numberOfLeadingZeros_i:
864 case vmIntrinsics::_numberOfLeadingZeros_l:
865 case vmIntrinsics::_numberOfTrailingZeros_i:
866 case vmIntrinsics::_numberOfTrailingZeros_l:
867 case vmIntrinsics::_bitCount_i:
868 case vmIntrinsics::_bitCount_l:
869 case vmIntrinsics::_reverseBytes_i:
870 case vmIntrinsics::_reverseBytes_l:
871 case vmIntrinsics::_reverseBytes_s:
872 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
873
874 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
875
876 case vmIntrinsics::_Reference_get: return inline_reference_get();
877
878 case vmIntrinsics::_aescrypt_encryptBlock:
879 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
880
881 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
882 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
883 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
884
885 case vmIntrinsics::_encodeISOArray:
886 return inline_encodeISOArray();
887
888 case vmIntrinsics::_updateCRC32:
889 return inline_updateCRC32();
890 case vmIntrinsics::_updateBytesCRC32:
891 return inline_updateBytesCRC32();
892 case vmIntrinsics::_updateByteBufferCRC32:
893 return inline_updateByteBufferCRC32();
894
895 default:
896 // If you get here, it may be that someone has added a new intrinsic
897 // to the list in vmSymbols.hpp without implementing it here.
898 #ifndef PRODUCT
899 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
900 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
901 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
902 }
903 #endif
904 return false;
905 }
906 }
907
908 Node* LibraryCallKit::try_to_predicate() {
909 if (!jvms()->has_method()) {
910 // Root JVMState has a null method.
911 assert(map()->memory()->Opcode() == Op_Parm, "");
912 // Insert the memory aliasing node
913 set_all_memory(reset_memory());
914 }
915 assert(merged_memory(), "");
916
917 switch (intrinsic_id()) {
918 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
919 return inline_cipherBlockChaining_AESCrypt_predicate(false);
920 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
921 return inline_cipherBlockChaining_AESCrypt_predicate(true);
922
923 default:
924 // If you get here, it may be that someone has added a new intrinsic
925 // to the list in vmSymbols.hpp without implementing it here.
926 #ifndef PRODUCT
927 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
928 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
929 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
930 }
931 #endif
932 Node* slow_ctl = control();
933 set_control(top()); // No fast path instrinsic
934 return slow_ctl;
935 }
936 }
937
938 //------------------------------set_result-------------------------------
939 // Helper function for finishing intrinsics.
940 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
941 record_for_igvn(region);
942 set_control(_gvn.transform(region));
943 set_result( _gvn.transform(value));
944 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
945 }
946
947 //------------------------------generate_guard---------------------------
948 // Helper function for generating guarded fast-slow graph structures.
949 // The given 'test', if true, guards a slow path. If the test fails
950 // then a fast path can be taken. (We generally hope it fails.)
951 // In all cases, GraphKit::control() is updated to the fast path.
952 // The returned value represents the control for the slow path.
953 // The return value is never 'top'; it is either a valid control
954 // or NULL if it is obvious that the slow path can never be taken.
955 // Also, if region and the slow control are not NULL, the slow edge
956 // is appended to the region.
957 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
958 if (stopped()) {
959 // Already short circuited.
960 return NULL;
961 }
962
963 // Build an if node and its projections.
964 // If test is true we take the slow path, which we assume is uncommon.
965 if (_gvn.type(test) == TypeInt::ZERO) {
966 // The slow branch is never taken. No need to build this guard.
967 return NULL;
968 }
969
970 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
971
972 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
973 if (if_slow == top()) {
974 // The slow branch is never taken. No need to build this guard.
975 return NULL;
976 }
977
978 if (region != NULL)
979 region->add_req(if_slow);
980
981 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
982 set_control(if_fast);
983
984 return if_slow;
985 }
986
987 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
988 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
989 }
990 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
991 return generate_guard(test, region, PROB_FAIR);
992 }
993
994 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
995 Node* *pos_index) {
996 if (stopped())
997 return NULL; // already stopped
998 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
999 return NULL; // index is already adequately typed
1000 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1001 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1002 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1003 if (is_neg != NULL && pos_index != NULL) {
1004 // Emulate effect of Parse::adjust_map_after_if.
1005 Node* ccast = new CastIINode(index, TypeInt::POS);
1006 ccast->set_req(0, control());
1007 (*pos_index) = _gvn.transform(ccast);
1008 }
1009 return is_neg;
1010 }
1011
1012 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1013 Node* *pos_index) {
1014 if (stopped())
1015 return NULL; // already stopped
1016 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1017 return NULL; // index is already adequately typed
1018 Node* cmp_le = _gvn.transform(new CmpINode(index, intcon(0)));
1019 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1020 Node* bol_le = _gvn.transform(new BoolNode(cmp_le, le_or_eq));
1021 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1022 if (is_notp != NULL && pos_index != NULL) {
1023 // Emulate effect of Parse::adjust_map_after_if.
1024 Node* ccast = new CastIINode(index, TypeInt::POS1);
1025 ccast->set_req(0, control());
1026 (*pos_index) = _gvn.transform(ccast);
1027 }
1028 return is_notp;
1029 }
1030
1031 // Make sure that 'position' is a valid limit index, in [0..length].
1032 // There are two equivalent plans for checking this:
1033 // A. (offset + copyLength) unsigned<= arrayLength
1034 // B. offset <= (arrayLength - copyLength)
1035 // We require that all of the values above, except for the sum and
1036 // difference, are already known to be non-negative.
1037 // Plan A is robust in the face of overflow, if offset and copyLength
1038 // are both hugely positive.
1039 //
1040 // Plan B is less direct and intuitive, but it does not overflow at
1041 // all, since the difference of two non-negatives is always
1042 // representable. Whenever Java methods must perform the equivalent
1043 // check they generally use Plan B instead of Plan A.
1044 // For the moment we use Plan A.
1045 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1046 Node* subseq_length,
1047 Node* array_length,
1048 RegionNode* region) {
1049 if (stopped())
1050 return NULL; // already stopped
1051 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1052 if (zero_offset && subseq_length->eqv_uncast(array_length))
1053 return NULL; // common case of whole-array copy
1054 Node* last = subseq_length;
1055 if (!zero_offset) // last += offset
1056 last = _gvn.transform(new AddINode(last, offset));
1057 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1058 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1059 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1060 return is_over;
1061 }
1062
1063
1064 //--------------------------generate_current_thread--------------------
1065 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1066 ciKlass* thread_klass = env()->Thread_klass();
1067 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1068 Node* thread = _gvn.transform(new ThreadLocalNode());
1069 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1070 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1071 tls_output = thread;
1072 return threadObj;
1073 }
1074
1075
1076 //------------------------------make_string_method_node------------------------
1077 // Helper method for String intrinsic functions. This version is called
1078 // with str1 and str2 pointing to String object nodes.
1079 //
1080 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1081 Node* no_ctrl = NULL;
1082
1083 // Get start addr of string
1084 Node* str1_value = load_String_value(no_ctrl, str1);
1085 Node* str1_offset = load_String_offset(no_ctrl, str1);
1086 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1087
1088 // Get length of string 1
1089 Node* str1_len = load_String_length(no_ctrl, str1);
1090
1091 Node* str2_value = load_String_value(no_ctrl, str2);
1092 Node* str2_offset = load_String_offset(no_ctrl, str2);
1093 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1094
1095 Node* str2_len = NULL;
1096 Node* result = NULL;
1097
1098 switch (opcode) {
1099 case Op_StrIndexOf:
1100 // Get length of string 2
1101 str2_len = load_String_length(no_ctrl, str2);
1102
1103 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1104 str1_start, str1_len, str2_start, str2_len);
1105 break;
1106 case Op_StrComp:
1107 // Get length of string 2
1108 str2_len = load_String_length(no_ctrl, str2);
1109
1110 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1111 str1_start, str1_len, str2_start, str2_len);
1112 break;
1113 case Op_StrEquals:
1114 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1115 str1_start, str2_start, str1_len);
1116 break;
1117 default:
1118 ShouldNotReachHere();
1119 return NULL;
1120 }
1121
1122 // All these intrinsics have checks.
1123 C->set_has_split_ifs(true); // Has chance for split-if optimization
1124
1125 return _gvn.transform(result);
1126 }
1127
1128 // Helper method for String intrinsic functions. This version is called
1129 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1130 // to Int nodes containing the lenghts of str1 and str2.
1131 //
1132 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1133 Node* result = NULL;
1134 switch (opcode) {
1135 case Op_StrIndexOf:
1136 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1137 str1_start, cnt1, str2_start, cnt2);
1138 break;
1139 case Op_StrComp:
1140 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1141 str1_start, cnt1, str2_start, cnt2);
1142 break;
1143 case Op_StrEquals:
1144 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1145 str1_start, str2_start, cnt1);
1146 break;
1147 default:
1148 ShouldNotReachHere();
1149 return NULL;
1150 }
1151
1152 // All these intrinsics have checks.
1153 C->set_has_split_ifs(true); // Has chance for split-if optimization
1154
1155 return _gvn.transform(result);
1156 }
1157
1158 //------------------------------inline_string_compareTo------------------------
1159 // public int java.lang.String.compareTo(String anotherString);
1160 bool LibraryCallKit::inline_string_compareTo() {
1161 Node* receiver = null_check(argument(0));
1162 Node* arg = null_check(argument(1));
1163 if (stopped()) {
1164 return true;
1165 }
1166 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1167 return true;
1168 }
1169
1170 //------------------------------inline_string_equals------------------------
1171 bool LibraryCallKit::inline_string_equals() {
1172 Node* receiver = null_check_receiver();
1173 // NOTE: Do not null check argument for String.equals() because spec
1174 // allows to specify NULL as argument.
1175 Node* argument = this->argument(1);
1176 if (stopped()) {
1177 return true;
1178 }
1179
1180 // paths (plus control) merge
1181 RegionNode* region = new RegionNode(5);
1182 Node* phi = new PhiNode(region, TypeInt::BOOL);
1183
1184 // does source == target string?
1185 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1186 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1187
1188 Node* if_eq = generate_slow_guard(bol, NULL);
1189 if (if_eq != NULL) {
1190 // receiver == argument
1191 phi->init_req(2, intcon(1));
1192 region->init_req(2, if_eq);
1193 }
1194
1195 // get String klass for instanceOf
1196 ciInstanceKlass* klass = env()->String_klass();
1197
1198 if (!stopped()) {
1199 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1200 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1201 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1202
1203 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1204 //instanceOf == true, fallthrough
1205
1206 if (inst_false != NULL) {
1207 phi->init_req(3, intcon(0));
1208 region->init_req(3, inst_false);
1209 }
1210 }
1211
1212 if (!stopped()) {
1213 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1214
1215 // Properly cast the argument to String
1216 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1217 // This path is taken only when argument's type is String:NotNull.
1218 argument = cast_not_null(argument, false);
1219
1220 Node* no_ctrl = NULL;
1221
1222 // Get start addr of receiver
1223 Node* receiver_val = load_String_value(no_ctrl, receiver);
1224 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1225 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1226
1227 // Get length of receiver
1228 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1229
1230 // Get start addr of argument
1231 Node* argument_val = load_String_value(no_ctrl, argument);
1232 Node* argument_offset = load_String_offset(no_ctrl, argument);
1233 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1234
1235 // Get length of argument
1236 Node* argument_cnt = load_String_length(no_ctrl, argument);
1237
1238 // Check for receiver count != argument count
1239 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1240 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1241 Node* if_ne = generate_slow_guard(bol, NULL);
1242 if (if_ne != NULL) {
1243 phi->init_req(4, intcon(0));
1244 region->init_req(4, if_ne);
1245 }
1246
1247 // Check for count == 0 is done by assembler code for StrEquals.
1248
1249 if (!stopped()) {
1250 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1251 phi->init_req(1, equals);
1252 region->init_req(1, control());
1253 }
1254 }
1255
1256 // post merge
1257 set_control(_gvn.transform(region));
1258 record_for_igvn(region);
1259
1260 set_result(_gvn.transform(phi));
1261 return true;
1262 }
1263
1264 //------------------------------inline_array_equals----------------------------
1265 bool LibraryCallKit::inline_array_equals() {
1266 Node* arg1 = argument(0);
1267 Node* arg2 = argument(1);
1268 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1269 return true;
1270 }
1271
1272 // Java version of String.indexOf(constant string)
1273 // class StringDecl {
1274 // StringDecl(char[] ca) {
1275 // offset = 0;
1276 // count = ca.length;
1277 // value = ca;
1278 // }
1279 // int offset;
1280 // int count;
1281 // char[] value;
1282 // }
1283 //
1284 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1285 // int targetOffset, int cache_i, int md2) {
1286 // int cache = cache_i;
1287 // int sourceOffset = string_object.offset;
1288 // int sourceCount = string_object.count;
1289 // int targetCount = target_object.length;
1290 //
1291 // int targetCountLess1 = targetCount - 1;
1292 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1293 //
1294 // char[] source = string_object.value;
1295 // char[] target = target_object;
1296 // int lastChar = target[targetCountLess1];
1297 //
1298 // outer_loop:
1299 // for (int i = sourceOffset; i < sourceEnd; ) {
1300 // int src = source[i + targetCountLess1];
1301 // if (src == lastChar) {
1302 // // With random strings and a 4-character alphabet,
1303 // // reverse matching at this point sets up 0.8% fewer
1304 // // frames, but (paradoxically) makes 0.3% more probes.
1305 // // Since those probes are nearer the lastChar probe,
1306 // // there is may be a net D$ win with reverse matching.
1307 // // But, reversing loop inhibits unroll of inner loop
1308 // // for unknown reason. So, does running outer loop from
1309 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1310 // for (int j = 0; j < targetCountLess1; j++) {
1311 // if (target[targetOffset + j] != source[i+j]) {
1312 // if ((cache & (1 << source[i+j])) == 0) {
1313 // if (md2 < j+1) {
1314 // i += j+1;
1315 // continue outer_loop;
1316 // }
1317 // }
1318 // i += md2;
1319 // continue outer_loop;
1320 // }
1321 // }
1322 // return i - sourceOffset;
1323 // }
1324 // if ((cache & (1 << src)) == 0) {
1325 // i += targetCountLess1;
1326 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1327 // i++;
1328 // }
1329 // return -1;
1330 // }
1331
1332 //------------------------------string_indexOf------------------------
1333 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1334 jint cache_i, jint md2_i) {
1335
1336 Node* no_ctrl = NULL;
1337 float likely = PROB_LIKELY(0.9);
1338 float unlikely = PROB_UNLIKELY(0.9);
1339
1340 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1341
1342 Node* source = load_String_value(no_ctrl, string_object);
1343 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1344 Node* sourceCount = load_String_length(no_ctrl, string_object);
1345
1346 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1347 jint target_length = target_array->length();
1348 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1349 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1350
1351 // String.value field is known to be @Stable.
1352 if (UseImplicitStableValues) {
1353 target = cast_array_to_stable(target, target_type);
1354 }
1355
1356 IdealKit kit(this, false, true);
1357 #define __ kit.
1358 Node* zero = __ ConI(0);
1359 Node* one = __ ConI(1);
1360 Node* cache = __ ConI(cache_i);
1361 Node* md2 = __ ConI(md2_i);
1362 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1363 Node* targetCount = __ ConI(target_length);
1364 Node* targetCountLess1 = __ ConI(target_length - 1);
1365 Node* targetOffset = __ ConI(targetOffset_i);
1366 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1367
1368 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1369 Node* outer_loop = __ make_label(2 /* goto */);
1370 Node* return_ = __ make_label(1);
1371
1372 __ set(rtn,__ ConI(-1));
1373 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1374 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1375 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1376 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1377 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1378 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1379 Node* tpj = __ AddI(targetOffset, __ value(j));
1380 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1381 Node* ipj = __ AddI(__ value(i), __ value(j));
1382 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1383 __ if_then(targ, BoolTest::ne, src2); {
1384 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1385 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1386 __ increment(i, __ AddI(__ value(j), one));
1387 __ goto_(outer_loop);
1388 } __ end_if(); __ dead(j);
1389 }__ end_if(); __ dead(j);
1390 __ increment(i, md2);
1391 __ goto_(outer_loop);
1392 }__ end_if();
1393 __ increment(j, one);
1394 }__ end_loop(); __ dead(j);
1395 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1396 __ goto_(return_);
1397 }__ end_if();
1398 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1399 __ increment(i, targetCountLess1);
1400 }__ end_if();
1401 __ increment(i, one);
1402 __ bind(outer_loop);
1403 }__ end_loop(); __ dead(i);
1404 __ bind(return_);
1405
1406 // Final sync IdealKit and GraphKit.
1407 final_sync(kit);
1408 Node* result = __ value(rtn);
1409 #undef __
1410 C->set_has_loops(true);
1411 return result;
1412 }
1413
1414 //------------------------------inline_string_indexOf------------------------
1415 bool LibraryCallKit::inline_string_indexOf() {
1416 Node* receiver = argument(0);
1417 Node* arg = argument(1);
1418
1419 Node* result;
1420 // Disable the use of pcmpestri until it can be guaranteed that
1421 // the load doesn't cross into the uncommited space.
1422 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1423 UseSSE42Intrinsics) {
1424 // Generate SSE4.2 version of indexOf
1425 // We currently only have match rules that use SSE4.2
1426
1427 receiver = null_check(receiver);
1428 arg = null_check(arg);
1429 if (stopped()) {
1430 return true;
1431 }
1432
1433 ciInstanceKlass* str_klass = env()->String_klass();
1434 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1435
1436 // Make the merge point
1437 RegionNode* result_rgn = new RegionNode(4);
1438 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1439 Node* no_ctrl = NULL;
1440
1441 // Get start addr of source string
1442 Node* source = load_String_value(no_ctrl, receiver);
1443 Node* source_offset = load_String_offset(no_ctrl, receiver);
1444 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1445
1446 // Get length of source string
1447 Node* source_cnt = load_String_length(no_ctrl, receiver);
1448
1449 // Get start addr of substring
1450 Node* substr = load_String_value(no_ctrl, arg);
1451 Node* substr_offset = load_String_offset(no_ctrl, arg);
1452 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1453
1454 // Get length of source string
1455 Node* substr_cnt = load_String_length(no_ctrl, arg);
1456
1457 // Check for substr count > string count
1458 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1459 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1460 Node* if_gt = generate_slow_guard(bol, NULL);
1461 if (if_gt != NULL) {
1462 result_phi->init_req(2, intcon(-1));
1463 result_rgn->init_req(2, if_gt);
1464 }
1465
1466 if (!stopped()) {
1467 // Check for substr count == 0
1468 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1469 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1470 Node* if_zero = generate_slow_guard(bol, NULL);
1471 if (if_zero != NULL) {
1472 result_phi->init_req(3, intcon(0));
1473 result_rgn->init_req(3, if_zero);
1474 }
1475 }
1476
1477 if (!stopped()) {
1478 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1479 result_phi->init_req(1, result);
1480 result_rgn->init_req(1, control());
1481 }
1482 set_control(_gvn.transform(result_rgn));
1483 record_for_igvn(result_rgn);
1484 result = _gvn.transform(result_phi);
1485
1486 } else { // Use LibraryCallKit::string_indexOf
1487 // don't intrinsify if argument isn't a constant string.
1488 if (!arg->is_Con()) {
1489 return false;
1490 }
1491 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1492 if (str_type == NULL) {
1493 return false;
1494 }
1495 ciInstanceKlass* klass = env()->String_klass();
1496 ciObject* str_const = str_type->const_oop();
1497 if (str_const == NULL || str_const->klass() != klass) {
1498 return false;
1499 }
1500 ciInstance* str = str_const->as_instance();
1501 assert(str != NULL, "must be instance");
1502
1503 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1504 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1505
1506 int o;
1507 int c;
1508 if (java_lang_String::has_offset_field()) {
1509 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1510 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1511 } else {
1512 o = 0;
1513 c = pat->length();
1514 }
1515
1516 // constant strings have no offset and count == length which
1517 // simplifies the resulting code somewhat so lets optimize for that.
1518 if (o != 0 || c != pat->length()) {
1519 return false;
1520 }
1521
1522 receiver = null_check(receiver, T_OBJECT);
1523 // NOTE: No null check on the argument is needed since it's a constant String oop.
1524 if (stopped()) {
1525 return true;
1526 }
1527
1528 // The null string as a pattern always returns 0 (match at beginning of string)
1529 if (c == 0) {
1530 set_result(intcon(0));
1531 return true;
1532 }
1533
1534 // Generate default indexOf
1535 jchar lastChar = pat->char_at(o + (c - 1));
1536 int cache = 0;
1537 int i;
1538 for (i = 0; i < c - 1; i++) {
1539 assert(i < pat->length(), "out of range");
1540 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1541 }
1542
1543 int md2 = c;
1544 for (i = 0; i < c - 1; i++) {
1545 assert(i < pat->length(), "out of range");
1546 if (pat->char_at(o + i) == lastChar) {
1547 md2 = (c - 1) - i;
1548 }
1549 }
1550
1551 result = string_indexOf(receiver, pat, o, cache, md2);
1552 }
1553 set_result(result);
1554 return true;
1555 }
1556
1557 //--------------------------round_double_node--------------------------------
1558 // Round a double node if necessary.
1559 Node* LibraryCallKit::round_double_node(Node* n) {
1560 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1561 n = _gvn.transform(new RoundDoubleNode(0, n));
1562 return n;
1563 }
1564
1565 //------------------------------inline_math-----------------------------------
1566 // public static double Math.abs(double)
1567 // public static double Math.sqrt(double)
1568 // public static double Math.log(double)
1569 // public static double Math.log10(double)
1570 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1571 Node* arg = round_double_node(argument(0));
1572 Node* n;
1573 switch (id) {
1574 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1575 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1576 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1577 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1578 default: fatal_unexpected_iid(id); break;
1579 }
1580 set_result(_gvn.transform(n));
1581 return true;
1582 }
1583
1584 //------------------------------inline_trig----------------------------------
1585 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1586 // argument reduction which will turn into a fast/slow diamond.
1587 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1588 Node* arg = round_double_node(argument(0));
1589 Node* n = NULL;
1590
1591 switch (id) {
1592 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1593 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1594 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1595 default: fatal_unexpected_iid(id); break;
1596 }
1597 n = _gvn.transform(n);
1598
1599 // Rounding required? Check for argument reduction!
1600 if (Matcher::strict_fp_requires_explicit_rounding) {
1601 static const double pi_4 = 0.7853981633974483;
1602 static const double neg_pi_4 = -0.7853981633974483;
1603 // pi/2 in 80-bit extended precision
1604 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1605 // -pi/2 in 80-bit extended precision
1606 // 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};
1607 // Cutoff value for using this argument reduction technique
1608 //static const double pi_2_minus_epsilon = 1.564660403643354;
1609 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1610
1611 // Pseudocode for sin:
1612 // if (x <= Math.PI / 4.0) {
1613 // if (x >= -Math.PI / 4.0) return fsin(x);
1614 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1615 // } else {
1616 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1617 // }
1618 // return StrictMath.sin(x);
1619
1620 // Pseudocode for cos:
1621 // if (x <= Math.PI / 4.0) {
1622 // if (x >= -Math.PI / 4.0) return fcos(x);
1623 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1624 // } else {
1625 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1626 // }
1627 // return StrictMath.cos(x);
1628
1629 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1630 // requires a special machine instruction to load it. Instead we'll try
1631 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1632 // probably do the math inside the SIN encoding.
1633
1634 // Make the merge point
1635 RegionNode* r = new RegionNode(3);
1636 Node* phi = new PhiNode(r, Type::DOUBLE);
1637
1638 // Flatten arg so we need only 1 test
1639 Node *abs = _gvn.transform(new AbsDNode(arg));
1640 // Node for PI/4 constant
1641 Node *pi4 = makecon(TypeD::make(pi_4));
1642 // Check PI/4 : abs(arg)
1643 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1644 // Check: If PI/4 < abs(arg) then go slow
1645 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1646 // Branch either way
1647 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1648 set_control(opt_iff(r,iff));
1649
1650 // Set fast path result
1651 phi->init_req(2, n);
1652
1653 // Slow path - non-blocking leaf call
1654 Node* call = NULL;
1655 switch (id) {
1656 case vmIntrinsics::_dsin:
1657 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1658 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1659 "Sin", NULL, arg, top());
1660 break;
1661 case vmIntrinsics::_dcos:
1662 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1663 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1664 "Cos", NULL, arg, top());
1665 break;
1666 case vmIntrinsics::_dtan:
1667 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1668 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1669 "Tan", NULL, arg, top());
1670 break;
1671 }
1672 assert(control()->in(0) == call, "");
1673 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1674 r->init_req(1, control());
1675 phi->init_req(1, slow_result);
1676
1677 // Post-merge
1678 set_control(_gvn.transform(r));
1679 record_for_igvn(r);
1680 n = _gvn.transform(phi);
1681
1682 C->set_has_split_ifs(true); // Has chance for split-if optimization
1683 }
1684 set_result(n);
1685 return true;
1686 }
1687
1688 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1689 //-------------------
1690 //result=(result.isNaN())? funcAddr():result;
1691 // Check: If isNaN() by checking result!=result? then either trap
1692 // or go to runtime
1693 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1694 // Build the boolean node
1695 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1696
1697 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1698 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1699 // The pow or exp intrinsic returned a NaN, which requires a call
1700 // to the runtime. Recompile with the runtime call.
1701 uncommon_trap(Deoptimization::Reason_intrinsic,
1702 Deoptimization::Action_make_not_entrant);
1703 }
1704 return result;
1705 } else {
1706 // If this inlining ever returned NaN in the past, we compile a call
1707 // to the runtime to properly handle corner cases
1708
1709 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1710 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1711 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1712
1713 if (!if_slow->is_top()) {
1714 RegionNode* result_region = new RegionNode(3);
1715 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1716
1717 result_region->init_req(1, if_fast);
1718 result_val->init_req(1, result);
1719
1720 set_control(if_slow);
1721
1722 const TypePtr* no_memory_effects = NULL;
1723 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1724 no_memory_effects,
1725 x, top(), y, y ? top() : NULL);
1726 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1727 #ifdef ASSERT
1728 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1729 assert(value_top == top(), "second value must be top");
1730 #endif
1731
1732 result_region->init_req(2, control());
1733 result_val->init_req(2, value);
1734 set_control(_gvn.transform(result_region));
1735 return _gvn.transform(result_val);
1736 } else {
1737 return result;
1738 }
1739 }
1740 }
1741
1742 //------------------------------inline_exp-------------------------------------
1743 // Inline exp instructions, if possible. The Intel hardware only misses
1744 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1745 bool LibraryCallKit::inline_exp() {
1746 Node* arg = round_double_node(argument(0));
1747 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1748
1749 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1750 set_result(n);
1751
1752 C->set_has_split_ifs(true); // Has chance for split-if optimization
1753 return true;
1754 }
1755
1756 //------------------------------inline_pow-------------------------------------
1757 // Inline power instructions, if possible.
1758 bool LibraryCallKit::inline_pow() {
1759 // Pseudocode for pow
1760 // if (y == 2) {
1761 // return x * x;
1762 // } else {
1763 // if (x <= 0.0) {
1764 // long longy = (long)y;
1765 // if ((double)longy == y) { // if y is long
1766 // if (y + 1 == y) longy = 0; // huge number: even
1767 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1768 // } else {
1769 // result = NaN;
1770 // }
1771 // } else {
1772 // result = DPow(x,y);
1773 // }
1774 // if (result != result)? {
1775 // result = uncommon_trap() or runtime_call();
1776 // }
1777 // return result;
1778 // }
1779
1780 Node* x = round_double_node(argument(0));
1781 Node* y = round_double_node(argument(2));
1782
1783 Node* result = NULL;
1784
1785 Node* const_two_node = makecon(TypeD::make(2.0));
1786 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1787 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1788 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1789 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1790 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1791
1792 RegionNode* region_node = new RegionNode(3);
1793 region_node->init_req(1, if_true);
1794
1795 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1796 // special case for x^y where y == 2, we can convert it to x * x
1797 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1798
1799 // set control to if_false since we will now process the false branch
1800 set_control(if_false);
1801
1802 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1803 // Short form: skip the fancy tests and just check for NaN result.
1804 result = _gvn.transform(new PowDNode(C, control(), x, y));
1805 } else {
1806 // If this inlining ever returned NaN in the past, include all
1807 // checks + call to the runtime.
1808
1809 // Set the merge point for If node with condition of (x <= 0.0)
1810 // There are four possible paths to region node and phi node
1811 RegionNode *r = new RegionNode(4);
1812 Node *phi = new PhiNode(r, Type::DOUBLE);
1813
1814 // Build the first if node: if (x <= 0.0)
1815 // Node for 0 constant
1816 Node *zeronode = makecon(TypeD::ZERO);
1817 // Check x:0
1818 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1819 // Check: If (x<=0) then go complex path
1820 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1821 // Branch either way
1822 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1823 // Fast path taken; set region slot 3
1824 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1825 r->init_req(3,fast_taken); // Capture fast-control
1826
1827 // Fast path not-taken, i.e. slow path
1828 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1829
1830 // Set fast path result
1831 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1832 phi->init_req(3, fast_result);
1833
1834 // Complex path
1835 // Build the second if node (if y is long)
1836 // Node for (long)y
1837 Node *longy = _gvn.transform(new ConvD2LNode(y));
1838 // Node for (double)((long) y)
1839 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1840 // Check (double)((long) y) : y
1841 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1842 // Check if (y isn't long) then go to slow path
1843
1844 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1845 // Branch either way
1846 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1847 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1848
1849 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1850
1851 // Calculate DPow(abs(x), y)*(1 & (long)y)
1852 // Node for constant 1
1853 Node *conone = longcon(1);
1854 // 1& (long)y
1855 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1856
1857 // A huge number is always even. Detect a huge number by checking
1858 // if y + 1 == y and set integer to be tested for parity to 0.
1859 // Required for corner case:
1860 // (long)9.223372036854776E18 = max_jlong
1861 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1862 // max_jlong is odd but 9.223372036854776E18 is even
1863 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1864 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1865 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1866 Node* correctedsign = NULL;
1867 if (ConditionalMoveLimit != 0) {
1868 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1869 } else {
1870 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1871 RegionNode *r = new RegionNode(3);
1872 Node *phi = new PhiNode(r, TypeLong::LONG);
1873 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1874 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1875 phi->init_req(1, signnode);
1876 phi->init_req(2, longcon(0));
1877 correctedsign = _gvn.transform(phi);
1878 ylong_path = _gvn.transform(r);
1879 record_for_igvn(r);
1880 }
1881
1882 // zero node
1883 Node *conzero = longcon(0);
1884 // Check (1&(long)y)==0?
1885 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1886 // Check if (1&(long)y)!=0?, if so the result is negative
1887 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1888 // abs(x)
1889 Node *absx=_gvn.transform(new AbsDNode(x));
1890 // abs(x)^y
1891 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1892 // -abs(x)^y
1893 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1894 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1895 Node *signresult = NULL;
1896 if (ConditionalMoveLimit != 0) {
1897 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1898 } else {
1899 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1900 RegionNode *r = new RegionNode(3);
1901 Node *phi = new PhiNode(r, Type::DOUBLE);
1902 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1903 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1904 phi->init_req(1, absxpowy);
1905 phi->init_req(2, negabsxpowy);
1906 signresult = _gvn.transform(phi);
1907 ylong_path = _gvn.transform(r);
1908 record_for_igvn(r);
1909 }
1910 // Set complex path fast result
1911 r->init_req(2, ylong_path);
1912 phi->init_req(2, signresult);
1913
1914 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1915 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1916 r->init_req(1,slow_path);
1917 phi->init_req(1,slow_result);
1918
1919 // Post merge
1920 set_control(_gvn.transform(r));
1921 record_for_igvn(r);
1922 result = _gvn.transform(phi);
1923 }
1924
1925 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1926
1927 // control from finish_pow_exp is now input to the region node
1928 region_node->set_req(2, control());
1929 // the result from finish_pow_exp is now input to the phi node
1930 phi_node->init_req(2, result);
1931 set_control(_gvn.transform(region_node));
1932 record_for_igvn(region_node);
1933 set_result(_gvn.transform(phi_node));
1934
1935 C->set_has_split_ifs(true); // Has chance for split-if optimization
1936 return true;
1937 }
1938
1939 //------------------------------runtime_math-----------------------------
1940 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1941 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1942 "must be (DD)D or (D)D type");
1943
1944 // Inputs
1945 Node* a = round_double_node(argument(0));
1946 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1947
1948 const TypePtr* no_memory_effects = NULL;
1949 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1950 no_memory_effects,
1951 a, top(), b, b ? top() : NULL);
1952 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1953 #ifdef ASSERT
1954 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1955 assert(value_top == top(), "second value must be top");
1956 #endif
1957
1958 set_result(value);
1959 return true;
1960 }
1961
1962 //------------------------------inline_math_native-----------------------------
1963 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1964 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1965 switch (id) {
1966 // These intrinsics are not properly supported on all hardware
1967 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1968 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1969 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1970 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1971 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1972 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1973
1974 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1975 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1976 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1977 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1978
1979 // These intrinsics are supported on all hardware
1980 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1981 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1982
1983 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1984 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1985 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1986 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1987 #undef FN_PTR
1988
1989 // These intrinsics are not yet correctly implemented
1990 case vmIntrinsics::_datan2:
1991 return false;
1992
1993 default:
1994 fatal_unexpected_iid(id);
1995 return false;
1996 }
1997 }
1998
1999 static bool is_simple_name(Node* n) {
2000 return (n->req() == 1 // constant
2001 || (n->is_Type() && n->as_Type()->type()->singleton())
2002 || n->is_Proj() // parameter or return value
2003 || n->is_Phi() // local of some sort
2004 );
2005 }
2006
2007 //----------------------------inline_min_max-----------------------------------
2008 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2009 set_result(generate_min_max(id, argument(0), argument(1)));
2010 return true;
2011 }
2012
2013 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2014 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2015 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2016 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2017 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2018
2019 {
2020 PreserveJVMState pjvms(this);
2021 PreserveReexecuteState preexecs(this);
2022 jvms()->set_should_reexecute(true);
2023
2024 set_control(slow_path);
2025 set_i_o(i_o());
2026
2027 uncommon_trap(Deoptimization::Reason_intrinsic,
2028 Deoptimization::Action_none);
2029 }
2030
2031 set_control(fast_path);
2032 set_result(math);
2033 }
2034
2035 template <typename OverflowOp>
2036 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2037 typedef typename OverflowOp::MathOp MathOp;
2038
2039 MathOp* mathOp = new MathOp(arg1, arg2);
2040 Node* operation = _gvn.transform( mathOp );
2041 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2042 inline_math_mathExact(operation, ofcheck);
2043 return true;
2044 }
2045
2046 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2047 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2048 }
2049
2050 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2051 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2052 }
2053
2054 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2055 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2056 }
2057
2058 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2059 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2060 }
2061
2062 bool LibraryCallKit::inline_math_negateExactI() {
2063 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2064 }
2065
2066 bool LibraryCallKit::inline_math_negateExactL() {
2067 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2068 }
2069
2070 bool LibraryCallKit::inline_math_multiplyExactI() {
2071 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2072 }
2073
2074 bool LibraryCallKit::inline_math_multiplyExactL() {
2075 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2076 }
2077
2078 Node*
2079 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2080 // These are the candidate return value:
2081 Node* xvalue = x0;
2082 Node* yvalue = y0;
2083
2084 if (xvalue == yvalue) {
2085 return xvalue;
2086 }
2087
2088 bool want_max = (id == vmIntrinsics::_max);
2089
2090 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2091 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2092 if (txvalue == NULL || tyvalue == NULL) return top();
2093 // This is not really necessary, but it is consistent with a
2094 // hypothetical MaxINode::Value method:
2095 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2096
2097 // %%% This folding logic should (ideally) be in a different place.
2098 // Some should be inside IfNode, and there to be a more reliable
2099 // transformation of ?: style patterns into cmoves. We also want
2100 // more powerful optimizations around cmove and min/max.
2101
2102 // Try to find a dominating comparison of these guys.
2103 // It can simplify the index computation for Arrays.copyOf
2104 // and similar uses of System.arraycopy.
2105 // First, compute the normalized version of CmpI(x, y).
2106 int cmp_op = Op_CmpI;
2107 Node* xkey = xvalue;
2108 Node* ykey = yvalue;
2109 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2110 if (ideal_cmpxy->is_Cmp()) {
2111 // E.g., if we have CmpI(length - offset, count),
2112 // it might idealize to CmpI(length, count + offset)
2113 cmp_op = ideal_cmpxy->Opcode();
2114 xkey = ideal_cmpxy->in(1);
2115 ykey = ideal_cmpxy->in(2);
2116 }
2117
2118 // Start by locating any relevant comparisons.
2119 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2120 Node* cmpxy = NULL;
2121 Node* cmpyx = NULL;
2122 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2123 Node* cmp = start_from->fast_out(k);
2124 if (cmp->outcnt() > 0 && // must have prior uses
2125 cmp->in(0) == NULL && // must be context-independent
2126 cmp->Opcode() == cmp_op) { // right kind of compare
2127 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2128 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2129 }
2130 }
2131
2132 const int NCMPS = 2;
2133 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2134 int cmpn;
2135 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2136 if (cmps[cmpn] != NULL) break; // find a result
2137 }
2138 if (cmpn < NCMPS) {
2139 // Look for a dominating test that tells us the min and max.
2140 int depth = 0; // Limit search depth for speed
2141 Node* dom = control();
2142 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2143 if (++depth >= 100) break;
2144 Node* ifproj = dom;
2145 if (!ifproj->is_Proj()) continue;
2146 Node* iff = ifproj->in(0);
2147 if (!iff->is_If()) continue;
2148 Node* bol = iff->in(1);
2149 if (!bol->is_Bool()) continue;
2150 Node* cmp = bol->in(1);
2151 if (cmp == NULL) continue;
2152 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2153 if (cmps[cmpn] == cmp) break;
2154 if (cmpn == NCMPS) continue;
2155 BoolTest::mask btest = bol->as_Bool()->_test._test;
2156 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2157 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2158 // At this point, we know that 'x btest y' is true.
2159 switch (btest) {
2160 case BoolTest::eq:
2161 // They are proven equal, so we can collapse the min/max.
2162 // Either value is the answer. Choose the simpler.
2163 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2164 return yvalue;
2165 return xvalue;
2166 case BoolTest::lt: // x < y
2167 case BoolTest::le: // x <= y
2168 return (want_max ? yvalue : xvalue);
2169 case BoolTest::gt: // x > y
2170 case BoolTest::ge: // x >= y
2171 return (want_max ? xvalue : yvalue);
2172 }
2173 }
2174 }
2175
2176 // We failed to find a dominating test.
2177 // Let's pick a test that might GVN with prior tests.
2178 Node* best_bol = NULL;
2179 BoolTest::mask best_btest = BoolTest::illegal;
2180 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2181 Node* cmp = cmps[cmpn];
2182 if (cmp == NULL) continue;
2183 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2184 Node* bol = cmp->fast_out(j);
2185 if (!bol->is_Bool()) continue;
2186 BoolTest::mask btest = bol->as_Bool()->_test._test;
2187 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2188 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2189 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2190 best_bol = bol->as_Bool();
2191 best_btest = btest;
2192 }
2193 }
2194 }
2195
2196 Node* answer_if_true = NULL;
2197 Node* answer_if_false = NULL;
2198 switch (best_btest) {
2199 default:
2200 if (cmpxy == NULL)
2201 cmpxy = ideal_cmpxy;
2202 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2203 // and fall through:
2204 case BoolTest::lt: // x < y
2205 case BoolTest::le: // x <= y
2206 answer_if_true = (want_max ? yvalue : xvalue);
2207 answer_if_false = (want_max ? xvalue : yvalue);
2208 break;
2209 case BoolTest::gt: // x > y
2210 case BoolTest::ge: // x >= y
2211 answer_if_true = (want_max ? xvalue : yvalue);
2212 answer_if_false = (want_max ? yvalue : xvalue);
2213 break;
2214 }
2215
2216 jint hi, lo;
2217 if (want_max) {
2218 // We can sharpen the minimum.
2219 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2220 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2221 } else {
2222 // We can sharpen the maximum.
2223 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2224 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2225 }
2226
2227 // Use a flow-free graph structure, to avoid creating excess control edges
2228 // which could hinder other optimizations.
2229 // Since Math.min/max is often used with arraycopy, we want
2230 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2231 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2232 answer_if_false, answer_if_true,
2233 TypeInt::make(lo, hi, widen));
2234
2235 return _gvn.transform(cmov);
2236
2237 /*
2238 // This is not as desirable as it may seem, since Min and Max
2239 // nodes do not have a full set of optimizations.
2240 // And they would interfere, anyway, with 'if' optimizations
2241 // and with CMoveI canonical forms.
2242 switch (id) {
2243 case vmIntrinsics::_min:
2244 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2245 case vmIntrinsics::_max:
2246 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2247 default:
2248 ShouldNotReachHere();
2249 }
2250 */
2251 }
2252
2253 inline int
2254 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2255 const TypePtr* base_type = TypePtr::NULL_PTR;
2256 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2257 if (base_type == NULL) {
2258 // Unknown type.
2259 return Type::AnyPtr;
2260 } else if (base_type == TypePtr::NULL_PTR) {
2261 // Since this is a NULL+long form, we have to switch to a rawptr.
2262 base = _gvn.transform(new CastX2PNode(offset));
2263 offset = MakeConX(0);
2264 return Type::RawPtr;
2265 } else if (base_type->base() == Type::RawPtr) {
2266 return Type::RawPtr;
2267 } else if (base_type->isa_oopptr()) {
2268 // Base is never null => always a heap address.
2269 if (base_type->ptr() == TypePtr::NotNull) {
2270 return Type::OopPtr;
2271 }
2272 // Offset is small => always a heap address.
2273 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2274 if (offset_type != NULL &&
2275 base_type->offset() == 0 && // (should always be?)
2276 offset_type->_lo >= 0 &&
2277 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2278 return Type::OopPtr;
2279 }
2280 // Otherwise, it might either be oop+off or NULL+addr.
2281 return Type::AnyPtr;
2282 } else {
2283 // No information:
2284 return Type::AnyPtr;
2285 }
2286 }
2287
2288 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2289 int kind = classify_unsafe_addr(base, offset);
2290 if (kind == Type::RawPtr) {
2291 return basic_plus_adr(top(), base, offset);
2292 } else {
2293 return basic_plus_adr(base, offset);
2294 }
2295 }
2296
2297 //--------------------------inline_number_methods-----------------------------
2298 // inline int Integer.numberOfLeadingZeros(int)
2299 // inline int Long.numberOfLeadingZeros(long)
2300 //
2301 // inline int Integer.numberOfTrailingZeros(int)
2302 // inline int Long.numberOfTrailingZeros(long)
2303 //
2304 // inline int Integer.bitCount(int)
2305 // inline int Long.bitCount(long)
2306 //
2307 // inline char Character.reverseBytes(char)
2308 // inline short Short.reverseBytes(short)
2309 // inline int Integer.reverseBytes(int)
2310 // inline long Long.reverseBytes(long)
2311 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2312 Node* arg = argument(0);
2313 Node* n;
2314 switch (id) {
2315 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2316 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2317 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2318 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2319 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2320 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2321 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2322 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2323 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2324 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2325 default: fatal_unexpected_iid(id); break;
2326 }
2327 set_result(_gvn.transform(n));
2328 return true;
2329 }
2330
2331 //----------------------------inline_unsafe_access----------------------------
2332
2333 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2334
2335 // Helper that guards and inserts a pre-barrier.
2336 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2337 Node* pre_val, bool need_mem_bar) {
2338 // We could be accessing the referent field of a reference object. If so, when G1
2339 // is enabled, we need to log the value in the referent field in an SATB buffer.
2340 // This routine performs some compile time filters and generates suitable
2341 // runtime filters that guard the pre-barrier code.
2342 // Also add memory barrier for non volatile load from the referent field
2343 // to prevent commoning of loads across safepoint.
2344 if (!UseG1GC && !need_mem_bar)
2345 return;
2346
2347 // Some compile time checks.
2348
2349 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2350 const TypeX* otype = offset->find_intptr_t_type();
2351 if (otype != NULL && otype->is_con() &&
2352 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2353 // Constant offset but not the reference_offset so just return
2354 return;
2355 }
2356
2357 // We only need to generate the runtime guards for instances.
2358 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2359 if (btype != NULL) {
2360 if (btype->isa_aryptr()) {
2361 // Array type so nothing to do
2362 return;
2363 }
2364
2365 const TypeInstPtr* itype = btype->isa_instptr();
2366 if (itype != NULL) {
2367 // Can the klass of base_oop be statically determined to be
2368 // _not_ a sub-class of Reference and _not_ Object?
2369 ciKlass* klass = itype->klass();
2370 if ( klass->is_loaded() &&
2371 !klass->is_subtype_of(env()->Reference_klass()) &&
2372 !env()->Object_klass()->is_subtype_of(klass)) {
2373 return;
2374 }
2375 }
2376 }
2377
2378 // The compile time filters did not reject base_oop/offset so
2379 // we need to generate the following runtime filters
2380 //
2381 // if (offset == java_lang_ref_Reference::_reference_offset) {
2382 // if (instance_of(base, java.lang.ref.Reference)) {
2383 // pre_barrier(_, pre_val, ...);
2384 // }
2385 // }
2386
2387 float likely = PROB_LIKELY( 0.999);
2388 float unlikely = PROB_UNLIKELY(0.999);
2389
2390 IdealKit ideal(this);
2391 #define __ ideal.
2392
2393 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2394
2395 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2396 // Update graphKit memory and control from IdealKit.
2397 sync_kit(ideal);
2398
2399 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2400 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2401
2402 // Update IdealKit memory and control from graphKit.
2403 __ sync_kit(this);
2404
2405 Node* one = __ ConI(1);
2406 // is_instof == 0 if base_oop == NULL
2407 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2408
2409 // Update graphKit from IdeakKit.
2410 sync_kit(ideal);
2411
2412 // Use the pre-barrier to record the value in the referent field
2413 pre_barrier(false /* do_load */,
2414 __ ctrl(),
2415 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2416 pre_val /* pre_val */,
2417 T_OBJECT);
2418 if (need_mem_bar) {
2419 // Add memory barrier to prevent commoning reads from this field
2420 // across safepoint since GC can change its value.
2421 insert_mem_bar(Op_MemBarCPUOrder);
2422 }
2423 // Update IdealKit from graphKit.
2424 __ sync_kit(this);
2425
2426 } __ end_if(); // _ref_type != ref_none
2427 } __ end_if(); // offset == referent_offset
2428
2429 // Final sync IdealKit and GraphKit.
2430 final_sync(ideal);
2431 #undef __
2432 }
2433
2434
2435 // Interpret Unsafe.fieldOffset cookies correctly:
2436 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2437
2438 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2439 // Attempt to infer a sharper value type from the offset and base type.
2440 ciKlass* sharpened_klass = NULL;
2441
2442 // See if it is an instance field, with an object type.
2443 if (alias_type->field() != NULL) {
2444 assert(!is_native_ptr, "native pointer op cannot use a java address");
2445 if (alias_type->field()->type()->is_klass()) {
2446 sharpened_klass = alias_type->field()->type()->as_klass();
2447 }
2448 }
2449
2450 // See if it is a narrow oop array.
2451 if (adr_type->isa_aryptr()) {
2452 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2453 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2454 if (elem_type != NULL) {
2455 sharpened_klass = elem_type->klass();
2456 }
2457 }
2458 }
2459
2460 // The sharpened class might be unloaded if there is no class loader
2461 // contraint in place.
2462 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2463 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2464
2465 #ifndef PRODUCT
2466 if (C->print_intrinsics() || C->print_inlining()) {
2467 tty->print(" from base type: "); adr_type->dump();
2468 tty->print(" sharpened value: "); tjp->dump();
2469 }
2470 #endif
2471 // Sharpen the value type.
2472 return tjp;
2473 }
2474 return NULL;
2475 }
2476
2477 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2478 if (callee()->is_static()) return false; // caller must have the capability!
2479
2480 #ifndef PRODUCT
2481 {
2482 ResourceMark rm;
2483 // Check the signatures.
2484 ciSignature* sig = callee()->signature();
2485 #ifdef ASSERT
2486 if (!is_store) {
2487 // Object getObject(Object base, int/long offset), etc.
2488 BasicType rtype = sig->return_type()->basic_type();
2489 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2490 rtype = T_ADDRESS; // it is really a C void*
2491 assert(rtype == type, "getter must return the expected value");
2492 if (!is_native_ptr) {
2493 assert(sig->count() == 2, "oop getter has 2 arguments");
2494 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2495 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2496 } else {
2497 assert(sig->count() == 1, "native getter has 1 argument");
2498 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2499 }
2500 } else {
2501 // void putObject(Object base, int/long offset, Object x), etc.
2502 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2503 if (!is_native_ptr) {
2504 assert(sig->count() == 3, "oop putter has 3 arguments");
2505 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2506 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2507 } else {
2508 assert(sig->count() == 2, "native putter has 2 arguments");
2509 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2510 }
2511 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2512 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2513 vtype = T_ADDRESS; // it is really a C void*
2514 assert(vtype == type, "putter must accept the expected value");
2515 }
2516 #endif // ASSERT
2517 }
2518 #endif //PRODUCT
2519
2520 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2521
2522 Node* receiver = argument(0); // type: oop
2523
2524 // Build address expression. See the code in inline_unsafe_prefetch.
2525 Node* adr;
2526 Node* heap_base_oop = top();
2527 Node* offset = top();
2528 Node* val;
2529
2530 if (!is_native_ptr) {
2531 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2532 Node* base = argument(1); // type: oop
2533 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2534 offset = argument(2); // type: long
2535 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2536 // to be plain byte offsets, which are also the same as those accepted
2537 // by oopDesc::field_base.
2538 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2539 "fieldOffset must be byte-scaled");
2540 // 32-bit machines ignore the high half!
2541 offset = ConvL2X(offset);
2542 adr = make_unsafe_address(base, offset);
2543 heap_base_oop = base;
2544 val = is_store ? argument(4) : NULL;
2545 } else {
2546 Node* ptr = argument(1); // type: long
2547 ptr = ConvL2X(ptr); // adjust Java long to machine word
2548 adr = make_unsafe_address(NULL, ptr);
2549 val = is_store ? argument(3) : NULL;
2550 }
2551
2552 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2553
2554 // First guess at the value type.
2555 const Type *value_type = Type::get_const_basic_type(type);
2556
2557 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2558 // there was not enough information to nail it down.
2559 Compile::AliasType* alias_type = C->alias_type(adr_type);
2560 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2561
2562 // We will need memory barriers unless we can determine a unique
2563 // alias category for this reference. (Note: If for some reason
2564 // the barriers get omitted and the unsafe reference begins to "pollute"
2565 // the alias analysis of the rest of the graph, either Compile::can_alias
2566 // or Compile::must_alias will throw a diagnostic assert.)
2567 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2568
2569 // If we are reading the value of the referent field of a Reference
2570 // object (either by using Unsafe directly or through reflection)
2571 // then, if G1 is enabled, we need to record the referent in an
2572 // SATB log buffer using the pre-barrier mechanism.
2573 // Also we need to add memory barrier to prevent commoning reads
2574 // from this field across safepoint since GC can change its value.
2575 bool need_read_barrier = !is_native_ptr && !is_store &&
2576 offset != top() && heap_base_oop != top();
2577
2578 if (!is_store && type == T_OBJECT) {
2579 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2580 if (tjp != NULL) {
2581 value_type = tjp;
2582 }
2583 }
2584
2585 receiver = null_check(receiver);
2586 if (stopped()) {
2587 return true;
2588 }
2589 // Heap pointers get a null-check from the interpreter,
2590 // as a courtesy. However, this is not guaranteed by Unsafe,
2591 // and it is not possible to fully distinguish unintended nulls
2592 // from intended ones in this API.
2593
2594 if (is_volatile) {
2595 // We need to emit leading and trailing CPU membars (see below) in
2596 // addition to memory membars when is_volatile. This is a little
2597 // too strong, but avoids the need to insert per-alias-type
2598 // volatile membars (for stores; compare Parse::do_put_xxx), which
2599 // we cannot do effectively here because we probably only have a
2600 // rough approximation of type.
2601 need_mem_bar = true;
2602 // For Stores, place a memory ordering barrier now.
2603 if (is_store) {
2604 insert_mem_bar(Op_MemBarRelease);
2605 } else {
2606 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2607 insert_mem_bar(Op_MemBarVolatile);
2608 }
2609 }
2610 }
2611
2612 // Memory barrier to prevent normal and 'unsafe' accesses from
2613 // bypassing each other. Happens after null checks, so the
2614 // exception paths do not take memory state from the memory barrier,
2615 // so there's no problems making a strong assert about mixing users
2616 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2617 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2618 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2619
2620 if (!is_store) {
2621 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2622 // load value
2623 switch (type) {
2624 case T_BOOLEAN:
2625 case T_CHAR:
2626 case T_BYTE:
2627 case T_SHORT:
2628 case T_INT:
2629 case T_LONG:
2630 case T_FLOAT:
2631 case T_DOUBLE:
2632 break;
2633 case T_OBJECT:
2634 if (need_read_barrier) {
2635 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2636 }
2637 break;
2638 case T_ADDRESS:
2639 // Cast to an int type.
2640 p = _gvn.transform(new CastP2XNode(NULL, p));
2641 p = ConvX2UL(p);
2642 break;
2643 default:
2644 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2645 break;
2646 }
2647 // The load node has the control of the preceding MemBarCPUOrder. All
2648 // following nodes will have the control of the MemBarCPUOrder inserted at
2649 // the end of this method. So, pushing the load onto the stack at a later
2650 // point is fine.
2651 set_result(p);
2652 } else {
2653 // place effect of store into memory
2654 switch (type) {
2655 case T_DOUBLE:
2656 val = dstore_rounding(val);
2657 break;
2658 case T_ADDRESS:
2659 // Repackage the long as a pointer.
2660 val = ConvL2X(val);
2661 val = _gvn.transform(new CastX2PNode(val));
2662 break;
2663 }
2664
2665 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2666 if (type != T_OBJECT ) {
2667 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2668 } else {
2669 // Possibly an oop being stored to Java heap or native memory
2670 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2671 // oop to Java heap.
2672 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2673 } else {
2674 // We can't tell at compile time if we are storing in the Java heap or outside
2675 // of it. So we need to emit code to conditionally do the proper type of
2676 // store.
2677
2678 IdealKit ideal(this);
2679 #define __ ideal.
2680 // QQQ who knows what probability is here??
2681 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2682 // Sync IdealKit and graphKit.
2683 sync_kit(ideal);
2684 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2685 // Update IdealKit memory.
2686 __ sync_kit(this);
2687 } __ else_(); {
2688 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2689 } __ end_if();
2690 // Final sync IdealKit and GraphKit.
2691 final_sync(ideal);
2692 #undef __
2693 }
2694 }
2695 }
2696
2697 if (is_volatile) {
2698 if (!is_store) {
2699 insert_mem_bar(Op_MemBarAcquire);
2700 } else {
2701 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2702 insert_mem_bar(Op_MemBarVolatile);
2703 }
2704 }
2705 }
2706
2707 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2708
2709 return true;
2710 }
2711
2712 //----------------------------inline_unsafe_prefetch----------------------------
2713
2714 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2715 #ifndef PRODUCT
2716 {
2717 ResourceMark rm;
2718 // Check the signatures.
2719 ciSignature* sig = callee()->signature();
2720 #ifdef ASSERT
2721 // Object getObject(Object base, int/long offset), etc.
2722 BasicType rtype = sig->return_type()->basic_type();
2723 if (!is_native_ptr) {
2724 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2725 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2726 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2727 } else {
2728 assert(sig->count() == 1, "native prefetch has 1 argument");
2729 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2730 }
2731 #endif // ASSERT
2732 }
2733 #endif // !PRODUCT
2734
2735 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2736
2737 const int idx = is_static ? 0 : 1;
2738 if (!is_static) {
2739 null_check_receiver();
2740 if (stopped()) {
2741 return true;
2742 }
2743 }
2744
2745 // Build address expression. See the code in inline_unsafe_access.
2746 Node *adr;
2747 if (!is_native_ptr) {
2748 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2749 Node* base = argument(idx + 0); // type: oop
2750 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2751 Node* offset = argument(idx + 1); // type: long
2752 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2753 // to be plain byte offsets, which are also the same as those accepted
2754 // by oopDesc::field_base.
2755 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2756 "fieldOffset must be byte-scaled");
2757 // 32-bit machines ignore the high half!
2758 offset = ConvL2X(offset);
2759 adr = make_unsafe_address(base, offset);
2760 } else {
2761 Node* ptr = argument(idx + 0); // type: long
2762 ptr = ConvL2X(ptr); // adjust Java long to machine word
2763 adr = make_unsafe_address(NULL, ptr);
2764 }
2765
2766 // Generate the read or write prefetch
2767 Node *prefetch;
2768 if (is_store) {
2769 prefetch = new PrefetchWriteNode(i_o(), adr);
2770 } else {
2771 prefetch = new PrefetchReadNode(i_o(), adr);
2772 }
2773 prefetch->init_req(0, control());
2774 set_i_o(_gvn.transform(prefetch));
2775
2776 return true;
2777 }
2778
2779 //----------------------------inline_unsafe_load_store----------------------------
2780 // This method serves a couple of different customers (depending on LoadStoreKind):
2781 //
2782 // LS_cmpxchg:
2783 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2784 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2785 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2786 //
2787 // LS_xadd:
2788 // public int getAndAddInt( Object o, long offset, int delta)
2789 // public long getAndAddLong(Object o, long offset, long delta)
2790 //
2791 // LS_xchg:
2792 // int getAndSet(Object o, long offset, int newValue)
2793 // long getAndSet(Object o, long offset, long newValue)
2794 // Object getAndSet(Object o, long offset, Object newValue)
2795 //
2796 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2797 // This basic scheme here is the same as inline_unsafe_access, but
2798 // differs in enough details that combining them would make the code
2799 // overly confusing. (This is a true fact! I originally combined
2800 // them, but even I was confused by it!) As much code/comments as
2801 // possible are retained from inline_unsafe_access though to make
2802 // the correspondences clearer. - dl
2803
2804 if (callee()->is_static()) return false; // caller must have the capability!
2805
2806 #ifndef PRODUCT
2807 BasicType rtype;
2808 {
2809 ResourceMark rm;
2810 // Check the signatures.
2811 ciSignature* sig = callee()->signature();
2812 rtype = sig->return_type()->basic_type();
2813 if (kind == LS_xadd || kind == LS_xchg) {
2814 // Check the signatures.
2815 #ifdef ASSERT
2816 assert(rtype == type, "get and set must return the expected type");
2817 assert(sig->count() == 3, "get and set has 3 arguments");
2818 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2819 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2820 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2821 #endif // ASSERT
2822 } else if (kind == LS_cmpxchg) {
2823 // Check the signatures.
2824 #ifdef ASSERT
2825 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2826 assert(sig->count() == 4, "CAS has 4 arguments");
2827 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2828 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2829 #endif // ASSERT
2830 } else {
2831 ShouldNotReachHere();
2832 }
2833 }
2834 #endif //PRODUCT
2835
2836 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2837
2838 // Get arguments:
2839 Node* receiver = NULL;
2840 Node* base = NULL;
2841 Node* offset = NULL;
2842 Node* oldval = NULL;
2843 Node* newval = NULL;
2844 if (kind == LS_cmpxchg) {
2845 const bool two_slot_type = type2size[type] == 2;
2846 receiver = argument(0); // type: oop
2847 base = argument(1); // type: oop
2848 offset = argument(2); // type: long
2849 oldval = argument(4); // type: oop, int, or long
2850 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2851 } else if (kind == LS_xadd || kind == LS_xchg){
2852 receiver = argument(0); // type: oop
2853 base = argument(1); // type: oop
2854 offset = argument(2); // type: long
2855 oldval = NULL;
2856 newval = argument(4); // type: oop, int, or long
2857 }
2858
2859 // Null check receiver.
2860 receiver = null_check(receiver);
2861 if (stopped()) {
2862 return true;
2863 }
2864
2865 // Build field offset expression.
2866 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2867 // to be plain byte offsets, which are also the same as those accepted
2868 // by oopDesc::field_base.
2869 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2870 // 32-bit machines ignore the high half of long offsets
2871 offset = ConvL2X(offset);
2872 Node* adr = make_unsafe_address(base, offset);
2873 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2874
2875 // For CAS, unlike inline_unsafe_access, there seems no point in
2876 // trying to refine types. Just use the coarse types here.
2877 const Type *value_type = Type::get_const_basic_type(type);
2878 Compile::AliasType* alias_type = C->alias_type(adr_type);
2879 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2880
2881 if (kind == LS_xchg && type == T_OBJECT) {
2882 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2883 if (tjp != NULL) {
2884 value_type = tjp;
2885 }
2886 }
2887
2888 int alias_idx = C->get_alias_index(adr_type);
2889
2890 // Memory-model-wise, a LoadStore acts like a little synchronized
2891 // block, so needs barriers on each side. These don't translate
2892 // into actual barriers on most machines, but we still need rest of
2893 // compiler to respect ordering.
2894
2895 insert_mem_bar(Op_MemBarRelease);
2896 insert_mem_bar(Op_MemBarCPUOrder);
2897
2898 // 4984716: MemBars must be inserted before this
2899 // memory node in order to avoid a false
2900 // dependency which will confuse the scheduler.
2901 Node *mem = memory(alias_idx);
2902
2903 // For now, we handle only those cases that actually exist: ints,
2904 // longs, and Object. Adding others should be straightforward.
2905 Node* load_store;
2906 switch(type) {
2907 case T_INT:
2908 if (kind == LS_xadd) {
2909 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2910 } else if (kind == LS_xchg) {
2911 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2912 } else if (kind == LS_cmpxchg) {
2913 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2914 } else {
2915 ShouldNotReachHere();
2916 }
2917 break;
2918 case T_LONG:
2919 if (kind == LS_xadd) {
2920 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2921 } else if (kind == LS_xchg) {
2922 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2923 } else if (kind == LS_cmpxchg) {
2924 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2925 } else {
2926 ShouldNotReachHere();
2927 }
2928 break;
2929 case T_OBJECT:
2930 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2931 // could be delayed during Parse (for example, in adjust_map_after_if()).
2932 // Execute transformation here to avoid barrier generation in such case.
2933 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2934 newval = _gvn.makecon(TypePtr::NULL_PTR);
2935
2936 // Reference stores need a store barrier.
2937 if (kind == LS_xchg) {
2938 // If pre-barrier must execute before the oop store, old value will require do_load here.
2939 if (!can_move_pre_barrier()) {
2940 pre_barrier(true /* do_load*/,
2941 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2942 NULL /* pre_val*/,
2943 T_OBJECT);
2944 } // Else move pre_barrier to use load_store value, see below.
2945 } else if (kind == LS_cmpxchg) {
2946 // Same as for newval above:
2947 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2948 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2949 }
2950 // The only known value which might get overwritten is oldval.
2951 pre_barrier(false /* do_load */,
2952 control(), NULL, NULL, max_juint, NULL, NULL,
2953 oldval /* pre_val */,
2954 T_OBJECT);
2955 } else {
2956 ShouldNotReachHere();
2957 }
2958
2959 #ifdef _LP64
2960 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2961 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2962 if (kind == LS_xchg) {
2963 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2964 newval_enc, adr_type, value_type->make_narrowoop()));
2965 } else {
2966 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2967 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2968 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2969 newval_enc, oldval_enc));
2970 }
2971 } else
2972 #endif
2973 {
2974 if (kind == LS_xchg) {
2975 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2976 } else {
2977 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2978 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2979 }
2980 }
2981 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2982 break;
2983 default:
2984 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2985 break;
2986 }
2987
2988 // SCMemProjNodes represent the memory state of a LoadStore. Their
2989 // main role is to prevent LoadStore nodes from being optimized away
2990 // when their results aren't used.
2991 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2992 set_memory(proj, alias_idx);
2993
2994 if (type == T_OBJECT && kind == LS_xchg) {
2995 #ifdef _LP64
2996 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2997 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2998 }
2999 #endif
3000 if (can_move_pre_barrier()) {
3001 // Don't need to load pre_val. The old value is returned by load_store.
3002 // The pre_barrier can execute after the xchg as long as no safepoint
3003 // gets inserted between them.
3004 pre_barrier(false /* do_load */,
3005 control(), NULL, NULL, max_juint, NULL, NULL,
3006 load_store /* pre_val */,
3007 T_OBJECT);
3008 }
3009 }
3010
3011 // Add the trailing membar surrounding the access
3012 insert_mem_bar(Op_MemBarCPUOrder);
3013 insert_mem_bar(Op_MemBarAcquire);
3014
3015 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3016 set_result(load_store);
3017 return true;
3018 }
3019
3020 //----------------------------inline_unsafe_ordered_store----------------------
3021 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3022 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3023 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3024 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3025 // This is another variant of inline_unsafe_access, differing in
3026 // that it always issues store-store ("release") barrier and ensures
3027 // store-atomicity (which only matters for "long").
3028
3029 if (callee()->is_static()) return false; // caller must have the capability!
3030
3031 #ifndef PRODUCT
3032 {
3033 ResourceMark rm;
3034 // Check the signatures.
3035 ciSignature* sig = callee()->signature();
3036 #ifdef ASSERT
3037 BasicType rtype = sig->return_type()->basic_type();
3038 assert(rtype == T_VOID, "must return void");
3039 assert(sig->count() == 3, "has 3 arguments");
3040 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3041 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3042 #endif // ASSERT
3043 }
3044 #endif //PRODUCT
3045
3046 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3047
3048 // Get arguments:
3049 Node* receiver = argument(0); // type: oop
3050 Node* base = argument(1); // type: oop
3051 Node* offset = argument(2); // type: long
3052 Node* val = argument(4); // type: oop, int, or long
3053
3054 // Null check receiver.
3055 receiver = null_check(receiver);
3056 if (stopped()) {
3057 return true;
3058 }
3059
3060 // Build field offset expression.
3061 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3062 // 32-bit machines ignore the high half of long offsets
3063 offset = ConvL2X(offset);
3064 Node* adr = make_unsafe_address(base, offset);
3065 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3066 const Type *value_type = Type::get_const_basic_type(type);
3067 Compile::AliasType* alias_type = C->alias_type(adr_type);
3068
3069 insert_mem_bar(Op_MemBarRelease);
3070 insert_mem_bar(Op_MemBarCPUOrder);
3071 // Ensure that the store is atomic for longs:
3072 const bool require_atomic_access = true;
3073 Node* store;
3074 if (type == T_OBJECT) // reference stores need a store barrier.
3075 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3076 else {
3077 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3078 }
3079 insert_mem_bar(Op_MemBarCPUOrder);
3080 return true;
3081 }
3082
3083 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3084 // Regardless of form, don't allow previous ld/st to move down,
3085 // then issue acquire, release, or volatile mem_bar.
3086 insert_mem_bar(Op_MemBarCPUOrder);
3087 switch(id) {
3088 case vmIntrinsics::_loadFence:
3089 insert_mem_bar(Op_LoadFence);
3090 return true;
3091 case vmIntrinsics::_storeFence:
3092 insert_mem_bar(Op_StoreFence);
3093 return true;
3094 case vmIntrinsics::_fullFence:
3095 insert_mem_bar(Op_MemBarVolatile);
3096 return true;
3097 default:
3098 fatal_unexpected_iid(id);
3099 return false;
3100 }
3101 }
3102
3103 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3104 if (!kls->is_Con()) {
3105 return true;
3106 }
3107 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3108 if (klsptr == NULL) {
3109 return true;
3110 }
3111 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3112 // don't need a guard for a klass that is already initialized
3113 return !ik->is_initialized();
3114 }
3115
3116 //----------------------------inline_unsafe_allocate---------------------------
3117 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3118 bool LibraryCallKit::inline_unsafe_allocate() {
3119 if (callee()->is_static()) return false; // caller must have the capability!
3120
3121 null_check_receiver(); // null-check, then ignore
3122 Node* cls = null_check(argument(1));
3123 if (stopped()) return true;
3124
3125 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3126 kls = null_check(kls);
3127 if (stopped()) return true; // argument was like int.class
3128
3129 Node* test = NULL;
3130 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3131 // Note: The argument might still be an illegal value like
3132 // Serializable.class or Object[].class. The runtime will handle it.
3133 // But we must make an explicit check for initialization.
3134 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3135 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3136 // can generate code to load it as unsigned byte.
3137 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3138 Node* bits = intcon(InstanceKlass::fully_initialized);
3139 test = _gvn.transform(new SubINode(inst, bits));
3140 // The 'test' is non-zero if we need to take a slow path.
3141 }
3142
3143 Node* obj = new_instance(kls, test);
3144 set_result(obj);
3145 return true;
3146 }
3147
3148 #ifdef TRACE_HAVE_INTRINSICS
3149 /*
3150 * oop -> myklass
3151 * myklass->trace_id |= USED
3152 * return myklass->trace_id & ~0x3
3153 */
3154 bool LibraryCallKit::inline_native_classID() {
3155 null_check_receiver(); // null-check, then ignore
3156 Node* cls = null_check(argument(1), T_OBJECT);
3157 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3158 kls = null_check(kls, T_OBJECT);
3159 ByteSize offset = TRACE_ID_OFFSET;
3160 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3161 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3162 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3163 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3164 Node* clsused = longcon(0x01l); // set the class bit
3165 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3166
3167 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3168 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3169 set_result(andl);
3170 return true;
3171 }
3172
3173 bool LibraryCallKit::inline_native_threadID() {
3174 Node* tls_ptr = NULL;
3175 Node* cur_thr = generate_current_thread(tls_ptr);
3176 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3177 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3178 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3179
3180 Node* threadid = NULL;
3181 size_t thread_id_size = OSThread::thread_id_size();
3182 if (thread_id_size == (size_t) BytesPerLong) {
3183 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3184 } else if (thread_id_size == (size_t) BytesPerInt) {
3185 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3186 } else {
3187 ShouldNotReachHere();
3188 }
3189 set_result(threadid);
3190 return true;
3191 }
3192 #endif
3193
3194 //------------------------inline_native_time_funcs--------------
3195 // inline code for System.currentTimeMillis() and System.nanoTime()
3196 // these have the same type and signature
3197 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3198 const TypeFunc* tf = OptoRuntime::void_long_Type();
3199 const TypePtr* no_memory_effects = NULL;
3200 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3201 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3202 #ifdef ASSERT
3203 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3204 assert(value_top == top(), "second value must be top");
3205 #endif
3206 set_result(value);
3207 return true;
3208 }
3209
3210 //------------------------inline_native_currentThread------------------
3211 bool LibraryCallKit::inline_native_currentThread() {
3212 Node* junk = NULL;
3213 set_result(generate_current_thread(junk));
3214 return true;
3215 }
3216
3217 //------------------------inline_native_isInterrupted------------------
3218 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3219 bool LibraryCallKit::inline_native_isInterrupted() {
3220 // Add a fast path to t.isInterrupted(clear_int):
3221 // (t == Thread.current() &&
3222 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3223 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3224 // So, in the common case that the interrupt bit is false,
3225 // we avoid making a call into the VM. Even if the interrupt bit
3226 // is true, if the clear_int argument is false, we avoid the VM call.
3227 // However, if the receiver is not currentThread, we must call the VM,
3228 // because there must be some locking done around the operation.
3229
3230 // We only go to the fast case code if we pass two guards.
3231 // Paths which do not pass are accumulated in the slow_region.
3232
3233 enum {
3234 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3235 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3236 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3237 PATH_LIMIT
3238 };
3239
3240 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3241 // out of the function.
3242 insert_mem_bar(Op_MemBarCPUOrder);
3243
3244 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3245 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3246
3247 RegionNode* slow_region = new RegionNode(1);
3248 record_for_igvn(slow_region);
3249
3250 // (a) Receiving thread must be the current thread.
3251 Node* rec_thr = argument(0);
3252 Node* tls_ptr = NULL;
3253 Node* cur_thr = generate_current_thread(tls_ptr);
3254 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3255 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3256
3257 generate_slow_guard(bol_thr, slow_region);
3258
3259 // (b) Interrupt bit on TLS must be false.
3260 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3261 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3262 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3263
3264 // Set the control input on the field _interrupted read to prevent it floating up.
3265 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3266 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3267 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3268
3269 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3270
3271 // First fast path: if (!TLS._interrupted) return false;
3272 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3273 result_rgn->init_req(no_int_result_path, false_bit);
3274 result_val->init_req(no_int_result_path, intcon(0));
3275
3276 // drop through to next case
3277 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3278
3279 #ifndef TARGET_OS_FAMILY_windows
3280 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3281 Node* clr_arg = argument(1);
3282 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3283 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3284 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3285
3286 // Second fast path: ... else if (!clear_int) return true;
3287 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3288 result_rgn->init_req(no_clear_result_path, false_arg);
3289 result_val->init_req(no_clear_result_path, intcon(1));
3290
3291 // drop through to next case
3292 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3293 #else
3294 // To return true on Windows you must read the _interrupted field
3295 // and check the the event state i.e. take the slow path.
3296 #endif // TARGET_OS_FAMILY_windows
3297
3298 // (d) Otherwise, go to the slow path.
3299 slow_region->add_req(control());
3300 set_control( _gvn.transform(slow_region));
3301
3302 if (stopped()) {
3303 // There is no slow path.
3304 result_rgn->init_req(slow_result_path, top());
3305 result_val->init_req(slow_result_path, top());
3306 } else {
3307 // non-virtual because it is a private non-static
3308 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3309
3310 Node* slow_val = set_results_for_java_call(slow_call);
3311 // this->control() comes from set_results_for_java_call
3312
3313 Node* fast_io = slow_call->in(TypeFunc::I_O);
3314 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3315
3316 // These two phis are pre-filled with copies of of the fast IO and Memory
3317 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3318 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3319
3320 result_rgn->init_req(slow_result_path, control());
3321 result_io ->init_req(slow_result_path, i_o());
3322 result_mem->init_req(slow_result_path, reset_memory());
3323 result_val->init_req(slow_result_path, slow_val);
3324
3325 set_all_memory(_gvn.transform(result_mem));
3326 set_i_o( _gvn.transform(result_io));
3327 }
3328
3329 C->set_has_split_ifs(true); // Has chance for split-if optimization
3330 set_result(result_rgn, result_val);
3331 return true;
3332 }
3333
3334 //---------------------------load_mirror_from_klass----------------------------
3335 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3336 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3337 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3338 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3339 }
3340
3341 //-----------------------load_klass_from_mirror_common-------------------------
3342 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3343 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3344 // and branch to the given path on the region.
3345 // If never_see_null, take an uncommon trap on null, so we can optimistically
3346 // compile for the non-null case.
3347 // If the region is NULL, force never_see_null = true.
3348 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3349 bool never_see_null,
3350 RegionNode* region,
3351 int null_path,
3352 int offset) {
3353 if (region == NULL) never_see_null = true;
3354 Node* p = basic_plus_adr(mirror, offset);
3355 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3356 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3357 Node* null_ctl = top();
3358 kls = null_check_oop(kls, &null_ctl, never_see_null);
3359 if (region != NULL) {
3360 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3361 region->init_req(null_path, null_ctl);
3362 } else {
3363 assert(null_ctl == top(), "no loose ends");
3364 }
3365 return kls;
3366 }
3367
3368 //--------------------(inline_native_Class_query helpers)---------------------
3369 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3370 // Fall through if (mods & mask) == bits, take the guard otherwise.
3371 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3372 // Branch around if the given klass has the given modifier bit set.
3373 // Like generate_guard, adds a new path onto the region.
3374 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3375 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3376 Node* mask = intcon(modifier_mask);
3377 Node* bits = intcon(modifier_bits);
3378 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3379 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3380 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3381 return generate_fair_guard(bol, region);
3382 }
3383 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3384 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3385 }
3386
3387 //-------------------------inline_native_Class_query-------------------
3388 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3389 const Type* return_type = TypeInt::BOOL;
3390 Node* prim_return_value = top(); // what happens if it's a primitive class?
3391 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3392 bool expect_prim = false; // most of these guys expect to work on refs
3393
3394 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3395
3396 Node* mirror = argument(0);
3397 Node* obj = top();
3398
3399 switch (id) {
3400 case vmIntrinsics::_isInstance:
3401 // nothing is an instance of a primitive type
3402 prim_return_value = intcon(0);
3403 obj = argument(1);
3404 break;
3405 case vmIntrinsics::_getModifiers:
3406 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3407 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3408 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3409 break;
3410 case vmIntrinsics::_isInterface:
3411 prim_return_value = intcon(0);
3412 break;
3413 case vmIntrinsics::_isArray:
3414 prim_return_value = intcon(0);
3415 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3416 break;
3417 case vmIntrinsics::_isPrimitive:
3418 prim_return_value = intcon(1);
3419 expect_prim = true; // obviously
3420 break;
3421 case vmIntrinsics::_getSuperclass:
3422 prim_return_value = null();
3423 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3424 break;
3425 case vmIntrinsics::_getComponentType:
3426 prim_return_value = null();
3427 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3428 break;
3429 case vmIntrinsics::_getClassAccessFlags:
3430 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3431 return_type = TypeInt::INT; // not bool! 6297094
3432 break;
3433 default:
3434 fatal_unexpected_iid(id);
3435 break;
3436 }
3437
3438 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3439 if (mirror_con == NULL) return false; // cannot happen?
3440
3441 #ifndef PRODUCT
3442 if (C->print_intrinsics() || C->print_inlining()) {
3443 ciType* k = mirror_con->java_mirror_type();
3444 if (k) {
3445 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3446 k->print_name();
3447 tty->cr();
3448 }
3449 }
3450 #endif
3451
3452 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3453 RegionNode* region = new RegionNode(PATH_LIMIT);
3454 record_for_igvn(region);
3455 PhiNode* phi = new PhiNode(region, return_type);
3456
3457 // The mirror will never be null of Reflection.getClassAccessFlags, however
3458 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3459 // if it is. See bug 4774291.
3460
3461 // For Reflection.getClassAccessFlags(), the null check occurs in
3462 // the wrong place; see inline_unsafe_access(), above, for a similar
3463 // situation.
3464 mirror = null_check(mirror);
3465 // If mirror or obj is dead, only null-path is taken.
3466 if (stopped()) return true;
3467
3468 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3469
3470 // Now load the mirror's klass metaobject, and null-check it.
3471 // Side-effects region with the control path if the klass is null.
3472 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3473 // If kls is null, we have a primitive mirror.
3474 phi->init_req(_prim_path, prim_return_value);
3475 if (stopped()) { set_result(region, phi); return true; }
3476 bool safe_for_replace = (region->in(_prim_path) == top());
3477
3478 Node* p; // handy temp
3479 Node* null_ctl;
3480
3481 // Now that we have the non-null klass, we can perform the real query.
3482 // For constant classes, the query will constant-fold in LoadNode::Value.
3483 Node* query_value = top();
3484 switch (id) {
3485 case vmIntrinsics::_isInstance:
3486 // nothing is an instance of a primitive type
3487 query_value = gen_instanceof(obj, kls, safe_for_replace);
3488 break;
3489
3490 case vmIntrinsics::_getModifiers:
3491 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3492 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3493 break;
3494
3495 case vmIntrinsics::_isInterface:
3496 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3497 if (generate_interface_guard(kls, region) != NULL)
3498 // A guard was added. If the guard is taken, it was an interface.
3499 phi->add_req(intcon(1));
3500 // If we fall through, it's a plain class.
3501 query_value = intcon(0);
3502 break;
3503
3504 case vmIntrinsics::_isArray:
3505 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3506 if (generate_array_guard(kls, region) != NULL)
3507 // A guard was added. If the guard is taken, it was an array.
3508 phi->add_req(intcon(1));
3509 // If we fall through, it's a plain class.
3510 query_value = intcon(0);
3511 break;
3512
3513 case vmIntrinsics::_isPrimitive:
3514 query_value = intcon(0); // "normal" path produces false
3515 break;
3516
3517 case vmIntrinsics::_getSuperclass:
3518 // The rules here are somewhat unfortunate, but we can still do better
3519 // with random logic than with a JNI call.
3520 // Interfaces store null or Object as _super, but must report null.
3521 // Arrays store an intermediate super as _super, but must report Object.
3522 // Other types can report the actual _super.
3523 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3524 if (generate_interface_guard(kls, region) != NULL)
3525 // A guard was added. If the guard is taken, it was an interface.
3526 phi->add_req(null());
3527 if (generate_array_guard(kls, region) != NULL)
3528 // A guard was added. If the guard is taken, it was an array.
3529 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3530 // If we fall through, it's a plain class. Get its _super.
3531 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3532 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3533 null_ctl = top();
3534 kls = null_check_oop(kls, &null_ctl);
3535 if (null_ctl != top()) {
3536 // If the guard is taken, Object.superClass is null (both klass and mirror).
3537 region->add_req(null_ctl);
3538 phi ->add_req(null());
3539 }
3540 if (!stopped()) {
3541 query_value = load_mirror_from_klass(kls);
3542 }
3543 break;
3544
3545 case vmIntrinsics::_getComponentType:
3546 if (generate_array_guard(kls, region) != NULL) {
3547 // Be sure to pin the oop load to the guard edge just created:
3548 Node* is_array_ctrl = region->in(region->req()-1);
3549 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3550 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3551 phi->add_req(cmo);
3552 }
3553 query_value = null(); // non-array case is null
3554 break;
3555
3556 case vmIntrinsics::_getClassAccessFlags:
3557 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3558 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3559 break;
3560
3561 default:
3562 fatal_unexpected_iid(id);
3563 break;
3564 }
3565
3566 // Fall-through is the normal case of a query to a real class.
3567 phi->init_req(1, query_value);
3568 region->init_req(1, control());
3569
3570 C->set_has_split_ifs(true); // Has chance for split-if optimization
3571 set_result(region, phi);
3572 return true;
3573 }
3574
3575 //--------------------------inline_native_subtype_check------------------------
3576 // This intrinsic takes the JNI calls out of the heart of
3577 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3578 bool LibraryCallKit::inline_native_subtype_check() {
3579 // Pull both arguments off the stack.
3580 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3581 args[0] = argument(0);
3582 args[1] = argument(1);
3583 Node* klasses[2]; // corresponding Klasses: superk, subk
3584 klasses[0] = klasses[1] = top();
3585
3586 enum {
3587 // A full decision tree on {superc is prim, subc is prim}:
3588 _prim_0_path = 1, // {P,N} => false
3589 // {P,P} & superc!=subc => false
3590 _prim_same_path, // {P,P} & superc==subc => true
3591 _prim_1_path, // {N,P} => false
3592 _ref_subtype_path, // {N,N} & subtype check wins => true
3593 _both_ref_path, // {N,N} & subtype check loses => false
3594 PATH_LIMIT
3595 };
3596
3597 RegionNode* region = new RegionNode(PATH_LIMIT);
3598 Node* phi = new PhiNode(region, TypeInt::BOOL);
3599 record_for_igvn(region);
3600
3601 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3602 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3603 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3604
3605 // First null-check both mirrors and load each mirror's klass metaobject.
3606 int which_arg;
3607 for (which_arg = 0; which_arg <= 1; which_arg++) {
3608 Node* arg = args[which_arg];
3609 arg = null_check(arg);
3610 if (stopped()) break;
3611 args[which_arg] = arg;
3612
3613 Node* p = basic_plus_adr(arg, class_klass_offset);
3614 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3615 klasses[which_arg] = _gvn.transform(kls);
3616 }
3617
3618 // Having loaded both klasses, test each for null.
3619 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3620 for (which_arg = 0; which_arg <= 1; which_arg++) {
3621 Node* kls = klasses[which_arg];
3622 Node* null_ctl = top();
3623 kls = null_check_oop(kls, &null_ctl, never_see_null);
3624 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3625 region->init_req(prim_path, null_ctl);
3626 if (stopped()) break;
3627 klasses[which_arg] = kls;
3628 }
3629
3630 if (!stopped()) {
3631 // now we have two reference types, in klasses[0..1]
3632 Node* subk = klasses[1]; // the argument to isAssignableFrom
3633 Node* superk = klasses[0]; // the receiver
3634 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3635 // now we have a successful reference subtype check
3636 region->set_req(_ref_subtype_path, control());
3637 }
3638
3639 // If both operands are primitive (both klasses null), then
3640 // we must return true when they are identical primitives.
3641 // It is convenient to test this after the first null klass check.
3642 set_control(region->in(_prim_0_path)); // go back to first null check
3643 if (!stopped()) {
3644 // Since superc is primitive, make a guard for the superc==subc case.
3645 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3646 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3647 generate_guard(bol_eq, region, PROB_FAIR);
3648 if (region->req() == PATH_LIMIT+1) {
3649 // A guard was added. If the added guard is taken, superc==subc.
3650 region->swap_edges(PATH_LIMIT, _prim_same_path);
3651 region->del_req(PATH_LIMIT);
3652 }
3653 region->set_req(_prim_0_path, control()); // Not equal after all.
3654 }
3655
3656 // these are the only paths that produce 'true':
3657 phi->set_req(_prim_same_path, intcon(1));
3658 phi->set_req(_ref_subtype_path, intcon(1));
3659
3660 // pull together the cases:
3661 assert(region->req() == PATH_LIMIT, "sane region");
3662 for (uint i = 1; i < region->req(); i++) {
3663 Node* ctl = region->in(i);
3664 if (ctl == NULL || ctl == top()) {
3665 region->set_req(i, top());
3666 phi ->set_req(i, top());
3667 } else if (phi->in(i) == NULL) {
3668 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3669 }
3670 }
3671
3672 set_control(_gvn.transform(region));
3673 set_result(_gvn.transform(phi));
3674 return true;
3675 }
3676
3677 //---------------------generate_array_guard_common------------------------
3678 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3679 bool obj_array, bool not_array) {
3680 // If obj_array/non_array==false/false:
3681 // Branch around if the given klass is in fact an array (either obj or prim).
3682 // If obj_array/non_array==false/true:
3683 // Branch around if the given klass is not an array klass of any kind.
3684 // If obj_array/non_array==true/true:
3685 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3686 // If obj_array/non_array==true/false:
3687 // Branch around if the kls is an oop array (Object[] or subtype)
3688 //
3689 // Like generate_guard, adds a new path onto the region.
3690 jint layout_con = 0;
3691 Node* layout_val = get_layout_helper(kls, layout_con);
3692 if (layout_val == NULL) {
3693 bool query = (obj_array
3694 ? Klass::layout_helper_is_objArray(layout_con)
3695 : Klass::layout_helper_is_array(layout_con));
3696 if (query == not_array) {
3697 return NULL; // never a branch
3698 } else { // always a branch
3699 Node* always_branch = control();
3700 if (region != NULL)
3701 region->add_req(always_branch);
3702 set_control(top());
3703 return always_branch;
3704 }
3705 }
3706 // Now test the correct condition.
3707 jint nval = (obj_array
3708 ? ((jint)Klass::_lh_array_tag_type_value
3709 << Klass::_lh_array_tag_shift)
3710 : Klass::_lh_neutral_value);
3711 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3712 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3713 // invert the test if we are looking for a non-array
3714 if (not_array) btest = BoolTest(btest).negate();
3715 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3716 return generate_fair_guard(bol, region);
3717 }
3718
3719
3720 //-----------------------inline_native_newArray--------------------------
3721 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3722 bool LibraryCallKit::inline_native_newArray() {
3723 Node* mirror = argument(0);
3724 Node* count_val = argument(1);
3725
3726 mirror = null_check(mirror);
3727 // If mirror or obj is dead, only null-path is taken.
3728 if (stopped()) return true;
3729
3730 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3731 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3732 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3733 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3734 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3735
3736 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3737 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3738 result_reg, _slow_path);
3739 Node* normal_ctl = control();
3740 Node* no_array_ctl = result_reg->in(_slow_path);
3741
3742 // Generate code for the slow case. We make a call to newArray().
3743 set_control(no_array_ctl);
3744 if (!stopped()) {
3745 // Either the input type is void.class, or else the
3746 // array klass has not yet been cached. Either the
3747 // ensuing call will throw an exception, or else it
3748 // will cache the array klass for next time.
3749 PreserveJVMState pjvms(this);
3750 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3751 Node* slow_result = set_results_for_java_call(slow_call);
3752 // this->control() comes from set_results_for_java_call
3753 result_reg->set_req(_slow_path, control());
3754 result_val->set_req(_slow_path, slow_result);
3755 result_io ->set_req(_slow_path, i_o());
3756 result_mem->set_req(_slow_path, reset_memory());
3757 }
3758
3759 set_control(normal_ctl);
3760 if (!stopped()) {
3761 // Normal case: The array type has been cached in the java.lang.Class.
3762 // The following call works fine even if the array type is polymorphic.
3763 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3764 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3765 result_reg->init_req(_normal_path, control());
3766 result_val->init_req(_normal_path, obj);
3767 result_io ->init_req(_normal_path, i_o());
3768 result_mem->init_req(_normal_path, reset_memory());
3769 }
3770
3771 // Return the combined state.
3772 set_i_o( _gvn.transform(result_io) );
3773 set_all_memory( _gvn.transform(result_mem));
3774
3775 C->set_has_split_ifs(true); // Has chance for split-if optimization
3776 set_result(result_reg, result_val);
3777 return true;
3778 }
3779
3780 //----------------------inline_native_getLength--------------------------
3781 // public static native int java.lang.reflect.Array.getLength(Object array);
3782 bool LibraryCallKit::inline_native_getLength() {
3783 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3784
3785 Node* array = null_check(argument(0));
3786 // If array is dead, only null-path is taken.
3787 if (stopped()) return true;
3788
3789 // Deoptimize if it is a non-array.
3790 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3791
3792 if (non_array != NULL) {
3793 PreserveJVMState pjvms(this);
3794 set_control(non_array);
3795 uncommon_trap(Deoptimization::Reason_intrinsic,
3796 Deoptimization::Action_maybe_recompile);
3797 }
3798
3799 // If control is dead, only non-array-path is taken.
3800 if (stopped()) return true;
3801
3802 // The works fine even if the array type is polymorphic.
3803 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3804 Node* result = load_array_length(array);
3805
3806 C->set_has_split_ifs(true); // Has chance for split-if optimization
3807 set_result(result);
3808 return true;
3809 }
3810
3811 //------------------------inline_array_copyOf----------------------------
3812 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3813 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3814 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3815 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3816
3817 // Get the arguments.
3818 Node* original = argument(0);
3819 Node* start = is_copyOfRange? argument(1): intcon(0);
3820 Node* end = is_copyOfRange? argument(2): argument(1);
3821 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3822
3823 Node* newcopy;
3824
3825 // Set the original stack and the reexecute bit for the interpreter to reexecute
3826 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3827 { PreserveReexecuteState preexecs(this);
3828 jvms()->set_should_reexecute(true);
3829
3830 array_type_mirror = null_check(array_type_mirror);
3831 original = null_check(original);
3832
3833 // Check if a null path was taken unconditionally.
3834 if (stopped()) return true;
3835
3836 Node* orig_length = load_array_length(original);
3837
3838 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3839 klass_node = null_check(klass_node);
3840
3841 RegionNode* bailout = new RegionNode(1);
3842 record_for_igvn(bailout);
3843
3844 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3845 // Bail out if that is so.
3846 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3847 if (not_objArray != NULL) {
3848 // Improve the klass node's type from the new optimistic assumption:
3849 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3850 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3851 Node* cast = new CastPPNode(klass_node, akls);
3852 cast->init_req(0, control());
3853 klass_node = _gvn.transform(cast);
3854 }
3855
3856 // Bail out if either start or end is negative.
3857 generate_negative_guard(start, bailout, &start);
3858 generate_negative_guard(end, bailout, &end);
3859
3860 Node* length = end;
3861 if (_gvn.type(start) != TypeInt::ZERO) {
3862 length = _gvn.transform(new SubINode(end, start));
3863 }
3864
3865 // Bail out if length is negative.
3866 // Without this the new_array would throw
3867 // NegativeArraySizeException but IllegalArgumentException is what
3868 // should be thrown
3869 generate_negative_guard(length, bailout, &length);
3870
3871 if (bailout->req() > 1) {
3872 PreserveJVMState pjvms(this);
3873 set_control(_gvn.transform(bailout));
3874 uncommon_trap(Deoptimization::Reason_intrinsic,
3875 Deoptimization::Action_maybe_recompile);
3876 }
3877
3878 if (!stopped()) {
3879 // How many elements will we copy from the original?
3880 // The answer is MinI(orig_length - start, length).
3881 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3882 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3883
3884 newcopy = new_array(klass_node, length, 0); // no argments to push
3885
3886 // Generate a direct call to the right arraycopy function(s).
3887 // We know the copy is disjoint but we might not know if the
3888 // oop stores need checking.
3889 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3890 // This will fail a store-check if x contains any non-nulls.
3891 bool disjoint_bases = true;
3892 // if start > orig_length then the length of the copy may be
3893 // negative.
3894 bool length_never_negative = !is_copyOfRange;
3895 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3896 original, start, newcopy, intcon(0), moved,
3897 disjoint_bases, length_never_negative);
3898 }
3899 } // original reexecute is set back here
3900
3901 C->set_has_split_ifs(true); // Has chance for split-if optimization
3902 if (!stopped()) {
3903 set_result(newcopy);
3904 }
3905 return true;
3906 }
3907
3908
3909 //----------------------generate_virtual_guard---------------------------
3910 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3911 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3912 RegionNode* slow_region) {
3913 ciMethod* method = callee();
3914 int vtable_index = method->vtable_index();
3915 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3916 err_msg_res("bad index %d", vtable_index));
3917 // Get the Method* out of the appropriate vtable entry.
3918 int entry_offset = (InstanceKlass::vtable_start_offset() +
3919 vtable_index*vtableEntry::size()) * wordSize +
3920 vtableEntry::method_offset_in_bytes();
3921 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3922 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3923
3924 // Compare the target method with the expected method (e.g., Object.hashCode).
3925 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3926
3927 Node* native_call = makecon(native_call_addr);
3928 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3929 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3930
3931 return generate_slow_guard(test_native, slow_region);
3932 }
3933
3934 //-----------------------generate_method_call----------------------------
3935 // Use generate_method_call to make a slow-call to the real
3936 // method if the fast path fails. An alternative would be to
3937 // use a stub like OptoRuntime::slow_arraycopy_Java.
3938 // This only works for expanding the current library call,
3939 // not another intrinsic. (E.g., don't use this for making an
3940 // arraycopy call inside of the copyOf intrinsic.)
3941 CallJavaNode*
3942 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3943 // When compiling the intrinsic method itself, do not use this technique.
3944 guarantee(callee() != C->method(), "cannot make slow-call to self");
3945
3946 ciMethod* method = callee();
3947 // ensure the JVMS we have will be correct for this call
3948 guarantee(method_id == method->intrinsic_id(), "must match");
3949
3950 const TypeFunc* tf = TypeFunc::make(method);
3951 CallJavaNode* slow_call;
3952 if (is_static) {
3953 assert(!is_virtual, "");
3954 slow_call = new CallStaticJavaNode(C, tf,
3955 SharedRuntime::get_resolve_static_call_stub(),
3956 method, bci());
3957 } else if (is_virtual) {
3958 null_check_receiver();
3959 int vtable_index = Method::invalid_vtable_index;
3960 if (UseInlineCaches) {
3961 // Suppress the vtable call
3962 } else {
3963 // hashCode and clone are not a miranda methods,
3964 // so the vtable index is fixed.
3965 // No need to use the linkResolver to get it.
3966 vtable_index = method->vtable_index();
3967 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3968 err_msg_res("bad index %d", vtable_index));
3969 }
3970 slow_call = new CallDynamicJavaNode(tf,
3971 SharedRuntime::get_resolve_virtual_call_stub(),
3972 method, vtable_index, bci());
3973 } else { // neither virtual nor static: opt_virtual
3974 null_check_receiver();
3975 slow_call = new CallStaticJavaNode(C, tf,
3976 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3977 method, bci());
3978 slow_call->set_optimized_virtual(true);
3979 }
3980 set_arguments_for_java_call(slow_call);
3981 set_edges_for_java_call(slow_call);
3982 return slow_call;
3983 }
3984
3985
3986 /**
3987 * Build special case code for calls to hashCode on an object. This call may
3988 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3989 * slightly different code.
3990 */
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 RegionNode(PATH_LIMIT);
3998 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3999 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4000 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4001 Node* obj = NULL;
4002 if (!is_static) {
4003 // Check for hashing null object
4004 obj = null_check_receiver();
4005 if (stopped()) return true; // unconditionally null
4006 result_reg->init_req(_null_path, top());
4007 result_val->init_req(_null_path, top());
4008 } else {
4009 // Do a null check, and return zero if null.
4010 // System.identityHashCode(null) == 0
4011 obj = argument(0);
4012 Node* null_ctl = top();
4013 obj = null_check_oop(obj, &null_ctl);
4014 result_reg->init_req(_null_path, null_ctl);
4015 result_val->init_req(_null_path, _gvn.intcon(0));
4016 }
4017
4018 // Unconditionally null? Then return right away.
4019 if (stopped()) {
4020 set_control( result_reg->in(_null_path));
4021 if (!stopped())
4022 set_result(result_val->in(_null_path));
4023 return true;
4024 }
4025
4026 // We only go to the fast case code if we pass a number of guards. The
4027 // paths which do not pass are accumulated in the slow_region.
4028 RegionNode* slow_region = new RegionNode(1);
4029 record_for_igvn(slow_region);
4030
4031 // If this is a virtual call, we generate a funny guard. We pull out
4032 // the vtable entry corresponding to hashCode() from the target object.
4033 // If the target method which we are calling happens to be the native
4034 // Object hashCode() method, we pass the guard. We do not need this
4035 // guard for non-virtual calls -- the caller is known to be the native
4036 // Object hashCode().
4037 if (is_virtual) {
4038 // After null check, get the object's klass.
4039 Node* obj_klass = load_object_klass(obj);
4040 generate_virtual_guard(obj_klass, slow_region);
4041 }
4042
4043 // Get the header out of the object, use LoadMarkNode when available
4044 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4045 // The control of the load must be NULL. Otherwise, the load can move before
4046 // the null check after castPP removal.
4047 Node* no_ctrl = NULL;
4048 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4049
4050 // Test the header to see if it is unlocked.
4051 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4052 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4053 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4054 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4055 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4056
4057 generate_slow_guard(test_unlocked, slow_region);
4058
4059 // Get the hash value and check to see that it has been properly assigned.
4060 // We depend on hash_mask being at most 32 bits and avoid the use of
4061 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4062 // vm: see markOop.hpp.
4063 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4064 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4065 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4066 // This hack lets the hash bits live anywhere in the mark object now, as long
4067 // as the shift drops the relevant bits into the low 32 bits. Note that
4068 // Java spec says that HashCode is an int so there's no point in capturing
4069 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4070 hshifted_header = ConvX2I(hshifted_header);
4071 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4072
4073 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4074 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4075 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4076
4077 generate_slow_guard(test_assigned, slow_region);
4078
4079 Node* init_mem = reset_memory();
4080 // fill in the rest of the null path:
4081 result_io ->init_req(_null_path, i_o());
4082 result_mem->init_req(_null_path, init_mem);
4083
4084 result_val->init_req(_fast_path, hash_val);
4085 result_reg->init_req(_fast_path, control());
4086 result_io ->init_req(_fast_path, i_o());
4087 result_mem->init_req(_fast_path, init_mem);
4088
4089 // Generate code for the slow case. We make a call to hashCode().
4090 set_control(_gvn.transform(slow_region));
4091 if (!stopped()) {
4092 // No need for PreserveJVMState, because we're using up the present state.
4093 set_all_memory(init_mem);
4094 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4095 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4096 Node* slow_result = set_results_for_java_call(slow_call);
4097 // this->control() comes from set_results_for_java_call
4098 result_reg->init_req(_slow_path, control());
4099 result_val->init_req(_slow_path, slow_result);
4100 result_io ->set_req(_slow_path, i_o());
4101 result_mem ->set_req(_slow_path, reset_memory());
4102 }
4103
4104 // Return the combined state.
4105 set_i_o( _gvn.transform(result_io) );
4106 set_all_memory( _gvn.transform(result_mem));
4107
4108 set_result(result_reg, result_val);
4109 return true;
4110 }
4111
4112 //---------------------------inline_native_getClass----------------------------
4113 // public final native Class<?> java.lang.Object.getClass();
4114 //
4115 // Build special case code for calls to getClass on an object.
4116 bool LibraryCallKit::inline_native_getClass() {
4117 Node* obj = null_check_receiver();
4118 if (stopped()) return true;
4119 set_result(load_mirror_from_klass(load_object_klass(obj)));
4120 return true;
4121 }
4122
4123 //-----------------inline_native_Reflection_getCallerClass---------------------
4124 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4125 //
4126 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4127 //
4128 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4129 // in that it must skip particular security frames and checks for
4130 // caller sensitive methods.
4131 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4132 #ifndef PRODUCT
4133 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4134 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4135 }
4136 #endif
4137
4138 if (!jvms()->has_method()) {
4139 #ifndef PRODUCT
4140 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4141 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4142 }
4143 #endif
4144 return false;
4145 }
4146
4147 // Walk back up the JVM state to find the caller at the required
4148 // depth.
4149 JVMState* caller_jvms = jvms();
4150
4151 // Cf. JVM_GetCallerClass
4152 // NOTE: Start the loop at depth 1 because the current JVM state does
4153 // not include the Reflection.getCallerClass() frame.
4154 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4155 ciMethod* m = caller_jvms->method();
4156 switch (n) {
4157 case 0:
4158 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4159 break;
4160 case 1:
4161 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4162 if (!m->caller_sensitive()) {
4163 #ifndef PRODUCT
4164 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4165 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4166 }
4167 #endif
4168 return false; // bail-out; let JVM_GetCallerClass do the work
4169 }
4170 break;
4171 default:
4172 if (!m->is_ignored_by_security_stack_walk()) {
4173 // We have reached the desired frame; return the holder class.
4174 // Acquire method holder as java.lang.Class and push as constant.
4175 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4176 ciInstance* caller_mirror = caller_klass->java_mirror();
4177 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4178
4179 #ifndef PRODUCT
4180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4181 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());
4182 tty->print_cr(" JVM state at this point:");
4183 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4184 ciMethod* m = jvms()->of_depth(i)->method();
4185 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4186 }
4187 }
4188 #endif
4189 return true;
4190 }
4191 break;
4192 }
4193 }
4194
4195 #ifndef PRODUCT
4196 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4197 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4198 tty->print_cr(" JVM state at this point:");
4199 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4200 ciMethod* m = jvms()->of_depth(i)->method();
4201 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4202 }
4203 }
4204 #endif
4205
4206 return false; // bail-out; let JVM_GetCallerClass do the work
4207 }
4208
4209 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4210 Node* arg = argument(0);
4211 Node* result;
4212
4213 switch (id) {
4214 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4215 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4216 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4217 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4218
4219 case vmIntrinsics::_doubleToLongBits: {
4220 // two paths (plus control) merge in a wood
4221 RegionNode *r = new RegionNode(3);
4222 Node *phi = new PhiNode(r, TypeLong::LONG);
4223
4224 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4225 // Build the boolean node
4226 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4227
4228 // Branch either way.
4229 // NaN case is less traveled, which makes all the difference.
4230 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4231 Node *opt_isnan = _gvn.transform(ifisnan);
4232 assert( opt_isnan->is_If(), "Expect an IfNode");
4233 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4234 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4235
4236 set_control(iftrue);
4237
4238 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4239 Node *slow_result = longcon(nan_bits); // return NaN
4240 phi->init_req(1, _gvn.transform( slow_result ));
4241 r->init_req(1, iftrue);
4242
4243 // Else fall through
4244 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4245 set_control(iffalse);
4246
4247 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4248 r->init_req(2, iffalse);
4249
4250 // Post merge
4251 set_control(_gvn.transform(r));
4252 record_for_igvn(r);
4253
4254 C->set_has_split_ifs(true); // Has chance for split-if optimization
4255 result = phi;
4256 assert(result->bottom_type()->isa_long(), "must be");
4257 break;
4258 }
4259
4260 case vmIntrinsics::_floatToIntBits: {
4261 // two paths (plus control) merge in a wood
4262 RegionNode *r = new RegionNode(3);
4263 Node *phi = new PhiNode(r, TypeInt::INT);
4264
4265 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4266 // Build the boolean node
4267 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4268
4269 // Branch either way.
4270 // NaN case is less traveled, which makes all the difference.
4271 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4272 Node *opt_isnan = _gvn.transform(ifisnan);
4273 assert( opt_isnan->is_If(), "Expect an IfNode");
4274 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4275 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4276
4277 set_control(iftrue);
4278
4279 static const jint nan_bits = 0x7fc00000;
4280 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4281 phi->init_req(1, _gvn.transform( slow_result ));
4282 r->init_req(1, iftrue);
4283
4284 // Else fall through
4285 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4286 set_control(iffalse);
4287
4288 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4289 r->init_req(2, iffalse);
4290
4291 // Post merge
4292 set_control(_gvn.transform(r));
4293 record_for_igvn(r);
4294
4295 C->set_has_split_ifs(true); // Has chance for split-if optimization
4296 result = phi;
4297 assert(result->bottom_type()->isa_int(), "must be");
4298 break;
4299 }
4300
4301 default:
4302 fatal_unexpected_iid(id);
4303 break;
4304 }
4305 set_result(_gvn.transform(result));
4306 return true;
4307 }
4308
4309 #ifdef _LP64
4310 #define XTOP ,top() /*additional argument*/
4311 #else //_LP64
4312 #define XTOP /*no additional argument*/
4313 #endif //_LP64
4314
4315 //----------------------inline_unsafe_copyMemory-------------------------
4316 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4317 bool LibraryCallKit::inline_unsafe_copyMemory() {
4318 if (callee()->is_static()) return false; // caller must have the capability!
4319 null_check_receiver(); // null-check receiver
4320 if (stopped()) return true;
4321
4322 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4323
4324 Node* src_ptr = argument(1); // type: oop
4325 Node* src_off = ConvL2X(argument(2)); // type: long
4326 Node* dst_ptr = argument(4); // type: oop
4327 Node* dst_off = ConvL2X(argument(5)); // type: long
4328 Node* size = ConvL2X(argument(7)); // type: long
4329
4330 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4331 "fieldOffset must be byte-scaled");
4332
4333 Node* src = make_unsafe_address(src_ptr, src_off);
4334 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4335
4336 // Conservatively insert a memory barrier on all memory slices.
4337 // Do not let writes of the copy source or destination float below the copy.
4338 insert_mem_bar(Op_MemBarCPUOrder);
4339
4340 // Call it. Note that the length argument is not scaled.
4341 make_runtime_call(RC_LEAF|RC_NO_FP,
4342 OptoRuntime::fast_arraycopy_Type(),
4343 StubRoutines::unsafe_arraycopy(),
4344 "unsafe_arraycopy",
4345 TypeRawPtr::BOTTOM,
4346 src, dst, size XTOP);
4347
4348 // Do not let reads of the copy destination float above the copy.
4349 insert_mem_bar(Op_MemBarCPUOrder);
4350
4351 return true;
4352 }
4353
4354 //------------------------clone_coping-----------------------------------
4355 // Helper function for inline_native_clone.
4356 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4357 assert(obj_size != NULL, "");
4358 Node* raw_obj = alloc_obj->in(1);
4359 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4360
4361 AllocateNode* alloc = NULL;
4362 if (ReduceBulkZeroing) {
4363 // We will be completely responsible for initializing this object -
4364 // mark Initialize node as complete.
4365 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4366 // The object was just allocated - there should be no any stores!
4367 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4368 // Mark as complete_with_arraycopy so that on AllocateNode
4369 // expansion, we know this AllocateNode is initialized by an array
4370 // copy and a StoreStore barrier exists after the array copy.
4371 alloc->initialization()->set_complete_with_arraycopy();
4372 }
4373
4374 // Copy the fastest available way.
4375 // TODO: generate fields copies for small objects instead.
4376 Node* src = obj;
4377 Node* dest = alloc_obj;
4378 Node* size = _gvn.transform(obj_size);
4379
4380 // Exclude the header but include array length to copy by 8 bytes words.
4381 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4382 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4383 instanceOopDesc::base_offset_in_bytes();
4384 // base_off:
4385 // 8 - 32-bit VM
4386 // 12 - 64-bit VM, compressed klass
4387 // 16 - 64-bit VM, normal klass
4388 if (base_off % BytesPerLong != 0) {
4389 assert(UseCompressedClassPointers, "");
4390 if (is_array) {
4391 // Exclude length to copy by 8 bytes words.
4392 base_off += sizeof(int);
4393 } else {
4394 // Include klass to copy by 8 bytes words.
4395 base_off = instanceOopDesc::klass_offset_in_bytes();
4396 }
4397 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4398 }
4399 src = basic_plus_adr(src, base_off);
4400 dest = basic_plus_adr(dest, base_off);
4401
4402 // Compute the length also, if needed:
4403 Node* countx = size;
4404 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4405 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4406
4407 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4408 bool disjoint_bases = true;
4409 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4410 src, NULL, dest, NULL, countx,
4411 /*dest_uninitialized*/true);
4412
4413 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4414 if (card_mark) {
4415 assert(!is_array, "");
4416 // Put in store barrier for any and all oops we are sticking
4417 // into this object. (We could avoid this if we could prove
4418 // that the object type contains no oop fields at all.)
4419 Node* no_particular_value = NULL;
4420 Node* no_particular_field = NULL;
4421 int raw_adr_idx = Compile::AliasIdxRaw;
4422 post_barrier(control(),
4423 memory(raw_adr_type),
4424 alloc_obj,
4425 no_particular_field,
4426 raw_adr_idx,
4427 no_particular_value,
4428 T_OBJECT,
4429 false);
4430 }
4431
4432 // Do not let reads from the cloned object float above the arraycopy.
4433 if (alloc != NULL) {
4434 // Do not let stores that initialize this object be reordered with
4435 // a subsequent store that would make this object accessible by
4436 // other threads.
4437 // Record what AllocateNode this StoreStore protects so that
4438 // escape analysis can go from the MemBarStoreStoreNode to the
4439 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4440 // based on the escape status of the AllocateNode.
4441 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4442 } else {
4443 insert_mem_bar(Op_MemBarCPUOrder);
4444 }
4445 }
4446
4447 //------------------------inline_native_clone----------------------------
4448 // protected native Object java.lang.Object.clone();
4449 //
4450 // Here are the simple edge cases:
4451 // null receiver => normal trap
4452 // virtual and clone was overridden => slow path to out-of-line clone
4453 // not cloneable or finalizer => slow path to out-of-line Object.clone
4454 //
4455 // The general case has two steps, allocation and copying.
4456 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4457 //
4458 // Copying also has two cases, oop arrays and everything else.
4459 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4460 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4461 //
4462 // These steps fold up nicely if and when the cloned object's klass
4463 // can be sharply typed as an object array, a type array, or an instance.
4464 //
4465 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4466 PhiNode* result_val;
4467
4468 // Set the reexecute bit for the interpreter to reexecute
4469 // the bytecode that invokes Object.clone if deoptimization happens.
4470 { PreserveReexecuteState preexecs(this);
4471 jvms()->set_should_reexecute(true);
4472
4473 Node* obj = null_check_receiver();
4474 if (stopped()) return true;
4475
4476 Node* obj_klass = load_object_klass(obj);
4477 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4478 const TypeOopPtr* toop = ((tklass != NULL)
4479 ? tklass->as_instance_type()
4480 : TypeInstPtr::NOTNULL);
4481
4482 // Conservatively insert a memory barrier on all memory slices.
4483 // Do not let writes into the original float below the clone.
4484 insert_mem_bar(Op_MemBarCPUOrder);
4485
4486 // paths into result_reg:
4487 enum {
4488 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4489 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4490 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4491 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4492 PATH_LIMIT
4493 };
4494 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4495 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4496 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4497 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4498 record_for_igvn(result_reg);
4499
4500 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4501 int raw_adr_idx = Compile::AliasIdxRaw;
4502
4503 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4504 if (array_ctl != NULL) {
4505 // It's an array.
4506 PreserveJVMState pjvms(this);
4507 set_control(array_ctl);
4508 Node* obj_length = load_array_length(obj);
4509 Node* obj_size = NULL;
4510 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4511
4512 if (!use_ReduceInitialCardMarks()) {
4513 // If it is an oop array, it requires very special treatment,
4514 // because card marking is required on each card of the array.
4515 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4516 if (is_obja != NULL) {
4517 PreserveJVMState pjvms2(this);
4518 set_control(is_obja);
4519 // Generate a direct call to the right arraycopy function(s).
4520 bool disjoint_bases = true;
4521 bool length_never_negative = true;
4522 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4523 obj, intcon(0), alloc_obj, intcon(0),
4524 obj_length,
4525 disjoint_bases, length_never_negative);
4526 result_reg->init_req(_objArray_path, control());
4527 result_val->init_req(_objArray_path, alloc_obj);
4528 result_i_o ->set_req(_objArray_path, i_o());
4529 result_mem ->set_req(_objArray_path, reset_memory());
4530 }
4531 }
4532 // Otherwise, there are no card marks to worry about.
4533 // (We can dispense with card marks if we know the allocation
4534 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4535 // causes the non-eden paths to take compensating steps to
4536 // simulate a fresh allocation, so that no further
4537 // card marks are required in compiled code to initialize
4538 // the object.)
4539
4540 if (!stopped()) {
4541 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4542
4543 // Present the results of the copy.
4544 result_reg->init_req(_array_path, control());
4545 result_val->init_req(_array_path, alloc_obj);
4546 result_i_o ->set_req(_array_path, i_o());
4547 result_mem ->set_req(_array_path, reset_memory());
4548 }
4549 }
4550
4551 // We only go to the instance fast case code if we pass a number of guards.
4552 // The paths which do not pass are accumulated in the slow_region.
4553 RegionNode* slow_region = new RegionNode(1);
4554 record_for_igvn(slow_region);
4555 if (!stopped()) {
4556 // It's an instance (we did array above). Make the slow-path tests.
4557 // If this is a virtual call, we generate a funny guard. We grab
4558 // the vtable entry corresponding to clone() from the target object.
4559 // If the target method which we are calling happens to be the
4560 // Object clone() method, we pass the guard. We do not need this
4561 // guard for non-virtual calls; the caller is known to be the native
4562 // Object clone().
4563 if (is_virtual) {
4564 generate_virtual_guard(obj_klass, slow_region);
4565 }
4566
4567 // The object must be cloneable and must not have a finalizer.
4568 // Both of these conditions may be checked in a single test.
4569 // We could optimize the cloneable test further, but we don't care.
4570 generate_access_flags_guard(obj_klass,
4571 // Test both conditions:
4572 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4573 // Must be cloneable but not finalizer:
4574 JVM_ACC_IS_CLONEABLE,
4575 slow_region);
4576 }
4577
4578 if (!stopped()) {
4579 // It's an instance, and it passed the slow-path tests.
4580 PreserveJVMState pjvms(this);
4581 Node* obj_size = NULL;
4582 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4583
4584 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4585
4586 // Present the results of the slow call.
4587 result_reg->init_req(_instance_path, control());
4588 result_val->init_req(_instance_path, alloc_obj);
4589 result_i_o ->set_req(_instance_path, i_o());
4590 result_mem ->set_req(_instance_path, reset_memory());
4591 }
4592
4593 // Generate code for the slow case. We make a call to clone().
4594 set_control(_gvn.transform(slow_region));
4595 if (!stopped()) {
4596 PreserveJVMState pjvms(this);
4597 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4598 Node* slow_result = set_results_for_java_call(slow_call);
4599 // this->control() comes from set_results_for_java_call
4600 result_reg->init_req(_slow_path, control());
4601 result_val->init_req(_slow_path, slow_result);
4602 result_i_o ->set_req(_slow_path, i_o());
4603 result_mem ->set_req(_slow_path, reset_memory());
4604 }
4605
4606 // Return the combined state.
4607 set_control( _gvn.transform(result_reg));
4608 set_i_o( _gvn.transform(result_i_o));
4609 set_all_memory( _gvn.transform(result_mem));
4610 } // original reexecute is set back here
4611
4612 set_result(_gvn.transform(result_val));
4613 return true;
4614 }
4615
4616 //------------------------------basictype2arraycopy----------------------------
4617 address LibraryCallKit::basictype2arraycopy(BasicType t,
4618 Node* src_offset,
4619 Node* dest_offset,
4620 bool disjoint_bases,
4621 const char* &name,
4622 bool dest_uninitialized) {
4623 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4624 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4625
4626 bool aligned = false;
4627 bool disjoint = disjoint_bases;
4628
4629 // if the offsets are the same, we can treat the memory regions as
4630 // disjoint, because either the memory regions are in different arrays,
4631 // or they are identical (which we can treat as disjoint.) We can also
4632 // treat a copy with a destination index less that the source index
4633 // as disjoint since a low->high copy will work correctly in this case.
4634 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4635 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4636 // both indices are constants
4637 int s_offs = src_offset_inttype->get_con();
4638 int d_offs = dest_offset_inttype->get_con();
4639 int element_size = type2aelembytes(t);
4640 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4641 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4642 if (s_offs >= d_offs) disjoint = true;
4643 } else if (src_offset == dest_offset && src_offset != NULL) {
4644 // This can occur if the offsets are identical non-constants.
4645 disjoint = true;
4646 }
4647
4648 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4649 }
4650
4651
4652 //------------------------------inline_arraycopy-----------------------
4653 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4654 // Object dest, int destPos,
4655 // int length);
4656 bool LibraryCallKit::inline_arraycopy() {
4657 // Get the arguments.
4658 Node* src = argument(0); // type: oop
4659 Node* src_offset = argument(1); // type: int
4660 Node* dest = argument(2); // type: oop
4661 Node* dest_offset = argument(3); // type: int
4662 Node* length = argument(4); // type: int
4663
4664 // Compile time checks. If any of these checks cannot be verified at compile time,
4665 // we do not make a fast path for this call. Instead, we let the call remain as it
4666 // is. The checks we choose to mandate at compile time are:
4667 //
4668 // (1) src and dest are arrays.
4669 const Type* src_type = src->Value(&_gvn);
4670 const Type* dest_type = dest->Value(&_gvn);
4671 const TypeAryPtr* top_src = src_type->isa_aryptr();
4672 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4673
4674 // Do we have the type of src?
4675 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4676 // Do we have the type of dest?
4677 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4678 // Is the type for src from speculation?
4679 bool src_spec = false;
4680 // Is the type for dest from speculation?
4681 bool dest_spec = false;
4682
4683 if (!has_src || !has_dest) {
4684 // We don't have sufficient type information, let's see if
4685 // speculative types can help. We need to have types for both src
4686 // and dest so that it pays off.
4687
4688 // Do we already have or could we have type information for src
4689 bool could_have_src = has_src;
4690 // Do we already have or could we have type information for dest
4691 bool could_have_dest = has_dest;
4692
4693 ciKlass* src_k = NULL;
4694 if (!has_src) {
4695 src_k = src_type->speculative_type_not_null();
4696 if (src_k != NULL && src_k->is_array_klass()) {
4697 could_have_src = true;
4698 }
4699 }
4700
4701 ciKlass* dest_k = NULL;
4702 if (!has_dest) {
4703 dest_k = dest_type->speculative_type_not_null();
4704 if (dest_k != NULL && dest_k->is_array_klass()) {
4705 could_have_dest = true;
4706 }
4707 }
4708
4709 if (could_have_src && could_have_dest) {
4710 // This is going to pay off so emit the required guards
4711 if (!has_src) {
4712 src = maybe_cast_profiled_obj(src, src_k);
4713 src_type = _gvn.type(src);
4714 top_src = src_type->isa_aryptr();
4715 has_src = (top_src != NULL && top_src->klass() != NULL);
4716 src_spec = true;
4717 }
4718 if (!has_dest) {
4719 dest = maybe_cast_profiled_obj(dest, dest_k);
4720 dest_type = _gvn.type(dest);
4721 top_dest = dest_type->isa_aryptr();
4722 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4723 dest_spec = true;
4724 }
4725 }
4726 }
4727
4728 if (!has_src || !has_dest) {
4729 // Conservatively insert a memory barrier on all memory slices.
4730 // Do not let writes into the source float below the arraycopy.
4731 insert_mem_bar(Op_MemBarCPUOrder);
4732
4733 // Call StubRoutines::generic_arraycopy stub.
4734 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4735 src, src_offset, dest, dest_offset, length);
4736
4737 // Do not let reads from the destination float above the arraycopy.
4738 // Since we cannot type the arrays, we don't know which slices
4739 // might be affected. We could restrict this barrier only to those
4740 // memory slices which pertain to array elements--but don't bother.
4741 if (!InsertMemBarAfterArraycopy)
4742 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4743 insert_mem_bar(Op_MemBarCPUOrder);
4744 return true;
4745 }
4746
4747 // (2) src and dest arrays must have elements of the same BasicType
4748 // Figure out the size and type of the elements we will be copying.
4749 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4750 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4751 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4752 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4753
4754 if (src_elem != dest_elem || dest_elem == T_VOID) {
4755 // The component types are not the same or are not recognized. Punt.
4756 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4757 generate_slow_arraycopy(TypePtr::BOTTOM,
4758 src, src_offset, dest, dest_offset, length,
4759 /*dest_uninitialized*/false);
4760 return true;
4761 }
4762
4763 if (src_elem == T_OBJECT) {
4764 // If both arrays are object arrays then having the exact types
4765 // for both will remove the need for a subtype check at runtime
4766 // before the call and may make it possible to pick a faster copy
4767 // routine (without a subtype check on every element)
4768 // Do we have the exact type of src?
4769 bool could_have_src = src_spec;
4770 // Do we have the exact type of dest?
4771 bool could_have_dest = dest_spec;
4772 ciKlass* src_k = top_src->klass();
4773 ciKlass* dest_k = top_dest->klass();
4774 if (!src_spec) {
4775 src_k = src_type->speculative_type_not_null();
4776 if (src_k != NULL && src_k->is_array_klass()) {
4777 could_have_src = true;
4778 }
4779 }
4780 if (!dest_spec) {
4781 dest_k = dest_type->speculative_type_not_null();
4782 if (dest_k != NULL && dest_k->is_array_klass()) {
4783 could_have_dest = true;
4784 }
4785 }
4786 if (could_have_src && could_have_dest) {
4787 // If we can have both exact types, emit the missing guards
4788 if (could_have_src && !src_spec) {
4789 src = maybe_cast_profiled_obj(src, src_k);
4790 }
4791 if (could_have_dest && !dest_spec) {
4792 dest = maybe_cast_profiled_obj(dest, dest_k);
4793 }
4794 }
4795 }
4796
4797 //---------------------------------------------------------------------------
4798 // We will make a fast path for this call to arraycopy.
4799
4800 // We have the following tests left to perform:
4801 //
4802 // (3) src and dest must not be null.
4803 // (4) src_offset must not be negative.
4804 // (5) dest_offset must not be negative.
4805 // (6) length must not be negative.
4806 // (7) src_offset + length must not exceed length of src.
4807 // (8) dest_offset + length must not exceed length of dest.
4808 // (9) each element of an oop array must be assignable
4809
4810 RegionNode* slow_region = new RegionNode(1);
4811 record_for_igvn(slow_region);
4812
4813 // (3) operands must not be null
4814 // We currently perform our null checks with the null_check routine.
4815 // This means that the null exceptions will be reported in the caller
4816 // rather than (correctly) reported inside of the native arraycopy call.
4817 // This should be corrected, given time. We do our null check with the
4818 // stack pointer restored.
4819 src = null_check(src, T_ARRAY);
4820 dest = null_check(dest, T_ARRAY);
4821
4822 // (4) src_offset must not be negative.
4823 generate_negative_guard(src_offset, slow_region);
4824
4825 // (5) dest_offset must not be negative.
4826 generate_negative_guard(dest_offset, slow_region);
4827
4828 // (6) length must not be negative (moved to generate_arraycopy()).
4829 // generate_negative_guard(length, slow_region);
4830
4831 // (7) src_offset + length must not exceed length of src.
4832 generate_limit_guard(src_offset, length,
4833 load_array_length(src),
4834 slow_region);
4835
4836 // (8) dest_offset + length must not exceed length of dest.
4837 generate_limit_guard(dest_offset, length,
4838 load_array_length(dest),
4839 slow_region);
4840
4841 // (9) each element of an oop array must be assignable
4842 // The generate_arraycopy subroutine checks this.
4843
4844 // This is where the memory effects are placed:
4845 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4846 generate_arraycopy(adr_type, dest_elem,
4847 src, src_offset, dest, dest_offset, length,
4848 false, false, slow_region);
4849
4850 return true;
4851 }
4852
4853 //-----------------------------generate_arraycopy----------------------
4854 // Generate an optimized call to arraycopy.
4855 // Caller must guard against non-arrays.
4856 // Caller must determine a common array basic-type for both arrays.
4857 // Caller must validate offsets against array bounds.
4858 // The slow_region has already collected guard failure paths
4859 // (such as out of bounds length or non-conformable array types).
4860 // The generated code has this shape, in general:
4861 //
4862 // if (length == 0) return // via zero_path
4863 // slowval = -1
4864 // if (types unknown) {
4865 // slowval = call generic copy loop
4866 // if (slowval == 0) return // via checked_path
4867 // } else if (indexes in bounds) {
4868 // if ((is object array) && !(array type check)) {
4869 // slowval = call checked copy loop
4870 // if (slowval == 0) return // via checked_path
4871 // } else {
4872 // call bulk copy loop
4873 // return // via fast_path
4874 // }
4875 // }
4876 // // adjust params for remaining work:
4877 // if (slowval != -1) {
4878 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4879 // }
4880 // slow_region:
4881 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4882 // return // via slow_call_path
4883 //
4884 // This routine is used from several intrinsics: System.arraycopy,
4885 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4886 //
4887 void
4888 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4889 BasicType basic_elem_type,
4890 Node* src, Node* src_offset,
4891 Node* dest, Node* dest_offset,
4892 Node* copy_length,
4893 bool disjoint_bases,
4894 bool length_never_negative,
4895 RegionNode* slow_region) {
4896
4897 if (slow_region == NULL) {
4898 slow_region = new RegionNode(1);
4899 record_for_igvn(slow_region);
4900 }
4901
4902 Node* original_dest = dest;
4903 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4904 bool dest_uninitialized = false;
4905
4906 // See if this is the initialization of a newly-allocated array.
4907 // If so, we will take responsibility here for initializing it to zero.
4908 // (Note: Because tightly_coupled_allocation performs checks on the
4909 // out-edges of the dest, we need to avoid making derived pointers
4910 // from it until we have checked its uses.)
4911 if (ReduceBulkZeroing
4912 && !ZeroTLAB // pointless if already zeroed
4913 && basic_elem_type != T_CONFLICT // avoid corner case
4914 && !src->eqv_uncast(dest)
4915 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4916 != NULL)
4917 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4918 && alloc->maybe_set_complete(&_gvn)) {
4919 // "You break it, you buy it."
4920 InitializeNode* init = alloc->initialization();
4921 assert(init->is_complete(), "we just did this");
4922 init->set_complete_with_arraycopy();
4923 assert(dest->is_CheckCastPP(), "sanity");
4924 assert(dest->in(0)->in(0) == init, "dest pinned");
4925 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4926 // From this point on, every exit path is responsible for
4927 // initializing any non-copied parts of the object to zero.
4928 // Also, if this flag is set we make sure that arraycopy interacts properly
4929 // with G1, eliding pre-barriers. See CR 6627983.
4930 dest_uninitialized = true;
4931 } else {
4932 // No zeroing elimination here.
4933 alloc = NULL;
4934 //original_dest = dest;
4935 //dest_uninitialized = false;
4936 }
4937
4938 // Results are placed here:
4939 enum { fast_path = 1, // normal void-returning assembly stub
4940 checked_path = 2, // special assembly stub with cleanup
4941 slow_call_path = 3, // something went wrong; call the VM
4942 zero_path = 4, // bypass when length of copy is zero
4943 bcopy_path = 5, // copy primitive array by 64-bit blocks
4944 PATH_LIMIT = 6
4945 };
4946 RegionNode* result_region = new RegionNode(PATH_LIMIT);
4947 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
4948 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
4949 record_for_igvn(result_region);
4950 _gvn.set_type_bottom(result_i_o);
4951 _gvn.set_type_bottom(result_memory);
4952 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4953
4954 // The slow_control path:
4955 Node* slow_control;
4956 Node* slow_i_o = i_o();
4957 Node* slow_mem = memory(adr_type);
4958 debug_only(slow_control = (Node*) badAddress);
4959
4960 // Checked control path:
4961 Node* checked_control = top();
4962 Node* checked_mem = NULL;
4963 Node* checked_i_o = NULL;
4964 Node* checked_value = NULL;
4965
4966 if (basic_elem_type == T_CONFLICT) {
4967 assert(!dest_uninitialized, "");
4968 Node* cv = generate_generic_arraycopy(adr_type,
4969 src, src_offset, dest, dest_offset,
4970 copy_length, dest_uninitialized);
4971 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4972 checked_control = control();
4973 checked_i_o = i_o();
4974 checked_mem = memory(adr_type);
4975 checked_value = cv;
4976 set_control(top()); // no fast path
4977 }
4978
4979 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4980 if (not_pos != NULL) {
4981 PreserveJVMState pjvms(this);
4982 set_control(not_pos);
4983
4984 // (6) length must not be negative.
4985 if (!length_never_negative) {
4986 generate_negative_guard(copy_length, slow_region);
4987 }
4988
4989 // copy_length is 0.
4990 if (!stopped() && dest_uninitialized) {
4991 Node* dest_length = alloc->in(AllocateNode::ALength);
4992 if (copy_length->eqv_uncast(dest_length)
4993 || _gvn.find_int_con(dest_length, 1) <= 0) {
4994 // There is no zeroing to do. No need for a secondary raw memory barrier.
4995 } else {
4996 // Clear the whole thing since there are no source elements to copy.
4997 generate_clear_array(adr_type, dest, basic_elem_type,
4998 intcon(0), NULL,
4999 alloc->in(AllocateNode::AllocSize));
5000 // Use a secondary InitializeNode as raw memory barrier.
5001 // Currently it is needed only on this path since other
5002 // paths have stub or runtime calls as raw memory barriers.
5003 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5004 Compile::AliasIdxRaw,
5005 top())->as_Initialize();
5006 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5007 }
5008 }
5009
5010 // Present the results of the fast call.
5011 result_region->init_req(zero_path, control());
5012 result_i_o ->init_req(zero_path, i_o());
5013 result_memory->init_req(zero_path, memory(adr_type));
5014 }
5015
5016 if (!stopped() && dest_uninitialized) {
5017 // We have to initialize the *uncopied* part of the array to zero.
5018 // The copy destination is the slice dest[off..off+len]. The other slices
5019 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5020 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5021 Node* dest_length = alloc->in(AllocateNode::ALength);
5022 Node* dest_tail = _gvn.transform(new AddINode(dest_offset, copy_length));
5023
5024 // If there is a head section that needs zeroing, do it now.
5025 if (find_int_con(dest_offset, -1) != 0) {
5026 generate_clear_array(adr_type, dest, basic_elem_type,
5027 intcon(0), dest_offset,
5028 NULL);
5029 }
5030
5031 // Next, perform a dynamic check on the tail length.
5032 // It is often zero, and we can win big if we prove this.
5033 // There are two wins: Avoid generating the ClearArray
5034 // with its attendant messy index arithmetic, and upgrade
5035 // the copy to a more hardware-friendly word size of 64 bits.
5036 Node* tail_ctl = NULL;
5037 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5038 Node* cmp_lt = _gvn.transform(new CmpINode(dest_tail, dest_length));
5039 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
5040 tail_ctl = generate_slow_guard(bol_lt, NULL);
5041 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5042 }
5043
5044 // At this point, let's assume there is no tail.
5045 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5046 // There is no tail. Try an upgrade to a 64-bit copy.
5047 bool didit = false;
5048 { PreserveJVMState pjvms(this);
5049 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5050 src, src_offset, dest, dest_offset,
5051 dest_size, dest_uninitialized);
5052 if (didit) {
5053 // Present the results of the block-copying fast call.
5054 result_region->init_req(bcopy_path, control());
5055 result_i_o ->init_req(bcopy_path, i_o());
5056 result_memory->init_req(bcopy_path, memory(adr_type));
5057 }
5058 }
5059 if (didit)
5060 set_control(top()); // no regular fast path
5061 }
5062
5063 // Clear the tail, if any.
5064 if (tail_ctl != NULL) {
5065 Node* notail_ctl = stopped() ? NULL : control();
5066 set_control(tail_ctl);
5067 if (notail_ctl == NULL) {
5068 generate_clear_array(adr_type, dest, basic_elem_type,
5069 dest_tail, NULL,
5070 dest_size);
5071 } else {
5072 // Make a local merge.
5073 Node* done_ctl = new RegionNode(3);
5074 Node* done_mem = new PhiNode(done_ctl, Type::MEMORY, adr_type);
5075 done_ctl->init_req(1, notail_ctl);
5076 done_mem->init_req(1, memory(adr_type));
5077 generate_clear_array(adr_type, dest, basic_elem_type,
5078 dest_tail, NULL,
5079 dest_size);
5080 done_ctl->init_req(2, control());
5081 done_mem->init_req(2, memory(adr_type));
5082 set_control( _gvn.transform(done_ctl));
5083 set_memory( _gvn.transform(done_mem), adr_type );
5084 }
5085 }
5086 }
5087
5088 BasicType copy_type = basic_elem_type;
5089 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5090 if (!stopped() && copy_type == T_OBJECT) {
5091 // If src and dest have compatible element types, we can copy bits.
5092 // Types S[] and D[] are compatible if D is a supertype of S.
5093 //
5094 // If they are not, we will use checked_oop_disjoint_arraycopy,
5095 // which performs a fast optimistic per-oop check, and backs off
5096 // further to JVM_ArrayCopy on the first per-oop check that fails.
5097 // (Actually, we don't move raw bits only; the GC requires card marks.)
5098
5099 // Get the Klass* for both src and dest
5100 Node* src_klass = load_object_klass(src);
5101 Node* dest_klass = load_object_klass(dest);
5102
5103 // Generate the subtype check.
5104 // This might fold up statically, or then again it might not.
5105 //
5106 // Non-static example: Copying List<String>.elements to a new String[].
5107 // The backing store for a List<String> is always an Object[],
5108 // but its elements are always type String, if the generic types
5109 // are correct at the source level.
5110 //
5111 // Test S[] against D[], not S against D, because (probably)
5112 // the secondary supertype cache is less busy for S[] than S.
5113 // This usually only matters when D is an interface.
5114 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5115 // Plug failing path into checked_oop_disjoint_arraycopy
5116 if (not_subtype_ctrl != top()) {
5117 PreserveJVMState pjvms(this);
5118 set_control(not_subtype_ctrl);
5119 // (At this point we can assume disjoint_bases, since types differ.)
5120 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5121 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5122 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5123 Node* dest_elem_klass = _gvn.transform(n1);
5124 Node* cv = generate_checkcast_arraycopy(adr_type,
5125 dest_elem_klass,
5126 src, src_offset, dest, dest_offset,
5127 ConvI2X(copy_length), dest_uninitialized);
5128 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5129 checked_control = control();
5130 checked_i_o = i_o();
5131 checked_mem = memory(adr_type);
5132 checked_value = cv;
5133 }
5134 // At this point we know we do not need type checks on oop stores.
5135
5136 // Let's see if we need card marks:
5137 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5138 // If we do not need card marks, copy using the jint or jlong stub.
5139 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5140 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5141 "sizes agree");
5142 }
5143 }
5144
5145 if (!stopped()) {
5146 // Generate the fast path, if possible.
5147 PreserveJVMState pjvms(this);
5148 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5149 src, src_offset, dest, dest_offset,
5150 ConvI2X(copy_length), dest_uninitialized);
5151
5152 // Present the results of the fast call.
5153 result_region->init_req(fast_path, control());
5154 result_i_o ->init_req(fast_path, i_o());
5155 result_memory->init_req(fast_path, memory(adr_type));
5156 }
5157
5158 // Here are all the slow paths up to this point, in one bundle:
5159 slow_control = top();
5160 if (slow_region != NULL)
5161 slow_control = _gvn.transform(slow_region);
5162 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5163
5164 set_control(checked_control);
5165 if (!stopped()) {
5166 // Clean up after the checked call.
5167 // The returned value is either 0 or -1^K,
5168 // where K = number of partially transferred array elements.
5169 Node* cmp = _gvn.transform(new CmpINode(checked_value, intcon(0)));
5170 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
5171 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5172
5173 // If it is 0, we are done, so transfer to the end.
5174 Node* checks_done = _gvn.transform(new IfTrueNode(iff));
5175 result_region->init_req(checked_path, checks_done);
5176 result_i_o ->init_req(checked_path, checked_i_o);
5177 result_memory->init_req(checked_path, checked_mem);
5178
5179 // If it is not zero, merge into the slow call.
5180 set_control( _gvn.transform(new IfFalseNode(iff) ));
5181 RegionNode* slow_reg2 = new RegionNode(3);
5182 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
5183 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
5184 record_for_igvn(slow_reg2);
5185 slow_reg2 ->init_req(1, slow_control);
5186 slow_i_o2 ->init_req(1, slow_i_o);
5187 slow_mem2 ->init_req(1, slow_mem);
5188 slow_reg2 ->init_req(2, control());
5189 slow_i_o2 ->init_req(2, checked_i_o);
5190 slow_mem2 ->init_req(2, checked_mem);
5191
5192 slow_control = _gvn.transform(slow_reg2);
5193 slow_i_o = _gvn.transform(slow_i_o2);
5194 slow_mem = _gvn.transform(slow_mem2);
5195
5196 if (alloc != NULL) {
5197 // We'll restart from the very beginning, after zeroing the whole thing.
5198 // This can cause double writes, but that's OK since dest is brand new.
5199 // So we ignore the low 31 bits of the value returned from the stub.
5200 } else {
5201 // We must continue the copy exactly where it failed, or else
5202 // another thread might see the wrong number of writes to dest.
5203 Node* checked_offset = _gvn.transform(new XorINode(checked_value, intcon(-1)));
5204 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
5205 slow_offset->init_req(1, intcon(0));
5206 slow_offset->init_req(2, checked_offset);
5207 slow_offset = _gvn.transform(slow_offset);
5208
5209 // Adjust the arguments by the conditionally incoming offset.
5210 Node* src_off_plus = _gvn.transform(new AddINode(src_offset, slow_offset));
5211 Node* dest_off_plus = _gvn.transform(new AddINode(dest_offset, slow_offset));
5212 Node* length_minus = _gvn.transform(new SubINode(copy_length, slow_offset));
5213
5214 // Tweak the node variables to adjust the code produced below:
5215 src_offset = src_off_plus;
5216 dest_offset = dest_off_plus;
5217 copy_length = length_minus;
5218 }
5219 }
5220
5221 set_control(slow_control);
5222 if (!stopped()) {
5223 // Generate the slow path, if needed.
5224 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5225
5226 set_memory(slow_mem, adr_type);
5227 set_i_o(slow_i_o);
5228
5229 if (dest_uninitialized) {
5230 generate_clear_array(adr_type, dest, basic_elem_type,
5231 intcon(0), NULL,
5232 alloc->in(AllocateNode::AllocSize));
5233 }
5234
5235 generate_slow_arraycopy(adr_type,
5236 src, src_offset, dest, dest_offset,
5237 copy_length, /*dest_uninitialized*/false);
5238
5239 result_region->init_req(slow_call_path, control());
5240 result_i_o ->init_req(slow_call_path, i_o());
5241 result_memory->init_req(slow_call_path, memory(adr_type));
5242 }
5243
5244 // Remove unused edges.
5245 for (uint i = 1; i < result_region->req(); i++) {
5246 if (result_region->in(i) == NULL)
5247 result_region->init_req(i, top());
5248 }
5249
5250 // Finished; return the combined state.
5251 set_control( _gvn.transform(result_region));
5252 set_i_o( _gvn.transform(result_i_o) );
5253 set_memory( _gvn.transform(result_memory), adr_type );
5254
5255 // The memory edges above are precise in order to model effects around
5256 // array copies accurately to allow value numbering of field loads around
5257 // arraycopy. Such field loads, both before and after, are common in Java
5258 // collections and similar classes involving header/array data structures.
5259 //
5260 // But with low number of register or when some registers are used or killed
5261 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5262 // The next memory barrier is added to avoid it. If the arraycopy can be
5263 // optimized away (which it can, sometimes) then we can manually remove
5264 // the membar also.
5265 //
5266 // Do not let reads from the cloned object float above the arraycopy.
5267 if (alloc != NULL) {
5268 // Do not let stores that initialize this object be reordered with
5269 // a subsequent store that would make this object accessible by
5270 // other threads.
5271 // Record what AllocateNode this StoreStore protects so that
5272 // escape analysis can go from the MemBarStoreStoreNode to the
5273 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5274 // based on the escape status of the AllocateNode.
5275 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5276 } else if (InsertMemBarAfterArraycopy)
5277 insert_mem_bar(Op_MemBarCPUOrder);
5278 }
5279
5280
5281 // Helper function which determines if an arraycopy immediately follows
5282 // an allocation, with no intervening tests or other escapes for the object.
5283 AllocateArrayNode*
5284 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5285 RegionNode* slow_region) {
5286 if (stopped()) return NULL; // no fast path
5287 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5288
5289 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5290 if (alloc == NULL) return NULL;
5291
5292 Node* rawmem = memory(Compile::AliasIdxRaw);
5293 // Is the allocation's memory state untouched?
5294 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5295 // Bail out if there have been raw-memory effects since the allocation.
5296 // (Example: There might have been a call or safepoint.)
5297 return NULL;
5298 }
5299 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5300 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5301 return NULL;
5302 }
5303
5304 // There must be no unexpected observers of this allocation.
5305 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5306 Node* obs = ptr->fast_out(i);
5307 if (obs != this->map()) {
5308 return NULL;
5309 }
5310 }
5311
5312 // This arraycopy must unconditionally follow the allocation of the ptr.
5313 Node* alloc_ctl = ptr->in(0);
5314 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5315
5316 Node* ctl = control();
5317 while (ctl != alloc_ctl) {
5318 // There may be guards which feed into the slow_region.
5319 // Any other control flow means that we might not get a chance
5320 // to finish initializing the allocated object.
5321 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5322 IfNode* iff = ctl->in(0)->as_If();
5323 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5324 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5325 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5326 ctl = iff->in(0); // This test feeds the known slow_region.
5327 continue;
5328 }
5329 // One more try: Various low-level checks bottom out in
5330 // uncommon traps. If the debug-info of the trap omits
5331 // any reference to the allocation, as we've already
5332 // observed, then there can be no objection to the trap.
5333 bool found_trap = false;
5334 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5335 Node* obs = not_ctl->fast_out(j);
5336 if (obs->in(0) == not_ctl && obs->is_Call() &&
5337 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5338 found_trap = true; break;
5339 }
5340 }
5341 if (found_trap) {
5342 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5343 continue;
5344 }
5345 }
5346 return NULL;
5347 }
5348
5349 // If we get this far, we have an allocation which immediately
5350 // precedes the arraycopy, and we can take over zeroing the new object.
5351 // The arraycopy will finish the initialization, and provide
5352 // a new control state to which we will anchor the destination pointer.
5353
5354 return alloc;
5355 }
5356
5357 // Helper for initialization of arrays, creating a ClearArray.
5358 // It writes zero bits in [start..end), within the body of an array object.
5359 // The memory effects are all chained onto the 'adr_type' alias category.
5360 //
5361 // Since the object is otherwise uninitialized, we are free
5362 // to put a little "slop" around the edges of the cleared area,
5363 // as long as it does not go back into the array's header,
5364 // or beyond the array end within the heap.
5365 //
5366 // The lower edge can be rounded down to the nearest jint and the
5367 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5368 //
5369 // Arguments:
5370 // adr_type memory slice where writes are generated
5371 // dest oop of the destination array
5372 // basic_elem_type element type of the destination
5373 // slice_idx array index of first element to store
5374 // slice_len number of elements to store (or NULL)
5375 // dest_size total size in bytes of the array object
5376 //
5377 // Exactly one of slice_len or dest_size must be non-NULL.
5378 // If dest_size is non-NULL, zeroing extends to the end of the object.
5379 // If slice_len is non-NULL, the slice_idx value must be a constant.
5380 void
5381 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5382 Node* dest,
5383 BasicType basic_elem_type,
5384 Node* slice_idx,
5385 Node* slice_len,
5386 Node* dest_size) {
5387 // one or the other but not both of slice_len and dest_size:
5388 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5389 if (slice_len == NULL) slice_len = top();
5390 if (dest_size == NULL) dest_size = top();
5391
5392 // operate on this memory slice:
5393 Node* mem = memory(adr_type); // memory slice to operate on
5394
5395 // scaling and rounding of indexes:
5396 int scale = exact_log2(type2aelembytes(basic_elem_type));
5397 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5398 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5399 int bump_bit = (-1 << scale) & BytesPerInt;
5400
5401 // determine constant starts and ends
5402 const intptr_t BIG_NEG = -128;
5403 assert(BIG_NEG + 2*abase < 0, "neg enough");
5404 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5405 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5406 if (slice_len_con == 0) {
5407 return; // nothing to do here
5408 }
5409 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5410 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5411 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5412 assert(end_con < 0, "not two cons");
5413 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5414 BytesPerLong);
5415 }
5416
5417 if (start_con >= 0 && end_con >= 0) {
5418 // Constant start and end. Simple.
5419 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5420 start_con, end_con, &_gvn);
5421 } else if (start_con >= 0 && dest_size != top()) {
5422 // Constant start, pre-rounded end after the tail of the array.
5423 Node* end = dest_size;
5424 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5425 start_con, end, &_gvn);
5426 } else if (start_con >= 0 && slice_len != top()) {
5427 // Constant start, non-constant end. End needs rounding up.
5428 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5429 intptr_t end_base = abase + (slice_idx_con << scale);
5430 int end_round = (-1 << scale) & (BytesPerLong - 1);
5431 Node* end = ConvI2X(slice_len);
5432 if (scale != 0)
5433 end = _gvn.transform(new LShiftXNode(end, intcon(scale) ));
5434 end_base += end_round;
5435 end = _gvn.transform(new AddXNode(end, MakeConX(end_base)));
5436 end = _gvn.transform(new AndXNode(end, MakeConX(~end_round)));
5437 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5438 start_con, end, &_gvn);
5439 } else if (start_con < 0 && dest_size != top()) {
5440 // Non-constant start, pre-rounded end after the tail of the array.
5441 // This is almost certainly a "round-to-end" operation.
5442 Node* start = slice_idx;
5443 start = ConvI2X(start);
5444 if (scale != 0)
5445 start = _gvn.transform(new LShiftXNode( start, intcon(scale) ));
5446 start = _gvn.transform(new AddXNode(start, MakeConX(abase)));
5447 if ((bump_bit | clear_low) != 0) {
5448 int to_clear = (bump_bit | clear_low);
5449 // Align up mod 8, then store a jint zero unconditionally
5450 // just before the mod-8 boundary.
5451 if (((abase + bump_bit) & ~to_clear) - bump_bit
5452 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5453 bump_bit = 0;
5454 assert((abase & to_clear) == 0, "array base must be long-aligned");
5455 } else {
5456 // Bump 'start' up to (or past) the next jint boundary:
5457 start = _gvn.transform(new AddXNode(start, MakeConX(bump_bit)));
5458 assert((abase & clear_low) == 0, "array base must be int-aligned");
5459 }
5460 // Round bumped 'start' down to jlong boundary in body of array.
5461 start = _gvn.transform(new AndXNode(start, MakeConX(~to_clear)));
5462 if (bump_bit != 0) {
5463 // Store a zero to the immediately preceding jint:
5464 Node* x1 = _gvn.transform(new AddXNode(start, MakeConX(-bump_bit)));
5465 Node* p1 = basic_plus_adr(dest, x1);
5466 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5467 mem = _gvn.transform(mem);
5468 }
5469 }
5470 Node* end = dest_size; // pre-rounded
5471 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5472 start, end, &_gvn);
5473 } else {
5474 // Non-constant start, unrounded non-constant end.
5475 // (Nobody zeroes a random midsection of an array using this routine.)
5476 ShouldNotReachHere(); // fix caller
5477 }
5478
5479 // Done.
5480 set_memory(mem, adr_type);
5481 }
5482
5483
5484 bool
5485 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5486 BasicType basic_elem_type,
5487 AllocateNode* alloc,
5488 Node* src, Node* src_offset,
5489 Node* dest, Node* dest_offset,
5490 Node* dest_size, bool dest_uninitialized) {
5491 // See if there is an advantage from block transfer.
5492 int scale = exact_log2(type2aelembytes(basic_elem_type));
5493 if (scale >= LogBytesPerLong)
5494 return false; // it is already a block transfer
5495
5496 // Look at the alignment of the starting offsets.
5497 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5498
5499 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5500 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5501 if (src_off_con < 0 || dest_off_con < 0)
5502 // At present, we can only understand constants.
5503 return false;
5504
5505 intptr_t src_off = abase + (src_off_con << scale);
5506 intptr_t dest_off = abase + (dest_off_con << scale);
5507
5508 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5509 // Non-aligned; too bad.
5510 // One more chance: Pick off an initial 32-bit word.
5511 // This is a common case, since abase can be odd mod 8.
5512 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5513 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5514 Node* sptr = basic_plus_adr(src, src_off);
5515 Node* dptr = basic_plus_adr(dest, dest_off);
5516 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5517 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5518 src_off += BytesPerInt;
5519 dest_off += BytesPerInt;
5520 } else {
5521 return false;
5522 }
5523 }
5524 assert(src_off % BytesPerLong == 0, "");
5525 assert(dest_off % BytesPerLong == 0, "");
5526
5527 // Do this copy by giant steps.
5528 Node* sptr = basic_plus_adr(src, src_off);
5529 Node* dptr = basic_plus_adr(dest, dest_off);
5530 Node* countx = dest_size;
5531 countx = _gvn.transform(new SubXNode(countx, MakeConX(dest_off)));
5532 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong)));
5533
5534 bool disjoint_bases = true; // since alloc != NULL
5535 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5536 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5537
5538 return true;
5539 }
5540
5541
5542 // Helper function; generates code for the slow case.
5543 // We make a call to a runtime method which emulates the native method,
5544 // but without the native wrapper overhead.
5545 void
5546 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5547 Node* src, Node* src_offset,
5548 Node* dest, Node* dest_offset,
5549 Node* copy_length, bool dest_uninitialized) {
5550 assert(!dest_uninitialized, "Invariant");
5551 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5552 OptoRuntime::slow_arraycopy_Type(),
5553 OptoRuntime::slow_arraycopy_Java(),
5554 "slow_arraycopy", adr_type,
5555 src, src_offset, dest, dest_offset,
5556 copy_length);
5557
5558 // Handle exceptions thrown by this fellow:
5559 make_slow_call_ex(call, env()->Throwable_klass(), false);
5560 }
5561
5562 // Helper function; generates code for cases requiring runtime checks.
5563 Node*
5564 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5565 Node* dest_elem_klass,
5566 Node* src, Node* src_offset,
5567 Node* dest, Node* dest_offset,
5568 Node* copy_length, bool dest_uninitialized) {
5569 if (stopped()) return NULL;
5570
5571 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5572 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5573 return NULL;
5574 }
5575
5576 // Pick out the parameters required to perform a store-check
5577 // for the target array. This is an optimistic check. It will
5578 // look in each non-null element's class, at the desired klass's
5579 // super_check_offset, for the desired klass.
5580 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5581 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5582 Node* n3 = new LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5583 Node* check_offset = ConvI2X(_gvn.transform(n3));
5584 Node* check_value = dest_elem_klass;
5585
5586 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5587 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5588
5589 // (We know the arrays are never conjoint, because their types differ.)
5590 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5591 OptoRuntime::checkcast_arraycopy_Type(),
5592 copyfunc_addr, "checkcast_arraycopy", adr_type,
5593 // five arguments, of which two are
5594 // intptr_t (jlong in LP64)
5595 src_start, dest_start,
5596 copy_length XTOP,
5597 check_offset XTOP,
5598 check_value);
5599
5600 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5601 }
5602
5603
5604 // Helper function; generates code for cases requiring runtime checks.
5605 Node*
5606 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5607 Node* src, Node* src_offset,
5608 Node* dest, Node* dest_offset,
5609 Node* copy_length, bool dest_uninitialized) {
5610 assert(!dest_uninitialized, "Invariant");
5611 if (stopped()) return NULL;
5612 address copyfunc_addr = StubRoutines::generic_arraycopy();
5613 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5614 return NULL;
5615 }
5616
5617 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5618 OptoRuntime::generic_arraycopy_Type(),
5619 copyfunc_addr, "generic_arraycopy", adr_type,
5620 src, src_offset, dest, dest_offset, copy_length);
5621
5622 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5623 }
5624
5625 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5626 void
5627 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5628 BasicType basic_elem_type,
5629 bool disjoint_bases,
5630 Node* src, Node* src_offset,
5631 Node* dest, Node* dest_offset,
5632 Node* copy_length, bool dest_uninitialized) {
5633 if (stopped()) return; // nothing to do
5634
5635 Node* src_start = src;
5636 Node* dest_start = dest;
5637 if (src_offset != NULL || dest_offset != NULL) {
5638 assert(src_offset != NULL && dest_offset != NULL, "");
5639 src_start = array_element_address(src, src_offset, basic_elem_type);
5640 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5641 }
5642
5643 // Figure out which arraycopy runtime method to call.
5644 const char* copyfunc_name = "arraycopy";
5645 address copyfunc_addr =
5646 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5647 disjoint_bases, copyfunc_name, dest_uninitialized);
5648
5649 // Call it. Note that the count_ix value is not scaled to a byte-size.
5650 make_runtime_call(RC_LEAF|RC_NO_FP,
5651 OptoRuntime::fast_arraycopy_Type(),
5652 copyfunc_addr, copyfunc_name, adr_type,
5653 src_start, dest_start, copy_length XTOP);
5654 }
5655
5656 //-------------inline_encodeISOArray-----------------------------------
5657 // encode char[] to byte[] in ISO_8859_1
5658 bool LibraryCallKit::inline_encodeISOArray() {
5659 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5660 // no receiver since it is static method
5661 Node *src = argument(0);
5662 Node *src_offset = argument(1);
5663 Node *dst = argument(2);
5664 Node *dst_offset = argument(3);
5665 Node *length = argument(4);
5666
5667 const Type* src_type = src->Value(&_gvn);
5668 const Type* dst_type = dst->Value(&_gvn);
5669 const TypeAryPtr* top_src = src_type->isa_aryptr();
5670 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5671 if (top_src == NULL || top_src->klass() == NULL ||
5672 top_dest == NULL || top_dest->klass() == NULL) {
5673 // failed array check
5674 return false;
5675 }
5676
5677 // Figure out the size and type of the elements we will be copying.
5678 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5679 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5680 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5681 return false;
5682 }
5683 Node* src_start = array_element_address(src, src_offset, src_elem);
5684 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5685 // 'src_start' points to src array + scaled offset
5686 // 'dst_start' points to dst array + scaled offset
5687
5688 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5689 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5690 enc = _gvn.transform(enc);
5691 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5692 set_memory(res_mem, mtype);
5693 set_result(enc);
5694 return true;
5695 }
5696
5697 /**
5698 * Calculate CRC32 for byte.
5699 * int java.util.zip.CRC32.update(int crc, int b)
5700 */
5701 bool LibraryCallKit::inline_updateCRC32() {
5702 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5703 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5704 // no receiver since it is static method
5705 Node* crc = argument(0); // type: int
5706 Node* b = argument(1); // type: int
5707
5708 /*
5709 * int c = ~ crc;
5710 * b = timesXtoThe32[(b ^ c) & 0xFF];
5711 * b = b ^ (c >>> 8);
5712 * crc = ~b;
5713 */
5714
5715 Node* M1 = intcon(-1);
5716 crc = _gvn.transform(new XorINode(crc, M1));
5717 Node* result = _gvn.transform(new XorINode(crc, b));
5718 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5719
5720 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5721 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5722 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5723 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5724
5725 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5726 result = _gvn.transform(new XorINode(crc, result));
5727 result = _gvn.transform(new XorINode(result, M1));
5728 set_result(result);
5729 return true;
5730 }
5731
5732 /**
5733 * Calculate CRC32 for byte[] array.
5734 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5735 */
5736 bool LibraryCallKit::inline_updateBytesCRC32() {
5737 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5738 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5739 // no receiver since it is static method
5740 Node* crc = argument(0); // type: int
5741 Node* src = argument(1); // type: oop
5742 Node* offset = argument(2); // type: int
5743 Node* length = argument(3); // type: int
5744
5745 const Type* src_type = src->Value(&_gvn);
5746 const TypeAryPtr* top_src = src_type->isa_aryptr();
5747 if (top_src == NULL || top_src->klass() == NULL) {
5748 // failed array check
5749 return false;
5750 }
5751
5752 // Figure out the size and type of the elements we will be copying.
5753 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5754 if (src_elem != T_BYTE) {
5755 return false;
5756 }
5757
5758 // 'src_start' points to src array + scaled offset
5759 Node* src_start = array_element_address(src, offset, src_elem);
5760
5761 // We assume that range check is done by caller.
5762 // TODO: generate range check (offset+length < src.length) in debug VM.
5763
5764 // Call the stub.
5765 address stubAddr = StubRoutines::updateBytesCRC32();
5766 const char *stubName = "updateBytesCRC32";
5767
5768 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5769 stubAddr, stubName, TypePtr::BOTTOM,
5770 crc, src_start, length);
5771 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5772 set_result(result);
5773 return true;
5774 }
5775
5776 /**
5777 * Calculate CRC32 for ByteBuffer.
5778 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5779 */
5780 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5781 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5782 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5783 // no receiver since it is static method
5784 Node* crc = argument(0); // type: int
5785 Node* src = argument(1); // type: long
5786 Node* offset = argument(3); // type: int
5787 Node* length = argument(4); // type: int
5788
5789 src = ConvL2X(src); // adjust Java long to machine word
5790 Node* base = _gvn.transform(new CastX2PNode(src));
5791 offset = ConvI2X(offset);
5792
5793 // 'src_start' points to src array + scaled offset
5794 Node* src_start = basic_plus_adr(top(), base, offset);
5795
5796 // Call the stub.
5797 address stubAddr = StubRoutines::updateBytesCRC32();
5798 const char *stubName = "updateBytesCRC32";
5799
5800 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5801 stubAddr, stubName, TypePtr::BOTTOM,
5802 crc, src_start, length);
5803 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5804 set_result(result);
5805 return true;
5806 }
5807
5808 //----------------------------inline_reference_get----------------------------
5809 // public T java.lang.ref.Reference.get();
5810 bool LibraryCallKit::inline_reference_get() {
5811 const int referent_offset = java_lang_ref_Reference::referent_offset;
5812 guarantee(referent_offset > 0, "should have already been set");
5813
5814 // Get the argument:
5815 Node* reference_obj = null_check_receiver();
5816 if (stopped()) return true;
5817
5818 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5819
5820 ciInstanceKlass* klass = env()->Object_klass();
5821 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5822
5823 Node* no_ctrl = NULL;
5824 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5825
5826 // Use the pre-barrier to record the value in the referent field
5827 pre_barrier(false /* do_load */,
5828 control(),
5829 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5830 result /* pre_val */,
5831 T_OBJECT);
5832
5833 // Add memory barrier to prevent commoning reads from this field
5834 // across safepoint since GC can change its value.
5835 insert_mem_bar(Op_MemBarCPUOrder);
5836
5837 set_result(result);
5838 return true;
5839 }
5840
5841
5842 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5843 bool is_exact=true, bool is_static=false) {
5844
5845 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5846 assert(tinst != NULL, "obj is null");
5847 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5848 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5849
5850 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5851 ciSymbol::make(fieldTypeString),
5852 is_static);
5853 if (field == NULL) return (Node *) NULL;
5854 assert (field != NULL, "undefined field");
5855
5856 // Next code copied from Parse::do_get_xxx():
5857
5858 // Compute address and memory type.
5859 int offset = field->offset_in_bytes();
5860 bool is_vol = field->is_volatile();
5861 ciType* field_klass = field->type();
5862 assert(field_klass->is_loaded(), "should be loaded");
5863 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5864 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5865 BasicType bt = field->layout_type();
5866
5867 // Build the resultant type of the load
5868 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5869
5870 // Build the load.
5871 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5872 return loadedField;
5873 }
5874
5875
5876 //------------------------------inline_aescrypt_Block-----------------------
5877 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5878 address stubAddr;
5879 const char *stubName;
5880 assert(UseAES, "need AES instruction support");
5881
5882 switch(id) {
5883 case vmIntrinsics::_aescrypt_encryptBlock:
5884 stubAddr = StubRoutines::aescrypt_encryptBlock();
5885 stubName = "aescrypt_encryptBlock";
5886 break;
5887 case vmIntrinsics::_aescrypt_decryptBlock:
5888 stubAddr = StubRoutines::aescrypt_decryptBlock();
5889 stubName = "aescrypt_decryptBlock";
5890 break;
5891 }
5892 if (stubAddr == NULL) return false;
5893
5894 Node* aescrypt_object = argument(0);
5895 Node* src = argument(1);
5896 Node* src_offset = argument(2);
5897 Node* dest = argument(3);
5898 Node* dest_offset = argument(4);
5899
5900 // (1) src and dest are arrays.
5901 const Type* src_type = src->Value(&_gvn);
5902 const Type* dest_type = dest->Value(&_gvn);
5903 const TypeAryPtr* top_src = src_type->isa_aryptr();
5904 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5905 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5906
5907 // for the quick and dirty code we will skip all the checks.
5908 // we are just trying to get the call to be generated.
5909 Node* src_start = src;
5910 Node* dest_start = dest;
5911 if (src_offset != NULL || dest_offset != NULL) {
5912 assert(src_offset != NULL && dest_offset != NULL, "");
5913 src_start = array_element_address(src, src_offset, T_BYTE);
5914 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5915 }
5916
5917 // now need to get the start of its expanded key array
5918 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5919 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5920 if (k_start == NULL) return false;
5921
5922 if (Matcher::pass_original_key_for_aes()) {
5923 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5924 // compatibility issues between Java key expansion and SPARC crypto instructions
5925 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5926 if (original_k_start == NULL) return false;
5927
5928 // Call the stub.
5929 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5930 stubAddr, stubName, TypePtr::BOTTOM,
5931 src_start, dest_start, k_start, original_k_start);
5932 } else {
5933 // Call the stub.
5934 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5935 stubAddr, stubName, TypePtr::BOTTOM,
5936 src_start, dest_start, k_start);
5937 }
5938
5939 return true;
5940 }
5941
5942 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5943 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5944 address stubAddr;
5945 const char *stubName;
5946
5947 assert(UseAES, "need AES instruction support");
5948
5949 switch(id) {
5950 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5951 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5952 stubName = "cipherBlockChaining_encryptAESCrypt";
5953 break;
5954 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5955 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5956 stubName = "cipherBlockChaining_decryptAESCrypt";
5957 break;
5958 }
5959 if (stubAddr == NULL) return false;
5960
5961 Node* cipherBlockChaining_object = argument(0);
5962 Node* src = argument(1);
5963 Node* src_offset = argument(2);
5964 Node* len = argument(3);
5965 Node* dest = argument(4);
5966 Node* dest_offset = argument(5);
5967
5968 // (1) src and dest are arrays.
5969 const Type* src_type = src->Value(&_gvn);
5970 const Type* dest_type = dest->Value(&_gvn);
5971 const TypeAryPtr* top_src = src_type->isa_aryptr();
5972 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5973 assert (top_src != NULL && top_src->klass() != NULL
5974 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5975
5976 // checks are the responsibility of the caller
5977 Node* src_start = src;
5978 Node* dest_start = dest;
5979 if (src_offset != NULL || dest_offset != NULL) {
5980 assert(src_offset != NULL && dest_offset != NULL, "");
5981 src_start = array_element_address(src, src_offset, T_BYTE);
5982 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5983 }
5984
5985 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5986 // (because of the predicated logic executed earlier).
5987 // so we cast it here safely.
5988 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5989
5990 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5991 if (embeddedCipherObj == NULL) return false;
5992
5993 // cast it to what we know it will be at runtime
5994 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5995 assert(tinst != NULL, "CBC obj is null");
5996 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5997 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5998 if (!klass_AESCrypt->is_loaded()) return false;
5999
6000 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6001 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6002 const TypeOopPtr* xtype = aklass->as_instance_type();
6003 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6004 aescrypt_object = _gvn.transform(aescrypt_object);
6005
6006 // we need to get the start of the aescrypt_object's expanded key array
6007 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6008 if (k_start == NULL) return false;
6009
6010 // similarly, get the start address of the r vector
6011 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6012 if (objRvec == NULL) return false;
6013 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6014
6015 Node* cbcCrypt;
6016 if (Matcher::pass_original_key_for_aes()) {
6017 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6018 // compatibility issues between Java key expansion and SPARC crypto instructions
6019 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6020 if (original_k_start == NULL) return false;
6021
6022 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6023 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6024 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6025 stubAddr, stubName, TypePtr::BOTTOM,
6026 src_start, dest_start, k_start, r_start, len, original_k_start);
6027 } else {
6028 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
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);
6033 }
6034
6035 // return cipher length (int)
6036 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6037 set_result(retvalue);
6038 return true;
6039 }
6040
6041 //------------------------------get_key_start_from_aescrypt_object-----------------------
6042 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6043 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6044 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6045 if (objAESCryptKey == NULL) return (Node *) NULL;
6046
6047 // now have the array, need to get the start address of the K array
6048 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6049 return k_start;
6050 }
6051
6052 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6053 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6054 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6055 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6056 if (objAESCryptKey == NULL) return (Node *) NULL;
6057
6058 // now have the array, need to get the start address of the lastKey array
6059 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6060 return original_k_start;
6061 }
6062
6063 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6064 // Return node representing slow path of predicate check.
6065 // the pseudo code we want to emulate with this predicate is:
6066 // for encryption:
6067 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6068 // for decryption:
6069 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6070 // note cipher==plain is more conservative than the original java code but that's OK
6071 //
6072 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6073 // First, check receiver for NULL since it is virtual method.
6074 Node* objCBC = argument(0);
6075 objCBC = null_check(objCBC);
6076
6077 if (stopped()) return NULL; // Always NULL
6078
6079 // Load embeddedCipher field of CipherBlockChaining object.
6080 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6081
6082 // get AESCrypt klass for instanceOf check
6083 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6084 // will have same classloader as CipherBlockChaining object
6085 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6086 assert(tinst != NULL, "CBCobj is null");
6087 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6088
6089 // we want to do an instanceof comparison against the AESCrypt class
6090 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6091 if (!klass_AESCrypt->is_loaded()) {
6092 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6093 Node* ctrl = control();
6094 set_control(top()); // no regular fast path
6095 return ctrl;
6096 }
6097 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6098
6099 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6100 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6101 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6102
6103 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6104
6105 // for encryption, we are done
6106 if (!decrypting)
6107 return instof_false; // even if it is NULL
6108
6109 // for decryption, we need to add a further check to avoid
6110 // taking the intrinsic path when cipher and plain are the same
6111 // see the original java code for why.
6112 RegionNode* region = new RegionNode(3);
6113 region->init_req(1, instof_false);
6114 Node* src = argument(1);
6115 Node* dest = argument(4);
6116 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6117 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6118 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6119 region->init_req(2, src_dest_conjoint);
6120
6121 record_for_igvn(region);
6122 return _gvn.transform(region);
6123 }