1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/countbitsnode.hpp"
37 #include "opto/intrinsicnode.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/mathexactnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "prims/nativeLookup.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "trace/traceMacros.hpp"
49
50 class LibraryIntrinsic : public InlineCallGenerator {
51 // Extend the set of intrinsics known to the runtime:
52 public:
53 private:
54 bool _is_virtual;
55 bool _is_predicted;
56 bool _does_virtual_dispatch;
57 vmIntrinsics::ID _intrinsic_id;
58
59 public:
60 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
61 : InlineCallGenerator(m),
62 _is_virtual(is_virtual),
63 _is_predicted(is_predicted),
64 _does_virtual_dispatch(does_virtual_dispatch),
65 _intrinsic_id(id)
66 {
67 }
68 virtual bool is_intrinsic() const { return true; }
69 virtual bool is_virtual() const { return _is_virtual; }
70 virtual bool is_predicted() const { return _is_predicted; }
71 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
72 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
73 virtual Node* generate_predicate(JVMState* jvms);
74 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
75 };
76
77
78 // Local helper class for LibraryIntrinsic:
79 class LibraryCallKit : public GraphKit {
80 private:
81 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
82 Node* _result; // the result node, if any
83 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
84
85 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
86
87 public:
88 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
89 : GraphKit(jvms),
90 _intrinsic(intrinsic),
91 _result(NULL)
92 {
93 // Check if this is a root compile. In that case we don't have a caller.
94 if (!jvms->has_method()) {
95 _reexecute_sp = sp();
96 } else {
97 // Find out how many arguments the interpreter needs when deoptimizing
98 // and save the stack pointer value so it can used by uncommon_trap.
99 // We find the argument count by looking at the declared signature.
100 bool ignored_will_link;
101 ciSignature* declared_signature = NULL;
102 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
103 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
104 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
105 }
106 }
107
108 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
109
110 ciMethod* caller() const { return jvms()->method(); }
111 int bci() const { return jvms()->bci(); }
112 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
113 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
114 ciMethod* callee() const { return _intrinsic->method(); }
115
116 bool try_to_inline();
117 Node* try_to_predicate();
118
119 void push_result() {
120 // Push the result onto the stack.
121 if (!stopped() && result() != NULL) {
122 BasicType bt = result()->bottom_type()->basic_type();
123 push_node(bt, result());
124 }
125 }
126
127 private:
128 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
129 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
130 }
131
132 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
133 void set_result(RegionNode* region, PhiNode* value);
134 Node* result() { return _result; }
135
136 virtual int reexecute_sp() { return _reexecute_sp; }
137
138 // Helper functions to inline natives
139 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
140 Node* generate_slow_guard(Node* test, RegionNode* region);
141 Node* generate_fair_guard(Node* test, RegionNode* region);
142 Node* generate_negative_guard(Node* index, RegionNode* region,
143 // resulting CastII of index:
144 Node* *pos_index = NULL);
145 Node* generate_nonpositive_guard(Node* index, bool never_negative,
146 // resulting CastII of index:
147 Node* *pos_index = NULL);
148 Node* generate_limit_guard(Node* offset, Node* subseq_length,
149 Node* array_length,
150 RegionNode* region);
151 Node* generate_current_thread(Node* &tls_output);
152 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
153 bool disjoint_bases, const char* &name, bool dest_uninitialized);
154 Node* load_mirror_from_klass(Node* klass);
155 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
156 RegionNode* region, int null_path,
157 int offset);
158 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
159 RegionNode* region, int null_path) {
160 int offset = java_lang_Class::klass_offset_in_bytes();
161 return load_klass_from_mirror_common(mirror, never_see_null,
162 region, null_path,
163 offset);
164 }
165 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
166 RegionNode* region, int null_path) {
167 int offset = java_lang_Class::array_klass_offset_in_bytes();
168 return load_klass_from_mirror_common(mirror, never_see_null,
169 region, null_path,
170 offset);
171 }
172 Node* generate_access_flags_guard(Node* kls,
173 int modifier_mask, int modifier_bits,
174 RegionNode* region);
175 Node* generate_interface_guard(Node* kls, RegionNode* region);
176 Node* generate_array_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, false, false);
178 }
179 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, false, true);
181 }
182 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
183 return generate_array_guard_common(kls, region, true, false);
184 }
185 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, true, true);
187 }
188 Node* generate_array_guard_common(Node* kls, RegionNode* region,
189 bool obj_array, bool not_array);
190 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
191 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
192 bool is_virtual = false, bool is_static = false);
193 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
194 return generate_method_call(method_id, false, true);
195 }
196 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
197 return generate_method_call(method_id, true, false);
198 }
199 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
200
201 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
202 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
203 bool inline_string_compareTo();
204 bool inline_string_indexOf();
205 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
206 bool inline_string_equals();
207 Node* round_double_node(Node* n);
208 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
209 bool inline_math_native(vmIntrinsics::ID id);
210 bool inline_trig(vmIntrinsics::ID id);
211 bool inline_math(vmIntrinsics::ID id);
212 template <typename OverflowOp>
213 bool inline_math_overflow(Node* arg1, Node* arg2);
214 void inline_math_mathExact(Node* math, Node* test);
215 bool inline_math_addExactI(bool is_increment);
216 bool inline_math_addExactL(bool is_increment);
217 bool inline_math_multiplyExactI();
218 bool inline_math_multiplyExactL();
219 bool inline_math_negateExactI();
220 bool inline_math_negateExactL();
221 bool inline_math_subtractExactI(bool is_decrement);
222 bool inline_math_subtractExactL(bool is_decrement);
223 bool inline_exp();
224 bool inline_pow();
225 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
226 bool inline_min_max(vmIntrinsics::ID id);
227 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
228 // This returns Type::AnyPtr, RawPtr, or OopPtr.
229 int classify_unsafe_addr(Node* &base, Node* &offset);
230 Node* make_unsafe_address(Node* base, Node* offset);
231 // Helper for inline_unsafe_access.
232 // Generates the guards that check whether the result of
233 // Unsafe.getObject should be recorded in an SATB log buffer.
234 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
235 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
236 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
237 static bool klass_needs_init_guard(Node* kls);
238 bool inline_unsafe_allocate();
239 bool inline_unsafe_copyMemory();
240 bool inline_native_currentThread();
241 #ifdef TRACE_HAVE_INTRINSICS
242 bool inline_native_classID();
243 bool inline_native_threadID();
244 #endif
245 bool inline_native_time_funcs(address method, const char* funcName);
246 bool inline_native_isInterrupted();
247 bool inline_native_Class_query(vmIntrinsics::ID id);
248 bool inline_native_subtype_check();
249
250 bool inline_native_newArray();
251 bool inline_native_getLength();
252 bool inline_array_copyOf(bool is_copyOfRange);
253 bool inline_array_equals();
254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
255 bool inline_native_clone(bool is_virtual);
256 bool inline_native_Reflection_getCallerClass();
257 // Helper function for inlining native object hash method
258 bool inline_native_hashcode(bool is_virtual, bool is_static);
259 bool inline_native_getClass();
260
261 // Helper functions for inlining arraycopy
262 bool inline_arraycopy();
263 void generate_arraycopy(const TypePtr* adr_type,
264 BasicType basic_elem_type,
265 Node* src, Node* src_offset,
266 Node* dest, Node* dest_offset,
267 Node* copy_length,
268 bool disjoint_bases = false,
269 bool length_never_negative = false,
270 RegionNode* slow_region = NULL);
271 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
272 RegionNode* slow_region);
273 void generate_clear_array(const TypePtr* adr_type,
274 Node* dest,
275 BasicType basic_elem_type,
276 Node* slice_off,
277 Node* slice_len,
278 Node* slice_end);
279 bool generate_block_arraycopy(const TypePtr* adr_type,
280 BasicType basic_elem_type,
281 AllocateNode* alloc,
282 Node* src, Node* src_offset,
283 Node* dest, Node* dest_offset,
284 Node* dest_size, bool dest_uninitialized);
285 void generate_slow_arraycopy(const TypePtr* adr_type,
286 Node* src, Node* src_offset,
287 Node* dest, Node* dest_offset,
288 Node* copy_length, bool dest_uninitialized);
289 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
290 Node* dest_elem_klass,
291 Node* src, Node* src_offset,
292 Node* dest, Node* dest_offset,
293 Node* copy_length, bool dest_uninitialized);
294 Node* generate_generic_arraycopy(const TypePtr* adr_type,
295 Node* src, Node* src_offset,
296 Node* dest, Node* dest_offset,
297 Node* copy_length, bool dest_uninitialized);
298 void generate_unchecked_arraycopy(const TypePtr* adr_type,
299 BasicType basic_elem_type,
300 bool disjoint_bases,
301 Node* src, Node* src_offset,
302 Node* dest, Node* dest_offset,
303 Node* copy_length, bool dest_uninitialized);
304 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
305 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
306 bool inline_unsafe_ordered_store(BasicType type);
307 bool inline_unsafe_fence(vmIntrinsics::ID id);
308 bool inline_fp_conversions(vmIntrinsics::ID id);
309 bool inline_number_methods(vmIntrinsics::ID id);
310 bool inline_reference_get();
311 bool inline_aescrypt_Block(vmIntrinsics::ID id);
312 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
313 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
314 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
315 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
316 bool inline_encodeISOArray();
317 bool inline_updateCRC32();
318 bool inline_updateBytesCRC32();
319 bool inline_updateByteBufferCRC32();
320 };
321
322
323 //---------------------------make_vm_intrinsic----------------------------
324 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
325 vmIntrinsics::ID id = m->intrinsic_id();
326 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
327
328 if (DisableIntrinsic[0] != '\0'
329 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
330 // disabled by a user request on the command line:
331 // example: -XX:DisableIntrinsic=_hashCode,_getClass
332 return NULL;
333 }
334
335 if (!m->is_loaded()) {
336 // do not attempt to inline unloaded methods
337 return NULL;
338 }
339
340 // Only a few intrinsics implement a virtual dispatch.
341 // They are expensive calls which are also frequently overridden.
342 if (is_virtual) {
343 switch (id) {
344 case vmIntrinsics::_hashCode:
345 case vmIntrinsics::_clone:
346 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
347 break;
348 default:
349 return NULL;
350 }
351 }
352
353 // -XX:-InlineNatives disables nearly all intrinsics:
354 if (!InlineNatives) {
355 switch (id) {
356 case vmIntrinsics::_indexOf:
357 case vmIntrinsics::_compareTo:
358 case vmIntrinsics::_equals:
359 case vmIntrinsics::_equalsC:
360 case vmIntrinsics::_getAndAddInt:
361 case vmIntrinsics::_getAndAddLong:
362 case vmIntrinsics::_getAndSetInt:
363 case vmIntrinsics::_getAndSetLong:
364 case vmIntrinsics::_getAndSetObject:
365 case vmIntrinsics::_loadFence:
366 case vmIntrinsics::_storeFence:
367 case vmIntrinsics::_fullFence:
368 break; // InlineNatives does not control String.compareTo
369 case vmIntrinsics::_Reference_get:
370 break; // InlineNatives does not control Reference.get
371 default:
372 return NULL;
373 }
374 }
375
376 bool is_predicted = false;
377 bool does_virtual_dispatch = false;
378
379 switch (id) {
380 case vmIntrinsics::_compareTo:
381 if (!SpecialStringCompareTo) return NULL;
382 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
383 break;
384 case vmIntrinsics::_indexOf:
385 if (!SpecialStringIndexOf) return NULL;
386 break;
387 case vmIntrinsics::_equals:
388 if (!SpecialStringEquals) return NULL;
389 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
390 break;
391 case vmIntrinsics::_equalsC:
392 if (!SpecialArraysEquals) return NULL;
393 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
394 break;
395 case vmIntrinsics::_arraycopy:
396 if (!InlineArrayCopy) return NULL;
397 break;
398 case vmIntrinsics::_copyMemory:
399 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
400 if (!InlineArrayCopy) return NULL;
401 break;
402 case vmIntrinsics::_hashCode:
403 if (!InlineObjectHash) return NULL;
404 does_virtual_dispatch = true;
405 break;
406 case vmIntrinsics::_clone:
407 does_virtual_dispatch = true;
408 case vmIntrinsics::_copyOf:
409 case vmIntrinsics::_copyOfRange:
410 if (!InlineObjectCopy) return NULL;
411 // These also use the arraycopy intrinsic mechanism:
412 if (!InlineArrayCopy) return NULL;
413 break;
414 case vmIntrinsics::_encodeISOArray:
415 if (!SpecialEncodeISOArray) return NULL;
416 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
417 break;
418 case vmIntrinsics::_checkIndex:
419 // We do not intrinsify this. The optimizer does fine with it.
420 return NULL;
421
422 case vmIntrinsics::_getCallerClass:
423 if (!UseNewReflection) return NULL;
424 if (!InlineReflectionGetCallerClass) return NULL;
425 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
426 break;
427
428 case vmIntrinsics::_bitCount_i:
429 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
430 break;
431
432 case vmIntrinsics::_bitCount_l:
433 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
434 break;
435
436 case vmIntrinsics::_numberOfLeadingZeros_i:
437 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
438 break;
439
440 case vmIntrinsics::_numberOfLeadingZeros_l:
441 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
442 break;
443
444 case vmIntrinsics::_numberOfTrailingZeros_i:
445 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
446 break;
447
448 case vmIntrinsics::_numberOfTrailingZeros_l:
449 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
450 break;
451
452 case vmIntrinsics::_reverseBytes_c:
453 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
454 break;
455 case vmIntrinsics::_reverseBytes_s:
456 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
457 break;
458 case vmIntrinsics::_reverseBytes_i:
459 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
460 break;
461 case vmIntrinsics::_reverseBytes_l:
462 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
463 break;
464
465 case vmIntrinsics::_Reference_get:
466 // Use the intrinsic version of Reference.get() so that the value in
467 // the referent field can be registered by the G1 pre-barrier code.
468 // Also add memory barrier to prevent commoning reads from this field
469 // across safepoint since GC can change it value.
470 break;
471
472 case vmIntrinsics::_compareAndSwapObject:
473 #ifdef _LP64
474 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
475 #endif
476 break;
477
478 case vmIntrinsics::_compareAndSwapLong:
479 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndAddInt:
483 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
484 break;
485
486 case vmIntrinsics::_getAndAddLong:
487 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
488 break;
489
490 case vmIntrinsics::_getAndSetInt:
491 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
492 break;
493
494 case vmIntrinsics::_getAndSetLong:
495 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
496 break;
497
498 case vmIntrinsics::_getAndSetObject:
499 #ifdef _LP64
500 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
501 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
502 break;
503 #else
504 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
505 break;
506 #endif
507
508 case vmIntrinsics::_aescrypt_encryptBlock:
509 case vmIntrinsics::_aescrypt_decryptBlock:
510 if (!UseAESIntrinsics) return NULL;
511 break;
512
513 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
514 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
515 if (!UseAESIntrinsics) return NULL;
516 // these two require the predicated logic
517 is_predicted = true;
518 break;
519
520 case vmIntrinsics::_updateCRC32:
521 case vmIntrinsics::_updateBytesCRC32:
522 case vmIntrinsics::_updateByteBufferCRC32:
523 if (!UseCRC32Intrinsics) return NULL;
524 break;
525
526 case vmIntrinsics::_incrementExactI:
527 case vmIntrinsics::_addExactI:
528 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
529 break;
530 case vmIntrinsics::_incrementExactL:
531 case vmIntrinsics::_addExactL:
532 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
533 break;
534 case vmIntrinsics::_decrementExactI:
535 case vmIntrinsics::_subtractExactI:
536 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
537 break;
538 case vmIntrinsics::_decrementExactL:
539 case vmIntrinsics::_subtractExactL:
540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
541 break;
542 case vmIntrinsics::_negateExactI:
543 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
544 break;
545 case vmIntrinsics::_negateExactL:
546 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
547 break;
548 case vmIntrinsics::_multiplyExactI:
549 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
550 break;
551 case vmIntrinsics::_multiplyExactL:
552 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
553 break;
554
555 default:
556 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
557 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
558 break;
559 }
560
561 // -XX:-InlineClassNatives disables natives from the Class class.
562 // The flag applies to all reflective calls, notably Array.newArray
563 // (visible to Java programmers as Array.newInstance).
564 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
565 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
566 if (!InlineClassNatives) return NULL;
567 }
568
569 // -XX:-InlineThreadNatives disables natives from the Thread class.
570 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
571 if (!InlineThreadNatives) return NULL;
572 }
573
574 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
575 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
576 m->holder()->name() == ciSymbol::java_lang_Float() ||
577 m->holder()->name() == ciSymbol::java_lang_Double()) {
578 if (!InlineMathNatives) return NULL;
579 }
580
581 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
582 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
583 if (!InlineUnsafeOps) return NULL;
584 }
585
586 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
587 }
588
589 //----------------------register_library_intrinsics-----------------------
590 // Initialize this file's data structures, for each Compile instance.
591 void Compile::register_library_intrinsics() {
592 // Nothing to do here.
593 }
594
595 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
596 LibraryCallKit kit(jvms, this);
597 Compile* C = kit.C;
598 int nodes = C->unique();
599 #ifndef PRODUCT
600 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
601 char buf[1000];
602 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
603 tty->print_cr("Intrinsic %s", str);
604 }
605 #endif
606 ciMethod* callee = kit.callee();
607 const int bci = kit.bci();
608
609 // Try to inline the intrinsic.
610 if (kit.try_to_inline()) {
611 if (C->print_intrinsics() || C->print_inlining()) {
612 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
613 }
614 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
615 if (C->log()) {
616 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
617 vmIntrinsics::name_at(intrinsic_id()),
618 (is_virtual() ? " virtual='1'" : ""),
619 C->unique() - nodes);
620 }
621 // Push the result from the inlined method onto the stack.
622 kit.push_result();
623 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 return NULL;
641 }
642
643 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
644 LibraryCallKit kit(jvms, this);
645 Compile* C = kit.C;
646 int nodes = C->unique();
647 #ifndef PRODUCT
648 assert(is_predicted(), "sanity");
649 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
650 char buf[1000];
651 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
652 tty->print_cr("Predicate for intrinsic %s", str);
653 }
654 #endif
655 ciMethod* callee = kit.callee();
656 const int bci = kit.bci();
657
658 Node* slow_ctl = kit.try_to_predicate();
659 if (!kit.failing()) {
660 if (C->print_intrinsics() || C->print_inlining()) {
661 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
662 }
663 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
664 if (C->log()) {
665 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
666 vmIntrinsics::name_at(intrinsic_id()),
667 (is_virtual() ? " virtual='1'" : ""),
668 C->unique() - nodes);
669 }
670 return slow_ctl; // Could be NULL if the check folds.
671 }
672
673 // The intrinsic bailed out
674 if (C->print_intrinsics() || C->print_inlining()) {
675 if (jvms->has_method()) {
676 // Not a root compile.
677 const char* msg = "failed to generate predicate for intrinsic";
678 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
679 } else {
680 // Root compile
681 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
682 vmIntrinsics::name_at(intrinsic_id()),
683 (is_virtual() ? " (virtual)" : ""), bci);
684 }
685 }
686 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
687 return NULL;
688 }
689
690 bool LibraryCallKit::try_to_inline() {
691 // Handle symbolic names for otherwise undistinguished boolean switches:
692 const bool is_store = true;
693 const bool is_native_ptr = true;
694 const bool is_static = true;
695 const bool is_volatile = true;
696
697 if (!jvms()->has_method()) {
698 // Root JVMState has a null method.
699 assert(map()->memory()->Opcode() == Op_Parm, "");
700 // Insert the memory aliasing node
701 set_all_memory(reset_memory());
702 }
703 assert(merged_memory(), "");
704
705
706 switch (intrinsic_id()) {
707 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
708 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
709 case vmIntrinsics::_getClass: return inline_native_getClass();
710
711 case vmIntrinsics::_dsin:
712 case vmIntrinsics::_dcos:
713 case vmIntrinsics::_dtan:
714 case vmIntrinsics::_dabs:
715 case vmIntrinsics::_datan2:
716 case vmIntrinsics::_dsqrt:
717 case vmIntrinsics::_dexp:
718 case vmIntrinsics::_dlog:
719 case vmIntrinsics::_dlog10:
720 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
721
722 case vmIntrinsics::_min:
723 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
724
725 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
726 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
727 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
728 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
729 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
730 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
731 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
732 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
733 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
734 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
735 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
736 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
737
738 case vmIntrinsics::_arraycopy: return inline_arraycopy();
739
740 case vmIntrinsics::_compareTo: return inline_string_compareTo();
741 case vmIntrinsics::_indexOf: return inline_string_indexOf();
742 case vmIntrinsics::_equals: return inline_string_equals();
743
744 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
745 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
746 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
747 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
748 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
749 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
750 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
751 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
752 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
753
754 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
755 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
756 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
757 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
758 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
759 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
760 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
761 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
762 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
763
764 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
765 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
766 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
767 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
768 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
769 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
770 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
771 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
772
773 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
774 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
775 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
776 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
777 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
778 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
779 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
780 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
781
782 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
783 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
784 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
785 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
786 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
787 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
788 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
789 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
790 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
791
792 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
793 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
794 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
795 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
796 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
797 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
798 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
799 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
800 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
801
802 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
803 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
804 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
805 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
806
807 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
808 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
809 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
810
811 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
812 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
813 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
814
815 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
816 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
817 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
818 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
819 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
820
821 case vmIntrinsics::_loadFence:
822 case vmIntrinsics::_storeFence:
823 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
824
825 case vmIntrinsics::_currentThread: return inline_native_currentThread();
826 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
827
828 #ifdef TRACE_HAVE_INTRINSICS
829 case vmIntrinsics::_classID: return inline_native_classID();
830 case vmIntrinsics::_threadID: return inline_native_threadID();
831 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
832 #endif
833 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
834 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
835 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
836 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
837 case vmIntrinsics::_newArray: return inline_native_newArray();
838 case vmIntrinsics::_getLength: return inline_native_getLength();
839 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
840 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
841 case vmIntrinsics::_equalsC: return inline_array_equals();
842 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
843
844 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
845
846 case vmIntrinsics::_isInstance:
847 case vmIntrinsics::_getModifiers:
848 case vmIntrinsics::_isInterface:
849 case vmIntrinsics::_isArray:
850 case vmIntrinsics::_isPrimitive:
851 case vmIntrinsics::_getSuperclass:
852 case vmIntrinsics::_getComponentType:
853 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
854
855 case vmIntrinsics::_floatToRawIntBits:
856 case vmIntrinsics::_floatToIntBits:
857 case vmIntrinsics::_intBitsToFloat:
858 case vmIntrinsics::_doubleToRawLongBits:
859 case vmIntrinsics::_doubleToLongBits:
860 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
861
862 case vmIntrinsics::_numberOfLeadingZeros_i:
863 case vmIntrinsics::_numberOfLeadingZeros_l:
864 case vmIntrinsics::_numberOfTrailingZeros_i:
865 case vmIntrinsics::_numberOfTrailingZeros_l:
866 case vmIntrinsics::_bitCount_i:
867 case vmIntrinsics::_bitCount_l:
868 case vmIntrinsics::_reverseBytes_i:
869 case vmIntrinsics::_reverseBytes_l:
870 case vmIntrinsics::_reverseBytes_s:
871 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
872
873 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
874
875 case vmIntrinsics::_Reference_get: return inline_reference_get();
876
877 case vmIntrinsics::_aescrypt_encryptBlock:
878 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
879
880 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
881 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
882 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
883
884 case vmIntrinsics::_encodeISOArray:
885 return inline_encodeISOArray();
886
887 case vmIntrinsics::_updateCRC32:
888 return inline_updateCRC32();
889 case vmIntrinsics::_updateBytesCRC32:
890 return inline_updateBytesCRC32();
891 case vmIntrinsics::_updateByteBufferCRC32:
892 return inline_updateByteBufferCRC32();
893
894 default:
895 // If you get here, it may be that someone has added a new intrinsic
896 // to the list in vmSymbols.hpp without implementing it here.
897 #ifndef PRODUCT
898 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
899 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
900 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
901 }
902 #endif
903 return false;
904 }
905 }
906
907 Node* LibraryCallKit::try_to_predicate() {
908 if (!jvms()->has_method()) {
909 // Root JVMState has a null method.
910 assert(map()->memory()->Opcode() == Op_Parm, "");
911 // Insert the memory aliasing node
912 set_all_memory(reset_memory());
913 }
914 assert(merged_memory(), "");
915
916 switch (intrinsic_id()) {
917 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
918 return inline_cipherBlockChaining_AESCrypt_predicate(false);
919 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
920 return inline_cipherBlockChaining_AESCrypt_predicate(true);
921
922 default:
923 // If you get here, it may be that someone has added a new intrinsic
924 // to the list in vmSymbols.hpp without implementing it here.
925 #ifndef PRODUCT
926 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
927 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
928 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
929 }
930 #endif
931 Node* slow_ctl = control();
932 set_control(top()); // No fast path instrinsic
933 return slow_ctl;
934 }
935 }
936
937 //------------------------------set_result-------------------------------
938 // Helper function for finishing intrinsics.
939 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
940 record_for_igvn(region);
941 set_control(_gvn.transform(region));
942 set_result( _gvn.transform(value));
943 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
944 }
945
946 //------------------------------generate_guard---------------------------
947 // Helper function for generating guarded fast-slow graph structures.
948 // The given 'test', if true, guards a slow path. If the test fails
949 // then a fast path can be taken. (We generally hope it fails.)
950 // In all cases, GraphKit::control() is updated to the fast path.
951 // The returned value represents the control for the slow path.
952 // The return value is never 'top'; it is either a valid control
953 // or NULL if it is obvious that the slow path can never be taken.
954 // Also, if region and the slow control are not NULL, the slow edge
955 // is appended to the region.
956 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
957 if (stopped()) {
958 // Already short circuited.
959 return NULL;
960 }
961
962 // Build an if node and its projections.
963 // If test is true we take the slow path, which we assume is uncommon.
964 if (_gvn.type(test) == TypeInt::ZERO) {
965 // The slow branch is never taken. No need to build this guard.
966 return NULL;
967 }
968
969 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
970
971 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
972 if (if_slow == top()) {
973 // The slow branch is never taken. No need to build this guard.
974 return NULL;
975 }
976
977 if (region != NULL)
978 region->add_req(if_slow);
979
980 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
981 set_control(if_fast);
982
983 return if_slow;
984 }
985
986 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
987 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
988 }
989 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
990 return generate_guard(test, region, PROB_FAIR);
991 }
992
993 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
994 Node* *pos_index) {
995 if (stopped())
996 return NULL; // already stopped
997 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
998 return NULL; // index is already adequately typed
999 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1000 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1001 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1002 if (is_neg != NULL && pos_index != NULL) {
1003 // Emulate effect of Parse::adjust_map_after_if.
1004 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1005 ccast->set_req(0, control());
1006 (*pos_index) = _gvn.transform(ccast);
1007 }
1008 return is_neg;
1009 }
1010
1011 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1012 Node* *pos_index) {
1013 if (stopped())
1014 return NULL; // already stopped
1015 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1016 return NULL; // index is already adequately typed
1017 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1018 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1019 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1020 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1021 if (is_notp != NULL && pos_index != NULL) {
1022 // Emulate effect of Parse::adjust_map_after_if.
1023 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1024 ccast->set_req(0, control());
1025 (*pos_index) = _gvn.transform(ccast);
1026 }
1027 return is_notp;
1028 }
1029
1030 // Make sure that 'position' is a valid limit index, in [0..length].
1031 // There are two equivalent plans for checking this:
1032 // A. (offset + copyLength) unsigned<= arrayLength
1033 // B. offset <= (arrayLength - copyLength)
1034 // We require that all of the values above, except for the sum and
1035 // difference, are already known to be non-negative.
1036 // Plan A is robust in the face of overflow, if offset and copyLength
1037 // are both hugely positive.
1038 //
1039 // Plan B is less direct and intuitive, but it does not overflow at
1040 // all, since the difference of two non-negatives is always
1041 // representable. Whenever Java methods must perform the equivalent
1042 // check they generally use Plan B instead of Plan A.
1043 // For the moment we use Plan A.
1044 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1045 Node* subseq_length,
1046 Node* array_length,
1047 RegionNode* region) {
1048 if (stopped())
1049 return NULL; // already stopped
1050 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1051 if (zero_offset && subseq_length->eqv_uncast(array_length))
1052 return NULL; // common case of whole-array copy
1053 Node* last = subseq_length;
1054 if (!zero_offset) // last += offset
1055 last = _gvn.transform(new (C) AddINode(last, offset));
1056 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1057 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1058 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1059 return is_over;
1060 }
1061
1062
1063 //--------------------------generate_current_thread--------------------
1064 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1065 ciKlass* thread_klass = env()->Thread_klass();
1066 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1067 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1068 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1069 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1070 tls_output = thread;
1071 return threadObj;
1072 }
1073
1074
1075 //------------------------------make_string_method_node------------------------
1076 // Helper method for String intrinsic functions. This version is called
1077 // with str1 and str2 pointing to String object nodes.
1078 //
1079 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1080 Node* no_ctrl = NULL;
1081
1082 // Get start addr of string
1083 Node* str1_value = load_String_value(no_ctrl, str1);
1084 Node* str1_offset = load_String_offset(no_ctrl, str1);
1085 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1086
1087 // Get length of string 1
1088 Node* str1_len = load_String_length(no_ctrl, str1);
1089
1090 Node* str2_value = load_String_value(no_ctrl, str2);
1091 Node* str2_offset = load_String_offset(no_ctrl, str2);
1092 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1093
1094 Node* str2_len = NULL;
1095 Node* result = NULL;
1096
1097 switch (opcode) {
1098 case Op_StrIndexOf:
1099 // Get length of string 2
1100 str2_len = load_String_length(no_ctrl, str2);
1101
1102 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1103 str1_start, str1_len, str2_start, str2_len);
1104 break;
1105 case Op_StrComp:
1106 // Get length of string 2
1107 str2_len = load_String_length(no_ctrl, str2);
1108
1109 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1110 str1_start, str1_len, str2_start, str2_len);
1111 break;
1112 case Op_StrEquals:
1113 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1114 str1_start, str2_start, str1_len);
1115 break;
1116 default:
1117 ShouldNotReachHere();
1118 return NULL;
1119 }
1120
1121 // All these intrinsics have checks.
1122 C->set_has_split_ifs(true); // Has chance for split-if optimization
1123
1124 return _gvn.transform(result);
1125 }
1126
1127 // Helper method for String intrinsic functions. This version is called
1128 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1129 // to Int nodes containing the lenghts of str1 and str2.
1130 //
1131 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1132 Node* result = NULL;
1133 switch (opcode) {
1134 case Op_StrIndexOf:
1135 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1136 str1_start, cnt1, str2_start, cnt2);
1137 break;
1138 case Op_StrComp:
1139 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1140 str1_start, cnt1, str2_start, cnt2);
1141 break;
1142 case Op_StrEquals:
1143 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1144 str1_start, str2_start, cnt1);
1145 break;
1146 default:
1147 ShouldNotReachHere();
1148 return NULL;
1149 }
1150
1151 // All these intrinsics have checks.
1152 C->set_has_split_ifs(true); // Has chance for split-if optimization
1153
1154 return _gvn.transform(result);
1155 }
1156
1157 //------------------------------inline_string_compareTo------------------------
1158 // public int java.lang.String.compareTo(String anotherString);
1159 bool LibraryCallKit::inline_string_compareTo() {
1160 Node* receiver = null_check(argument(0));
1161 Node* arg = null_check(argument(1));
1162 if (stopped()) {
1163 return true;
1164 }
1165 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1166 return true;
1167 }
1168
1169 //------------------------------inline_string_equals------------------------
1170 bool LibraryCallKit::inline_string_equals() {
1171 Node* receiver = null_check_receiver();
1172 // NOTE: Do not null check argument for String.equals() because spec
1173 // allows to specify NULL as argument.
1174 Node* argument = this->argument(1);
1175 if (stopped()) {
1176 return true;
1177 }
1178
1179 // paths (plus control) merge
1180 RegionNode* region = new (C) RegionNode(5);
1181 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1182
1183 // does source == target string?
1184 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1185 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1186
1187 Node* if_eq = generate_slow_guard(bol, NULL);
1188 if (if_eq != NULL) {
1189 // receiver == argument
1190 phi->init_req(2, intcon(1));
1191 region->init_req(2, if_eq);
1192 }
1193
1194 // get String klass for instanceOf
1195 ciInstanceKlass* klass = env()->String_klass();
1196
1197 if (!stopped()) {
1198 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1199 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1200 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1201
1202 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1203 //instanceOf == true, fallthrough
1204
1205 if (inst_false != NULL) {
1206 phi->init_req(3, intcon(0));
1207 region->init_req(3, inst_false);
1208 }
1209 }
1210
1211 if (!stopped()) {
1212 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1213
1214 // Properly cast the argument to String
1215 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1216 // This path is taken only when argument's type is String:NotNull.
1217 argument = cast_not_null(argument, false);
1218
1219 Node* no_ctrl = NULL;
1220
1221 // Get start addr of receiver
1222 Node* receiver_val = load_String_value(no_ctrl, receiver);
1223 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1224 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1225
1226 // Get length of receiver
1227 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1228
1229 // Get start addr of argument
1230 Node* argument_val = load_String_value(no_ctrl, argument);
1231 Node* argument_offset = load_String_offset(no_ctrl, argument);
1232 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1233
1234 // Get length of argument
1235 Node* argument_cnt = load_String_length(no_ctrl, argument);
1236
1237 // Check for receiver count != argument count
1238 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1239 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1240 Node* if_ne = generate_slow_guard(bol, NULL);
1241 if (if_ne != NULL) {
1242 phi->init_req(4, intcon(0));
1243 region->init_req(4, if_ne);
1244 }
1245
1246 // Check for count == 0 is done by assembler code for StrEquals.
1247
1248 if (!stopped()) {
1249 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1250 phi->init_req(1, equals);
1251 region->init_req(1, control());
1252 }
1253 }
1254
1255 // post merge
1256 set_control(_gvn.transform(region));
1257 record_for_igvn(region);
1258
1259 set_result(_gvn.transform(phi));
1260 return true;
1261 }
1262
1263 //------------------------------inline_array_equals----------------------------
1264 bool LibraryCallKit::inline_array_equals() {
1265 Node* arg1 = argument(0);
1266 Node* arg2 = argument(1);
1267 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1268 return true;
1269 }
1270
1271 // Java version of String.indexOf(constant string)
1272 // class StringDecl {
1273 // StringDecl(char[] ca) {
1274 // offset = 0;
1275 // count = ca.length;
1276 // value = ca;
1277 // }
1278 // int offset;
1279 // int count;
1280 // char[] value;
1281 // }
1282 //
1283 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1284 // int targetOffset, int cache_i, int md2) {
1285 // int cache = cache_i;
1286 // int sourceOffset = string_object.offset;
1287 // int sourceCount = string_object.count;
1288 // int targetCount = target_object.length;
1289 //
1290 // int targetCountLess1 = targetCount - 1;
1291 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1292 //
1293 // char[] source = string_object.value;
1294 // char[] target = target_object;
1295 // int lastChar = target[targetCountLess1];
1296 //
1297 // outer_loop:
1298 // for (int i = sourceOffset; i < sourceEnd; ) {
1299 // int src = source[i + targetCountLess1];
1300 // if (src == lastChar) {
1301 // // With random strings and a 4-character alphabet,
1302 // // reverse matching at this point sets up 0.8% fewer
1303 // // frames, but (paradoxically) makes 0.3% more probes.
1304 // // Since those probes are nearer the lastChar probe,
1305 // // there is may be a net D$ win with reverse matching.
1306 // // But, reversing loop inhibits unroll of inner loop
1307 // // for unknown reason. So, does running outer loop from
1308 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1309 // for (int j = 0; j < targetCountLess1; j++) {
1310 // if (target[targetOffset + j] != source[i+j]) {
1311 // if ((cache & (1 << source[i+j])) == 0) {
1312 // if (md2 < j+1) {
1313 // i += j+1;
1314 // continue outer_loop;
1315 // }
1316 // }
1317 // i += md2;
1318 // continue outer_loop;
1319 // }
1320 // }
1321 // return i - sourceOffset;
1322 // }
1323 // if ((cache & (1 << src)) == 0) {
1324 // i += targetCountLess1;
1325 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1326 // i++;
1327 // }
1328 // return -1;
1329 // }
1330
1331 //------------------------------string_indexOf------------------------
1332 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1333 jint cache_i, jint md2_i) {
1334
1335 Node* no_ctrl = NULL;
1336 float likely = PROB_LIKELY(0.9);
1337 float unlikely = PROB_UNLIKELY(0.9);
1338
1339 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1340
1341 Node* source = load_String_value(no_ctrl, string_object);
1342 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1343 Node* sourceCount = load_String_length(no_ctrl, string_object);
1344
1345 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1346 jint target_length = target_array->length();
1347 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1348 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1349
1350 // String.value field is known to be @Stable.
1351 if (UseImplicitStableValues) {
1352 target = cast_array_to_stable(target, target_type);
1353 }
1354
1355 IdealKit kit(this, false, true);
1356 #define __ kit.
1357 Node* zero = __ ConI(0);
1358 Node* one = __ ConI(1);
1359 Node* cache = __ ConI(cache_i);
1360 Node* md2 = __ ConI(md2_i);
1361 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1362 Node* targetCount = __ ConI(target_length);
1363 Node* targetCountLess1 = __ ConI(target_length - 1);
1364 Node* targetOffset = __ ConI(targetOffset_i);
1365 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1366
1367 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1368 Node* outer_loop = __ make_label(2 /* goto */);
1369 Node* return_ = __ make_label(1);
1370
1371 __ set(rtn,__ ConI(-1));
1372 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1373 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1374 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1375 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1376 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1377 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1378 Node* tpj = __ AddI(targetOffset, __ value(j));
1379 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1380 Node* ipj = __ AddI(__ value(i), __ value(j));
1381 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1382 __ if_then(targ, BoolTest::ne, src2); {
1383 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1384 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1385 __ increment(i, __ AddI(__ value(j), one));
1386 __ goto_(outer_loop);
1387 } __ end_if(); __ dead(j);
1388 }__ end_if(); __ dead(j);
1389 __ increment(i, md2);
1390 __ goto_(outer_loop);
1391 }__ end_if();
1392 __ increment(j, one);
1393 }__ end_loop(); __ dead(j);
1394 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1395 __ goto_(return_);
1396 }__ end_if();
1397 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1398 __ increment(i, targetCountLess1);
1399 }__ end_if();
1400 __ increment(i, one);
1401 __ bind(outer_loop);
1402 }__ end_loop(); __ dead(i);
1403 __ bind(return_);
1404
1405 // Final sync IdealKit and GraphKit.
1406 final_sync(kit);
1407 Node* result = __ value(rtn);
1408 #undef __
1409 C->set_has_loops(true);
1410 return result;
1411 }
1412
1413 //------------------------------inline_string_indexOf------------------------
1414 bool LibraryCallKit::inline_string_indexOf() {
1415 Node* receiver = argument(0);
1416 Node* arg = argument(1);
1417
1418 Node* result;
1419 // Disable the use of pcmpestri until it can be guaranteed that
1420 // the load doesn't cross into the uncommited space.
1421 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1422 UseSSE42Intrinsics) {
1423 // Generate SSE4.2 version of indexOf
1424 // We currently only have match rules that use SSE4.2
1425
1426 receiver = null_check(receiver);
1427 arg = null_check(arg);
1428 if (stopped()) {
1429 return true;
1430 }
1431
1432 ciInstanceKlass* str_klass = env()->String_klass();
1433 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1434
1435 // Make the merge point
1436 RegionNode* result_rgn = new (C) RegionNode(4);
1437 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1438 Node* no_ctrl = NULL;
1439
1440 // Get start addr of source string
1441 Node* source = load_String_value(no_ctrl, receiver);
1442 Node* source_offset = load_String_offset(no_ctrl, receiver);
1443 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1444
1445 // Get length of source string
1446 Node* source_cnt = load_String_length(no_ctrl, receiver);
1447
1448 // Get start addr of substring
1449 Node* substr = load_String_value(no_ctrl, arg);
1450 Node* substr_offset = load_String_offset(no_ctrl, arg);
1451 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1452
1453 // Get length of source string
1454 Node* substr_cnt = load_String_length(no_ctrl, arg);
1455
1456 // Check for substr count > string count
1457 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1458 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1459 Node* if_gt = generate_slow_guard(bol, NULL);
1460 if (if_gt != NULL) {
1461 result_phi->init_req(2, intcon(-1));
1462 result_rgn->init_req(2, if_gt);
1463 }
1464
1465 if (!stopped()) {
1466 // Check for substr count == 0
1467 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1468 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1469 Node* if_zero = generate_slow_guard(bol, NULL);
1470 if (if_zero != NULL) {
1471 result_phi->init_req(3, intcon(0));
1472 result_rgn->init_req(3, if_zero);
1473 }
1474 }
1475
1476 if (!stopped()) {
1477 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1478 result_phi->init_req(1, result);
1479 result_rgn->init_req(1, control());
1480 }
1481 set_control(_gvn.transform(result_rgn));
1482 record_for_igvn(result_rgn);
1483 result = _gvn.transform(result_phi);
1484
1485 } else { // Use LibraryCallKit::string_indexOf
1486 // don't intrinsify if argument isn't a constant string.
1487 if (!arg->is_Con()) {
1488 return false;
1489 }
1490 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1491 if (str_type == NULL) {
1492 return false;
1493 }
1494 ciInstanceKlass* klass = env()->String_klass();
1495 ciObject* str_const = str_type->const_oop();
1496 if (str_const == NULL || str_const->klass() != klass) {
1497 return false;
1498 }
1499 ciInstance* str = str_const->as_instance();
1500 assert(str != NULL, "must be instance");
1501
1502 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1503 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1504
1505 int o;
1506 int c;
1507 if (java_lang_String::has_offset_field()) {
1508 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1509 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1510 } else {
1511 o = 0;
1512 c = pat->length();
1513 }
1514
1515 // constant strings have no offset and count == length which
1516 // simplifies the resulting code somewhat so lets optimize for that.
1517 if (o != 0 || c != pat->length()) {
1518 return false;
1519 }
1520
1521 receiver = null_check(receiver, T_OBJECT);
1522 // NOTE: No null check on the argument is needed since it's a constant String oop.
1523 if (stopped()) {
1524 return true;
1525 }
1526
1527 // The null string as a pattern always returns 0 (match at beginning of string)
1528 if (c == 0) {
1529 set_result(intcon(0));
1530 return true;
1531 }
1532
1533 // Generate default indexOf
1534 jchar lastChar = pat->char_at(o + (c - 1));
1535 int cache = 0;
1536 int i;
1537 for (i = 0; i < c - 1; i++) {
1538 assert(i < pat->length(), "out of range");
1539 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1540 }
1541
1542 int md2 = c;
1543 for (i = 0; i < c - 1; i++) {
1544 assert(i < pat->length(), "out of range");
1545 if (pat->char_at(o + i) == lastChar) {
1546 md2 = (c - 1) - i;
1547 }
1548 }
1549
1550 result = string_indexOf(receiver, pat, o, cache, md2);
1551 }
1552 set_result(result);
1553 return true;
1554 }
1555
1556 //--------------------------round_double_node--------------------------------
1557 // Round a double node if necessary.
1558 Node* LibraryCallKit::round_double_node(Node* n) {
1559 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1560 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1561 return n;
1562 }
1563
1564 //------------------------------inline_math-----------------------------------
1565 // public static double Math.abs(double)
1566 // public static double Math.sqrt(double)
1567 // public static double Math.log(double)
1568 // public static double Math.log10(double)
1569 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1570 Node* arg = round_double_node(argument(0));
1571 Node* n;
1572 switch (id) {
1573 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1574 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1575 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1576 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1577 default: fatal_unexpected_iid(id); break;
1578 }
1579 set_result(_gvn.transform(n));
1580 return true;
1581 }
1582
1583 //------------------------------inline_trig----------------------------------
1584 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1585 // argument reduction which will turn into a fast/slow diamond.
1586 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1587 Node* arg = round_double_node(argument(0));
1588 Node* n = NULL;
1589
1590 switch (id) {
1591 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1592 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1593 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1594 default: fatal_unexpected_iid(id); break;
1595 }
1596 n = _gvn.transform(n);
1597
1598 // Rounding required? Check for argument reduction!
1599 if (Matcher::strict_fp_requires_explicit_rounding) {
1600 static const double pi_4 = 0.7853981633974483;
1601 static const double neg_pi_4 = -0.7853981633974483;
1602 // pi/2 in 80-bit extended precision
1603 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1604 // -pi/2 in 80-bit extended precision
1605 // 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};
1606 // Cutoff value for using this argument reduction technique
1607 //static const double pi_2_minus_epsilon = 1.564660403643354;
1608 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1609
1610 // Pseudocode for sin:
1611 // if (x <= Math.PI / 4.0) {
1612 // if (x >= -Math.PI / 4.0) return fsin(x);
1613 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1614 // } else {
1615 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1616 // }
1617 // return StrictMath.sin(x);
1618
1619 // Pseudocode for cos:
1620 // if (x <= Math.PI / 4.0) {
1621 // if (x >= -Math.PI / 4.0) return fcos(x);
1622 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1623 // } else {
1624 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1625 // }
1626 // return StrictMath.cos(x);
1627
1628 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1629 // requires a special machine instruction to load it. Instead we'll try
1630 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1631 // probably do the math inside the SIN encoding.
1632
1633 // Make the merge point
1634 RegionNode* r = new (C) RegionNode(3);
1635 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1636
1637 // Flatten arg so we need only 1 test
1638 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1639 // Node for PI/4 constant
1640 Node *pi4 = makecon(TypeD::make(pi_4));
1641 // Check PI/4 : abs(arg)
1642 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1643 // Check: If PI/4 < abs(arg) then go slow
1644 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1645 // Branch either way
1646 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1647 set_control(opt_iff(r,iff));
1648
1649 // Set fast path result
1650 phi->init_req(2, n);
1651
1652 // Slow path - non-blocking leaf call
1653 Node* call = NULL;
1654 switch (id) {
1655 case vmIntrinsics::_dsin:
1656 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1657 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1658 "Sin", NULL, arg, top());
1659 break;
1660 case vmIntrinsics::_dcos:
1661 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1662 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1663 "Cos", NULL, arg, top());
1664 break;
1665 case vmIntrinsics::_dtan:
1666 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1667 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1668 "Tan", NULL, arg, top());
1669 break;
1670 }
1671 assert(control()->in(0) == call, "");
1672 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1673 r->init_req(1, control());
1674 phi->init_req(1, slow_result);
1675
1676 // Post-merge
1677 set_control(_gvn.transform(r));
1678 record_for_igvn(r);
1679 n = _gvn.transform(phi);
1680
1681 C->set_has_split_ifs(true); // Has chance for split-if optimization
1682 }
1683 set_result(n);
1684 return true;
1685 }
1686
1687 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1688 //-------------------
1689 //result=(result.isNaN())? funcAddr():result;
1690 // Check: If isNaN() by checking result!=result? then either trap
1691 // or go to runtime
1692 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1693 // Build the boolean node
1694 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1695
1696 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1697 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1698 // The pow or exp intrinsic returned a NaN, which requires a call
1699 // to the runtime. Recompile with the runtime call.
1700 uncommon_trap(Deoptimization::Reason_intrinsic,
1701 Deoptimization::Action_make_not_entrant);
1702 }
1703 set_result(result);
1704 } else {
1705 // If this inlining ever returned NaN in the past, we compile a call
1706 // to the runtime to properly handle corner cases
1707
1708 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1709 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1710 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1711
1712 if (!if_slow->is_top()) {
1713 RegionNode* result_region = new (C) RegionNode(3);
1714 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1715
1716 result_region->init_req(1, if_fast);
1717 result_val->init_req(1, result);
1718
1719 set_control(if_slow);
1720
1721 const TypePtr* no_memory_effects = NULL;
1722 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1723 no_memory_effects,
1724 x, top(), y, y ? top() : NULL);
1725 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1726 #ifdef ASSERT
1727 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1728 assert(value_top == top(), "second value must be top");
1729 #endif
1730
1731 result_region->init_req(2, control());
1732 result_val->init_req(2, value);
1733 set_result(result_region, result_val);
1734 } else {
1735 set_result(result);
1736 }
1737 }
1738 }
1739
1740 //------------------------------inline_exp-------------------------------------
1741 // Inline exp instructions, if possible. The Intel hardware only misses
1742 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1743 bool LibraryCallKit::inline_exp() {
1744 Node* arg = round_double_node(argument(0));
1745 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1746
1747 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1748
1749 C->set_has_split_ifs(true); // Has chance for split-if optimization
1750 return true;
1751 }
1752
1753 //------------------------------inline_pow-------------------------------------
1754 // Inline power instructions, if possible.
1755 bool LibraryCallKit::inline_pow() {
1756 // Pseudocode for pow
1757 // if (x <= 0.0) {
1758 // long longy = (long)y;
1759 // if ((double)longy == y) { // if y is long
1760 // if (y + 1 == y) longy = 0; // huge number: even
1761 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1762 // } else {
1763 // result = NaN;
1764 // }
1765 // } else {
1766 // result = DPow(x,y);
1767 // }
1768 // if (result != result)? {
1769 // result = uncommon_trap() or runtime_call();
1770 // }
1771 // return result;
1772
1773 Node* x = round_double_node(argument(0));
1774 Node* y = round_double_node(argument(2));
1775
1776 Node* result = NULL;
1777
1778 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1779 // Short form: skip the fancy tests and just check for NaN result.
1780 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1781 } else {
1782 // If this inlining ever returned NaN in the past, include all
1783 // checks + call to the runtime.
1784
1785 // Set the merge point for If node with condition of (x <= 0.0)
1786 // There are four possible paths to region node and phi node
1787 RegionNode *r = new (C) RegionNode(4);
1788 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1789
1790 // Build the first if node: if (x <= 0.0)
1791 // Node for 0 constant
1792 Node *zeronode = makecon(TypeD::ZERO);
1793 // Check x:0
1794 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1795 // Check: If (x<=0) then go complex path
1796 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1797 // Branch either way
1798 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1799 // Fast path taken; set region slot 3
1800 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1801 r->init_req(3,fast_taken); // Capture fast-control
1802
1803 // Fast path not-taken, i.e. slow path
1804 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1805
1806 // Set fast path result
1807 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1808 phi->init_req(3, fast_result);
1809
1810 // Complex path
1811 // Build the second if node (if y is long)
1812 // Node for (long)y
1813 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1814 // Node for (double)((long) y)
1815 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1816 // Check (double)((long) y) : y
1817 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1818 // Check if (y isn't long) then go to slow path
1819
1820 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1821 // Branch either way
1822 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1823 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1824
1825 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1826
1827 // Calculate DPow(abs(x), y)*(1 & (long)y)
1828 // Node for constant 1
1829 Node *conone = longcon(1);
1830 // 1& (long)y
1831 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1832
1833 // A huge number is always even. Detect a huge number by checking
1834 // if y + 1 == y and set integer to be tested for parity to 0.
1835 // Required for corner case:
1836 // (long)9.223372036854776E18 = max_jlong
1837 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1838 // max_jlong is odd but 9.223372036854776E18 is even
1839 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1840 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1841 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1842 Node* correctedsign = NULL;
1843 if (ConditionalMoveLimit != 0) {
1844 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1845 } else {
1846 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1847 RegionNode *r = new (C) RegionNode(3);
1848 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1849 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1850 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1851 phi->init_req(1, signnode);
1852 phi->init_req(2, longcon(0));
1853 correctedsign = _gvn.transform(phi);
1854 ylong_path = _gvn.transform(r);
1855 record_for_igvn(r);
1856 }
1857
1858 // zero node
1859 Node *conzero = longcon(0);
1860 // Check (1&(long)y)==0?
1861 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1862 // Check if (1&(long)y)!=0?, if so the result is negative
1863 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1864 // abs(x)
1865 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1866 // abs(x)^y
1867 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1868 // -abs(x)^y
1869 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1870 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1871 Node *signresult = NULL;
1872 if (ConditionalMoveLimit != 0) {
1873 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1874 } else {
1875 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1876 RegionNode *r = new (C) RegionNode(3);
1877 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1878 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1879 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1880 phi->init_req(1, absxpowy);
1881 phi->init_req(2, negabsxpowy);
1882 signresult = _gvn.transform(phi);
1883 ylong_path = _gvn.transform(r);
1884 record_for_igvn(r);
1885 }
1886 // Set complex path fast result
1887 r->init_req(2, ylong_path);
1888 phi->init_req(2, signresult);
1889
1890 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1891 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1892 r->init_req(1,slow_path);
1893 phi->init_req(1,slow_result);
1894
1895 // Post merge
1896 set_control(_gvn.transform(r));
1897 record_for_igvn(r);
1898 result = _gvn.transform(phi);
1899 }
1900
1901 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1902
1903 C->set_has_split_ifs(true); // Has chance for split-if optimization
1904 return true;
1905 }
1906
1907 //------------------------------runtime_math-----------------------------
1908 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1909 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1910 "must be (DD)D or (D)D type");
1911
1912 // Inputs
1913 Node* a = round_double_node(argument(0));
1914 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1915
1916 const TypePtr* no_memory_effects = NULL;
1917 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1918 no_memory_effects,
1919 a, top(), b, b ? top() : NULL);
1920 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1921 #ifdef ASSERT
1922 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1923 assert(value_top == top(), "second value must be top");
1924 #endif
1925
1926 set_result(value);
1927 return true;
1928 }
1929
1930 //------------------------------inline_math_native-----------------------------
1931 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1932 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1933 switch (id) {
1934 // These intrinsics are not properly supported on all hardware
1935 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1936 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1937 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1938 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1939 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1940 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1941
1942 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1943 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1944 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1945 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1946
1947 // These intrinsics are supported on all hardware
1948 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1949 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1950
1951 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1952 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1953 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1954 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1955 #undef FN_PTR
1956
1957 // These intrinsics are not yet correctly implemented
1958 case vmIntrinsics::_datan2:
1959 return false;
1960
1961 default:
1962 fatal_unexpected_iid(id);
1963 return false;
1964 }
1965 }
1966
1967 static bool is_simple_name(Node* n) {
1968 return (n->req() == 1 // constant
1969 || (n->is_Type() && n->as_Type()->type()->singleton())
1970 || n->is_Proj() // parameter or return value
1971 || n->is_Phi() // local of some sort
1972 );
1973 }
1974
1975 //----------------------------inline_min_max-----------------------------------
1976 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1977 set_result(generate_min_max(id, argument(0), argument(1)));
1978 return true;
1979 }
1980
1981 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1982 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
1983 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1984 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1985 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1986
1987 {
1988 PreserveJVMState pjvms(this);
1989 PreserveReexecuteState preexecs(this);
1990 jvms()->set_should_reexecute(true);
1991
1992 set_control(slow_path);
1993 set_i_o(i_o());
1994
1995 uncommon_trap(Deoptimization::Reason_intrinsic,
1996 Deoptimization::Action_none);
1997 }
1998
1999 set_control(fast_path);
2000 set_result(math);
2001 }
2002
2003 template <typename OverflowOp>
2004 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2005 typedef typename OverflowOp::MathOp MathOp;
2006
2007 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2008 Node* operation = _gvn.transform( mathOp );
2009 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2010 inline_math_mathExact(operation, ofcheck);
2011 return true;
2012 }
2013
2014 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2015 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2016 }
2017
2018 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2019 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2020 }
2021
2022 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2023 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2024 }
2025
2026 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2027 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2028 }
2029
2030 bool LibraryCallKit::inline_math_negateExactI() {
2031 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2032 }
2033
2034 bool LibraryCallKit::inline_math_negateExactL() {
2035 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2036 }
2037
2038 bool LibraryCallKit::inline_math_multiplyExactI() {
2039 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2040 }
2041
2042 bool LibraryCallKit::inline_math_multiplyExactL() {
2043 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2044 }
2045
2046 Node*
2047 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2048 // These are the candidate return value:
2049 Node* xvalue = x0;
2050 Node* yvalue = y0;
2051
2052 if (xvalue == yvalue) {
2053 return xvalue;
2054 }
2055
2056 bool want_max = (id == vmIntrinsics::_max);
2057
2058 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2059 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2060 if (txvalue == NULL || tyvalue == NULL) return top();
2061 // This is not really necessary, but it is consistent with a
2062 // hypothetical MaxINode::Value method:
2063 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2064
2065 // %%% This folding logic should (ideally) be in a different place.
2066 // Some should be inside IfNode, and there to be a more reliable
2067 // transformation of ?: style patterns into cmoves. We also want
2068 // more powerful optimizations around cmove and min/max.
2069
2070 // Try to find a dominating comparison of these guys.
2071 // It can simplify the index computation for Arrays.copyOf
2072 // and similar uses of System.arraycopy.
2073 // First, compute the normalized version of CmpI(x, y).
2074 int cmp_op = Op_CmpI;
2075 Node* xkey = xvalue;
2076 Node* ykey = yvalue;
2077 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2078 if (ideal_cmpxy->is_Cmp()) {
2079 // E.g., if we have CmpI(length - offset, count),
2080 // it might idealize to CmpI(length, count + offset)
2081 cmp_op = ideal_cmpxy->Opcode();
2082 xkey = ideal_cmpxy->in(1);
2083 ykey = ideal_cmpxy->in(2);
2084 }
2085
2086 // Start by locating any relevant comparisons.
2087 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2088 Node* cmpxy = NULL;
2089 Node* cmpyx = NULL;
2090 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2091 Node* cmp = start_from->fast_out(k);
2092 if (cmp->outcnt() > 0 && // must have prior uses
2093 cmp->in(0) == NULL && // must be context-independent
2094 cmp->Opcode() == cmp_op) { // right kind of compare
2095 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2096 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2097 }
2098 }
2099
2100 const int NCMPS = 2;
2101 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2102 int cmpn;
2103 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2104 if (cmps[cmpn] != NULL) break; // find a result
2105 }
2106 if (cmpn < NCMPS) {
2107 // Look for a dominating test that tells us the min and max.
2108 int depth = 0; // Limit search depth for speed
2109 Node* dom = control();
2110 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2111 if (++depth >= 100) break;
2112 Node* ifproj = dom;
2113 if (!ifproj->is_Proj()) continue;
2114 Node* iff = ifproj->in(0);
2115 if (!iff->is_If()) continue;
2116 Node* bol = iff->in(1);
2117 if (!bol->is_Bool()) continue;
2118 Node* cmp = bol->in(1);
2119 if (cmp == NULL) continue;
2120 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2121 if (cmps[cmpn] == cmp) break;
2122 if (cmpn == NCMPS) continue;
2123 BoolTest::mask btest = bol->as_Bool()->_test._test;
2124 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2125 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2126 // At this point, we know that 'x btest y' is true.
2127 switch (btest) {
2128 case BoolTest::eq:
2129 // They are proven equal, so we can collapse the min/max.
2130 // Either value is the answer. Choose the simpler.
2131 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2132 return yvalue;
2133 return xvalue;
2134 case BoolTest::lt: // x < y
2135 case BoolTest::le: // x <= y
2136 return (want_max ? yvalue : xvalue);
2137 case BoolTest::gt: // x > y
2138 case BoolTest::ge: // x >= y
2139 return (want_max ? xvalue : yvalue);
2140 }
2141 }
2142 }
2143
2144 // We failed to find a dominating test.
2145 // Let's pick a test that might GVN with prior tests.
2146 Node* best_bol = NULL;
2147 BoolTest::mask best_btest = BoolTest::illegal;
2148 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2149 Node* cmp = cmps[cmpn];
2150 if (cmp == NULL) continue;
2151 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2152 Node* bol = cmp->fast_out(j);
2153 if (!bol->is_Bool()) continue;
2154 BoolTest::mask btest = bol->as_Bool()->_test._test;
2155 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2156 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2157 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2158 best_bol = bol->as_Bool();
2159 best_btest = btest;
2160 }
2161 }
2162 }
2163
2164 Node* answer_if_true = NULL;
2165 Node* answer_if_false = NULL;
2166 switch (best_btest) {
2167 default:
2168 if (cmpxy == NULL)
2169 cmpxy = ideal_cmpxy;
2170 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2171 // and fall through:
2172 case BoolTest::lt: // x < y
2173 case BoolTest::le: // x <= y
2174 answer_if_true = (want_max ? yvalue : xvalue);
2175 answer_if_false = (want_max ? xvalue : yvalue);
2176 break;
2177 case BoolTest::gt: // x > y
2178 case BoolTest::ge: // x >= y
2179 answer_if_true = (want_max ? xvalue : yvalue);
2180 answer_if_false = (want_max ? yvalue : xvalue);
2181 break;
2182 }
2183
2184 jint hi, lo;
2185 if (want_max) {
2186 // We can sharpen the minimum.
2187 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2188 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2189 } else {
2190 // We can sharpen the maximum.
2191 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2192 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2193 }
2194
2195 // Use a flow-free graph structure, to avoid creating excess control edges
2196 // which could hinder other optimizations.
2197 // Since Math.min/max is often used with arraycopy, we want
2198 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2199 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2200 answer_if_false, answer_if_true,
2201 TypeInt::make(lo, hi, widen));
2202
2203 return _gvn.transform(cmov);
2204
2205 /*
2206 // This is not as desirable as it may seem, since Min and Max
2207 // nodes do not have a full set of optimizations.
2208 // And they would interfere, anyway, with 'if' optimizations
2209 // and with CMoveI canonical forms.
2210 switch (id) {
2211 case vmIntrinsics::_min:
2212 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2213 case vmIntrinsics::_max:
2214 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2215 default:
2216 ShouldNotReachHere();
2217 }
2218 */
2219 }
2220
2221 inline int
2222 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2223 const TypePtr* base_type = TypePtr::NULL_PTR;
2224 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2225 if (base_type == NULL) {
2226 // Unknown type.
2227 return Type::AnyPtr;
2228 } else if (base_type == TypePtr::NULL_PTR) {
2229 // Since this is a NULL+long form, we have to switch to a rawptr.
2230 base = _gvn.transform(new (C) CastX2PNode(offset));
2231 offset = MakeConX(0);
2232 return Type::RawPtr;
2233 } else if (base_type->base() == Type::RawPtr) {
2234 return Type::RawPtr;
2235 } else if (base_type->isa_oopptr()) {
2236 // Base is never null => always a heap address.
2237 if (base_type->ptr() == TypePtr::NotNull) {
2238 return Type::OopPtr;
2239 }
2240 // Offset is small => always a heap address.
2241 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2242 if (offset_type != NULL &&
2243 base_type->offset() == 0 && // (should always be?)
2244 offset_type->_lo >= 0 &&
2245 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2246 return Type::OopPtr;
2247 }
2248 // Otherwise, it might either be oop+off or NULL+addr.
2249 return Type::AnyPtr;
2250 } else {
2251 // No information:
2252 return Type::AnyPtr;
2253 }
2254 }
2255
2256 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2257 int kind = classify_unsafe_addr(base, offset);
2258 if (kind == Type::RawPtr) {
2259 return basic_plus_adr(top(), base, offset);
2260 } else {
2261 return basic_plus_adr(base, offset);
2262 }
2263 }
2264
2265 //--------------------------inline_number_methods-----------------------------
2266 // inline int Integer.numberOfLeadingZeros(int)
2267 // inline int Long.numberOfLeadingZeros(long)
2268 //
2269 // inline int Integer.numberOfTrailingZeros(int)
2270 // inline int Long.numberOfTrailingZeros(long)
2271 //
2272 // inline int Integer.bitCount(int)
2273 // inline int Long.bitCount(long)
2274 //
2275 // inline char Character.reverseBytes(char)
2276 // inline short Short.reverseBytes(short)
2277 // inline int Integer.reverseBytes(int)
2278 // inline long Long.reverseBytes(long)
2279 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2280 Node* arg = argument(0);
2281 Node* n;
2282 switch (id) {
2283 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2284 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2285 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2286 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2287 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2288 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2289 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2290 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2291 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2292 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2293 default: fatal_unexpected_iid(id); break;
2294 }
2295 set_result(_gvn.transform(n));
2296 return true;
2297 }
2298
2299 //----------------------------inline_unsafe_access----------------------------
2300
2301 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2302
2303 // Helper that guards and inserts a pre-barrier.
2304 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2305 Node* pre_val, bool need_mem_bar) {
2306 // We could be accessing the referent field of a reference object. If so, when G1
2307 // is enabled, we need to log the value in the referent field in an SATB buffer.
2308 // This routine performs some compile time filters and generates suitable
2309 // runtime filters that guard the pre-barrier code.
2310 // Also add memory barrier for non volatile load from the referent field
2311 // to prevent commoning of loads across safepoint.
2312 if (!UseG1GC && !need_mem_bar)
2313 return;
2314
2315 // Some compile time checks.
2316
2317 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2318 const TypeX* otype = offset->find_intptr_t_type();
2319 if (otype != NULL && otype->is_con() &&
2320 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2321 // Constant offset but not the reference_offset so just return
2322 return;
2323 }
2324
2325 // We only need to generate the runtime guards for instances.
2326 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2327 if (btype != NULL) {
2328 if (btype->isa_aryptr()) {
2329 // Array type so nothing to do
2330 return;
2331 }
2332
2333 const TypeInstPtr* itype = btype->isa_instptr();
2334 if (itype != NULL) {
2335 // Can the klass of base_oop be statically determined to be
2336 // _not_ a sub-class of Reference and _not_ Object?
2337 ciKlass* klass = itype->klass();
2338 if ( klass->is_loaded() &&
2339 !klass->is_subtype_of(env()->Reference_klass()) &&
2340 !env()->Object_klass()->is_subtype_of(klass)) {
2341 return;
2342 }
2343 }
2344 }
2345
2346 // The compile time filters did not reject base_oop/offset so
2347 // we need to generate the following runtime filters
2348 //
2349 // if (offset == java_lang_ref_Reference::_reference_offset) {
2350 // if (instance_of(base, java.lang.ref.Reference)) {
2351 // pre_barrier(_, pre_val, ...);
2352 // }
2353 // }
2354
2355 float likely = PROB_LIKELY( 0.999);
2356 float unlikely = PROB_UNLIKELY(0.999);
2357
2358 IdealKit ideal(this);
2359 #define __ ideal.
2360
2361 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2362
2363 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2364 // Update graphKit memory and control from IdealKit.
2365 sync_kit(ideal);
2366
2367 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2368 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2369
2370 // Update IdealKit memory and control from graphKit.
2371 __ sync_kit(this);
2372
2373 Node* one = __ ConI(1);
2374 // is_instof == 0 if base_oop == NULL
2375 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2376
2377 // Update graphKit from IdeakKit.
2378 sync_kit(ideal);
2379
2380 // Use the pre-barrier to record the value in the referent field
2381 pre_barrier(false /* do_load */,
2382 __ ctrl(),
2383 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2384 pre_val /* pre_val */,
2385 T_OBJECT);
2386 if (need_mem_bar) {
2387 // Add memory barrier to prevent commoning reads from this field
2388 // across safepoint since GC can change its value.
2389 insert_mem_bar(Op_MemBarCPUOrder);
2390 }
2391 // Update IdealKit from graphKit.
2392 __ sync_kit(this);
2393
2394 } __ end_if(); // _ref_type != ref_none
2395 } __ end_if(); // offset == referent_offset
2396
2397 // Final sync IdealKit and GraphKit.
2398 final_sync(ideal);
2399 #undef __
2400 }
2401
2402
2403 // Interpret Unsafe.fieldOffset cookies correctly:
2404 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2405
2406 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2407 // Attempt to infer a sharper value type from the offset and base type.
2408 ciKlass* sharpened_klass = NULL;
2409
2410 // See if it is an instance field, with an object type.
2411 if (alias_type->field() != NULL) {
2412 assert(!is_native_ptr, "native pointer op cannot use a java address");
2413 if (alias_type->field()->type()->is_klass()) {
2414 sharpened_klass = alias_type->field()->type()->as_klass();
2415 }
2416 }
2417
2418 // See if it is a narrow oop array.
2419 if (adr_type->isa_aryptr()) {
2420 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2421 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2422 if (elem_type != NULL) {
2423 sharpened_klass = elem_type->klass();
2424 }
2425 }
2426 }
2427
2428 // The sharpened class might be unloaded if there is no class loader
2429 // contraint in place.
2430 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2431 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2432
2433 #ifndef PRODUCT
2434 if (C->print_intrinsics() || C->print_inlining()) {
2435 tty->print(" from base type: "); adr_type->dump();
2436 tty->print(" sharpened value: "); tjp->dump();
2437 }
2438 #endif
2439 // Sharpen the value type.
2440 return tjp;
2441 }
2442 return NULL;
2443 }
2444
2445 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2446 if (callee()->is_static()) return false; // caller must have the capability!
2447
2448 #ifndef PRODUCT
2449 {
2450 ResourceMark rm;
2451 // Check the signatures.
2452 ciSignature* sig = callee()->signature();
2453 #ifdef ASSERT
2454 if (!is_store) {
2455 // Object getObject(Object base, int/long offset), etc.
2456 BasicType rtype = sig->return_type()->basic_type();
2457 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2458 rtype = T_ADDRESS; // it is really a C void*
2459 assert(rtype == type, "getter must return the expected value");
2460 if (!is_native_ptr) {
2461 assert(sig->count() == 2, "oop getter has 2 arguments");
2462 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2463 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2464 } else {
2465 assert(sig->count() == 1, "native getter has 1 argument");
2466 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2467 }
2468 } else {
2469 // void putObject(Object base, int/long offset, Object x), etc.
2470 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2471 if (!is_native_ptr) {
2472 assert(sig->count() == 3, "oop putter has 3 arguments");
2473 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2474 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2475 } else {
2476 assert(sig->count() == 2, "native putter has 2 arguments");
2477 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2478 }
2479 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2480 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2481 vtype = T_ADDRESS; // it is really a C void*
2482 assert(vtype == type, "putter must accept the expected value");
2483 }
2484 #endif // ASSERT
2485 }
2486 #endif //PRODUCT
2487
2488 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2489
2490 Node* receiver = argument(0); // type: oop
2491
2492 // Build address expression. See the code in inline_unsafe_prefetch.
2493 Node* adr;
2494 Node* heap_base_oop = top();
2495 Node* offset = top();
2496 Node* val;
2497
2498 if (!is_native_ptr) {
2499 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2500 Node* base = argument(1); // type: oop
2501 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2502 offset = argument(2); // type: long
2503 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2504 // to be plain byte offsets, which are also the same as those accepted
2505 // by oopDesc::field_base.
2506 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2507 "fieldOffset must be byte-scaled");
2508 // 32-bit machines ignore the high half!
2509 offset = ConvL2X(offset);
2510 adr = make_unsafe_address(base, offset);
2511 heap_base_oop = base;
2512 val = is_store ? argument(4) : NULL;
2513 } else {
2514 Node* ptr = argument(1); // type: long
2515 ptr = ConvL2X(ptr); // adjust Java long to machine word
2516 adr = make_unsafe_address(NULL, ptr);
2517 val = is_store ? argument(3) : NULL;
2518 }
2519
2520 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2521
2522 // First guess at the value type.
2523 const Type *value_type = Type::get_const_basic_type(type);
2524
2525 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2526 // there was not enough information to nail it down.
2527 Compile::AliasType* alias_type = C->alias_type(adr_type);
2528 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2529
2530 // We will need memory barriers unless we can determine a unique
2531 // alias category for this reference. (Note: If for some reason
2532 // the barriers get omitted and the unsafe reference begins to "pollute"
2533 // the alias analysis of the rest of the graph, either Compile::can_alias
2534 // or Compile::must_alias will throw a diagnostic assert.)
2535 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2536
2537 // If we are reading the value of the referent field of a Reference
2538 // object (either by using Unsafe directly or through reflection)
2539 // then, if G1 is enabled, we need to record the referent in an
2540 // SATB log buffer using the pre-barrier mechanism.
2541 // Also we need to add memory barrier to prevent commoning reads
2542 // from this field across safepoint since GC can change its value.
2543 bool need_read_barrier = !is_native_ptr && !is_store &&
2544 offset != top() && heap_base_oop != top();
2545
2546 if (!is_store && type == T_OBJECT) {
2547 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2548 if (tjp != NULL) {
2549 value_type = tjp;
2550 }
2551 }
2552
2553 receiver = null_check(receiver);
2554 if (stopped()) {
2555 return true;
2556 }
2557 // Heap pointers get a null-check from the interpreter,
2558 // as a courtesy. However, this is not guaranteed by Unsafe,
2559 // and it is not possible to fully distinguish unintended nulls
2560 // from intended ones in this API.
2561
2562 if (is_volatile) {
2563 // We need to emit leading and trailing CPU membars (see below) in
2564 // addition to memory membars when is_volatile. This is a little
2565 // too strong, but avoids the need to insert per-alias-type
2566 // volatile membars (for stores; compare Parse::do_put_xxx), which
2567 // we cannot do effectively here because we probably only have a
2568 // rough approximation of type.
2569 need_mem_bar = true;
2570 // For Stores, place a memory ordering barrier now.
2571 if (is_store) {
2572 insert_mem_bar(Op_MemBarRelease);
2573 } else {
2574 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2575 insert_mem_bar(Op_MemBarVolatile);
2576 }
2577 }
2578 }
2579
2580 // Memory barrier to prevent normal and 'unsafe' accesses from
2581 // bypassing each other. Happens after null checks, so the
2582 // exception paths do not take memory state from the memory barrier,
2583 // so there's no problems making a strong assert about mixing users
2584 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2585 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2586 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2587
2588 if (!is_store) {
2589 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2590 // load value
2591 switch (type) {
2592 case T_BOOLEAN:
2593 case T_CHAR:
2594 case T_BYTE:
2595 case T_SHORT:
2596 case T_INT:
2597 case T_LONG:
2598 case T_FLOAT:
2599 case T_DOUBLE:
2600 break;
2601 case T_OBJECT:
2602 if (need_read_barrier) {
2603 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2604 }
2605 break;
2606 case T_ADDRESS:
2607 // Cast to an int type.
2608 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2609 p = ConvX2L(p);
2610 break;
2611 default:
2612 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2613 break;
2614 }
2615 // The load node has the control of the preceding MemBarCPUOrder. All
2616 // following nodes will have the control of the MemBarCPUOrder inserted at
2617 // the end of this method. So, pushing the load onto the stack at a later
2618 // point is fine.
2619 set_result(p);
2620 } else {
2621 // place effect of store into memory
2622 switch (type) {
2623 case T_DOUBLE:
2624 val = dstore_rounding(val);
2625 break;
2626 case T_ADDRESS:
2627 // Repackage the long as a pointer.
2628 val = ConvL2X(val);
2629 val = _gvn.transform(new (C) CastX2PNode(val));
2630 break;
2631 }
2632
2633 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2634 if (type != T_OBJECT ) {
2635 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2636 } else {
2637 // Possibly an oop being stored to Java heap or native memory
2638 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2639 // oop to Java heap.
2640 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2641 } else {
2642 // We can't tell at compile time if we are storing in the Java heap or outside
2643 // of it. So we need to emit code to conditionally do the proper type of
2644 // store.
2645
2646 IdealKit ideal(this);
2647 #define __ ideal.
2648 // QQQ who knows what probability is here??
2649 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2650 // Sync IdealKit and graphKit.
2651 sync_kit(ideal);
2652 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2653 // Update IdealKit memory.
2654 __ sync_kit(this);
2655 } __ else_(); {
2656 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2657 } __ end_if();
2658 // Final sync IdealKit and GraphKit.
2659 final_sync(ideal);
2660 #undef __
2661 }
2662 }
2663 }
2664
2665 if (is_volatile) {
2666 if (!is_store) {
2667 insert_mem_bar(Op_MemBarAcquire);
2668 } else {
2669 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2670 insert_mem_bar(Op_MemBarVolatile);
2671 }
2672 }
2673 }
2674
2675 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2676
2677 return true;
2678 }
2679
2680 //----------------------------inline_unsafe_prefetch----------------------------
2681
2682 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2683 #ifndef PRODUCT
2684 {
2685 ResourceMark rm;
2686 // Check the signatures.
2687 ciSignature* sig = callee()->signature();
2688 #ifdef ASSERT
2689 // Object getObject(Object base, int/long offset), etc.
2690 BasicType rtype = sig->return_type()->basic_type();
2691 if (!is_native_ptr) {
2692 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2693 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2694 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2695 } else {
2696 assert(sig->count() == 1, "native prefetch has 1 argument");
2697 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2698 }
2699 #endif // ASSERT
2700 }
2701 #endif // !PRODUCT
2702
2703 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2704
2705 const int idx = is_static ? 0 : 1;
2706 if (!is_static) {
2707 null_check_receiver();
2708 if (stopped()) {
2709 return true;
2710 }
2711 }
2712
2713 // Build address expression. See the code in inline_unsafe_access.
2714 Node *adr;
2715 if (!is_native_ptr) {
2716 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2717 Node* base = argument(idx + 0); // type: oop
2718 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2719 Node* offset = argument(idx + 1); // type: long
2720 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2721 // to be plain byte offsets, which are also the same as those accepted
2722 // by oopDesc::field_base.
2723 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2724 "fieldOffset must be byte-scaled");
2725 // 32-bit machines ignore the high half!
2726 offset = ConvL2X(offset);
2727 adr = make_unsafe_address(base, offset);
2728 } else {
2729 Node* ptr = argument(idx + 0); // type: long
2730 ptr = ConvL2X(ptr); // adjust Java long to machine word
2731 adr = make_unsafe_address(NULL, ptr);
2732 }
2733
2734 // Generate the read or write prefetch
2735 Node *prefetch;
2736 if (is_store) {
2737 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2738 } else {
2739 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2740 }
2741 prefetch->init_req(0, control());
2742 set_i_o(_gvn.transform(prefetch));
2743
2744 return true;
2745 }
2746
2747 //----------------------------inline_unsafe_load_store----------------------------
2748 // This method serves a couple of different customers (depending on LoadStoreKind):
2749 //
2750 // LS_cmpxchg:
2751 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2752 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2753 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2754 //
2755 // LS_xadd:
2756 // public int getAndAddInt( Object o, long offset, int delta)
2757 // public long getAndAddLong(Object o, long offset, long delta)
2758 //
2759 // LS_xchg:
2760 // int getAndSet(Object o, long offset, int newValue)
2761 // long getAndSet(Object o, long offset, long newValue)
2762 // Object getAndSet(Object o, long offset, Object newValue)
2763 //
2764 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2765 // This basic scheme here is the same as inline_unsafe_access, but
2766 // differs in enough details that combining them would make the code
2767 // overly confusing. (This is a true fact! I originally combined
2768 // them, but even I was confused by it!) As much code/comments as
2769 // possible are retained from inline_unsafe_access though to make
2770 // the correspondences clearer. - dl
2771
2772 if (callee()->is_static()) return false; // caller must have the capability!
2773
2774 #ifndef PRODUCT
2775 BasicType rtype;
2776 {
2777 ResourceMark rm;
2778 // Check the signatures.
2779 ciSignature* sig = callee()->signature();
2780 rtype = sig->return_type()->basic_type();
2781 if (kind == LS_xadd || kind == LS_xchg) {
2782 // Check the signatures.
2783 #ifdef ASSERT
2784 assert(rtype == type, "get and set must return the expected type");
2785 assert(sig->count() == 3, "get and set has 3 arguments");
2786 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2787 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2788 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2789 #endif // ASSERT
2790 } else if (kind == LS_cmpxchg) {
2791 // Check the signatures.
2792 #ifdef ASSERT
2793 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2794 assert(sig->count() == 4, "CAS has 4 arguments");
2795 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2796 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2797 #endif // ASSERT
2798 } else {
2799 ShouldNotReachHere();
2800 }
2801 }
2802 #endif //PRODUCT
2803
2804 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2805
2806 // Get arguments:
2807 Node* receiver = NULL;
2808 Node* base = NULL;
2809 Node* offset = NULL;
2810 Node* oldval = NULL;
2811 Node* newval = NULL;
2812 if (kind == LS_cmpxchg) {
2813 const bool two_slot_type = type2size[type] == 2;
2814 receiver = argument(0); // type: oop
2815 base = argument(1); // type: oop
2816 offset = argument(2); // type: long
2817 oldval = argument(4); // type: oop, int, or long
2818 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2819 } else if (kind == LS_xadd || kind == LS_xchg){
2820 receiver = argument(0); // type: oop
2821 base = argument(1); // type: oop
2822 offset = argument(2); // type: long
2823 oldval = NULL;
2824 newval = argument(4); // type: oop, int, or long
2825 }
2826
2827 // Null check receiver.
2828 receiver = null_check(receiver);
2829 if (stopped()) {
2830 return true;
2831 }
2832
2833 // Build field offset expression.
2834 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2835 // to be plain byte offsets, which are also the same as those accepted
2836 // by oopDesc::field_base.
2837 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2838 // 32-bit machines ignore the high half of long offsets
2839 offset = ConvL2X(offset);
2840 Node* adr = make_unsafe_address(base, offset);
2841 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2842
2843 // For CAS, unlike inline_unsafe_access, there seems no point in
2844 // trying to refine types. Just use the coarse types here.
2845 const Type *value_type = Type::get_const_basic_type(type);
2846 Compile::AliasType* alias_type = C->alias_type(adr_type);
2847 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2848
2849 if (kind == LS_xchg && type == T_OBJECT) {
2850 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2851 if (tjp != NULL) {
2852 value_type = tjp;
2853 }
2854 }
2855
2856 int alias_idx = C->get_alias_index(adr_type);
2857
2858 // Memory-model-wise, a LoadStore acts like a little synchronized
2859 // block, so needs barriers on each side. These don't translate
2860 // into actual barriers on most machines, but we still need rest of
2861 // compiler to respect ordering.
2862
2863 insert_mem_bar(Op_MemBarRelease);
2864 insert_mem_bar(Op_MemBarCPUOrder);
2865
2866 // 4984716: MemBars must be inserted before this
2867 // memory node in order to avoid a false
2868 // dependency which will confuse the scheduler.
2869 Node *mem = memory(alias_idx);
2870
2871 // For now, we handle only those cases that actually exist: ints,
2872 // longs, and Object. Adding others should be straightforward.
2873 Node* load_store;
2874 switch(type) {
2875 case T_INT:
2876 if (kind == LS_xadd) {
2877 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2878 } else if (kind == LS_xchg) {
2879 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2880 } else if (kind == LS_cmpxchg) {
2881 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2882 } else {
2883 ShouldNotReachHere();
2884 }
2885 break;
2886 case T_LONG:
2887 if (kind == LS_xadd) {
2888 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2889 } else if (kind == LS_xchg) {
2890 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2891 } else if (kind == LS_cmpxchg) {
2892 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2893 } else {
2894 ShouldNotReachHere();
2895 }
2896 break;
2897 case T_OBJECT:
2898 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2899 // could be delayed during Parse (for example, in adjust_map_after_if()).
2900 // Execute transformation here to avoid barrier generation in such case.
2901 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2902 newval = _gvn.makecon(TypePtr::NULL_PTR);
2903
2904 // Reference stores need a store barrier.
2905 if (kind == LS_xchg) {
2906 // If pre-barrier must execute before the oop store, old value will require do_load here.
2907 if (!can_move_pre_barrier()) {
2908 pre_barrier(true /* do_load*/,
2909 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2910 NULL /* pre_val*/,
2911 T_OBJECT);
2912 } // Else move pre_barrier to use load_store value, see below.
2913 } else if (kind == LS_cmpxchg) {
2914 // Same as for newval above:
2915 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2916 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2917 }
2918 // The only known value which might get overwritten is oldval.
2919 pre_barrier(false /* do_load */,
2920 control(), NULL, NULL, max_juint, NULL, NULL,
2921 oldval /* pre_val */,
2922 T_OBJECT);
2923 } else {
2924 ShouldNotReachHere();
2925 }
2926
2927 #ifdef _LP64
2928 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2929 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2930 if (kind == LS_xchg) {
2931 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2932 newval_enc, adr_type, value_type->make_narrowoop()));
2933 } else {
2934 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2935 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2936 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2937 newval_enc, oldval_enc));
2938 }
2939 } else
2940 #endif
2941 {
2942 if (kind == LS_xchg) {
2943 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2944 } else {
2945 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2946 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2947 }
2948 }
2949 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2950 break;
2951 default:
2952 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2953 break;
2954 }
2955
2956 // SCMemProjNodes represent the memory state of a LoadStore. Their
2957 // main role is to prevent LoadStore nodes from being optimized away
2958 // when their results aren't used.
2959 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2960 set_memory(proj, alias_idx);
2961
2962 if (type == T_OBJECT && kind == LS_xchg) {
2963 #ifdef _LP64
2964 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2965 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2966 }
2967 #endif
2968 if (can_move_pre_barrier()) {
2969 // Don't need to load pre_val. The old value is returned by load_store.
2970 // The pre_barrier can execute after the xchg as long as no safepoint
2971 // gets inserted between them.
2972 pre_barrier(false /* do_load */,
2973 control(), NULL, NULL, max_juint, NULL, NULL,
2974 load_store /* pre_val */,
2975 T_OBJECT);
2976 }
2977 }
2978
2979 // Add the trailing membar surrounding the access
2980 insert_mem_bar(Op_MemBarCPUOrder);
2981 insert_mem_bar(Op_MemBarAcquire);
2982
2983 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2984 set_result(load_store);
2985 return true;
2986 }
2987
2988 //----------------------------inline_unsafe_ordered_store----------------------
2989 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2990 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2991 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2992 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2993 // This is another variant of inline_unsafe_access, differing in
2994 // that it always issues store-store ("release") barrier and ensures
2995 // store-atomicity (which only matters for "long").
2996
2997 if (callee()->is_static()) return false; // caller must have the capability!
2998
2999 #ifndef PRODUCT
3000 {
3001 ResourceMark rm;
3002 // Check the signatures.
3003 ciSignature* sig = callee()->signature();
3004 #ifdef ASSERT
3005 BasicType rtype = sig->return_type()->basic_type();
3006 assert(rtype == T_VOID, "must return void");
3007 assert(sig->count() == 3, "has 3 arguments");
3008 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3009 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3010 #endif // ASSERT
3011 }
3012 #endif //PRODUCT
3013
3014 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3015
3016 // Get arguments:
3017 Node* receiver = argument(0); // type: oop
3018 Node* base = argument(1); // type: oop
3019 Node* offset = argument(2); // type: long
3020 Node* val = argument(4); // type: oop, int, or long
3021
3022 // Null check receiver.
3023 receiver = null_check(receiver);
3024 if (stopped()) {
3025 return true;
3026 }
3027
3028 // Build field offset expression.
3029 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3030 // 32-bit machines ignore the high half of long offsets
3031 offset = ConvL2X(offset);
3032 Node* adr = make_unsafe_address(base, offset);
3033 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3034 const Type *value_type = Type::get_const_basic_type(type);
3035 Compile::AliasType* alias_type = C->alias_type(adr_type);
3036
3037 insert_mem_bar(Op_MemBarRelease);
3038 insert_mem_bar(Op_MemBarCPUOrder);
3039 // Ensure that the store is atomic for longs:
3040 const bool require_atomic_access = true;
3041 Node* store;
3042 if (type == T_OBJECT) // reference stores need a store barrier.
3043 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3044 else {
3045 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3046 }
3047 insert_mem_bar(Op_MemBarCPUOrder);
3048 return true;
3049 }
3050
3051 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3052 // Regardless of form, don't allow previous ld/st to move down,
3053 // then issue acquire, release, or volatile mem_bar.
3054 insert_mem_bar(Op_MemBarCPUOrder);
3055 switch(id) {
3056 case vmIntrinsics::_loadFence:
3057 insert_mem_bar(Op_LoadFence);
3058 return true;
3059 case vmIntrinsics::_storeFence:
3060 insert_mem_bar(Op_StoreFence);
3061 return true;
3062 case vmIntrinsics::_fullFence:
3063 insert_mem_bar(Op_MemBarVolatile);
3064 return true;
3065 default:
3066 fatal_unexpected_iid(id);
3067 return false;
3068 }
3069 }
3070
3071 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3072 if (!kls->is_Con()) {
3073 return true;
3074 }
3075 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3076 if (klsptr == NULL) {
3077 return true;
3078 }
3079 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3080 // don't need a guard for a klass that is already initialized
3081 return !ik->is_initialized();
3082 }
3083
3084 //----------------------------inline_unsafe_allocate---------------------------
3085 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3086 bool LibraryCallKit::inline_unsafe_allocate() {
3087 if (callee()->is_static()) return false; // caller must have the capability!
3088
3089 null_check_receiver(); // null-check, then ignore
3090 Node* cls = null_check(argument(1));
3091 if (stopped()) return true;
3092
3093 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3094 kls = null_check(kls);
3095 if (stopped()) return true; // argument was like int.class
3096
3097 Node* test = NULL;
3098 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3099 // Note: The argument might still be an illegal value like
3100 // Serializable.class or Object[].class. The runtime will handle it.
3101 // But we must make an explicit check for initialization.
3102 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3103 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3104 // can generate code to load it as unsigned byte.
3105 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3106 Node* bits = intcon(InstanceKlass::fully_initialized);
3107 test = _gvn.transform(new (C) SubINode(inst, bits));
3108 // The 'test' is non-zero if we need to take a slow path.
3109 }
3110
3111 Node* obj = new_instance(kls, test);
3112 set_result(obj);
3113 return true;
3114 }
3115
3116 #ifdef TRACE_HAVE_INTRINSICS
3117 /*
3118 * oop -> myklass
3119 * myklass->trace_id |= USED
3120 * return myklass->trace_id & ~0x3
3121 */
3122 bool LibraryCallKit::inline_native_classID() {
3123 null_check_receiver(); // null-check, then ignore
3124 Node* cls = null_check(argument(1), T_OBJECT);
3125 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3126 kls = null_check(kls, T_OBJECT);
3127 ByteSize offset = TRACE_ID_OFFSET;
3128 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3129 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3130 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3131 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3132 Node* clsused = longcon(0x01l); // set the class bit
3133 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3134
3135 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3136 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3137 set_result(andl);
3138 return true;
3139 }
3140
3141 bool LibraryCallKit::inline_native_threadID() {
3142 Node* tls_ptr = NULL;
3143 Node* cur_thr = generate_current_thread(tls_ptr);
3144 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3145 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3146 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3147
3148 Node* threadid = NULL;
3149 size_t thread_id_size = OSThread::thread_id_size();
3150 if (thread_id_size == (size_t) BytesPerLong) {
3151 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3152 } else if (thread_id_size == (size_t) BytesPerInt) {
3153 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3154 } else {
3155 ShouldNotReachHere();
3156 }
3157 set_result(threadid);
3158 return true;
3159 }
3160 #endif
3161
3162 //------------------------inline_native_time_funcs--------------
3163 // inline code for System.currentTimeMillis() and System.nanoTime()
3164 // these have the same type and signature
3165 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3166 const TypeFunc* tf = OptoRuntime::void_long_Type();
3167 const TypePtr* no_memory_effects = NULL;
3168 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3169 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3170 #ifdef ASSERT
3171 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3172 assert(value_top == top(), "second value must be top");
3173 #endif
3174 set_result(value);
3175 return true;
3176 }
3177
3178 //------------------------inline_native_currentThread------------------
3179 bool LibraryCallKit::inline_native_currentThread() {
3180 Node* junk = NULL;
3181 set_result(generate_current_thread(junk));
3182 return true;
3183 }
3184
3185 //------------------------inline_native_isInterrupted------------------
3186 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3187 bool LibraryCallKit::inline_native_isInterrupted() {
3188 // Add a fast path to t.isInterrupted(clear_int):
3189 // (t == Thread.current() &&
3190 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3191 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3192 // So, in the common case that the interrupt bit is false,
3193 // we avoid making a call into the VM. Even if the interrupt bit
3194 // is true, if the clear_int argument is false, we avoid the VM call.
3195 // However, if the receiver is not currentThread, we must call the VM,
3196 // because there must be some locking done around the operation.
3197
3198 // We only go to the fast case code if we pass two guards.
3199 // Paths which do not pass are accumulated in the slow_region.
3200
3201 enum {
3202 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3203 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3204 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3205 PATH_LIMIT
3206 };
3207
3208 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3209 // out of the function.
3210 insert_mem_bar(Op_MemBarCPUOrder);
3211
3212 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3213 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3214
3215 RegionNode* slow_region = new (C) RegionNode(1);
3216 record_for_igvn(slow_region);
3217
3218 // (a) Receiving thread must be the current thread.
3219 Node* rec_thr = argument(0);
3220 Node* tls_ptr = NULL;
3221 Node* cur_thr = generate_current_thread(tls_ptr);
3222 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3223 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3224
3225 generate_slow_guard(bol_thr, slow_region);
3226
3227 // (b) Interrupt bit on TLS must be false.
3228 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3229 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3230 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3231
3232 // Set the control input on the field _interrupted read to prevent it floating up.
3233 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3234 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3235 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3236
3237 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3238
3239 // First fast path: if (!TLS._interrupted) return false;
3240 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3241 result_rgn->init_req(no_int_result_path, false_bit);
3242 result_val->init_req(no_int_result_path, intcon(0));
3243
3244 // drop through to next case
3245 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3246
3247 #ifndef TARGET_OS_FAMILY_windows
3248 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3249 Node* clr_arg = argument(1);
3250 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3251 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3252 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3253
3254 // Second fast path: ... else if (!clear_int) return true;
3255 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3256 result_rgn->init_req(no_clear_result_path, false_arg);
3257 result_val->init_req(no_clear_result_path, intcon(1));
3258
3259 // drop through to next case
3260 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3261 #else
3262 // To return true on Windows you must read the _interrupted field
3263 // and check the the event state i.e. take the slow path.
3264 #endif // TARGET_OS_FAMILY_windows
3265
3266 // (d) Otherwise, go to the slow path.
3267 slow_region->add_req(control());
3268 set_control( _gvn.transform(slow_region));
3269
3270 if (stopped()) {
3271 // There is no slow path.
3272 result_rgn->init_req(slow_result_path, top());
3273 result_val->init_req(slow_result_path, top());
3274 } else {
3275 // non-virtual because it is a private non-static
3276 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3277
3278 Node* slow_val = set_results_for_java_call(slow_call);
3279 // this->control() comes from set_results_for_java_call
3280
3281 Node* fast_io = slow_call->in(TypeFunc::I_O);
3282 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3283
3284 // These two phis are pre-filled with copies of of the fast IO and Memory
3285 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3286 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3287
3288 result_rgn->init_req(slow_result_path, control());
3289 result_io ->init_req(slow_result_path, i_o());
3290 result_mem->init_req(slow_result_path, reset_memory());
3291 result_val->init_req(slow_result_path, slow_val);
3292
3293 set_all_memory(_gvn.transform(result_mem));
3294 set_i_o( _gvn.transform(result_io));
3295 }
3296
3297 C->set_has_split_ifs(true); // Has chance for split-if optimization
3298 set_result(result_rgn, result_val);
3299 return true;
3300 }
3301
3302 //---------------------------load_mirror_from_klass----------------------------
3303 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3304 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3305 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3306 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3307 }
3308
3309 //-----------------------load_klass_from_mirror_common-------------------------
3310 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3311 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3312 // and branch to the given path on the region.
3313 // If never_see_null, take an uncommon trap on null, so we can optimistically
3314 // compile for the non-null case.
3315 // If the region is NULL, force never_see_null = true.
3316 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3317 bool never_see_null,
3318 RegionNode* region,
3319 int null_path,
3320 int offset) {
3321 if (region == NULL) never_see_null = true;
3322 Node* p = basic_plus_adr(mirror, offset);
3323 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3324 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3325 Node* null_ctl = top();
3326 kls = null_check_oop(kls, &null_ctl, never_see_null);
3327 if (region != NULL) {
3328 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3329 region->init_req(null_path, null_ctl);
3330 } else {
3331 assert(null_ctl == top(), "no loose ends");
3332 }
3333 return kls;
3334 }
3335
3336 //--------------------(inline_native_Class_query helpers)---------------------
3337 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3338 // Fall through if (mods & mask) == bits, take the guard otherwise.
3339 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3340 // Branch around if the given klass has the given modifier bit set.
3341 // Like generate_guard, adds a new path onto the region.
3342 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3343 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3344 Node* mask = intcon(modifier_mask);
3345 Node* bits = intcon(modifier_bits);
3346 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3347 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3348 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3349 return generate_fair_guard(bol, region);
3350 }
3351 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3352 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3353 }
3354
3355 //-------------------------inline_native_Class_query-------------------
3356 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3357 const Type* return_type = TypeInt::BOOL;
3358 Node* prim_return_value = top(); // what happens if it's a primitive class?
3359 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3360 bool expect_prim = false; // most of these guys expect to work on refs
3361
3362 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3363
3364 Node* mirror = argument(0);
3365 Node* obj = top();
3366
3367 switch (id) {
3368 case vmIntrinsics::_isInstance:
3369 // nothing is an instance of a primitive type
3370 prim_return_value = intcon(0);
3371 obj = argument(1);
3372 break;
3373 case vmIntrinsics::_getModifiers:
3374 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3375 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3376 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3377 break;
3378 case vmIntrinsics::_isInterface:
3379 prim_return_value = intcon(0);
3380 break;
3381 case vmIntrinsics::_isArray:
3382 prim_return_value = intcon(0);
3383 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3384 break;
3385 case vmIntrinsics::_isPrimitive:
3386 prim_return_value = intcon(1);
3387 expect_prim = true; // obviously
3388 break;
3389 case vmIntrinsics::_getSuperclass:
3390 prim_return_value = null();
3391 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3392 break;
3393 case vmIntrinsics::_getComponentType:
3394 prim_return_value = null();
3395 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3396 break;
3397 case vmIntrinsics::_getClassAccessFlags:
3398 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3399 return_type = TypeInt::INT; // not bool! 6297094
3400 break;
3401 default:
3402 fatal_unexpected_iid(id);
3403 break;
3404 }
3405
3406 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3407 if (mirror_con == NULL) return false; // cannot happen?
3408
3409 #ifndef PRODUCT
3410 if (C->print_intrinsics() || C->print_inlining()) {
3411 ciType* k = mirror_con->java_mirror_type();
3412 if (k) {
3413 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3414 k->print_name();
3415 tty->cr();
3416 }
3417 }
3418 #endif
3419
3420 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3421 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3422 record_for_igvn(region);
3423 PhiNode* phi = new (C) PhiNode(region, return_type);
3424
3425 // The mirror will never be null of Reflection.getClassAccessFlags, however
3426 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3427 // if it is. See bug 4774291.
3428
3429 // For Reflection.getClassAccessFlags(), the null check occurs in
3430 // the wrong place; see inline_unsafe_access(), above, for a similar
3431 // situation.
3432 mirror = null_check(mirror);
3433 // If mirror or obj is dead, only null-path is taken.
3434 if (stopped()) return true;
3435
3436 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3437
3438 // Now load the mirror's klass metaobject, and null-check it.
3439 // Side-effects region with the control path if the klass is null.
3440 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3441 // If kls is null, we have a primitive mirror.
3442 phi->init_req(_prim_path, prim_return_value);
3443 if (stopped()) { set_result(region, phi); return true; }
3444 bool safe_for_replace = (region->in(_prim_path) == top());
3445
3446 Node* p; // handy temp
3447 Node* null_ctl;
3448
3449 // Now that we have the non-null klass, we can perform the real query.
3450 // For constant classes, the query will constant-fold in LoadNode::Value.
3451 Node* query_value = top();
3452 switch (id) {
3453 case vmIntrinsics::_isInstance:
3454 // nothing is an instance of a primitive type
3455 query_value = gen_instanceof(obj, kls, safe_for_replace);
3456 break;
3457
3458 case vmIntrinsics::_getModifiers:
3459 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3460 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3461 break;
3462
3463 case vmIntrinsics::_isInterface:
3464 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3465 if (generate_interface_guard(kls, region) != NULL)
3466 // A guard was added. If the guard is taken, it was an interface.
3467 phi->add_req(intcon(1));
3468 // If we fall through, it's a plain class.
3469 query_value = intcon(0);
3470 break;
3471
3472 case vmIntrinsics::_isArray:
3473 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3474 if (generate_array_guard(kls, region) != NULL)
3475 // A guard was added. If the guard is taken, it was an array.
3476 phi->add_req(intcon(1));
3477 // If we fall through, it's a plain class.
3478 query_value = intcon(0);
3479 break;
3480
3481 case vmIntrinsics::_isPrimitive:
3482 query_value = intcon(0); // "normal" path produces false
3483 break;
3484
3485 case vmIntrinsics::_getSuperclass:
3486 // The rules here are somewhat unfortunate, but we can still do better
3487 // with random logic than with a JNI call.
3488 // Interfaces store null or Object as _super, but must report null.
3489 // Arrays store an intermediate super as _super, but must report Object.
3490 // Other types can report the actual _super.
3491 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3492 if (generate_interface_guard(kls, region) != NULL)
3493 // A guard was added. If the guard is taken, it was an interface.
3494 phi->add_req(null());
3495 if (generate_array_guard(kls, region) != NULL)
3496 // A guard was added. If the guard is taken, it was an array.
3497 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3498 // If we fall through, it's a plain class. Get its _super.
3499 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3500 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3501 null_ctl = top();
3502 kls = null_check_oop(kls, &null_ctl);
3503 if (null_ctl != top()) {
3504 // If the guard is taken, Object.superClass is null (both klass and mirror).
3505 region->add_req(null_ctl);
3506 phi ->add_req(null());
3507 }
3508 if (!stopped()) {
3509 query_value = load_mirror_from_klass(kls);
3510 }
3511 break;
3512
3513 case vmIntrinsics::_getComponentType:
3514 if (generate_array_guard(kls, region) != NULL) {
3515 // Be sure to pin the oop load to the guard edge just created:
3516 Node* is_array_ctrl = region->in(region->req()-1);
3517 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3518 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3519 phi->add_req(cmo);
3520 }
3521 query_value = null(); // non-array case is null
3522 break;
3523
3524 case vmIntrinsics::_getClassAccessFlags:
3525 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3526 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3527 break;
3528
3529 default:
3530 fatal_unexpected_iid(id);
3531 break;
3532 }
3533
3534 // Fall-through is the normal case of a query to a real class.
3535 phi->init_req(1, query_value);
3536 region->init_req(1, control());
3537
3538 C->set_has_split_ifs(true); // Has chance for split-if optimization
3539 set_result(region, phi);
3540 return true;
3541 }
3542
3543 //--------------------------inline_native_subtype_check------------------------
3544 // This intrinsic takes the JNI calls out of the heart of
3545 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3546 bool LibraryCallKit::inline_native_subtype_check() {
3547 // Pull both arguments off the stack.
3548 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3549 args[0] = argument(0);
3550 args[1] = argument(1);
3551 Node* klasses[2]; // corresponding Klasses: superk, subk
3552 klasses[0] = klasses[1] = top();
3553
3554 enum {
3555 // A full decision tree on {superc is prim, subc is prim}:
3556 _prim_0_path = 1, // {P,N} => false
3557 // {P,P} & superc!=subc => false
3558 _prim_same_path, // {P,P} & superc==subc => true
3559 _prim_1_path, // {N,P} => false
3560 _ref_subtype_path, // {N,N} & subtype check wins => true
3561 _both_ref_path, // {N,N} & subtype check loses => false
3562 PATH_LIMIT
3563 };
3564
3565 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3566 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3567 record_for_igvn(region);
3568
3569 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3570 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3571 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3572
3573 // First null-check both mirrors and load each mirror's klass metaobject.
3574 int which_arg;
3575 for (which_arg = 0; which_arg <= 1; which_arg++) {
3576 Node* arg = args[which_arg];
3577 arg = null_check(arg);
3578 if (stopped()) break;
3579 args[which_arg] = arg;
3580
3581 Node* p = basic_plus_adr(arg, class_klass_offset);
3582 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3583 klasses[which_arg] = _gvn.transform(kls);
3584 }
3585
3586 // Having loaded both klasses, test each for null.
3587 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3588 for (which_arg = 0; which_arg <= 1; which_arg++) {
3589 Node* kls = klasses[which_arg];
3590 Node* null_ctl = top();
3591 kls = null_check_oop(kls, &null_ctl, never_see_null);
3592 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3593 region->init_req(prim_path, null_ctl);
3594 if (stopped()) break;
3595 klasses[which_arg] = kls;
3596 }
3597
3598 if (!stopped()) {
3599 // now we have two reference types, in klasses[0..1]
3600 Node* subk = klasses[1]; // the argument to isAssignableFrom
3601 Node* superk = klasses[0]; // the receiver
3602 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3603 // now we have a successful reference subtype check
3604 region->set_req(_ref_subtype_path, control());
3605 }
3606
3607 // If both operands are primitive (both klasses null), then
3608 // we must return true when they are identical primitives.
3609 // It is convenient to test this after the first null klass check.
3610 set_control(region->in(_prim_0_path)); // go back to first null check
3611 if (!stopped()) {
3612 // Since superc is primitive, make a guard for the superc==subc case.
3613 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3614 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3615 generate_guard(bol_eq, region, PROB_FAIR);
3616 if (region->req() == PATH_LIMIT+1) {
3617 // A guard was added. If the added guard is taken, superc==subc.
3618 region->swap_edges(PATH_LIMIT, _prim_same_path);
3619 region->del_req(PATH_LIMIT);
3620 }
3621 region->set_req(_prim_0_path, control()); // Not equal after all.
3622 }
3623
3624 // these are the only paths that produce 'true':
3625 phi->set_req(_prim_same_path, intcon(1));
3626 phi->set_req(_ref_subtype_path, intcon(1));
3627
3628 // pull together the cases:
3629 assert(region->req() == PATH_LIMIT, "sane region");
3630 for (uint i = 1; i < region->req(); i++) {
3631 Node* ctl = region->in(i);
3632 if (ctl == NULL || ctl == top()) {
3633 region->set_req(i, top());
3634 phi ->set_req(i, top());
3635 } else if (phi->in(i) == NULL) {
3636 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3637 }
3638 }
3639
3640 set_control(_gvn.transform(region));
3641 set_result(_gvn.transform(phi));
3642 return true;
3643 }
3644
3645 //---------------------generate_array_guard_common------------------------
3646 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3647 bool obj_array, bool not_array) {
3648 // If obj_array/non_array==false/false:
3649 // Branch around if the given klass is in fact an array (either obj or prim).
3650 // If obj_array/non_array==false/true:
3651 // Branch around if the given klass is not an array klass of any kind.
3652 // If obj_array/non_array==true/true:
3653 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3654 // If obj_array/non_array==true/false:
3655 // Branch around if the kls is an oop array (Object[] or subtype)
3656 //
3657 // Like generate_guard, adds a new path onto the region.
3658 jint layout_con = 0;
3659 Node* layout_val = get_layout_helper(kls, layout_con);
3660 if (layout_val == NULL) {
3661 bool query = (obj_array
3662 ? Klass::layout_helper_is_objArray(layout_con)
3663 : Klass::layout_helper_is_array(layout_con));
3664 if (query == not_array) {
3665 return NULL; // never a branch
3666 } else { // always a branch
3667 Node* always_branch = control();
3668 if (region != NULL)
3669 region->add_req(always_branch);
3670 set_control(top());
3671 return always_branch;
3672 }
3673 }
3674 // Now test the correct condition.
3675 jint nval = (obj_array
3676 ? ((jint)Klass::_lh_array_tag_type_value
3677 << Klass::_lh_array_tag_shift)
3678 : Klass::_lh_neutral_value);
3679 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3680 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3681 // invert the test if we are looking for a non-array
3682 if (not_array) btest = BoolTest(btest).negate();
3683 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3684 return generate_fair_guard(bol, region);
3685 }
3686
3687
3688 //-----------------------inline_native_newArray--------------------------
3689 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3690 bool LibraryCallKit::inline_native_newArray() {
3691 Node* mirror = argument(0);
3692 Node* count_val = argument(1);
3693
3694 mirror = null_check(mirror);
3695 // If mirror or obj is dead, only null-path is taken.
3696 if (stopped()) return true;
3697
3698 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3699 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3700 PhiNode* result_val = new(C) PhiNode(result_reg,
3701 TypeInstPtr::NOTNULL);
3702 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3703 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3704 TypePtr::BOTTOM);
3705
3706 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3707 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3708 result_reg, _slow_path);
3709 Node* normal_ctl = control();
3710 Node* no_array_ctl = result_reg->in(_slow_path);
3711
3712 // Generate code for the slow case. We make a call to newArray().
3713 set_control(no_array_ctl);
3714 if (!stopped()) {
3715 // Either the input type is void.class, or else the
3716 // array klass has not yet been cached. Either the
3717 // ensuing call will throw an exception, or else it
3718 // will cache the array klass for next time.
3719 PreserveJVMState pjvms(this);
3720 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3721 Node* slow_result = set_results_for_java_call(slow_call);
3722 // this->control() comes from set_results_for_java_call
3723 result_reg->set_req(_slow_path, control());
3724 result_val->set_req(_slow_path, slow_result);
3725 result_io ->set_req(_slow_path, i_o());
3726 result_mem->set_req(_slow_path, reset_memory());
3727 }
3728
3729 set_control(normal_ctl);
3730 if (!stopped()) {
3731 // Normal case: The array type has been cached in the java.lang.Class.
3732 // The following call works fine even if the array type is polymorphic.
3733 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3734 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3735 result_reg->init_req(_normal_path, control());
3736 result_val->init_req(_normal_path, obj);
3737 result_io ->init_req(_normal_path, i_o());
3738 result_mem->init_req(_normal_path, reset_memory());
3739 }
3740
3741 // Return the combined state.
3742 set_i_o( _gvn.transform(result_io) );
3743 set_all_memory( _gvn.transform(result_mem));
3744
3745 C->set_has_split_ifs(true); // Has chance for split-if optimization
3746 set_result(result_reg, result_val);
3747 return true;
3748 }
3749
3750 //----------------------inline_native_getLength--------------------------
3751 // public static native int java.lang.reflect.Array.getLength(Object array);
3752 bool LibraryCallKit::inline_native_getLength() {
3753 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3754
3755 Node* array = null_check(argument(0));
3756 // If array is dead, only null-path is taken.
3757 if (stopped()) return true;
3758
3759 // Deoptimize if it is a non-array.
3760 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3761
3762 if (non_array != NULL) {
3763 PreserveJVMState pjvms(this);
3764 set_control(non_array);
3765 uncommon_trap(Deoptimization::Reason_intrinsic,
3766 Deoptimization::Action_maybe_recompile);
3767 }
3768
3769 // If control is dead, only non-array-path is taken.
3770 if (stopped()) return true;
3771
3772 // The works fine even if the array type is polymorphic.
3773 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3774 Node* result = load_array_length(array);
3775
3776 C->set_has_split_ifs(true); // Has chance for split-if optimization
3777 set_result(result);
3778 return true;
3779 }
3780
3781 //------------------------inline_array_copyOf----------------------------
3782 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3783 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3784 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3785 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3786
3787 // Get the arguments.
3788 Node* original = argument(0);
3789 Node* start = is_copyOfRange? argument(1): intcon(0);
3790 Node* end = is_copyOfRange? argument(2): argument(1);
3791 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3792
3793 Node* newcopy;
3794
3795 // Set the original stack and the reexecute bit for the interpreter to reexecute
3796 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3797 { PreserveReexecuteState preexecs(this);
3798 jvms()->set_should_reexecute(true);
3799
3800 array_type_mirror = null_check(array_type_mirror);
3801 original = null_check(original);
3802
3803 // Check if a null path was taken unconditionally.
3804 if (stopped()) return true;
3805
3806 Node* orig_length = load_array_length(original);
3807
3808 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3809 klass_node = null_check(klass_node);
3810
3811 RegionNode* bailout = new (C) RegionNode(1);
3812 record_for_igvn(bailout);
3813
3814 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3815 // Bail out if that is so.
3816 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3817 if (not_objArray != NULL) {
3818 // Improve the klass node's type from the new optimistic assumption:
3819 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3820 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3821 Node* cast = new (C) CastPPNode(klass_node, akls);
3822 cast->init_req(0, control());
3823 klass_node = _gvn.transform(cast);
3824 }
3825
3826 // Bail out if either start or end is negative.
3827 generate_negative_guard(start, bailout, &start);
3828 generate_negative_guard(end, bailout, &end);
3829
3830 Node* length = end;
3831 if (_gvn.type(start) != TypeInt::ZERO) {
3832 length = _gvn.transform(new (C) SubINode(end, start));
3833 }
3834
3835 // Bail out if length is negative.
3836 // Without this the new_array would throw
3837 // NegativeArraySizeException but IllegalArgumentException is what
3838 // should be thrown
3839 generate_negative_guard(length, bailout, &length);
3840
3841 if (bailout->req() > 1) {
3842 PreserveJVMState pjvms(this);
3843 set_control(_gvn.transform(bailout));
3844 uncommon_trap(Deoptimization::Reason_intrinsic,
3845 Deoptimization::Action_maybe_recompile);
3846 }
3847
3848 if (!stopped()) {
3849 // How many elements will we copy from the original?
3850 // The answer is MinI(orig_length - start, length).
3851 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3852 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3853
3854 newcopy = new_array(klass_node, length, 0); // no argments to push
3855
3856 // Generate a direct call to the right arraycopy function(s).
3857 // We know the copy is disjoint but we might not know if the
3858 // oop stores need checking.
3859 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3860 // This will fail a store-check if x contains any non-nulls.
3861 bool disjoint_bases = true;
3862 // if start > orig_length then the length of the copy may be
3863 // negative.
3864 bool length_never_negative = !is_copyOfRange;
3865 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3866 original, start, newcopy, intcon(0), moved,
3867 disjoint_bases, length_never_negative);
3868 }
3869 } // original reexecute is set back here
3870
3871 C->set_has_split_ifs(true); // Has chance for split-if optimization
3872 if (!stopped()) {
3873 set_result(newcopy);
3874 }
3875 return true;
3876 }
3877
3878
3879 //----------------------generate_virtual_guard---------------------------
3880 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3881 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3882 RegionNode* slow_region) {
3883 ciMethod* method = callee();
3884 int vtable_index = method->vtable_index();
3885 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3886 err_msg_res("bad index %d", vtable_index));
3887 // Get the Method* out of the appropriate vtable entry.
3888 int entry_offset = (InstanceKlass::vtable_start_offset() +
3889 vtable_index*vtableEntry::size()) * wordSize +
3890 vtableEntry::method_offset_in_bytes();
3891 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3892 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3893
3894 // Compare the target method with the expected method (e.g., Object.hashCode).
3895 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3896
3897 Node* native_call = makecon(native_call_addr);
3898 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3899 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3900
3901 return generate_slow_guard(test_native, slow_region);
3902 }
3903
3904 //-----------------------generate_method_call----------------------------
3905 // Use generate_method_call to make a slow-call to the real
3906 // method if the fast path fails. An alternative would be to
3907 // use a stub like OptoRuntime::slow_arraycopy_Java.
3908 // This only works for expanding the current library call,
3909 // not another intrinsic. (E.g., don't use this for making an
3910 // arraycopy call inside of the copyOf intrinsic.)
3911 CallJavaNode*
3912 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3913 // When compiling the intrinsic method itself, do not use this technique.
3914 guarantee(callee() != C->method(), "cannot make slow-call to self");
3915
3916 ciMethod* method = callee();
3917 // ensure the JVMS we have will be correct for this call
3918 guarantee(method_id == method->intrinsic_id(), "must match");
3919
3920 const TypeFunc* tf = TypeFunc::make(method);
3921 CallJavaNode* slow_call;
3922 if (is_static) {
3923 assert(!is_virtual, "");
3924 slow_call = new(C) CallStaticJavaNode(C, tf,
3925 SharedRuntime::get_resolve_static_call_stub(),
3926 method, bci());
3927 } else if (is_virtual) {
3928 null_check_receiver();
3929 int vtable_index = Method::invalid_vtable_index;
3930 if (UseInlineCaches) {
3931 // Suppress the vtable call
3932 } else {
3933 // hashCode and clone are not a miranda methods,
3934 // so the vtable index is fixed.
3935 // No need to use the linkResolver to get it.
3936 vtable_index = method->vtable_index();
3937 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3938 err_msg_res("bad index %d", vtable_index));
3939 }
3940 slow_call = new(C) CallDynamicJavaNode(tf,
3941 SharedRuntime::get_resolve_virtual_call_stub(),
3942 method, vtable_index, bci());
3943 } else { // neither virtual nor static: opt_virtual
3944 null_check_receiver();
3945 slow_call = new(C) CallStaticJavaNode(C, tf,
3946 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3947 method, bci());
3948 slow_call->set_optimized_virtual(true);
3949 }
3950 set_arguments_for_java_call(slow_call);
3951 set_edges_for_java_call(slow_call);
3952 return slow_call;
3953 }
3954
3955
3956 //------------------------------inline_native_hashcode--------------------
3957 // Build special case code for calls to hashCode on an object.
3958 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3959 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3960 assert(!(is_virtual && is_static), "either virtual, special, or static");
3961
3962 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3963
3964 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3965 PhiNode* result_val = new(C) PhiNode(result_reg,
3966 TypeInt::INT);
3967 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3968 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3969 TypePtr::BOTTOM);
3970 Node* obj = NULL;
3971 if (!is_static) {
3972 // Check for hashing null object
3973 obj = null_check_receiver();
3974 if (stopped()) return true; // unconditionally null
3975 result_reg->init_req(_null_path, top());
3976 result_val->init_req(_null_path, top());
3977 } else {
3978 // Do a null check, and return zero if null.
3979 // System.identityHashCode(null) == 0
3980 obj = argument(0);
3981 Node* null_ctl = top();
3982 obj = null_check_oop(obj, &null_ctl);
3983 result_reg->init_req(_null_path, null_ctl);
3984 result_val->init_req(_null_path, _gvn.intcon(0));
3985 }
3986
3987 // Unconditionally null? Then return right away.
3988 if (stopped()) {
3989 set_control( result_reg->in(_null_path));
3990 if (!stopped())
3991 set_result(result_val->in(_null_path));
3992 return true;
3993 }
3994
3995 // After null check, get the object's klass.
3996 Node* obj_klass = load_object_klass(obj);
3997
3998 // This call may be virtual (invokevirtual) or bound (invokespecial).
3999 // For each case we generate slightly different code.
4000
4001 // We only go to the fast case code if we pass a number of guards. The
4002 // paths which do not pass are accumulated in the slow_region.
4003 RegionNode* slow_region = new (C) RegionNode(1);
4004 record_for_igvn(slow_region);
4005
4006 // If this is a virtual call, we generate a funny guard. We pull out
4007 // the vtable entry corresponding to hashCode() from the target object.
4008 // If the target method which we are calling happens to be the native
4009 // Object hashCode() method, we pass the guard. We do not need this
4010 // guard for non-virtual calls -- the caller is known to be the native
4011 // Object hashCode().
4012 if (is_virtual) {
4013 generate_virtual_guard(obj_klass, slow_region);
4014 }
4015
4016 // Get the header out of the object, use LoadMarkNode when available
4017 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4018 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4019
4020 // Test the header to see if it is unlocked.
4021 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4022 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4023 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4024 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4025 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4026
4027 generate_slow_guard(test_unlocked, slow_region);
4028
4029 // Get the hash value and check to see that it has been properly assigned.
4030 // We depend on hash_mask being at most 32 bits and avoid the use of
4031 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4032 // vm: see markOop.hpp.
4033 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4034 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4035 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4036 // This hack lets the hash bits live anywhere in the mark object now, as long
4037 // as the shift drops the relevant bits into the low 32 bits. Note that
4038 // Java spec says that HashCode is an int so there's no point in capturing
4039 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4040 hshifted_header = ConvX2I(hshifted_header);
4041 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4042
4043 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4044 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4045 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4046
4047 generate_slow_guard(test_assigned, slow_region);
4048
4049 Node* init_mem = reset_memory();
4050 // fill in the rest of the null path:
4051 result_io ->init_req(_null_path, i_o());
4052 result_mem->init_req(_null_path, init_mem);
4053
4054 result_val->init_req(_fast_path, hash_val);
4055 result_reg->init_req(_fast_path, control());
4056 result_io ->init_req(_fast_path, i_o());
4057 result_mem->init_req(_fast_path, init_mem);
4058
4059 // Generate code for the slow case. We make a call to hashCode().
4060 set_control(_gvn.transform(slow_region));
4061 if (!stopped()) {
4062 // No need for PreserveJVMState, because we're using up the present state.
4063 set_all_memory(init_mem);
4064 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4065 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4066 Node* slow_result = set_results_for_java_call(slow_call);
4067 // this->control() comes from set_results_for_java_call
4068 result_reg->init_req(_slow_path, control());
4069 result_val->init_req(_slow_path, slow_result);
4070 result_io ->set_req(_slow_path, i_o());
4071 result_mem ->set_req(_slow_path, reset_memory());
4072 }
4073
4074 // Return the combined state.
4075 set_i_o( _gvn.transform(result_io) );
4076 set_all_memory( _gvn.transform(result_mem));
4077
4078 set_result(result_reg, result_val);
4079 return true;
4080 }
4081
4082 //---------------------------inline_native_getClass----------------------------
4083 // public final native Class<?> java.lang.Object.getClass();
4084 //
4085 // Build special case code for calls to getClass on an object.
4086 bool LibraryCallKit::inline_native_getClass() {
4087 Node* obj = null_check_receiver();
4088 if (stopped()) return true;
4089 set_result(load_mirror_from_klass(load_object_klass(obj)));
4090 return true;
4091 }
4092
4093 //-----------------inline_native_Reflection_getCallerClass---------------------
4094 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4095 //
4096 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4097 //
4098 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4099 // in that it must skip particular security frames and checks for
4100 // caller sensitive methods.
4101 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4102 #ifndef PRODUCT
4103 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4104 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4105 }
4106 #endif
4107
4108 if (!jvms()->has_method()) {
4109 #ifndef PRODUCT
4110 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4111 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4112 }
4113 #endif
4114 return false;
4115 }
4116
4117 // Walk back up the JVM state to find the caller at the required
4118 // depth.
4119 JVMState* caller_jvms = jvms();
4120
4121 // Cf. JVM_GetCallerClass
4122 // NOTE: Start the loop at depth 1 because the current JVM state does
4123 // not include the Reflection.getCallerClass() frame.
4124 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4125 ciMethod* m = caller_jvms->method();
4126 switch (n) {
4127 case 0:
4128 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4129 break;
4130 case 1:
4131 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4132 if (!m->caller_sensitive()) {
4133 #ifndef PRODUCT
4134 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4135 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4136 }
4137 #endif
4138 return false; // bail-out; let JVM_GetCallerClass do the work
4139 }
4140 break;
4141 default:
4142 if (!m->is_ignored_by_security_stack_walk()) {
4143 // We have reached the desired frame; return the holder class.
4144 // Acquire method holder as java.lang.Class and push as constant.
4145 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4146 ciInstance* caller_mirror = caller_klass->java_mirror();
4147 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4148
4149 #ifndef PRODUCT
4150 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4151 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());
4152 tty->print_cr(" JVM state at this point:");
4153 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4154 ciMethod* m = jvms()->of_depth(i)->method();
4155 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4156 }
4157 }
4158 #endif
4159 return true;
4160 }
4161 break;
4162 }
4163 }
4164
4165 #ifndef PRODUCT
4166 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4167 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4168 tty->print_cr(" JVM state at this point:");
4169 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4170 ciMethod* m = jvms()->of_depth(i)->method();
4171 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4172 }
4173 }
4174 #endif
4175
4176 return false; // bail-out; let JVM_GetCallerClass do the work
4177 }
4178
4179 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4180 Node* arg = argument(0);
4181 Node* result;
4182
4183 switch (id) {
4184 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4185 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4186 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4187 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4188
4189 case vmIntrinsics::_doubleToLongBits: {
4190 // two paths (plus control) merge in a wood
4191 RegionNode *r = new (C) RegionNode(3);
4192 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4193
4194 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4195 // Build the boolean node
4196 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4197
4198 // Branch either way.
4199 // NaN case is less traveled, which makes all the difference.
4200 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4201 Node *opt_isnan = _gvn.transform(ifisnan);
4202 assert( opt_isnan->is_If(), "Expect an IfNode");
4203 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4204 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4205
4206 set_control(iftrue);
4207
4208 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4209 Node *slow_result = longcon(nan_bits); // return NaN
4210 phi->init_req(1, _gvn.transform( slow_result ));
4211 r->init_req(1, iftrue);
4212
4213 // Else fall through
4214 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4215 set_control(iffalse);
4216
4217 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4218 r->init_req(2, iffalse);
4219
4220 // Post merge
4221 set_control(_gvn.transform(r));
4222 record_for_igvn(r);
4223
4224 C->set_has_split_ifs(true); // Has chance for split-if optimization
4225 result = phi;
4226 assert(result->bottom_type()->isa_long(), "must be");
4227 break;
4228 }
4229
4230 case vmIntrinsics::_floatToIntBits: {
4231 // two paths (plus control) merge in a wood
4232 RegionNode *r = new (C) RegionNode(3);
4233 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4234
4235 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4236 // Build the boolean node
4237 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4238
4239 // Branch either way.
4240 // NaN case is less traveled, which makes all the difference.
4241 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4242 Node *opt_isnan = _gvn.transform(ifisnan);
4243 assert( opt_isnan->is_If(), "Expect an IfNode");
4244 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4245 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4246
4247 set_control(iftrue);
4248
4249 static const jint nan_bits = 0x7fc00000;
4250 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4251 phi->init_req(1, _gvn.transform( slow_result ));
4252 r->init_req(1, iftrue);
4253
4254 // Else fall through
4255 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4256 set_control(iffalse);
4257
4258 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4259 r->init_req(2, iffalse);
4260
4261 // Post merge
4262 set_control(_gvn.transform(r));
4263 record_for_igvn(r);
4264
4265 C->set_has_split_ifs(true); // Has chance for split-if optimization
4266 result = phi;
4267 assert(result->bottom_type()->isa_int(), "must be");
4268 break;
4269 }
4270
4271 default:
4272 fatal_unexpected_iid(id);
4273 break;
4274 }
4275 set_result(_gvn.transform(result));
4276 return true;
4277 }
4278
4279 #ifdef _LP64
4280 #define XTOP ,top() /*additional argument*/
4281 #else //_LP64
4282 #define XTOP /*no additional argument*/
4283 #endif //_LP64
4284
4285 //----------------------inline_unsafe_copyMemory-------------------------
4286 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4287 bool LibraryCallKit::inline_unsafe_copyMemory() {
4288 if (callee()->is_static()) return false; // caller must have the capability!
4289 null_check_receiver(); // null-check receiver
4290 if (stopped()) return true;
4291
4292 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4293
4294 Node* src_ptr = argument(1); // type: oop
4295 Node* src_off = ConvL2X(argument(2)); // type: long
4296 Node* dst_ptr = argument(4); // type: oop
4297 Node* dst_off = ConvL2X(argument(5)); // type: long
4298 Node* size = ConvL2X(argument(7)); // type: long
4299
4300 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4301 "fieldOffset must be byte-scaled");
4302
4303 Node* src = make_unsafe_address(src_ptr, src_off);
4304 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4305
4306 // Conservatively insert a memory barrier on all memory slices.
4307 // Do not let writes of the copy source or destination float below the copy.
4308 insert_mem_bar(Op_MemBarCPUOrder);
4309
4310 // Call it. Note that the length argument is not scaled.
4311 make_runtime_call(RC_LEAF|RC_NO_FP,
4312 OptoRuntime::fast_arraycopy_Type(),
4313 StubRoutines::unsafe_arraycopy(),
4314 "unsafe_arraycopy",
4315 TypeRawPtr::BOTTOM,
4316 src, dst, size XTOP);
4317
4318 // Do not let reads of the copy destination float above the copy.
4319 insert_mem_bar(Op_MemBarCPUOrder);
4320
4321 return true;
4322 }
4323
4324 //------------------------clone_coping-----------------------------------
4325 // Helper function for inline_native_clone.
4326 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4327 assert(obj_size != NULL, "");
4328 Node* raw_obj = alloc_obj->in(1);
4329 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4330
4331 AllocateNode* alloc = NULL;
4332 if (ReduceBulkZeroing) {
4333 // We will be completely responsible for initializing this object -
4334 // mark Initialize node as complete.
4335 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4336 // The object was just allocated - there should be no any stores!
4337 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4338 // Mark as complete_with_arraycopy so that on AllocateNode
4339 // expansion, we know this AllocateNode is initialized by an array
4340 // copy and a StoreStore barrier exists after the array copy.
4341 alloc->initialization()->set_complete_with_arraycopy();
4342 }
4343
4344 // Copy the fastest available way.
4345 // TODO: generate fields copies for small objects instead.
4346 Node* src = obj;
4347 Node* dest = alloc_obj;
4348 Node* size = _gvn.transform(obj_size);
4349
4350 // Exclude the header but include array length to copy by 8 bytes words.
4351 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4352 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4353 instanceOopDesc::base_offset_in_bytes();
4354 // base_off:
4355 // 8 - 32-bit VM
4356 // 12 - 64-bit VM, compressed klass
4357 // 16 - 64-bit VM, normal klass
4358 if (base_off % BytesPerLong != 0) {
4359 assert(UseCompressedClassPointers, "");
4360 if (is_array) {
4361 // Exclude length to copy by 8 bytes words.
4362 base_off += sizeof(int);
4363 } else {
4364 // Include klass to copy by 8 bytes words.
4365 base_off = instanceOopDesc::klass_offset_in_bytes();
4366 }
4367 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4368 }
4369 src = basic_plus_adr(src, base_off);
4370 dest = basic_plus_adr(dest, base_off);
4371
4372 // Compute the length also, if needed:
4373 Node* countx = size;
4374 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4375 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4376
4377 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4378 bool disjoint_bases = true;
4379 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4380 src, NULL, dest, NULL, countx,
4381 /*dest_uninitialized*/true);
4382
4383 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4384 if (card_mark) {
4385 assert(!is_array, "");
4386 // Put in store barrier for any and all oops we are sticking
4387 // into this object. (We could avoid this if we could prove
4388 // that the object type contains no oop fields at all.)
4389 Node* no_particular_value = NULL;
4390 Node* no_particular_field = NULL;
4391 int raw_adr_idx = Compile::AliasIdxRaw;
4392 post_barrier(control(),
4393 memory(raw_adr_type),
4394 alloc_obj,
4395 no_particular_field,
4396 raw_adr_idx,
4397 no_particular_value,
4398 T_OBJECT,
4399 false);
4400 }
4401
4402 // Do not let reads from the cloned object float above the arraycopy.
4403 if (alloc != NULL) {
4404 // Do not let stores that initialize this object be reordered with
4405 // a subsequent store that would make this object accessible by
4406 // other threads.
4407 // Record what AllocateNode this StoreStore protects so that
4408 // escape analysis can go from the MemBarStoreStoreNode to the
4409 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4410 // based on the escape status of the AllocateNode.
4411 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4412 } else {
4413 insert_mem_bar(Op_MemBarCPUOrder);
4414 }
4415 }
4416
4417 //------------------------inline_native_clone----------------------------
4418 // protected native Object java.lang.Object.clone();
4419 //
4420 // Here are the simple edge cases:
4421 // null receiver => normal trap
4422 // virtual and clone was overridden => slow path to out-of-line clone
4423 // not cloneable or finalizer => slow path to out-of-line Object.clone
4424 //
4425 // The general case has two steps, allocation and copying.
4426 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4427 //
4428 // Copying also has two cases, oop arrays and everything else.
4429 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4430 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4431 //
4432 // These steps fold up nicely if and when the cloned object's klass
4433 // can be sharply typed as an object array, a type array, or an instance.
4434 //
4435 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4436 PhiNode* result_val;
4437
4438 // Set the reexecute bit for the interpreter to reexecute
4439 // the bytecode that invokes Object.clone if deoptimization happens.
4440 { PreserveReexecuteState preexecs(this);
4441 jvms()->set_should_reexecute(true);
4442
4443 Node* obj = null_check_receiver();
4444 if (stopped()) return true;
4445
4446 Node* obj_klass = load_object_klass(obj);
4447 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4448 const TypeOopPtr* toop = ((tklass != NULL)
4449 ? tklass->as_instance_type()
4450 : TypeInstPtr::NOTNULL);
4451
4452 // Conservatively insert a memory barrier on all memory slices.
4453 // Do not let writes into the original float below the clone.
4454 insert_mem_bar(Op_MemBarCPUOrder);
4455
4456 // paths into result_reg:
4457 enum {
4458 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4459 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4460 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4461 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4462 PATH_LIMIT
4463 };
4464 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4465 result_val = new(C) PhiNode(result_reg,
4466 TypeInstPtr::NOTNULL);
4467 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4468 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4469 TypePtr::BOTTOM);
4470 record_for_igvn(result_reg);
4471
4472 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4473 int raw_adr_idx = Compile::AliasIdxRaw;
4474
4475 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4476 if (array_ctl != NULL) {
4477 // It's an array.
4478 PreserveJVMState pjvms(this);
4479 set_control(array_ctl);
4480 Node* obj_length = load_array_length(obj);
4481 Node* obj_size = NULL;
4482 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4483
4484 if (!use_ReduceInitialCardMarks()) {
4485 // If it is an oop array, it requires very special treatment,
4486 // because card marking is required on each card of the array.
4487 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4488 if (is_obja != NULL) {
4489 PreserveJVMState pjvms2(this);
4490 set_control(is_obja);
4491 // Generate a direct call to the right arraycopy function(s).
4492 bool disjoint_bases = true;
4493 bool length_never_negative = true;
4494 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4495 obj, intcon(0), alloc_obj, intcon(0),
4496 obj_length,
4497 disjoint_bases, length_never_negative);
4498 result_reg->init_req(_objArray_path, control());
4499 result_val->init_req(_objArray_path, alloc_obj);
4500 result_i_o ->set_req(_objArray_path, i_o());
4501 result_mem ->set_req(_objArray_path, reset_memory());
4502 }
4503 }
4504 // Otherwise, there are no card marks to worry about.
4505 // (We can dispense with card marks if we know the allocation
4506 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4507 // causes the non-eden paths to take compensating steps to
4508 // simulate a fresh allocation, so that no further
4509 // card marks are required in compiled code to initialize
4510 // the object.)
4511
4512 if (!stopped()) {
4513 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4514
4515 // Present the results of the copy.
4516 result_reg->init_req(_array_path, control());
4517 result_val->init_req(_array_path, alloc_obj);
4518 result_i_o ->set_req(_array_path, i_o());
4519 result_mem ->set_req(_array_path, reset_memory());
4520 }
4521 }
4522
4523 // We only go to the instance fast case code if we pass a number of guards.
4524 // The paths which do not pass are accumulated in the slow_region.
4525 RegionNode* slow_region = new (C) RegionNode(1);
4526 record_for_igvn(slow_region);
4527 if (!stopped()) {
4528 // It's an instance (we did array above). Make the slow-path tests.
4529 // If this is a virtual call, we generate a funny guard. We grab
4530 // the vtable entry corresponding to clone() from the target object.
4531 // If the target method which we are calling happens to be the
4532 // Object clone() method, we pass the guard. We do not need this
4533 // guard for non-virtual calls; the caller is known to be the native
4534 // Object clone().
4535 if (is_virtual) {
4536 generate_virtual_guard(obj_klass, slow_region);
4537 }
4538
4539 // The object must be cloneable and must not have a finalizer.
4540 // Both of these conditions may be checked in a single test.
4541 // We could optimize the cloneable test further, but we don't care.
4542 generate_access_flags_guard(obj_klass,
4543 // Test both conditions:
4544 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4545 // Must be cloneable but not finalizer:
4546 JVM_ACC_IS_CLONEABLE,
4547 slow_region);
4548 }
4549
4550 if (!stopped()) {
4551 // It's an instance, and it passed the slow-path tests.
4552 PreserveJVMState pjvms(this);
4553 Node* obj_size = NULL;
4554 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4555
4556 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4557
4558 // Present the results of the slow call.
4559 result_reg->init_req(_instance_path, control());
4560 result_val->init_req(_instance_path, alloc_obj);
4561 result_i_o ->set_req(_instance_path, i_o());
4562 result_mem ->set_req(_instance_path, reset_memory());
4563 }
4564
4565 // Generate code for the slow case. We make a call to clone().
4566 set_control(_gvn.transform(slow_region));
4567 if (!stopped()) {
4568 PreserveJVMState pjvms(this);
4569 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4570 Node* slow_result = set_results_for_java_call(slow_call);
4571 // this->control() comes from set_results_for_java_call
4572 result_reg->init_req(_slow_path, control());
4573 result_val->init_req(_slow_path, slow_result);
4574 result_i_o ->set_req(_slow_path, i_o());
4575 result_mem ->set_req(_slow_path, reset_memory());
4576 }
4577
4578 // Return the combined state.
4579 set_control( _gvn.transform(result_reg));
4580 set_i_o( _gvn.transform(result_i_o));
4581 set_all_memory( _gvn.transform(result_mem));
4582 } // original reexecute is set back here
4583
4584 set_result(_gvn.transform(result_val));
4585 return true;
4586 }
4587
4588 //------------------------------basictype2arraycopy----------------------------
4589 address LibraryCallKit::basictype2arraycopy(BasicType t,
4590 Node* src_offset,
4591 Node* dest_offset,
4592 bool disjoint_bases,
4593 const char* &name,
4594 bool dest_uninitialized) {
4595 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4596 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4597
4598 bool aligned = false;
4599 bool disjoint = disjoint_bases;
4600
4601 // if the offsets are the same, we can treat the memory regions as
4602 // disjoint, because either the memory regions are in different arrays,
4603 // or they are identical (which we can treat as disjoint.) We can also
4604 // treat a copy with a destination index less that the source index
4605 // as disjoint since a low->high copy will work correctly in this case.
4606 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4607 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4608 // both indices are constants
4609 int s_offs = src_offset_inttype->get_con();
4610 int d_offs = dest_offset_inttype->get_con();
4611 int element_size = type2aelembytes(t);
4612 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4613 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4614 if (s_offs >= d_offs) disjoint = true;
4615 } else if (src_offset == dest_offset && src_offset != NULL) {
4616 // This can occur if the offsets are identical non-constants.
4617 disjoint = true;
4618 }
4619
4620 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4621 }
4622
4623
4624 //------------------------------inline_arraycopy-----------------------
4625 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4626 // Object dest, int destPos,
4627 // int length);
4628 bool LibraryCallKit::inline_arraycopy() {
4629 // Get the arguments.
4630 Node* src = argument(0); // type: oop
4631 Node* src_offset = argument(1); // type: int
4632 Node* dest = argument(2); // type: oop
4633 Node* dest_offset = argument(3); // type: int
4634 Node* length = argument(4); // type: int
4635
4636 // Compile time checks. If any of these checks cannot be verified at compile time,
4637 // we do not make a fast path for this call. Instead, we let the call remain as it
4638 // is. The checks we choose to mandate at compile time are:
4639 //
4640 // (1) src and dest are arrays.
4641 const Type* src_type = src->Value(&_gvn);
4642 const Type* dest_type = dest->Value(&_gvn);
4643 const TypeAryPtr* top_src = src_type->isa_aryptr();
4644 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4645
4646 // Do we have the type of src?
4647 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4648 // Do we have the type of dest?
4649 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4650 // Is the type for src from speculation?
4651 bool src_spec = false;
4652 // Is the type for dest from speculation?
4653 bool dest_spec = false;
4654
4655 if (!has_src || !has_dest) {
4656 // We don't have sufficient type information, let's see if
4657 // speculative types can help. We need to have types for both src
4658 // and dest so that it pays off.
4659
4660 // Do we already have or could we have type information for src
4661 bool could_have_src = has_src;
4662 // Do we already have or could we have type information for dest
4663 bool could_have_dest = has_dest;
4664
4665 ciKlass* src_k = NULL;
4666 if (!has_src) {
4667 src_k = src_type->speculative_type();
4668 if (src_k != NULL && src_k->is_array_klass()) {
4669 could_have_src = true;
4670 }
4671 }
4672
4673 ciKlass* dest_k = NULL;
4674 if (!has_dest) {
4675 dest_k = dest_type->speculative_type();
4676 if (dest_k != NULL && dest_k->is_array_klass()) {
4677 could_have_dest = true;
4678 }
4679 }
4680
4681 if (could_have_src && could_have_dest) {
4682 // This is going to pay off so emit the required guards
4683 if (!has_src) {
4684 src = maybe_cast_profiled_obj(src, src_k);
4685 src_type = _gvn.type(src);
4686 top_src = src_type->isa_aryptr();
4687 has_src = (top_src != NULL && top_src->klass() != NULL);
4688 src_spec = true;
4689 }
4690 if (!has_dest) {
4691 dest = maybe_cast_profiled_obj(dest, dest_k);
4692 dest_type = _gvn.type(dest);
4693 top_dest = dest_type->isa_aryptr();
4694 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4695 dest_spec = true;
4696 }
4697 }
4698 }
4699
4700 if (!has_src || !has_dest) {
4701 // Conservatively insert a memory barrier on all memory slices.
4702 // Do not let writes into the source float below the arraycopy.
4703 insert_mem_bar(Op_MemBarCPUOrder);
4704
4705 // Call StubRoutines::generic_arraycopy stub.
4706 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4707 src, src_offset, dest, dest_offset, length);
4708
4709 // Do not let reads from the destination float above the arraycopy.
4710 // Since we cannot type the arrays, we don't know which slices
4711 // might be affected. We could restrict this barrier only to those
4712 // memory slices which pertain to array elements--but don't bother.
4713 if (!InsertMemBarAfterArraycopy)
4714 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4715 insert_mem_bar(Op_MemBarCPUOrder);
4716 return true;
4717 }
4718
4719 // (2) src and dest arrays must have elements of the same BasicType
4720 // Figure out the size and type of the elements we will be copying.
4721 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4722 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4723 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4724 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4725
4726 if (src_elem != dest_elem || dest_elem == T_VOID) {
4727 // The component types are not the same or are not recognized. Punt.
4728 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4729 generate_slow_arraycopy(TypePtr::BOTTOM,
4730 src, src_offset, dest, dest_offset, length,
4731 /*dest_uninitialized*/false);
4732 return true;
4733 }
4734
4735 if (src_elem == T_OBJECT) {
4736 // If both arrays are object arrays then having the exact types
4737 // for both will remove the need for a subtype check at runtime
4738 // before the call and may make it possible to pick a faster copy
4739 // routine (without a subtype check on every element)
4740 // Do we have the exact type of src?
4741 bool could_have_src = src_spec;
4742 // Do we have the exact type of dest?
4743 bool could_have_dest = dest_spec;
4744 ciKlass* src_k = top_src->klass();
4745 ciKlass* dest_k = top_dest->klass();
4746 if (!src_spec) {
4747 src_k = src_type->speculative_type();
4748 if (src_k != NULL && src_k->is_array_klass()) {
4749 could_have_src = true;
4750 }
4751 }
4752 if (!dest_spec) {
4753 dest_k = dest_type->speculative_type();
4754 if (dest_k != NULL && dest_k->is_array_klass()) {
4755 could_have_dest = true;
4756 }
4757 }
4758 if (could_have_src && could_have_dest) {
4759 // If we can have both exact types, emit the missing guards
4760 if (could_have_src && !src_spec) {
4761 src = maybe_cast_profiled_obj(src, src_k);
4762 }
4763 if (could_have_dest && !dest_spec) {
4764 dest = maybe_cast_profiled_obj(dest, dest_k);
4765 }
4766 }
4767 }
4768
4769 //---------------------------------------------------------------------------
4770 // We will make a fast path for this call to arraycopy.
4771
4772 // We have the following tests left to perform:
4773 //
4774 // (3) src and dest must not be null.
4775 // (4) src_offset must not be negative.
4776 // (5) dest_offset must not be negative.
4777 // (6) length must not be negative.
4778 // (7) src_offset + length must not exceed length of src.
4779 // (8) dest_offset + length must not exceed length of dest.
4780 // (9) each element of an oop array must be assignable
4781
4782 RegionNode* slow_region = new (C) RegionNode(1);
4783 record_for_igvn(slow_region);
4784
4785 // (3) operands must not be null
4786 // We currently perform our null checks with the null_check routine.
4787 // This means that the null exceptions will be reported in the caller
4788 // rather than (correctly) reported inside of the native arraycopy call.
4789 // This should be corrected, given time. We do our null check with the
4790 // stack pointer restored.
4791 src = null_check(src, T_ARRAY);
4792 dest = null_check(dest, T_ARRAY);
4793
4794 // (4) src_offset must not be negative.
4795 generate_negative_guard(src_offset, slow_region);
4796
4797 // (5) dest_offset must not be negative.
4798 generate_negative_guard(dest_offset, slow_region);
4799
4800 // (6) length must not be negative (moved to generate_arraycopy()).
4801 // generate_negative_guard(length, slow_region);
4802
4803 // (7) src_offset + length must not exceed length of src.
4804 generate_limit_guard(src_offset, length,
4805 load_array_length(src),
4806 slow_region);
4807
4808 // (8) dest_offset + length must not exceed length of dest.
4809 generate_limit_guard(dest_offset, length,
4810 load_array_length(dest),
4811 slow_region);
4812
4813 // (9) each element of an oop array must be assignable
4814 // The generate_arraycopy subroutine checks this.
4815
4816 // This is where the memory effects are placed:
4817 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4818 generate_arraycopy(adr_type, dest_elem,
4819 src, src_offset, dest, dest_offset, length,
4820 false, false, slow_region);
4821
4822 return true;
4823 }
4824
4825 //-----------------------------generate_arraycopy----------------------
4826 // Generate an optimized call to arraycopy.
4827 // Caller must guard against non-arrays.
4828 // Caller must determine a common array basic-type for both arrays.
4829 // Caller must validate offsets against array bounds.
4830 // The slow_region has already collected guard failure paths
4831 // (such as out of bounds length or non-conformable array types).
4832 // The generated code has this shape, in general:
4833 //
4834 // if (length == 0) return // via zero_path
4835 // slowval = -1
4836 // if (types unknown) {
4837 // slowval = call generic copy loop
4838 // if (slowval == 0) return // via checked_path
4839 // } else if (indexes in bounds) {
4840 // if ((is object array) && !(array type check)) {
4841 // slowval = call checked copy loop
4842 // if (slowval == 0) return // via checked_path
4843 // } else {
4844 // call bulk copy loop
4845 // return // via fast_path
4846 // }
4847 // }
4848 // // adjust params for remaining work:
4849 // if (slowval != -1) {
4850 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4851 // }
4852 // slow_region:
4853 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4854 // return // via slow_call_path
4855 //
4856 // This routine is used from several intrinsics: System.arraycopy,
4857 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4858 //
4859 void
4860 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4861 BasicType basic_elem_type,
4862 Node* src, Node* src_offset,
4863 Node* dest, Node* dest_offset,
4864 Node* copy_length,
4865 bool disjoint_bases,
4866 bool length_never_negative,
4867 RegionNode* slow_region) {
4868
4869 if (slow_region == NULL) {
4870 slow_region = new(C) RegionNode(1);
4871 record_for_igvn(slow_region);
4872 }
4873
4874 Node* original_dest = dest;
4875 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4876 bool dest_uninitialized = false;
4877
4878 // See if this is the initialization of a newly-allocated array.
4879 // If so, we will take responsibility here for initializing it to zero.
4880 // (Note: Because tightly_coupled_allocation performs checks on the
4881 // out-edges of the dest, we need to avoid making derived pointers
4882 // from it until we have checked its uses.)
4883 if (ReduceBulkZeroing
4884 && !ZeroTLAB // pointless if already zeroed
4885 && basic_elem_type != T_CONFLICT // avoid corner case
4886 && !src->eqv_uncast(dest)
4887 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4888 != NULL)
4889 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4890 && alloc->maybe_set_complete(&_gvn)) {
4891 // "You break it, you buy it."
4892 InitializeNode* init = alloc->initialization();
4893 assert(init->is_complete(), "we just did this");
4894 init->set_complete_with_arraycopy();
4895 assert(dest->is_CheckCastPP(), "sanity");
4896 assert(dest->in(0)->in(0) == init, "dest pinned");
4897 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4898 // From this point on, every exit path is responsible for
4899 // initializing any non-copied parts of the object to zero.
4900 // Also, if this flag is set we make sure that arraycopy interacts properly
4901 // with G1, eliding pre-barriers. See CR 6627983.
4902 dest_uninitialized = true;
4903 } else {
4904 // No zeroing elimination here.
4905 alloc = NULL;
4906 //original_dest = dest;
4907 //dest_uninitialized = false;
4908 }
4909
4910 // Results are placed here:
4911 enum { fast_path = 1, // normal void-returning assembly stub
4912 checked_path = 2, // special assembly stub with cleanup
4913 slow_call_path = 3, // something went wrong; call the VM
4914 zero_path = 4, // bypass when length of copy is zero
4915 bcopy_path = 5, // copy primitive array by 64-bit blocks
4916 PATH_LIMIT = 6
4917 };
4918 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4919 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4920 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4921 record_for_igvn(result_region);
4922 _gvn.set_type_bottom(result_i_o);
4923 _gvn.set_type_bottom(result_memory);
4924 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4925
4926 // The slow_control path:
4927 Node* slow_control;
4928 Node* slow_i_o = i_o();
4929 Node* slow_mem = memory(adr_type);
4930 debug_only(slow_control = (Node*) badAddress);
4931
4932 // Checked control path:
4933 Node* checked_control = top();
4934 Node* checked_mem = NULL;
4935 Node* checked_i_o = NULL;
4936 Node* checked_value = NULL;
4937
4938 if (basic_elem_type == T_CONFLICT) {
4939 assert(!dest_uninitialized, "");
4940 Node* cv = generate_generic_arraycopy(adr_type,
4941 src, src_offset, dest, dest_offset,
4942 copy_length, dest_uninitialized);
4943 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4944 checked_control = control();
4945 checked_i_o = i_o();
4946 checked_mem = memory(adr_type);
4947 checked_value = cv;
4948 set_control(top()); // no fast path
4949 }
4950
4951 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4952 if (not_pos != NULL) {
4953 PreserveJVMState pjvms(this);
4954 set_control(not_pos);
4955
4956 // (6) length must not be negative.
4957 if (!length_never_negative) {
4958 generate_negative_guard(copy_length, slow_region);
4959 }
4960
4961 // copy_length is 0.
4962 if (!stopped() && dest_uninitialized) {
4963 Node* dest_length = alloc->in(AllocateNode::ALength);
4964 if (copy_length->eqv_uncast(dest_length)
4965 || _gvn.find_int_con(dest_length, 1) <= 0) {
4966 // There is no zeroing to do. No need for a secondary raw memory barrier.
4967 } else {
4968 // Clear the whole thing since there are no source elements to copy.
4969 generate_clear_array(adr_type, dest, basic_elem_type,
4970 intcon(0), NULL,
4971 alloc->in(AllocateNode::AllocSize));
4972 // Use a secondary InitializeNode as raw memory barrier.
4973 // Currently it is needed only on this path since other
4974 // paths have stub or runtime calls as raw memory barriers.
4975 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4976 Compile::AliasIdxRaw,
4977 top())->as_Initialize();
4978 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4979 }
4980 }
4981
4982 // Present the results of the fast call.
4983 result_region->init_req(zero_path, control());
4984 result_i_o ->init_req(zero_path, i_o());
4985 result_memory->init_req(zero_path, memory(adr_type));
4986 }
4987
4988 if (!stopped() && dest_uninitialized) {
4989 // We have to initialize the *uncopied* part of the array to zero.
4990 // The copy destination is the slice dest[off..off+len]. The other slices
4991 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4992 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4993 Node* dest_length = alloc->in(AllocateNode::ALength);
4994 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4995 copy_length));
4996
4997 // If there is a head section that needs zeroing, do it now.
4998 if (find_int_con(dest_offset, -1) != 0) {
4999 generate_clear_array(adr_type, dest, basic_elem_type,
5000 intcon(0), dest_offset,
5001 NULL);
5002 }
5003
5004 // Next, perform a dynamic check on the tail length.
5005 // It is often zero, and we can win big if we prove this.
5006 // There are two wins: Avoid generating the ClearArray
5007 // with its attendant messy index arithmetic, and upgrade
5008 // the copy to a more hardware-friendly word size of 64 bits.
5009 Node* tail_ctl = NULL;
5010 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5011 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5012 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5013 tail_ctl = generate_slow_guard(bol_lt, NULL);
5014 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5015 }
5016
5017 // At this point, let's assume there is no tail.
5018 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5019 // There is no tail. Try an upgrade to a 64-bit copy.
5020 bool didit = false;
5021 { PreserveJVMState pjvms(this);
5022 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5023 src, src_offset, dest, dest_offset,
5024 dest_size, dest_uninitialized);
5025 if (didit) {
5026 // Present the results of the block-copying fast call.
5027 result_region->init_req(bcopy_path, control());
5028 result_i_o ->init_req(bcopy_path, i_o());
5029 result_memory->init_req(bcopy_path, memory(adr_type));
5030 }
5031 }
5032 if (didit)
5033 set_control(top()); // no regular fast path
5034 }
5035
5036 // Clear the tail, if any.
5037 if (tail_ctl != NULL) {
5038 Node* notail_ctl = stopped() ? NULL : control();
5039 set_control(tail_ctl);
5040 if (notail_ctl == NULL) {
5041 generate_clear_array(adr_type, dest, basic_elem_type,
5042 dest_tail, NULL,
5043 dest_size);
5044 } else {
5045 // Make a local merge.
5046 Node* done_ctl = new(C) RegionNode(3);
5047 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5048 done_ctl->init_req(1, notail_ctl);
5049 done_mem->init_req(1, memory(adr_type));
5050 generate_clear_array(adr_type, dest, basic_elem_type,
5051 dest_tail, NULL,
5052 dest_size);
5053 done_ctl->init_req(2, control());
5054 done_mem->init_req(2, memory(adr_type));
5055 set_control( _gvn.transform(done_ctl));
5056 set_memory( _gvn.transform(done_mem), adr_type );
5057 }
5058 }
5059 }
5060
5061 BasicType copy_type = basic_elem_type;
5062 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5063 if (!stopped() && copy_type == T_OBJECT) {
5064 // If src and dest have compatible element types, we can copy bits.
5065 // Types S[] and D[] are compatible if D is a supertype of S.
5066 //
5067 // If they are not, we will use checked_oop_disjoint_arraycopy,
5068 // which performs a fast optimistic per-oop check, and backs off
5069 // further to JVM_ArrayCopy on the first per-oop check that fails.
5070 // (Actually, we don't move raw bits only; the GC requires card marks.)
5071
5072 // Get the Klass* for both src and dest
5073 Node* src_klass = load_object_klass(src);
5074 Node* dest_klass = load_object_klass(dest);
5075
5076 // Generate the subtype check.
5077 // This might fold up statically, or then again it might not.
5078 //
5079 // Non-static example: Copying List<String>.elements to a new String[].
5080 // The backing store for a List<String> is always an Object[],
5081 // but its elements are always type String, if the generic types
5082 // are correct at the source level.
5083 //
5084 // Test S[] against D[], not S against D, because (probably)
5085 // the secondary supertype cache is less busy for S[] than S.
5086 // This usually only matters when D is an interface.
5087 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5088 // Plug failing path into checked_oop_disjoint_arraycopy
5089 if (not_subtype_ctrl != top()) {
5090 PreserveJVMState pjvms(this);
5091 set_control(not_subtype_ctrl);
5092 // (At this point we can assume disjoint_bases, since types differ.)
5093 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5094 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5095 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5096 Node* dest_elem_klass = _gvn.transform(n1);
5097 Node* cv = generate_checkcast_arraycopy(adr_type,
5098 dest_elem_klass,
5099 src, src_offset, dest, dest_offset,
5100 ConvI2X(copy_length), dest_uninitialized);
5101 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5102 checked_control = control();
5103 checked_i_o = i_o();
5104 checked_mem = memory(adr_type);
5105 checked_value = cv;
5106 }
5107 // At this point we know we do not need type checks on oop stores.
5108
5109 // Let's see if we need card marks:
5110 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5111 // If we do not need card marks, copy using the jint or jlong stub.
5112 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5113 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5114 "sizes agree");
5115 }
5116 }
5117
5118 if (!stopped()) {
5119 // Generate the fast path, if possible.
5120 PreserveJVMState pjvms(this);
5121 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5122 src, src_offset, dest, dest_offset,
5123 ConvI2X(copy_length), dest_uninitialized);
5124
5125 // Present the results of the fast call.
5126 result_region->init_req(fast_path, control());
5127 result_i_o ->init_req(fast_path, i_o());
5128 result_memory->init_req(fast_path, memory(adr_type));
5129 }
5130
5131 // Here are all the slow paths up to this point, in one bundle:
5132 slow_control = top();
5133 if (slow_region != NULL)
5134 slow_control = _gvn.transform(slow_region);
5135 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5136
5137 set_control(checked_control);
5138 if (!stopped()) {
5139 // Clean up after the checked call.
5140 // The returned value is either 0 or -1^K,
5141 // where K = number of partially transferred array elements.
5142 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5143 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5144 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5145
5146 // If it is 0, we are done, so transfer to the end.
5147 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5148 result_region->init_req(checked_path, checks_done);
5149 result_i_o ->init_req(checked_path, checked_i_o);
5150 result_memory->init_req(checked_path, checked_mem);
5151
5152 // If it is not zero, merge into the slow call.
5153 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5154 RegionNode* slow_reg2 = new(C) RegionNode(3);
5155 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5156 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5157 record_for_igvn(slow_reg2);
5158 slow_reg2 ->init_req(1, slow_control);
5159 slow_i_o2 ->init_req(1, slow_i_o);
5160 slow_mem2 ->init_req(1, slow_mem);
5161 slow_reg2 ->init_req(2, control());
5162 slow_i_o2 ->init_req(2, checked_i_o);
5163 slow_mem2 ->init_req(2, checked_mem);
5164
5165 slow_control = _gvn.transform(slow_reg2);
5166 slow_i_o = _gvn.transform(slow_i_o2);
5167 slow_mem = _gvn.transform(slow_mem2);
5168
5169 if (alloc != NULL) {
5170 // We'll restart from the very beginning, after zeroing the whole thing.
5171 // This can cause double writes, but that's OK since dest is brand new.
5172 // So we ignore the low 31 bits of the value returned from the stub.
5173 } else {
5174 // We must continue the copy exactly where it failed, or else
5175 // another thread might see the wrong number of writes to dest.
5176 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5177 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5178 slow_offset->init_req(1, intcon(0));
5179 slow_offset->init_req(2, checked_offset);
5180 slow_offset = _gvn.transform(slow_offset);
5181
5182 // Adjust the arguments by the conditionally incoming offset.
5183 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5184 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5185 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5186
5187 // Tweak the node variables to adjust the code produced below:
5188 src_offset = src_off_plus;
5189 dest_offset = dest_off_plus;
5190 copy_length = length_minus;
5191 }
5192 }
5193
5194 set_control(slow_control);
5195 if (!stopped()) {
5196 // Generate the slow path, if needed.
5197 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5198
5199 set_memory(slow_mem, adr_type);
5200 set_i_o(slow_i_o);
5201
5202 if (dest_uninitialized) {
5203 generate_clear_array(adr_type, dest, basic_elem_type,
5204 intcon(0), NULL,
5205 alloc->in(AllocateNode::AllocSize));
5206 }
5207
5208 generate_slow_arraycopy(adr_type,
5209 src, src_offset, dest, dest_offset,
5210 copy_length, /*dest_uninitialized*/false);
5211
5212 result_region->init_req(slow_call_path, control());
5213 result_i_o ->init_req(slow_call_path, i_o());
5214 result_memory->init_req(slow_call_path, memory(adr_type));
5215 }
5216
5217 // Remove unused edges.
5218 for (uint i = 1; i < result_region->req(); i++) {
5219 if (result_region->in(i) == NULL)
5220 result_region->init_req(i, top());
5221 }
5222
5223 // Finished; return the combined state.
5224 set_control( _gvn.transform(result_region));
5225 set_i_o( _gvn.transform(result_i_o) );
5226 set_memory( _gvn.transform(result_memory), adr_type );
5227
5228 // The memory edges above are precise in order to model effects around
5229 // array copies accurately to allow value numbering of field loads around
5230 // arraycopy. Such field loads, both before and after, are common in Java
5231 // collections and similar classes involving header/array data structures.
5232 //
5233 // But with low number of register or when some registers are used or killed
5234 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5235 // The next memory barrier is added to avoid it. If the arraycopy can be
5236 // optimized away (which it can, sometimes) then we can manually remove
5237 // the membar also.
5238 //
5239 // Do not let reads from the cloned object float above the arraycopy.
5240 if (alloc != NULL) {
5241 // Do not let stores that initialize this object be reordered with
5242 // a subsequent store that would make this object accessible by
5243 // other threads.
5244 // Record what AllocateNode this StoreStore protects so that
5245 // escape analysis can go from the MemBarStoreStoreNode to the
5246 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5247 // based on the escape status of the AllocateNode.
5248 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5249 } else if (InsertMemBarAfterArraycopy)
5250 insert_mem_bar(Op_MemBarCPUOrder);
5251 }
5252
5253
5254 // Helper function which determines if an arraycopy immediately follows
5255 // an allocation, with no intervening tests or other escapes for the object.
5256 AllocateArrayNode*
5257 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5258 RegionNode* slow_region) {
5259 if (stopped()) return NULL; // no fast path
5260 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5261
5262 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5263 if (alloc == NULL) return NULL;
5264
5265 Node* rawmem = memory(Compile::AliasIdxRaw);
5266 // Is the allocation's memory state untouched?
5267 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5268 // Bail out if there have been raw-memory effects since the allocation.
5269 // (Example: There might have been a call or safepoint.)
5270 return NULL;
5271 }
5272 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5273 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5274 return NULL;
5275 }
5276
5277 // There must be no unexpected observers of this allocation.
5278 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5279 Node* obs = ptr->fast_out(i);
5280 if (obs != this->map()) {
5281 return NULL;
5282 }
5283 }
5284
5285 // This arraycopy must unconditionally follow the allocation of the ptr.
5286 Node* alloc_ctl = ptr->in(0);
5287 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5288
5289 Node* ctl = control();
5290 while (ctl != alloc_ctl) {
5291 // There may be guards which feed into the slow_region.
5292 // Any other control flow means that we might not get a chance
5293 // to finish initializing the allocated object.
5294 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5295 IfNode* iff = ctl->in(0)->as_If();
5296 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5297 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5298 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5299 ctl = iff->in(0); // This test feeds the known slow_region.
5300 continue;
5301 }
5302 // One more try: Various low-level checks bottom out in
5303 // uncommon traps. If the debug-info of the trap omits
5304 // any reference to the allocation, as we've already
5305 // observed, then there can be no objection to the trap.
5306 bool found_trap = false;
5307 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5308 Node* obs = not_ctl->fast_out(j);
5309 if (obs->in(0) == not_ctl && obs->is_Call() &&
5310 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5311 found_trap = true; break;
5312 }
5313 }
5314 if (found_trap) {
5315 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5316 continue;
5317 }
5318 }
5319 return NULL;
5320 }
5321
5322 // If we get this far, we have an allocation which immediately
5323 // precedes the arraycopy, and we can take over zeroing the new object.
5324 // The arraycopy will finish the initialization, and provide
5325 // a new control state to which we will anchor the destination pointer.
5326
5327 return alloc;
5328 }
5329
5330 // Helper for initialization of arrays, creating a ClearArray.
5331 // It writes zero bits in [start..end), within the body of an array object.
5332 // The memory effects are all chained onto the 'adr_type' alias category.
5333 //
5334 // Since the object is otherwise uninitialized, we are free
5335 // to put a little "slop" around the edges of the cleared area,
5336 // as long as it does not go back into the array's header,
5337 // or beyond the array end within the heap.
5338 //
5339 // The lower edge can be rounded down to the nearest jint and the
5340 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5341 //
5342 // Arguments:
5343 // adr_type memory slice where writes are generated
5344 // dest oop of the destination array
5345 // basic_elem_type element type of the destination
5346 // slice_idx array index of first element to store
5347 // slice_len number of elements to store (or NULL)
5348 // dest_size total size in bytes of the array object
5349 //
5350 // Exactly one of slice_len or dest_size must be non-NULL.
5351 // If dest_size is non-NULL, zeroing extends to the end of the object.
5352 // If slice_len is non-NULL, the slice_idx value must be a constant.
5353 void
5354 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5355 Node* dest,
5356 BasicType basic_elem_type,
5357 Node* slice_idx,
5358 Node* slice_len,
5359 Node* dest_size) {
5360 // one or the other but not both of slice_len and dest_size:
5361 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5362 if (slice_len == NULL) slice_len = top();
5363 if (dest_size == NULL) dest_size = top();
5364
5365 // operate on this memory slice:
5366 Node* mem = memory(adr_type); // memory slice to operate on
5367
5368 // scaling and rounding of indexes:
5369 int scale = exact_log2(type2aelembytes(basic_elem_type));
5370 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5371 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5372 int bump_bit = (-1 << scale) & BytesPerInt;
5373
5374 // determine constant starts and ends
5375 const intptr_t BIG_NEG = -128;
5376 assert(BIG_NEG + 2*abase < 0, "neg enough");
5377 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5378 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5379 if (slice_len_con == 0) {
5380 return; // nothing to do here
5381 }
5382 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5383 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5384 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5385 assert(end_con < 0, "not two cons");
5386 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5387 BytesPerLong);
5388 }
5389
5390 if (start_con >= 0 && end_con >= 0) {
5391 // Constant start and end. Simple.
5392 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5393 start_con, end_con, &_gvn);
5394 } else if (start_con >= 0 && dest_size != top()) {
5395 // Constant start, pre-rounded end after the tail of the array.
5396 Node* end = dest_size;
5397 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5398 start_con, end, &_gvn);
5399 } else if (start_con >= 0 && slice_len != top()) {
5400 // Constant start, non-constant end. End needs rounding up.
5401 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5402 intptr_t end_base = abase + (slice_idx_con << scale);
5403 int end_round = (-1 << scale) & (BytesPerLong - 1);
5404 Node* end = ConvI2X(slice_len);
5405 if (scale != 0)
5406 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5407 end_base += end_round;
5408 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5409 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5410 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5411 start_con, end, &_gvn);
5412 } else if (start_con < 0 && dest_size != top()) {
5413 // Non-constant start, pre-rounded end after the tail of the array.
5414 // This is almost certainly a "round-to-end" operation.
5415 Node* start = slice_idx;
5416 start = ConvI2X(start);
5417 if (scale != 0)
5418 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5419 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5420 if ((bump_bit | clear_low) != 0) {
5421 int to_clear = (bump_bit | clear_low);
5422 // Align up mod 8, then store a jint zero unconditionally
5423 // just before the mod-8 boundary.
5424 if (((abase + bump_bit) & ~to_clear) - bump_bit
5425 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5426 bump_bit = 0;
5427 assert((abase & to_clear) == 0, "array base must be long-aligned");
5428 } else {
5429 // Bump 'start' up to (or past) the next jint boundary:
5430 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5431 assert((abase & clear_low) == 0, "array base must be int-aligned");
5432 }
5433 // Round bumped 'start' down to jlong boundary in body of array.
5434 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5435 if (bump_bit != 0) {
5436 // Store a zero to the immediately preceding jint:
5437 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5438 Node* p1 = basic_plus_adr(dest, x1);
5439 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5440 mem = _gvn.transform(mem);
5441 }
5442 }
5443 Node* end = dest_size; // pre-rounded
5444 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5445 start, end, &_gvn);
5446 } else {
5447 // Non-constant start, unrounded non-constant end.
5448 // (Nobody zeroes a random midsection of an array using this routine.)
5449 ShouldNotReachHere(); // fix caller
5450 }
5451
5452 // Done.
5453 set_memory(mem, adr_type);
5454 }
5455
5456
5457 bool
5458 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5459 BasicType basic_elem_type,
5460 AllocateNode* alloc,
5461 Node* src, Node* src_offset,
5462 Node* dest, Node* dest_offset,
5463 Node* dest_size, bool dest_uninitialized) {
5464 // See if there is an advantage from block transfer.
5465 int scale = exact_log2(type2aelembytes(basic_elem_type));
5466 if (scale >= LogBytesPerLong)
5467 return false; // it is already a block transfer
5468
5469 // Look at the alignment of the starting offsets.
5470 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5471
5472 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5473 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5474 if (src_off_con < 0 || dest_off_con < 0)
5475 // At present, we can only understand constants.
5476 return false;
5477
5478 intptr_t src_off = abase + (src_off_con << scale);
5479 intptr_t dest_off = abase + (dest_off_con << scale);
5480
5481 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5482 // Non-aligned; too bad.
5483 // One more chance: Pick off an initial 32-bit word.
5484 // This is a common case, since abase can be odd mod 8.
5485 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5486 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5487 Node* sptr = basic_plus_adr(src, src_off);
5488 Node* dptr = basic_plus_adr(dest, dest_off);
5489 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5490 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5491 src_off += BytesPerInt;
5492 dest_off += BytesPerInt;
5493 } else {
5494 return false;
5495 }
5496 }
5497 assert(src_off % BytesPerLong == 0, "");
5498 assert(dest_off % BytesPerLong == 0, "");
5499
5500 // Do this copy by giant steps.
5501 Node* sptr = basic_plus_adr(src, src_off);
5502 Node* dptr = basic_plus_adr(dest, dest_off);
5503 Node* countx = dest_size;
5504 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5505 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5506
5507 bool disjoint_bases = true; // since alloc != NULL
5508 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5509 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5510
5511 return true;
5512 }
5513
5514
5515 // Helper function; generates code for the slow case.
5516 // We make a call to a runtime method which emulates the native method,
5517 // but without the native wrapper overhead.
5518 void
5519 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5520 Node* src, Node* src_offset,
5521 Node* dest, Node* dest_offset,
5522 Node* copy_length, bool dest_uninitialized) {
5523 assert(!dest_uninitialized, "Invariant");
5524 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5525 OptoRuntime::slow_arraycopy_Type(),
5526 OptoRuntime::slow_arraycopy_Java(),
5527 "slow_arraycopy", adr_type,
5528 src, src_offset, dest, dest_offset,
5529 copy_length);
5530
5531 // Handle exceptions thrown by this fellow:
5532 make_slow_call_ex(call, env()->Throwable_klass(), false);
5533 }
5534
5535 // Helper function; generates code for cases requiring runtime checks.
5536 Node*
5537 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5538 Node* dest_elem_klass,
5539 Node* src, Node* src_offset,
5540 Node* dest, Node* dest_offset,
5541 Node* copy_length, bool dest_uninitialized) {
5542 if (stopped()) return NULL;
5543
5544 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5545 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5546 return NULL;
5547 }
5548
5549 // Pick out the parameters required to perform a store-check
5550 // for the target array. This is an optimistic check. It will
5551 // look in each non-null element's class, at the desired klass's
5552 // super_check_offset, for the desired klass.
5553 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5554 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5555 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5556 Node* check_offset = ConvI2X(_gvn.transform(n3));
5557 Node* check_value = dest_elem_klass;
5558
5559 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5560 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5561
5562 // (We know the arrays are never conjoint, because their types differ.)
5563 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5564 OptoRuntime::checkcast_arraycopy_Type(),
5565 copyfunc_addr, "checkcast_arraycopy", adr_type,
5566 // five arguments, of which two are
5567 // intptr_t (jlong in LP64)
5568 src_start, dest_start,
5569 copy_length XTOP,
5570 check_offset XTOP,
5571 check_value);
5572
5573 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5574 }
5575
5576
5577 // Helper function; generates code for cases requiring runtime checks.
5578 Node*
5579 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5580 Node* src, Node* src_offset,
5581 Node* dest, Node* dest_offset,
5582 Node* copy_length, bool dest_uninitialized) {
5583 assert(!dest_uninitialized, "Invariant");
5584 if (stopped()) return NULL;
5585 address copyfunc_addr = StubRoutines::generic_arraycopy();
5586 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5587 return NULL;
5588 }
5589
5590 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5591 OptoRuntime::generic_arraycopy_Type(),
5592 copyfunc_addr, "generic_arraycopy", adr_type,
5593 src, src_offset, dest, dest_offset, copy_length);
5594
5595 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5596 }
5597
5598 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5599 void
5600 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5601 BasicType basic_elem_type,
5602 bool disjoint_bases,
5603 Node* src, Node* src_offset,
5604 Node* dest, Node* dest_offset,
5605 Node* copy_length, bool dest_uninitialized) {
5606 if (stopped()) return; // nothing to do
5607
5608 Node* src_start = src;
5609 Node* dest_start = dest;
5610 if (src_offset != NULL || dest_offset != NULL) {
5611 assert(src_offset != NULL && dest_offset != NULL, "");
5612 src_start = array_element_address(src, src_offset, basic_elem_type);
5613 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5614 }
5615
5616 // Figure out which arraycopy runtime method to call.
5617 const char* copyfunc_name = "arraycopy";
5618 address copyfunc_addr =
5619 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5620 disjoint_bases, copyfunc_name, dest_uninitialized);
5621
5622 // Call it. Note that the count_ix value is not scaled to a byte-size.
5623 make_runtime_call(RC_LEAF|RC_NO_FP,
5624 OptoRuntime::fast_arraycopy_Type(),
5625 copyfunc_addr, copyfunc_name, adr_type,
5626 src_start, dest_start, copy_length XTOP);
5627 }
5628
5629 //-------------inline_encodeISOArray-----------------------------------
5630 // encode char[] to byte[] in ISO_8859_1
5631 bool LibraryCallKit::inline_encodeISOArray() {
5632 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5633 // no receiver since it is static method
5634 Node *src = argument(0);
5635 Node *src_offset = argument(1);
5636 Node *dst = argument(2);
5637 Node *dst_offset = argument(3);
5638 Node *length = argument(4);
5639
5640 const Type* src_type = src->Value(&_gvn);
5641 const Type* dst_type = dst->Value(&_gvn);
5642 const TypeAryPtr* top_src = src_type->isa_aryptr();
5643 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5644 if (top_src == NULL || top_src->klass() == NULL ||
5645 top_dest == NULL || top_dest->klass() == NULL) {
5646 // failed array check
5647 return false;
5648 }
5649
5650 // Figure out the size and type of the elements we will be copying.
5651 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5652 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5653 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5654 return false;
5655 }
5656 Node* src_start = array_element_address(src, src_offset, src_elem);
5657 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5658 // 'src_start' points to src array + scaled offset
5659 // 'dst_start' points to dst array + scaled offset
5660
5661 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5662 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5663 enc = _gvn.transform(enc);
5664 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5665 set_memory(res_mem, mtype);
5666 set_result(enc);
5667 return true;
5668 }
5669
5670 /**
5671 * Calculate CRC32 for byte.
5672 * int java.util.zip.CRC32.update(int crc, int b)
5673 */
5674 bool LibraryCallKit::inline_updateCRC32() {
5675 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5676 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5677 // no receiver since it is static method
5678 Node* crc = argument(0); // type: int
5679 Node* b = argument(1); // type: int
5680
5681 /*
5682 * int c = ~ crc;
5683 * b = timesXtoThe32[(b ^ c) & 0xFF];
5684 * b = b ^ (c >>> 8);
5685 * crc = ~b;
5686 */
5687
5688 Node* M1 = intcon(-1);
5689 crc = _gvn.transform(new (C) XorINode(crc, M1));
5690 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5691 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5692
5693 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5694 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5695 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5696 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5697
5698 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5699 result = _gvn.transform(new (C) XorINode(crc, result));
5700 result = _gvn.transform(new (C) XorINode(result, M1));
5701 set_result(result);
5702 return true;
5703 }
5704
5705 /**
5706 * Calculate CRC32 for byte[] array.
5707 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5708 */
5709 bool LibraryCallKit::inline_updateBytesCRC32() {
5710 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5711 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5712 // no receiver since it is static method
5713 Node* crc = argument(0); // type: int
5714 Node* src = argument(1); // type: oop
5715 Node* offset = argument(2); // type: int
5716 Node* length = argument(3); // type: int
5717
5718 const Type* src_type = src->Value(&_gvn);
5719 const TypeAryPtr* top_src = src_type->isa_aryptr();
5720 if (top_src == NULL || top_src->klass() == NULL) {
5721 // failed array check
5722 return false;
5723 }
5724
5725 // Figure out the size and type of the elements we will be copying.
5726 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5727 if (src_elem != T_BYTE) {
5728 return false;
5729 }
5730
5731 // 'src_start' points to src array + scaled offset
5732 Node* src_start = array_element_address(src, offset, src_elem);
5733
5734 // We assume that range check is done by caller.
5735 // TODO: generate range check (offset+length < src.length) in debug VM.
5736
5737 // Call the stub.
5738 address stubAddr = StubRoutines::updateBytesCRC32();
5739 const char *stubName = "updateBytesCRC32";
5740
5741 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5742 stubAddr, stubName, TypePtr::BOTTOM,
5743 crc, src_start, length);
5744 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5745 set_result(result);
5746 return true;
5747 }
5748
5749 /**
5750 * Calculate CRC32 for ByteBuffer.
5751 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5752 */
5753 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5754 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5755 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5756 // no receiver since it is static method
5757 Node* crc = argument(0); // type: int
5758 Node* src = argument(1); // type: long
5759 Node* offset = argument(3); // type: int
5760 Node* length = argument(4); // type: int
5761
5762 src = ConvL2X(src); // adjust Java long to machine word
5763 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5764 offset = ConvI2X(offset);
5765
5766 // 'src_start' points to src array + scaled offset
5767 Node* src_start = basic_plus_adr(top(), base, offset);
5768
5769 // Call the stub.
5770 address stubAddr = StubRoutines::updateBytesCRC32();
5771 const char *stubName = "updateBytesCRC32";
5772
5773 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5774 stubAddr, stubName, TypePtr::BOTTOM,
5775 crc, src_start, length);
5776 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5777 set_result(result);
5778 return true;
5779 }
5780
5781 //----------------------------inline_reference_get----------------------------
5782 // public T java.lang.ref.Reference.get();
5783 bool LibraryCallKit::inline_reference_get() {
5784 const int referent_offset = java_lang_ref_Reference::referent_offset;
5785 guarantee(referent_offset > 0, "should have already been set");
5786
5787 // Get the argument:
5788 Node* reference_obj = null_check_receiver();
5789 if (stopped()) return true;
5790
5791 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5792
5793 ciInstanceKlass* klass = env()->Object_klass();
5794 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5795
5796 Node* no_ctrl = NULL;
5797 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5798
5799 // Use the pre-barrier to record the value in the referent field
5800 pre_barrier(false /* do_load */,
5801 control(),
5802 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5803 result /* pre_val */,
5804 T_OBJECT);
5805
5806 // Add memory barrier to prevent commoning reads from this field
5807 // across safepoint since GC can change its value.
5808 insert_mem_bar(Op_MemBarCPUOrder);
5809
5810 set_result(result);
5811 return true;
5812 }
5813
5814
5815 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5816 bool is_exact=true, bool is_static=false) {
5817
5818 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5819 assert(tinst != NULL, "obj is null");
5820 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5821 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5822
5823 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5824 ciSymbol::make(fieldTypeString),
5825 is_static);
5826 if (field == NULL) return (Node *) NULL;
5827 assert (field != NULL, "undefined field");
5828
5829 // Next code copied from Parse::do_get_xxx():
5830
5831 // Compute address and memory type.
5832 int offset = field->offset_in_bytes();
5833 bool is_vol = field->is_volatile();
5834 ciType* field_klass = field->type();
5835 assert(field_klass->is_loaded(), "should be loaded");
5836 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5837 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5838 BasicType bt = field->layout_type();
5839
5840 // Build the resultant type of the load
5841 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5842
5843 // Build the load.
5844 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5845 return loadedField;
5846 }
5847
5848
5849 //------------------------------inline_aescrypt_Block-----------------------
5850 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5851 address stubAddr;
5852 const char *stubName;
5853 assert(UseAES, "need AES instruction support");
5854
5855 switch(id) {
5856 case vmIntrinsics::_aescrypt_encryptBlock:
5857 stubAddr = StubRoutines::aescrypt_encryptBlock();
5858 stubName = "aescrypt_encryptBlock";
5859 break;
5860 case vmIntrinsics::_aescrypt_decryptBlock:
5861 stubAddr = StubRoutines::aescrypt_decryptBlock();
5862 stubName = "aescrypt_decryptBlock";
5863 break;
5864 }
5865 if (stubAddr == NULL) return false;
5866
5867 Node* aescrypt_object = argument(0);
5868 Node* src = argument(1);
5869 Node* src_offset = argument(2);
5870 Node* dest = argument(3);
5871 Node* dest_offset = argument(4);
5872
5873 // (1) src and dest are arrays.
5874 const Type* src_type = src->Value(&_gvn);
5875 const Type* dest_type = dest->Value(&_gvn);
5876 const TypeAryPtr* top_src = src_type->isa_aryptr();
5877 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5878 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5879
5880 // for the quick and dirty code we will skip all the checks.
5881 // we are just trying to get the call to be generated.
5882 Node* src_start = src;
5883 Node* dest_start = dest;
5884 if (src_offset != NULL || dest_offset != NULL) {
5885 assert(src_offset != NULL && dest_offset != NULL, "");
5886 src_start = array_element_address(src, src_offset, T_BYTE);
5887 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5888 }
5889
5890 // now need to get the start of its expanded key array
5891 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5892 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5893 if (k_start == NULL) return false;
5894
5895 if (Matcher::pass_original_key_for_aes()) {
5896 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5897 // compatibility issues between Java key expansion and SPARC crypto instructions
5898 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5899 if (original_k_start == NULL) return false;
5900
5901 // Call the stub.
5902 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5903 stubAddr, stubName, TypePtr::BOTTOM,
5904 src_start, dest_start, k_start, original_k_start);
5905 } else {
5906 // Call the stub.
5907 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5908 stubAddr, stubName, TypePtr::BOTTOM,
5909 src_start, dest_start, k_start);
5910 }
5911
5912 return true;
5913 }
5914
5915 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5916 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5917 address stubAddr;
5918 const char *stubName;
5919
5920 assert(UseAES, "need AES instruction support");
5921
5922 switch(id) {
5923 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5924 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5925 stubName = "cipherBlockChaining_encryptAESCrypt";
5926 break;
5927 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5928 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5929 stubName = "cipherBlockChaining_decryptAESCrypt";
5930 break;
5931 }
5932 if (stubAddr == NULL) return false;
5933
5934 Node* cipherBlockChaining_object = argument(0);
5935 Node* src = argument(1);
5936 Node* src_offset = argument(2);
5937 Node* len = argument(3);
5938 Node* dest = argument(4);
5939 Node* dest_offset = argument(5);
5940
5941 // (1) src and dest are arrays.
5942 const Type* src_type = src->Value(&_gvn);
5943 const Type* dest_type = dest->Value(&_gvn);
5944 const TypeAryPtr* top_src = src_type->isa_aryptr();
5945 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5946 assert (top_src != NULL && top_src->klass() != NULL
5947 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5948
5949 // checks are the responsibility of the caller
5950 Node* src_start = src;
5951 Node* dest_start = dest;
5952 if (src_offset != NULL || dest_offset != NULL) {
5953 assert(src_offset != NULL && dest_offset != NULL, "");
5954 src_start = array_element_address(src, src_offset, T_BYTE);
5955 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5956 }
5957
5958 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5959 // (because of the predicated logic executed earlier).
5960 // so we cast it here safely.
5961 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5962
5963 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5964 if (embeddedCipherObj == NULL) return false;
5965
5966 // cast it to what we know it will be at runtime
5967 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5968 assert(tinst != NULL, "CBC obj is null");
5969 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5970 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5971 if (!klass_AESCrypt->is_loaded()) return false;
5972
5973 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5974 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5975 const TypeOopPtr* xtype = aklass->as_instance_type();
5976 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5977 aescrypt_object = _gvn.transform(aescrypt_object);
5978
5979 // we need to get the start of the aescrypt_object's expanded key array
5980 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5981 if (k_start == NULL) return false;
5982
5983 // similarly, get the start address of the r vector
5984 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5985 if (objRvec == NULL) return false;
5986 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5987
5988 Node* cbcCrypt;
5989 if (Matcher::pass_original_key_for_aes()) {
5990 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5991 // compatibility issues between Java key expansion and SPARC crypto instructions
5992 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5993 if (original_k_start == NULL) return false;
5994
5995 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5996 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5997 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5998 stubAddr, stubName, TypePtr::BOTTOM,
5999 src_start, dest_start, k_start, r_start, len, original_k_start);
6000 } else {
6001 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6002 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6003 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6004 stubAddr, stubName, TypePtr::BOTTOM,
6005 src_start, dest_start, k_start, r_start, len);
6006 }
6007
6008 // return cipher length (int)
6009 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6010 set_result(retvalue);
6011 return true;
6012 }
6013
6014 //------------------------------get_key_start_from_aescrypt_object-----------------------
6015 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6016 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6017 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6018 if (objAESCryptKey == NULL) return (Node *) NULL;
6019
6020 // now have the array, need to get the start address of the K array
6021 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6022 return k_start;
6023 }
6024
6025 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6026 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6027 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6028 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6029 if (objAESCryptKey == NULL) return (Node *) NULL;
6030
6031 // now have the array, need to get the start address of the lastKey array
6032 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6033 return original_k_start;
6034 }
6035
6036 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6037 // Return node representing slow path of predicate check.
6038 // the pseudo code we want to emulate with this predicate is:
6039 // for encryption:
6040 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6041 // for decryption:
6042 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6043 // note cipher==plain is more conservative than the original java code but that's OK
6044 //
6045 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6046 // First, check receiver for NULL since it is virtual method.
6047 Node* objCBC = argument(0);
6048 objCBC = null_check(objCBC);
6049
6050 if (stopped()) return NULL; // Always NULL
6051
6052 // Load embeddedCipher field of CipherBlockChaining object.
6053 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6054
6055 // get AESCrypt klass for instanceOf check
6056 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6057 // will have same classloader as CipherBlockChaining object
6058 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6059 assert(tinst != NULL, "CBCobj is null");
6060 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6061
6062 // we want to do an instanceof comparison against the AESCrypt class
6063 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6064 if (!klass_AESCrypt->is_loaded()) {
6065 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6066 Node* ctrl = control();
6067 set_control(top()); // no regular fast path
6068 return ctrl;
6069 }
6070 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6071
6072 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6073 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6074 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6075
6076 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6077
6078 // for encryption, we are done
6079 if (!decrypting)
6080 return instof_false; // even if it is NULL
6081
6082 // for decryption, we need to add a further check to avoid
6083 // taking the intrinsic path when cipher and plain are the same
6084 // see the original java code for why.
6085 RegionNode* region = new(C) RegionNode(3);
6086 region->init_req(1, instof_false);
6087 Node* src = argument(1);
6088 Node* dest = argument(4);
6089 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6090 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6091 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6092 region->init_req(2, src_dest_conjoint);
6093
6094 record_for_igvn(region);
6095 return _gvn.transform(region);
6096 }