Fri, 11 Jul 2014 19:51:36 -0400
8036588: VerifyFieldClosure fails instanceKlass:3133
Summary: Changed deopt live-pointer test to use returns-object instead of live-and-returns-object
Reviewed-by: iveresov, kvn, jrose
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 */
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/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mathexactnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "prims/nativeLookup.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "trace/traceMacros.hpp"
44 class LibraryIntrinsic : public InlineCallGenerator {
45 // Extend the set of intrinsics known to the runtime:
46 public:
47 private:
48 bool _is_virtual;
49 bool _does_virtual_dispatch;
50 int8_t _predicates_count; // Intrinsic is predicated by several conditions
51 int8_t _last_predicate; // Last generated predicate
52 vmIntrinsics::ID _intrinsic_id;
54 public:
55 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
56 : InlineCallGenerator(m),
57 _is_virtual(is_virtual),
58 _does_virtual_dispatch(does_virtual_dispatch),
59 _predicates_count((int8_t)predicates_count),
60 _last_predicate((int8_t)-1),
61 _intrinsic_id(id)
62 {
63 }
64 virtual bool is_intrinsic() const { return true; }
65 virtual bool is_virtual() const { return _is_virtual; }
66 virtual bool is_predicated() const { return _predicates_count > 0; }
67 virtual int predicates_count() const { return _predicates_count; }
68 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
69 virtual JVMState* generate(JVMState* jvms);
70 virtual Node* generate_predicate(JVMState* jvms, int predicate);
71 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
72 };
75 // Local helper class for LibraryIntrinsic:
76 class LibraryCallKit : public GraphKit {
77 private:
78 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
79 Node* _result; // the result node, if any
80 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
82 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
84 public:
85 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
86 : GraphKit(jvms),
87 _intrinsic(intrinsic),
88 _result(NULL)
89 {
90 // Check if this is a root compile. In that case we don't have a caller.
91 if (!jvms->has_method()) {
92 _reexecute_sp = sp();
93 } else {
94 // Find out how many arguments the interpreter needs when deoptimizing
95 // and save the stack pointer value so it can used by uncommon_trap.
96 // We find the argument count by looking at the declared signature.
97 bool ignored_will_link;
98 ciSignature* declared_signature = NULL;
99 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
100 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
101 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
102 }
103 }
105 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
107 ciMethod* caller() const { return jvms()->method(); }
108 int bci() const { return jvms()->bci(); }
109 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
110 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
111 ciMethod* callee() const { return _intrinsic->method(); }
113 bool try_to_inline(int predicate);
114 Node* try_to_predicate(int predicate);
116 void push_result() {
117 // Push the result onto the stack.
118 if (!stopped() && result() != NULL) {
119 BasicType bt = result()->bottom_type()->basic_type();
120 push_node(bt, result());
121 }
122 }
124 private:
125 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
126 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
127 }
129 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
130 void set_result(RegionNode* region, PhiNode* value);
131 Node* result() { return _result; }
133 virtual int reexecute_sp() { return _reexecute_sp; }
135 // Helper functions to inline natives
136 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
137 Node* generate_slow_guard(Node* test, RegionNode* region);
138 Node* generate_fair_guard(Node* test, RegionNode* region);
139 Node* generate_negative_guard(Node* index, RegionNode* region,
140 // resulting CastII of index:
141 Node* *pos_index = NULL);
142 Node* generate_nonpositive_guard(Node* index, bool never_negative,
143 // resulting CastII of index:
144 Node* *pos_index = NULL);
145 Node* generate_limit_guard(Node* offset, Node* subseq_length,
146 Node* array_length,
147 RegionNode* region);
148 Node* generate_current_thread(Node* &tls_output);
149 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
150 bool disjoint_bases, const char* &name, bool dest_uninitialized);
151 Node* load_mirror_from_klass(Node* klass);
152 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
153 RegionNode* region, int null_path,
154 int offset);
155 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
156 RegionNode* region, int null_path) {
157 int offset = java_lang_Class::klass_offset_in_bytes();
158 return load_klass_from_mirror_common(mirror, never_see_null,
159 region, null_path,
160 offset);
161 }
162 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
163 RegionNode* region, int null_path) {
164 int offset = java_lang_Class::array_klass_offset_in_bytes();
165 return load_klass_from_mirror_common(mirror, never_see_null,
166 region, null_path,
167 offset);
168 }
169 Node* generate_access_flags_guard(Node* kls,
170 int modifier_mask, int modifier_bits,
171 RegionNode* region);
172 Node* generate_interface_guard(Node* kls, RegionNode* region);
173 Node* generate_array_guard(Node* kls, RegionNode* region) {
174 return generate_array_guard_common(kls, region, false, false);
175 }
176 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, false, true);
178 }
179 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, true, false);
181 }
182 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
183 return generate_array_guard_common(kls, region, true, true);
184 }
185 Node* generate_array_guard_common(Node* kls, RegionNode* region,
186 bool obj_array, bool not_array);
187 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
188 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
189 bool is_virtual = false, bool is_static = false);
190 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
191 return generate_method_call(method_id, false, true);
192 }
193 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
194 return generate_method_call(method_id, true, false);
195 }
196 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
198 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
199 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
200 bool inline_string_compareTo();
201 bool inline_string_indexOf();
202 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
203 bool inline_string_equals();
204 Node* round_double_node(Node* n);
205 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
206 bool inline_math_native(vmIntrinsics::ID id);
207 bool inline_trig(vmIntrinsics::ID id);
208 bool inline_math(vmIntrinsics::ID id);
209 template <typename OverflowOp>
210 bool inline_math_overflow(Node* arg1, Node* arg2);
211 void inline_math_mathExact(Node* math, Node* test);
212 bool inline_math_addExactI(bool is_increment);
213 bool inline_math_addExactL(bool is_increment);
214 bool inline_math_multiplyExactI();
215 bool inline_math_multiplyExactL();
216 bool inline_math_negateExactI();
217 bool inline_math_negateExactL();
218 bool inline_math_subtractExactI(bool is_decrement);
219 bool inline_math_subtractExactL(bool is_decrement);
220 bool inline_exp();
221 bool inline_pow();
222 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
223 bool inline_min_max(vmIntrinsics::ID id);
224 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
225 // This returns Type::AnyPtr, RawPtr, or OopPtr.
226 int classify_unsafe_addr(Node* &base, Node* &offset);
227 Node* make_unsafe_address(Node* base, Node* offset);
228 // Helper for inline_unsafe_access.
229 // Generates the guards that check whether the result of
230 // Unsafe.getObject should be recorded in an SATB log buffer.
231 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
232 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
233 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
234 static bool klass_needs_init_guard(Node* kls);
235 bool inline_unsafe_allocate();
236 bool inline_unsafe_copyMemory();
237 bool inline_native_currentThread();
238 #ifdef TRACE_HAVE_INTRINSICS
239 bool inline_native_classID();
240 bool inline_native_threadID();
241 #endif
242 bool inline_native_time_funcs(address method, const char* funcName);
243 bool inline_native_isInterrupted();
244 bool inline_native_Class_query(vmIntrinsics::ID id);
245 bool inline_native_subtype_check();
247 bool inline_native_newArray();
248 bool inline_native_getLength();
249 bool inline_array_copyOf(bool is_copyOfRange);
250 bool inline_array_equals();
251 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
252 bool inline_native_clone(bool is_virtual);
253 bool inline_native_Reflection_getCallerClass();
254 // Helper function for inlining native object hash method
255 bool inline_native_hashcode(bool is_virtual, bool is_static);
256 bool inline_native_getClass();
258 // Helper functions for inlining arraycopy
259 bool inline_arraycopy();
260 void generate_arraycopy(const TypePtr* adr_type,
261 BasicType basic_elem_type,
262 Node* src, Node* src_offset,
263 Node* dest, Node* dest_offset,
264 Node* copy_length,
265 bool disjoint_bases = false,
266 bool length_never_negative = false,
267 RegionNode* slow_region = NULL);
268 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
269 RegionNode* slow_region);
270 void generate_clear_array(const TypePtr* adr_type,
271 Node* dest,
272 BasicType basic_elem_type,
273 Node* slice_off,
274 Node* slice_len,
275 Node* slice_end);
276 bool generate_block_arraycopy(const TypePtr* adr_type,
277 BasicType basic_elem_type,
278 AllocateNode* alloc,
279 Node* src, Node* src_offset,
280 Node* dest, Node* dest_offset,
281 Node* dest_size, bool dest_uninitialized);
282 void generate_slow_arraycopy(const TypePtr* adr_type,
283 Node* src, Node* src_offset,
284 Node* dest, Node* dest_offset,
285 Node* copy_length, bool dest_uninitialized);
286 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
287 Node* dest_elem_klass,
288 Node* src, Node* src_offset,
289 Node* dest, Node* dest_offset,
290 Node* copy_length, bool dest_uninitialized);
291 Node* generate_generic_arraycopy(const TypePtr* adr_type,
292 Node* src, Node* src_offset,
293 Node* dest, Node* dest_offset,
294 Node* copy_length, bool dest_uninitialized);
295 void generate_unchecked_arraycopy(const TypePtr* adr_type,
296 BasicType basic_elem_type,
297 bool disjoint_bases,
298 Node* src, Node* src_offset,
299 Node* dest, Node* dest_offset,
300 Node* copy_length, bool dest_uninitialized);
301 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
302 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
303 bool inline_unsafe_ordered_store(BasicType type);
304 bool inline_unsafe_fence(vmIntrinsics::ID id);
305 bool inline_fp_conversions(vmIntrinsics::ID id);
306 bool inline_number_methods(vmIntrinsics::ID id);
307 bool inline_reference_get();
308 bool inline_aescrypt_Block(vmIntrinsics::ID id);
309 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
310 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
311 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
312 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
313 bool inline_sha_implCompress(vmIntrinsics::ID id);
314 bool inline_digestBase_implCompressMB(int predicate);
315 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
316 bool long_state, address stubAddr, const char *stubName,
317 Node* src_start, Node* ofs, Node* limit);
318 Node* get_state_from_sha_object(Node *sha_object);
319 Node* get_state_from_sha5_object(Node *sha_object);
320 Node* inline_digestBase_implCompressMB_predicate(int predicate);
321 bool inline_encodeISOArray();
322 bool inline_updateCRC32();
323 bool inline_updateBytesCRC32();
324 bool inline_updateByteBufferCRC32();
325 };
328 //---------------------------make_vm_intrinsic----------------------------
329 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
330 vmIntrinsics::ID id = m->intrinsic_id();
331 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
333 if (DisableIntrinsic[0] != '\0'
334 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
335 // disabled by a user request on the command line:
336 // example: -XX:DisableIntrinsic=_hashCode,_getClass
337 return NULL;
338 }
340 if (!m->is_loaded()) {
341 // do not attempt to inline unloaded methods
342 return NULL;
343 }
345 // Only a few intrinsics implement a virtual dispatch.
346 // They are expensive calls which are also frequently overridden.
347 if (is_virtual) {
348 switch (id) {
349 case vmIntrinsics::_hashCode:
350 case vmIntrinsics::_clone:
351 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
352 break;
353 default:
354 return NULL;
355 }
356 }
358 // -XX:-InlineNatives disables nearly all intrinsics:
359 if (!InlineNatives) {
360 switch (id) {
361 case vmIntrinsics::_indexOf:
362 case vmIntrinsics::_compareTo:
363 case vmIntrinsics::_equals:
364 case vmIntrinsics::_equalsC:
365 case vmIntrinsics::_getAndAddInt:
366 case vmIntrinsics::_getAndAddLong:
367 case vmIntrinsics::_getAndSetInt:
368 case vmIntrinsics::_getAndSetLong:
369 case vmIntrinsics::_getAndSetObject:
370 case vmIntrinsics::_loadFence:
371 case vmIntrinsics::_storeFence:
372 case vmIntrinsics::_fullFence:
373 break; // InlineNatives does not control String.compareTo
374 case vmIntrinsics::_Reference_get:
375 break; // InlineNatives does not control Reference.get
376 default:
377 return NULL;
378 }
379 }
381 int predicates = 0;
382 bool does_virtual_dispatch = false;
384 switch (id) {
385 case vmIntrinsics::_compareTo:
386 if (!SpecialStringCompareTo) return NULL;
387 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
388 break;
389 case vmIntrinsics::_indexOf:
390 if (!SpecialStringIndexOf) return NULL;
391 break;
392 case vmIntrinsics::_equals:
393 if (!SpecialStringEquals) return NULL;
394 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
395 break;
396 case vmIntrinsics::_equalsC:
397 if (!SpecialArraysEquals) return NULL;
398 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
399 break;
400 case vmIntrinsics::_arraycopy:
401 if (!InlineArrayCopy) return NULL;
402 break;
403 case vmIntrinsics::_copyMemory:
404 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
405 if (!InlineArrayCopy) return NULL;
406 break;
407 case vmIntrinsics::_hashCode:
408 if (!InlineObjectHash) return NULL;
409 does_virtual_dispatch = true;
410 break;
411 case vmIntrinsics::_clone:
412 does_virtual_dispatch = true;
413 case vmIntrinsics::_copyOf:
414 case vmIntrinsics::_copyOfRange:
415 if (!InlineObjectCopy) return NULL;
416 // These also use the arraycopy intrinsic mechanism:
417 if (!InlineArrayCopy) return NULL;
418 break;
419 case vmIntrinsics::_encodeISOArray:
420 if (!SpecialEncodeISOArray) return NULL;
421 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
422 break;
423 case vmIntrinsics::_checkIndex:
424 // We do not intrinsify this. The optimizer does fine with it.
425 return NULL;
427 case vmIntrinsics::_getCallerClass:
428 if (!UseNewReflection) return NULL;
429 if (!InlineReflectionGetCallerClass) return NULL;
430 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
431 break;
433 case vmIntrinsics::_bitCount_i:
434 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
435 break;
437 case vmIntrinsics::_bitCount_l:
438 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
439 break;
441 case vmIntrinsics::_numberOfLeadingZeros_i:
442 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
443 break;
445 case vmIntrinsics::_numberOfLeadingZeros_l:
446 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
447 break;
449 case vmIntrinsics::_numberOfTrailingZeros_i:
450 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
451 break;
453 case vmIntrinsics::_numberOfTrailingZeros_l:
454 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
455 break;
457 case vmIntrinsics::_reverseBytes_c:
458 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
459 break;
460 case vmIntrinsics::_reverseBytes_s:
461 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
462 break;
463 case vmIntrinsics::_reverseBytes_i:
464 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
465 break;
466 case vmIntrinsics::_reverseBytes_l:
467 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
468 break;
470 case vmIntrinsics::_Reference_get:
471 // Use the intrinsic version of Reference.get() so that the value in
472 // the referent field can be registered by the G1 pre-barrier code.
473 // Also add memory barrier to prevent commoning reads from this field
474 // across safepoint since GC can change it value.
475 break;
477 case vmIntrinsics::_compareAndSwapObject:
478 #ifdef _LP64
479 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
480 #endif
481 break;
483 case vmIntrinsics::_compareAndSwapLong:
484 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
485 break;
487 case vmIntrinsics::_getAndAddInt:
488 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
489 break;
491 case vmIntrinsics::_getAndAddLong:
492 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
493 break;
495 case vmIntrinsics::_getAndSetInt:
496 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
497 break;
499 case vmIntrinsics::_getAndSetLong:
500 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
501 break;
503 case vmIntrinsics::_getAndSetObject:
504 #ifdef _LP64
505 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
506 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
507 break;
508 #else
509 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
510 break;
511 #endif
513 case vmIntrinsics::_aescrypt_encryptBlock:
514 case vmIntrinsics::_aescrypt_decryptBlock:
515 if (!UseAESIntrinsics) return NULL;
516 break;
518 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
519 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
520 if (!UseAESIntrinsics) return NULL;
521 // these two require the predicated logic
522 predicates = 1;
523 break;
525 case vmIntrinsics::_sha_implCompress:
526 if (!UseSHA1Intrinsics) return NULL;
527 break;
529 case vmIntrinsics::_sha2_implCompress:
530 if (!UseSHA256Intrinsics) return NULL;
531 break;
533 case vmIntrinsics::_sha5_implCompress:
534 if (!UseSHA512Intrinsics) return NULL;
535 break;
537 case vmIntrinsics::_digestBase_implCompressMB:
538 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
539 predicates = 3;
540 break;
542 case vmIntrinsics::_updateCRC32:
543 case vmIntrinsics::_updateBytesCRC32:
544 case vmIntrinsics::_updateByteBufferCRC32:
545 if (!UseCRC32Intrinsics) return NULL;
546 break;
548 case vmIntrinsics::_incrementExactI:
549 case vmIntrinsics::_addExactI:
550 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
551 break;
552 case vmIntrinsics::_incrementExactL:
553 case vmIntrinsics::_addExactL:
554 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
555 break;
556 case vmIntrinsics::_decrementExactI:
557 case vmIntrinsics::_subtractExactI:
558 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
559 break;
560 case vmIntrinsics::_decrementExactL:
561 case vmIntrinsics::_subtractExactL:
562 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
563 break;
564 case vmIntrinsics::_negateExactI:
565 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
566 break;
567 case vmIntrinsics::_negateExactL:
568 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
569 break;
570 case vmIntrinsics::_multiplyExactI:
571 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
572 break;
573 case vmIntrinsics::_multiplyExactL:
574 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
575 break;
577 default:
578 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
579 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
580 break;
581 }
583 // -XX:-InlineClassNatives disables natives from the Class class.
584 // The flag applies to all reflective calls, notably Array.newArray
585 // (visible to Java programmers as Array.newInstance).
586 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
587 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
588 if (!InlineClassNatives) return NULL;
589 }
591 // -XX:-InlineThreadNatives disables natives from the Thread class.
592 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
593 if (!InlineThreadNatives) return NULL;
594 }
596 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
597 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
598 m->holder()->name() == ciSymbol::java_lang_Float() ||
599 m->holder()->name() == ciSymbol::java_lang_Double()) {
600 if (!InlineMathNatives) return NULL;
601 }
603 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
604 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
605 if (!InlineUnsafeOps) return NULL;
606 }
608 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
609 }
611 //----------------------register_library_intrinsics-----------------------
612 // Initialize this file's data structures, for each Compile instance.
613 void Compile::register_library_intrinsics() {
614 // Nothing to do here.
615 }
617 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
618 LibraryCallKit kit(jvms, this);
619 Compile* C = kit.C;
620 int nodes = C->unique();
621 #ifndef PRODUCT
622 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
623 char buf[1000];
624 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
625 tty->print_cr("Intrinsic %s", str);
626 }
627 #endif
628 ciMethod* callee = kit.callee();
629 const int bci = kit.bci();
631 // Try to inline the intrinsic.
632 if (kit.try_to_inline(_last_predicate)) {
633 if (C->print_intrinsics() || C->print_inlining()) {
634 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
635 }
636 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
637 if (C->log()) {
638 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
639 vmIntrinsics::name_at(intrinsic_id()),
640 (is_virtual() ? " virtual='1'" : ""),
641 C->unique() - nodes);
642 }
643 // Push the result from the inlined method onto the stack.
644 kit.push_result();
645 return kit.transfer_exceptions_into_jvms();
646 }
648 // The intrinsic bailed out
649 if (C->print_intrinsics() || C->print_inlining()) {
650 if (jvms->has_method()) {
651 // Not a root compile.
652 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
653 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
654 } else {
655 // Root compile
656 tty->print("Did not generate intrinsic %s%s at bci:%d in",
657 vmIntrinsics::name_at(intrinsic_id()),
658 (is_virtual() ? " (virtual)" : ""), bci);
659 }
660 }
661 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
662 return NULL;
663 }
665 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
666 LibraryCallKit kit(jvms, this);
667 Compile* C = kit.C;
668 int nodes = C->unique();
669 _last_predicate = predicate;
670 #ifndef PRODUCT
671 assert(is_predicated() && predicate < predicates_count(), "sanity");
672 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
673 char buf[1000];
674 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
675 tty->print_cr("Predicate for intrinsic %s", str);
676 }
677 #endif
678 ciMethod* callee = kit.callee();
679 const int bci = kit.bci();
681 Node* slow_ctl = kit.try_to_predicate(predicate);
682 if (!kit.failing()) {
683 if (C->print_intrinsics() || C->print_inlining()) {
684 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
685 }
686 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
687 if (C->log()) {
688 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
689 vmIntrinsics::name_at(intrinsic_id()),
690 (is_virtual() ? " virtual='1'" : ""),
691 C->unique() - nodes);
692 }
693 return slow_ctl; // Could be NULL if the check folds.
694 }
696 // The intrinsic bailed out
697 if (C->print_intrinsics() || C->print_inlining()) {
698 if (jvms->has_method()) {
699 // Not a root compile.
700 const char* msg = "failed to generate predicate for intrinsic";
701 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
702 } else {
703 // Root compile
704 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
705 vmIntrinsics::name_at(intrinsic_id()),
706 (is_virtual() ? " (virtual)" : ""), bci);
707 }
708 }
709 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
710 return NULL;
711 }
713 bool LibraryCallKit::try_to_inline(int predicate) {
714 // Handle symbolic names for otherwise undistinguished boolean switches:
715 const bool is_store = true;
716 const bool is_native_ptr = true;
717 const bool is_static = true;
718 const bool is_volatile = true;
720 if (!jvms()->has_method()) {
721 // Root JVMState has a null method.
722 assert(map()->memory()->Opcode() == Op_Parm, "");
723 // Insert the memory aliasing node
724 set_all_memory(reset_memory());
725 }
726 assert(merged_memory(), "");
729 switch (intrinsic_id()) {
730 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
731 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
732 case vmIntrinsics::_getClass: return inline_native_getClass();
734 case vmIntrinsics::_dsin:
735 case vmIntrinsics::_dcos:
736 case vmIntrinsics::_dtan:
737 case vmIntrinsics::_dabs:
738 case vmIntrinsics::_datan2:
739 case vmIntrinsics::_dsqrt:
740 case vmIntrinsics::_dexp:
741 case vmIntrinsics::_dlog:
742 case vmIntrinsics::_dlog10:
743 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
745 case vmIntrinsics::_min:
746 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
748 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
749 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
750 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
751 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
752 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
753 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
754 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
755 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
756 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
757 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
758 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
759 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
761 case vmIntrinsics::_arraycopy: return inline_arraycopy();
763 case vmIntrinsics::_compareTo: return inline_string_compareTo();
764 case vmIntrinsics::_indexOf: return inline_string_indexOf();
765 case vmIntrinsics::_equals: return inline_string_equals();
767 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
768 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
769 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
770 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
771 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
772 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
773 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
774 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
775 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
777 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
778 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
779 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
780 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
781 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
782 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
783 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
784 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
785 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
787 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
788 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
789 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
790 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
791 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
792 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
793 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
794 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
796 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
797 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
798 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
799 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
800 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
801 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
802 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
803 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
805 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
806 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
807 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
808 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
809 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
810 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
811 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
812 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
813 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
815 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
816 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
817 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
818 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
819 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
820 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
821 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
822 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
823 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
825 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
826 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
827 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
828 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
830 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
831 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
832 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
834 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
835 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
836 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
838 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
839 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
840 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
841 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
842 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
844 case vmIntrinsics::_loadFence:
845 case vmIntrinsics::_storeFence:
846 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
848 case vmIntrinsics::_currentThread: return inline_native_currentThread();
849 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
851 #ifdef TRACE_HAVE_INTRINSICS
852 case vmIntrinsics::_classID: return inline_native_classID();
853 case vmIntrinsics::_threadID: return inline_native_threadID();
854 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
855 #endif
856 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
857 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
858 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
859 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
860 case vmIntrinsics::_newArray: return inline_native_newArray();
861 case vmIntrinsics::_getLength: return inline_native_getLength();
862 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
863 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
864 case vmIntrinsics::_equalsC: return inline_array_equals();
865 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
867 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
869 case vmIntrinsics::_isInstance:
870 case vmIntrinsics::_getModifiers:
871 case vmIntrinsics::_isInterface:
872 case vmIntrinsics::_isArray:
873 case vmIntrinsics::_isPrimitive:
874 case vmIntrinsics::_getSuperclass:
875 case vmIntrinsics::_getComponentType:
876 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
878 case vmIntrinsics::_floatToRawIntBits:
879 case vmIntrinsics::_floatToIntBits:
880 case vmIntrinsics::_intBitsToFloat:
881 case vmIntrinsics::_doubleToRawLongBits:
882 case vmIntrinsics::_doubleToLongBits:
883 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
885 case vmIntrinsics::_numberOfLeadingZeros_i:
886 case vmIntrinsics::_numberOfLeadingZeros_l:
887 case vmIntrinsics::_numberOfTrailingZeros_i:
888 case vmIntrinsics::_numberOfTrailingZeros_l:
889 case vmIntrinsics::_bitCount_i:
890 case vmIntrinsics::_bitCount_l:
891 case vmIntrinsics::_reverseBytes_i:
892 case vmIntrinsics::_reverseBytes_l:
893 case vmIntrinsics::_reverseBytes_s:
894 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
896 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
898 case vmIntrinsics::_Reference_get: return inline_reference_get();
900 case vmIntrinsics::_aescrypt_encryptBlock:
901 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
903 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
904 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
905 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
907 case vmIntrinsics::_sha_implCompress:
908 case vmIntrinsics::_sha2_implCompress:
909 case vmIntrinsics::_sha5_implCompress:
910 return inline_sha_implCompress(intrinsic_id());
912 case vmIntrinsics::_digestBase_implCompressMB:
913 return inline_digestBase_implCompressMB(predicate);
915 case vmIntrinsics::_encodeISOArray:
916 return inline_encodeISOArray();
918 case vmIntrinsics::_updateCRC32:
919 return inline_updateCRC32();
920 case vmIntrinsics::_updateBytesCRC32:
921 return inline_updateBytesCRC32();
922 case vmIntrinsics::_updateByteBufferCRC32:
923 return inline_updateByteBufferCRC32();
925 default:
926 // If you get here, it may be that someone has added a new intrinsic
927 // to the list in vmSymbols.hpp without implementing it here.
928 #ifndef PRODUCT
929 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
930 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
931 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
932 }
933 #endif
934 return false;
935 }
936 }
938 Node* LibraryCallKit::try_to_predicate(int predicate) {
939 if (!jvms()->has_method()) {
940 // Root JVMState has a null method.
941 assert(map()->memory()->Opcode() == Op_Parm, "");
942 // Insert the memory aliasing node
943 set_all_memory(reset_memory());
944 }
945 assert(merged_memory(), "");
947 switch (intrinsic_id()) {
948 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
949 return inline_cipherBlockChaining_AESCrypt_predicate(false);
950 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
951 return inline_cipherBlockChaining_AESCrypt_predicate(true);
952 case vmIntrinsics::_digestBase_implCompressMB:
953 return inline_digestBase_implCompressMB_predicate(predicate);
955 default:
956 // If you get here, it may be that someone has added a new intrinsic
957 // to the list in vmSymbols.hpp without implementing it here.
958 #ifndef PRODUCT
959 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
960 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
961 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
962 }
963 #endif
964 Node* slow_ctl = control();
965 set_control(top()); // No fast path instrinsic
966 return slow_ctl;
967 }
968 }
970 //------------------------------set_result-------------------------------
971 // Helper function for finishing intrinsics.
972 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
973 record_for_igvn(region);
974 set_control(_gvn.transform(region));
975 set_result( _gvn.transform(value));
976 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
977 }
979 //------------------------------generate_guard---------------------------
980 // Helper function for generating guarded fast-slow graph structures.
981 // The given 'test', if true, guards a slow path. If the test fails
982 // then a fast path can be taken. (We generally hope it fails.)
983 // In all cases, GraphKit::control() is updated to the fast path.
984 // The returned value represents the control for the slow path.
985 // The return value is never 'top'; it is either a valid control
986 // or NULL if it is obvious that the slow path can never be taken.
987 // Also, if region and the slow control are not NULL, the slow edge
988 // is appended to the region.
989 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
990 if (stopped()) {
991 // Already short circuited.
992 return NULL;
993 }
995 // Build an if node and its projections.
996 // If test is true we take the slow path, which we assume is uncommon.
997 if (_gvn.type(test) == TypeInt::ZERO) {
998 // The slow branch is never taken. No need to build this guard.
999 return NULL;
1000 }
1002 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1004 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1005 if (if_slow == top()) {
1006 // The slow branch is never taken. No need to build this guard.
1007 return NULL;
1008 }
1010 if (region != NULL)
1011 region->add_req(if_slow);
1013 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1014 set_control(if_fast);
1016 return if_slow;
1017 }
1019 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1020 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1021 }
1022 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1023 return generate_guard(test, region, PROB_FAIR);
1024 }
1026 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1027 Node* *pos_index) {
1028 if (stopped())
1029 return NULL; // already stopped
1030 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1031 return NULL; // index is already adequately typed
1032 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1033 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1034 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1035 if (is_neg != NULL && pos_index != NULL) {
1036 // Emulate effect of Parse::adjust_map_after_if.
1037 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1038 ccast->set_req(0, control());
1039 (*pos_index) = _gvn.transform(ccast);
1040 }
1041 return is_neg;
1042 }
1044 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1045 Node* *pos_index) {
1046 if (stopped())
1047 return NULL; // already stopped
1048 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1049 return NULL; // index is already adequately typed
1050 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1051 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1052 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1053 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1054 if (is_notp != NULL && pos_index != NULL) {
1055 // Emulate effect of Parse::adjust_map_after_if.
1056 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1057 ccast->set_req(0, control());
1058 (*pos_index) = _gvn.transform(ccast);
1059 }
1060 return is_notp;
1061 }
1063 // Make sure that 'position' is a valid limit index, in [0..length].
1064 // There are two equivalent plans for checking this:
1065 // A. (offset + copyLength) unsigned<= arrayLength
1066 // B. offset <= (arrayLength - copyLength)
1067 // We require that all of the values above, except for the sum and
1068 // difference, are already known to be non-negative.
1069 // Plan A is robust in the face of overflow, if offset and copyLength
1070 // are both hugely positive.
1071 //
1072 // Plan B is less direct and intuitive, but it does not overflow at
1073 // all, since the difference of two non-negatives is always
1074 // representable. Whenever Java methods must perform the equivalent
1075 // check they generally use Plan B instead of Plan A.
1076 // For the moment we use Plan A.
1077 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1078 Node* subseq_length,
1079 Node* array_length,
1080 RegionNode* region) {
1081 if (stopped())
1082 return NULL; // already stopped
1083 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1084 if (zero_offset && subseq_length->eqv_uncast(array_length))
1085 return NULL; // common case of whole-array copy
1086 Node* last = subseq_length;
1087 if (!zero_offset) // last += offset
1088 last = _gvn.transform(new (C) AddINode(last, offset));
1089 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1090 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1091 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1092 return is_over;
1093 }
1096 //--------------------------generate_current_thread--------------------
1097 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1098 ciKlass* thread_klass = env()->Thread_klass();
1099 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1100 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1101 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1102 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1103 tls_output = thread;
1104 return threadObj;
1105 }
1108 //------------------------------make_string_method_node------------------------
1109 // Helper method for String intrinsic functions. This version is called
1110 // with str1 and str2 pointing to String object nodes.
1111 //
1112 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1113 Node* no_ctrl = NULL;
1115 // Get start addr of string
1116 Node* str1_value = load_String_value(no_ctrl, str1);
1117 Node* str1_offset = load_String_offset(no_ctrl, str1);
1118 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1120 // Get length of string 1
1121 Node* str1_len = load_String_length(no_ctrl, str1);
1123 Node* str2_value = load_String_value(no_ctrl, str2);
1124 Node* str2_offset = load_String_offset(no_ctrl, str2);
1125 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1127 Node* str2_len = NULL;
1128 Node* result = NULL;
1130 switch (opcode) {
1131 case Op_StrIndexOf:
1132 // Get length of string 2
1133 str2_len = load_String_length(no_ctrl, str2);
1135 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1136 str1_start, str1_len, str2_start, str2_len);
1137 break;
1138 case Op_StrComp:
1139 // Get length of string 2
1140 str2_len = load_String_length(no_ctrl, str2);
1142 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1143 str1_start, str1_len, str2_start, str2_len);
1144 break;
1145 case Op_StrEquals:
1146 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1147 str1_start, str2_start, str1_len);
1148 break;
1149 default:
1150 ShouldNotReachHere();
1151 return NULL;
1152 }
1154 // All these intrinsics have checks.
1155 C->set_has_split_ifs(true); // Has chance for split-if optimization
1157 return _gvn.transform(result);
1158 }
1160 // Helper method for String intrinsic functions. This version is called
1161 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1162 // to Int nodes containing the lenghts of str1 and str2.
1163 //
1164 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1165 Node* result = NULL;
1166 switch (opcode) {
1167 case Op_StrIndexOf:
1168 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1169 str1_start, cnt1, str2_start, cnt2);
1170 break;
1171 case Op_StrComp:
1172 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1173 str1_start, cnt1, str2_start, cnt2);
1174 break;
1175 case Op_StrEquals:
1176 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1177 str1_start, str2_start, cnt1);
1178 break;
1179 default:
1180 ShouldNotReachHere();
1181 return NULL;
1182 }
1184 // All these intrinsics have checks.
1185 C->set_has_split_ifs(true); // Has chance for split-if optimization
1187 return _gvn.transform(result);
1188 }
1190 //------------------------------inline_string_compareTo------------------------
1191 // public int java.lang.String.compareTo(String anotherString);
1192 bool LibraryCallKit::inline_string_compareTo() {
1193 Node* receiver = null_check(argument(0));
1194 Node* arg = null_check(argument(1));
1195 if (stopped()) {
1196 return true;
1197 }
1198 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1199 return true;
1200 }
1202 //------------------------------inline_string_equals------------------------
1203 bool LibraryCallKit::inline_string_equals() {
1204 Node* receiver = null_check_receiver();
1205 // NOTE: Do not null check argument for String.equals() because spec
1206 // allows to specify NULL as argument.
1207 Node* argument = this->argument(1);
1208 if (stopped()) {
1209 return true;
1210 }
1212 // paths (plus control) merge
1213 RegionNode* region = new (C) RegionNode(5);
1214 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1216 // does source == target string?
1217 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1218 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1220 Node* if_eq = generate_slow_guard(bol, NULL);
1221 if (if_eq != NULL) {
1222 // receiver == argument
1223 phi->init_req(2, intcon(1));
1224 region->init_req(2, if_eq);
1225 }
1227 // get String klass for instanceOf
1228 ciInstanceKlass* klass = env()->String_klass();
1230 if (!stopped()) {
1231 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1232 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1233 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1235 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1236 //instanceOf == true, fallthrough
1238 if (inst_false != NULL) {
1239 phi->init_req(3, intcon(0));
1240 region->init_req(3, inst_false);
1241 }
1242 }
1244 if (!stopped()) {
1245 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1247 // Properly cast the argument to String
1248 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1249 // This path is taken only when argument's type is String:NotNull.
1250 argument = cast_not_null(argument, false);
1252 Node* no_ctrl = NULL;
1254 // Get start addr of receiver
1255 Node* receiver_val = load_String_value(no_ctrl, receiver);
1256 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1257 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1259 // Get length of receiver
1260 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1262 // Get start addr of argument
1263 Node* argument_val = load_String_value(no_ctrl, argument);
1264 Node* argument_offset = load_String_offset(no_ctrl, argument);
1265 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1267 // Get length of argument
1268 Node* argument_cnt = load_String_length(no_ctrl, argument);
1270 // Check for receiver count != argument count
1271 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1272 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1273 Node* if_ne = generate_slow_guard(bol, NULL);
1274 if (if_ne != NULL) {
1275 phi->init_req(4, intcon(0));
1276 region->init_req(4, if_ne);
1277 }
1279 // Check for count == 0 is done by assembler code for StrEquals.
1281 if (!stopped()) {
1282 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1283 phi->init_req(1, equals);
1284 region->init_req(1, control());
1285 }
1286 }
1288 // post merge
1289 set_control(_gvn.transform(region));
1290 record_for_igvn(region);
1292 set_result(_gvn.transform(phi));
1293 return true;
1294 }
1296 //------------------------------inline_array_equals----------------------------
1297 bool LibraryCallKit::inline_array_equals() {
1298 Node* arg1 = argument(0);
1299 Node* arg2 = argument(1);
1300 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1301 return true;
1302 }
1304 // Java version of String.indexOf(constant string)
1305 // class StringDecl {
1306 // StringDecl(char[] ca) {
1307 // offset = 0;
1308 // count = ca.length;
1309 // value = ca;
1310 // }
1311 // int offset;
1312 // int count;
1313 // char[] value;
1314 // }
1315 //
1316 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1317 // int targetOffset, int cache_i, int md2) {
1318 // int cache = cache_i;
1319 // int sourceOffset = string_object.offset;
1320 // int sourceCount = string_object.count;
1321 // int targetCount = target_object.length;
1322 //
1323 // int targetCountLess1 = targetCount - 1;
1324 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1325 //
1326 // char[] source = string_object.value;
1327 // char[] target = target_object;
1328 // int lastChar = target[targetCountLess1];
1329 //
1330 // outer_loop:
1331 // for (int i = sourceOffset; i < sourceEnd; ) {
1332 // int src = source[i + targetCountLess1];
1333 // if (src == lastChar) {
1334 // // With random strings and a 4-character alphabet,
1335 // // reverse matching at this point sets up 0.8% fewer
1336 // // frames, but (paradoxically) makes 0.3% more probes.
1337 // // Since those probes are nearer the lastChar probe,
1338 // // there is may be a net D$ win with reverse matching.
1339 // // But, reversing loop inhibits unroll of inner loop
1340 // // for unknown reason. So, does running outer loop from
1341 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1342 // for (int j = 0; j < targetCountLess1; j++) {
1343 // if (target[targetOffset + j] != source[i+j]) {
1344 // if ((cache & (1 << source[i+j])) == 0) {
1345 // if (md2 < j+1) {
1346 // i += j+1;
1347 // continue outer_loop;
1348 // }
1349 // }
1350 // i += md2;
1351 // continue outer_loop;
1352 // }
1353 // }
1354 // return i - sourceOffset;
1355 // }
1356 // if ((cache & (1 << src)) == 0) {
1357 // i += targetCountLess1;
1358 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1359 // i++;
1360 // }
1361 // return -1;
1362 // }
1364 //------------------------------string_indexOf------------------------
1365 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1366 jint cache_i, jint md2_i) {
1368 Node* no_ctrl = NULL;
1369 float likely = PROB_LIKELY(0.9);
1370 float unlikely = PROB_UNLIKELY(0.9);
1372 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1374 Node* source = load_String_value(no_ctrl, string_object);
1375 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1376 Node* sourceCount = load_String_length(no_ctrl, string_object);
1378 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1379 jint target_length = target_array->length();
1380 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1381 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1383 // String.value field is known to be @Stable.
1384 if (UseImplicitStableValues) {
1385 target = cast_array_to_stable(target, target_type);
1386 }
1388 IdealKit kit(this, false, true);
1389 #define __ kit.
1390 Node* zero = __ ConI(0);
1391 Node* one = __ ConI(1);
1392 Node* cache = __ ConI(cache_i);
1393 Node* md2 = __ ConI(md2_i);
1394 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1395 Node* targetCount = __ ConI(target_length);
1396 Node* targetCountLess1 = __ ConI(target_length - 1);
1397 Node* targetOffset = __ ConI(targetOffset_i);
1398 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1400 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1401 Node* outer_loop = __ make_label(2 /* goto */);
1402 Node* return_ = __ make_label(1);
1404 __ set(rtn,__ ConI(-1));
1405 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1406 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1407 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1408 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1409 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1410 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1411 Node* tpj = __ AddI(targetOffset, __ value(j));
1412 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1413 Node* ipj = __ AddI(__ value(i), __ value(j));
1414 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1415 __ if_then(targ, BoolTest::ne, src2); {
1416 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1417 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1418 __ increment(i, __ AddI(__ value(j), one));
1419 __ goto_(outer_loop);
1420 } __ end_if(); __ dead(j);
1421 }__ end_if(); __ dead(j);
1422 __ increment(i, md2);
1423 __ goto_(outer_loop);
1424 }__ end_if();
1425 __ increment(j, one);
1426 }__ end_loop(); __ dead(j);
1427 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1428 __ goto_(return_);
1429 }__ end_if();
1430 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1431 __ increment(i, targetCountLess1);
1432 }__ end_if();
1433 __ increment(i, one);
1434 __ bind(outer_loop);
1435 }__ end_loop(); __ dead(i);
1436 __ bind(return_);
1438 // Final sync IdealKit and GraphKit.
1439 final_sync(kit);
1440 Node* result = __ value(rtn);
1441 #undef __
1442 C->set_has_loops(true);
1443 return result;
1444 }
1446 //------------------------------inline_string_indexOf------------------------
1447 bool LibraryCallKit::inline_string_indexOf() {
1448 Node* receiver = argument(0);
1449 Node* arg = argument(1);
1451 Node* result;
1452 // Disable the use of pcmpestri until it can be guaranteed that
1453 // the load doesn't cross into the uncommited space.
1454 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1455 UseSSE42Intrinsics) {
1456 // Generate SSE4.2 version of indexOf
1457 // We currently only have match rules that use SSE4.2
1459 receiver = null_check(receiver);
1460 arg = null_check(arg);
1461 if (stopped()) {
1462 return true;
1463 }
1465 ciInstanceKlass* str_klass = env()->String_klass();
1466 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1468 // Make the merge point
1469 RegionNode* result_rgn = new (C) RegionNode(4);
1470 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1471 Node* no_ctrl = NULL;
1473 // Get start addr of source string
1474 Node* source = load_String_value(no_ctrl, receiver);
1475 Node* source_offset = load_String_offset(no_ctrl, receiver);
1476 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1478 // Get length of source string
1479 Node* source_cnt = load_String_length(no_ctrl, receiver);
1481 // Get start addr of substring
1482 Node* substr = load_String_value(no_ctrl, arg);
1483 Node* substr_offset = load_String_offset(no_ctrl, arg);
1484 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1486 // Get length of source string
1487 Node* substr_cnt = load_String_length(no_ctrl, arg);
1489 // Check for substr count > string count
1490 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1491 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1492 Node* if_gt = generate_slow_guard(bol, NULL);
1493 if (if_gt != NULL) {
1494 result_phi->init_req(2, intcon(-1));
1495 result_rgn->init_req(2, if_gt);
1496 }
1498 if (!stopped()) {
1499 // Check for substr count == 0
1500 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1501 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1502 Node* if_zero = generate_slow_guard(bol, NULL);
1503 if (if_zero != NULL) {
1504 result_phi->init_req(3, intcon(0));
1505 result_rgn->init_req(3, if_zero);
1506 }
1507 }
1509 if (!stopped()) {
1510 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1511 result_phi->init_req(1, result);
1512 result_rgn->init_req(1, control());
1513 }
1514 set_control(_gvn.transform(result_rgn));
1515 record_for_igvn(result_rgn);
1516 result = _gvn.transform(result_phi);
1518 } else { // Use LibraryCallKit::string_indexOf
1519 // don't intrinsify if argument isn't a constant string.
1520 if (!arg->is_Con()) {
1521 return false;
1522 }
1523 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1524 if (str_type == NULL) {
1525 return false;
1526 }
1527 ciInstanceKlass* klass = env()->String_klass();
1528 ciObject* str_const = str_type->const_oop();
1529 if (str_const == NULL || str_const->klass() != klass) {
1530 return false;
1531 }
1532 ciInstance* str = str_const->as_instance();
1533 assert(str != NULL, "must be instance");
1535 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1536 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1538 int o;
1539 int c;
1540 if (java_lang_String::has_offset_field()) {
1541 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1542 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1543 } else {
1544 o = 0;
1545 c = pat->length();
1546 }
1548 // constant strings have no offset and count == length which
1549 // simplifies the resulting code somewhat so lets optimize for that.
1550 if (o != 0 || c != pat->length()) {
1551 return false;
1552 }
1554 receiver = null_check(receiver, T_OBJECT);
1555 // NOTE: No null check on the argument is needed since it's a constant String oop.
1556 if (stopped()) {
1557 return true;
1558 }
1560 // The null string as a pattern always returns 0 (match at beginning of string)
1561 if (c == 0) {
1562 set_result(intcon(0));
1563 return true;
1564 }
1566 // Generate default indexOf
1567 jchar lastChar = pat->char_at(o + (c - 1));
1568 int cache = 0;
1569 int i;
1570 for (i = 0; i < c - 1; i++) {
1571 assert(i < pat->length(), "out of range");
1572 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1573 }
1575 int md2 = c;
1576 for (i = 0; i < c - 1; i++) {
1577 assert(i < pat->length(), "out of range");
1578 if (pat->char_at(o + i) == lastChar) {
1579 md2 = (c - 1) - i;
1580 }
1581 }
1583 result = string_indexOf(receiver, pat, o, cache, md2);
1584 }
1585 set_result(result);
1586 return true;
1587 }
1589 //--------------------------round_double_node--------------------------------
1590 // Round a double node if necessary.
1591 Node* LibraryCallKit::round_double_node(Node* n) {
1592 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1593 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1594 return n;
1595 }
1597 //------------------------------inline_math-----------------------------------
1598 // public static double Math.abs(double)
1599 // public static double Math.sqrt(double)
1600 // public static double Math.log(double)
1601 // public static double Math.log10(double)
1602 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1603 Node* arg = round_double_node(argument(0));
1604 Node* n;
1605 switch (id) {
1606 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1607 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1608 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1609 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1610 default: fatal_unexpected_iid(id); break;
1611 }
1612 set_result(_gvn.transform(n));
1613 return true;
1614 }
1616 //------------------------------inline_trig----------------------------------
1617 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1618 // argument reduction which will turn into a fast/slow diamond.
1619 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1620 Node* arg = round_double_node(argument(0));
1621 Node* n = NULL;
1623 switch (id) {
1624 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1625 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1626 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1627 default: fatal_unexpected_iid(id); break;
1628 }
1629 n = _gvn.transform(n);
1631 // Rounding required? Check for argument reduction!
1632 if (Matcher::strict_fp_requires_explicit_rounding) {
1633 static const double pi_4 = 0.7853981633974483;
1634 static const double neg_pi_4 = -0.7853981633974483;
1635 // pi/2 in 80-bit extended precision
1636 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1637 // -pi/2 in 80-bit extended precision
1638 // 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};
1639 // Cutoff value for using this argument reduction technique
1640 //static const double pi_2_minus_epsilon = 1.564660403643354;
1641 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1643 // Pseudocode for sin:
1644 // if (x <= Math.PI / 4.0) {
1645 // if (x >= -Math.PI / 4.0) return fsin(x);
1646 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1647 // } else {
1648 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1649 // }
1650 // return StrictMath.sin(x);
1652 // Pseudocode for cos:
1653 // if (x <= Math.PI / 4.0) {
1654 // if (x >= -Math.PI / 4.0) return fcos(x);
1655 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1656 // } else {
1657 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1658 // }
1659 // return StrictMath.cos(x);
1661 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1662 // requires a special machine instruction to load it. Instead we'll try
1663 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1664 // probably do the math inside the SIN encoding.
1666 // Make the merge point
1667 RegionNode* r = new (C) RegionNode(3);
1668 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1670 // Flatten arg so we need only 1 test
1671 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1672 // Node for PI/4 constant
1673 Node *pi4 = makecon(TypeD::make(pi_4));
1674 // Check PI/4 : abs(arg)
1675 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1676 // Check: If PI/4 < abs(arg) then go slow
1677 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1678 // Branch either way
1679 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1680 set_control(opt_iff(r,iff));
1682 // Set fast path result
1683 phi->init_req(2, n);
1685 // Slow path - non-blocking leaf call
1686 Node* call = NULL;
1687 switch (id) {
1688 case vmIntrinsics::_dsin:
1689 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1690 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1691 "Sin", NULL, arg, top());
1692 break;
1693 case vmIntrinsics::_dcos:
1694 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1695 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1696 "Cos", NULL, arg, top());
1697 break;
1698 case vmIntrinsics::_dtan:
1699 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1700 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1701 "Tan", NULL, arg, top());
1702 break;
1703 }
1704 assert(control()->in(0) == call, "");
1705 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1706 r->init_req(1, control());
1707 phi->init_req(1, slow_result);
1709 // Post-merge
1710 set_control(_gvn.transform(r));
1711 record_for_igvn(r);
1712 n = _gvn.transform(phi);
1714 C->set_has_split_ifs(true); // Has chance for split-if optimization
1715 }
1716 set_result(n);
1717 return true;
1718 }
1720 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1721 //-------------------
1722 //result=(result.isNaN())? funcAddr():result;
1723 // Check: If isNaN() by checking result!=result? then either trap
1724 // or go to runtime
1725 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1726 // Build the boolean node
1727 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1729 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1730 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1731 // The pow or exp intrinsic returned a NaN, which requires a call
1732 // to the runtime. Recompile with the runtime call.
1733 uncommon_trap(Deoptimization::Reason_intrinsic,
1734 Deoptimization::Action_make_not_entrant);
1735 }
1736 return result;
1737 } else {
1738 // If this inlining ever returned NaN in the past, we compile a call
1739 // to the runtime to properly handle corner cases
1741 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1742 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1743 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1745 if (!if_slow->is_top()) {
1746 RegionNode* result_region = new (C) RegionNode(3);
1747 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1749 result_region->init_req(1, if_fast);
1750 result_val->init_req(1, result);
1752 set_control(if_slow);
1754 const TypePtr* no_memory_effects = NULL;
1755 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1756 no_memory_effects,
1757 x, top(), y, y ? top() : NULL);
1758 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1759 #ifdef ASSERT
1760 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1761 assert(value_top == top(), "second value must be top");
1762 #endif
1764 result_region->init_req(2, control());
1765 result_val->init_req(2, value);
1766 set_control(_gvn.transform(result_region));
1767 return _gvn.transform(result_val);
1768 } else {
1769 return result;
1770 }
1771 }
1772 }
1774 //------------------------------inline_exp-------------------------------------
1775 // Inline exp instructions, if possible. The Intel hardware only misses
1776 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1777 bool LibraryCallKit::inline_exp() {
1778 Node* arg = round_double_node(argument(0));
1779 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1781 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1782 set_result(n);
1784 C->set_has_split_ifs(true); // Has chance for split-if optimization
1785 return true;
1786 }
1788 //------------------------------inline_pow-------------------------------------
1789 // Inline power instructions, if possible.
1790 bool LibraryCallKit::inline_pow() {
1791 // Pseudocode for pow
1792 // if (y == 2) {
1793 // return x * x;
1794 // } else {
1795 // if (x <= 0.0) {
1796 // long longy = (long)y;
1797 // if ((double)longy == y) { // if y is long
1798 // if (y + 1 == y) longy = 0; // huge number: even
1799 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1800 // } else {
1801 // result = NaN;
1802 // }
1803 // } else {
1804 // result = DPow(x,y);
1805 // }
1806 // if (result != result)? {
1807 // result = uncommon_trap() or runtime_call();
1808 // }
1809 // return result;
1810 // }
1812 Node* x = round_double_node(argument(0));
1813 Node* y = round_double_node(argument(2));
1815 Node* result = NULL;
1817 Node* const_two_node = makecon(TypeD::make(2.0));
1818 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1819 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1820 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1821 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1822 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1824 RegionNode* region_node = new (C) RegionNode(3);
1825 region_node->init_req(1, if_true);
1827 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1828 // special case for x^y where y == 2, we can convert it to x * x
1829 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1831 // set control to if_false since we will now process the false branch
1832 set_control(if_false);
1834 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1835 // Short form: skip the fancy tests and just check for NaN result.
1836 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1837 } else {
1838 // If this inlining ever returned NaN in the past, include all
1839 // checks + call to the runtime.
1841 // Set the merge point for If node with condition of (x <= 0.0)
1842 // There are four possible paths to region node and phi node
1843 RegionNode *r = new (C) RegionNode(4);
1844 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1846 // Build the first if node: if (x <= 0.0)
1847 // Node for 0 constant
1848 Node *zeronode = makecon(TypeD::ZERO);
1849 // Check x:0
1850 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1851 // Check: If (x<=0) then go complex path
1852 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1853 // Branch either way
1854 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1855 // Fast path taken; set region slot 3
1856 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1857 r->init_req(3,fast_taken); // Capture fast-control
1859 // Fast path not-taken, i.e. slow path
1860 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1862 // Set fast path result
1863 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1864 phi->init_req(3, fast_result);
1866 // Complex path
1867 // Build the second if node (if y is long)
1868 // Node for (long)y
1869 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1870 // Node for (double)((long) y)
1871 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1872 // Check (double)((long) y) : y
1873 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1874 // Check if (y isn't long) then go to slow path
1876 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1877 // Branch either way
1878 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1879 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1881 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1883 // Calculate DPow(abs(x), y)*(1 & (long)y)
1884 // Node for constant 1
1885 Node *conone = longcon(1);
1886 // 1& (long)y
1887 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1889 // A huge number is always even. Detect a huge number by checking
1890 // if y + 1 == y and set integer to be tested for parity to 0.
1891 // Required for corner case:
1892 // (long)9.223372036854776E18 = max_jlong
1893 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1894 // max_jlong is odd but 9.223372036854776E18 is even
1895 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1896 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1897 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1898 Node* correctedsign = NULL;
1899 if (ConditionalMoveLimit != 0) {
1900 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1901 } else {
1902 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1903 RegionNode *r = new (C) RegionNode(3);
1904 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1905 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1906 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1907 phi->init_req(1, signnode);
1908 phi->init_req(2, longcon(0));
1909 correctedsign = _gvn.transform(phi);
1910 ylong_path = _gvn.transform(r);
1911 record_for_igvn(r);
1912 }
1914 // zero node
1915 Node *conzero = longcon(0);
1916 // Check (1&(long)y)==0?
1917 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1918 // Check if (1&(long)y)!=0?, if so the result is negative
1919 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1920 // abs(x)
1921 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1922 // abs(x)^y
1923 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1924 // -abs(x)^y
1925 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1926 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1927 Node *signresult = NULL;
1928 if (ConditionalMoveLimit != 0) {
1929 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1930 } else {
1931 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1932 RegionNode *r = new (C) RegionNode(3);
1933 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1934 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1935 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1936 phi->init_req(1, absxpowy);
1937 phi->init_req(2, negabsxpowy);
1938 signresult = _gvn.transform(phi);
1939 ylong_path = _gvn.transform(r);
1940 record_for_igvn(r);
1941 }
1942 // Set complex path fast result
1943 r->init_req(2, ylong_path);
1944 phi->init_req(2, signresult);
1946 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1947 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1948 r->init_req(1,slow_path);
1949 phi->init_req(1,slow_result);
1951 // Post merge
1952 set_control(_gvn.transform(r));
1953 record_for_igvn(r);
1954 result = _gvn.transform(phi);
1955 }
1957 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1959 // control from finish_pow_exp is now input to the region node
1960 region_node->set_req(2, control());
1961 // the result from finish_pow_exp is now input to the phi node
1962 phi_node->init_req(2, result);
1963 set_control(_gvn.transform(region_node));
1964 record_for_igvn(region_node);
1965 set_result(_gvn.transform(phi_node));
1967 C->set_has_split_ifs(true); // Has chance for split-if optimization
1968 return true;
1969 }
1971 //------------------------------runtime_math-----------------------------
1972 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1973 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1974 "must be (DD)D or (D)D type");
1976 // Inputs
1977 Node* a = round_double_node(argument(0));
1978 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1980 const TypePtr* no_memory_effects = NULL;
1981 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1982 no_memory_effects,
1983 a, top(), b, b ? top() : NULL);
1984 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1985 #ifdef ASSERT
1986 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1987 assert(value_top == top(), "second value must be top");
1988 #endif
1990 set_result(value);
1991 return true;
1992 }
1994 //------------------------------inline_math_native-----------------------------
1995 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1996 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1997 switch (id) {
1998 // These intrinsics are not properly supported on all hardware
1999 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2000 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2001 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2002 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2003 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2004 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2006 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2007 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2008 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2009 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2011 // These intrinsics are supported on all hardware
2012 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2013 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2015 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2016 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2017 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2018 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2019 #undef FN_PTR
2021 // These intrinsics are not yet correctly implemented
2022 case vmIntrinsics::_datan2:
2023 return false;
2025 default:
2026 fatal_unexpected_iid(id);
2027 return false;
2028 }
2029 }
2031 static bool is_simple_name(Node* n) {
2032 return (n->req() == 1 // constant
2033 || (n->is_Type() && n->as_Type()->type()->singleton())
2034 || n->is_Proj() // parameter or return value
2035 || n->is_Phi() // local of some sort
2036 );
2037 }
2039 //----------------------------inline_min_max-----------------------------------
2040 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2041 set_result(generate_min_max(id, argument(0), argument(1)));
2042 return true;
2043 }
2045 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2046 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2047 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2048 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2049 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2051 {
2052 PreserveJVMState pjvms(this);
2053 PreserveReexecuteState preexecs(this);
2054 jvms()->set_should_reexecute(true);
2056 set_control(slow_path);
2057 set_i_o(i_o());
2059 uncommon_trap(Deoptimization::Reason_intrinsic,
2060 Deoptimization::Action_none);
2061 }
2063 set_control(fast_path);
2064 set_result(math);
2065 }
2067 template <typename OverflowOp>
2068 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2069 typedef typename OverflowOp::MathOp MathOp;
2071 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2072 Node* operation = _gvn.transform( mathOp );
2073 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2074 inline_math_mathExact(operation, ofcheck);
2075 return true;
2076 }
2078 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2079 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2080 }
2082 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2083 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2084 }
2086 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2087 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2088 }
2090 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2091 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2092 }
2094 bool LibraryCallKit::inline_math_negateExactI() {
2095 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2096 }
2098 bool LibraryCallKit::inline_math_negateExactL() {
2099 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2100 }
2102 bool LibraryCallKit::inline_math_multiplyExactI() {
2103 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2104 }
2106 bool LibraryCallKit::inline_math_multiplyExactL() {
2107 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2108 }
2110 Node*
2111 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2112 // These are the candidate return value:
2113 Node* xvalue = x0;
2114 Node* yvalue = y0;
2116 if (xvalue == yvalue) {
2117 return xvalue;
2118 }
2120 bool want_max = (id == vmIntrinsics::_max);
2122 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2123 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2124 if (txvalue == NULL || tyvalue == NULL) return top();
2125 // This is not really necessary, but it is consistent with a
2126 // hypothetical MaxINode::Value method:
2127 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2129 // %%% This folding logic should (ideally) be in a different place.
2130 // Some should be inside IfNode, and there to be a more reliable
2131 // transformation of ?: style patterns into cmoves. We also want
2132 // more powerful optimizations around cmove and min/max.
2134 // Try to find a dominating comparison of these guys.
2135 // It can simplify the index computation for Arrays.copyOf
2136 // and similar uses of System.arraycopy.
2137 // First, compute the normalized version of CmpI(x, y).
2138 int cmp_op = Op_CmpI;
2139 Node* xkey = xvalue;
2140 Node* ykey = yvalue;
2141 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2142 if (ideal_cmpxy->is_Cmp()) {
2143 // E.g., if we have CmpI(length - offset, count),
2144 // it might idealize to CmpI(length, count + offset)
2145 cmp_op = ideal_cmpxy->Opcode();
2146 xkey = ideal_cmpxy->in(1);
2147 ykey = ideal_cmpxy->in(2);
2148 }
2150 // Start by locating any relevant comparisons.
2151 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2152 Node* cmpxy = NULL;
2153 Node* cmpyx = NULL;
2154 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2155 Node* cmp = start_from->fast_out(k);
2156 if (cmp->outcnt() > 0 && // must have prior uses
2157 cmp->in(0) == NULL && // must be context-independent
2158 cmp->Opcode() == cmp_op) { // right kind of compare
2159 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2160 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2161 }
2162 }
2164 const int NCMPS = 2;
2165 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2166 int cmpn;
2167 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2168 if (cmps[cmpn] != NULL) break; // find a result
2169 }
2170 if (cmpn < NCMPS) {
2171 // Look for a dominating test that tells us the min and max.
2172 int depth = 0; // Limit search depth for speed
2173 Node* dom = control();
2174 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2175 if (++depth >= 100) break;
2176 Node* ifproj = dom;
2177 if (!ifproj->is_Proj()) continue;
2178 Node* iff = ifproj->in(0);
2179 if (!iff->is_If()) continue;
2180 Node* bol = iff->in(1);
2181 if (!bol->is_Bool()) continue;
2182 Node* cmp = bol->in(1);
2183 if (cmp == NULL) continue;
2184 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2185 if (cmps[cmpn] == cmp) break;
2186 if (cmpn == NCMPS) continue;
2187 BoolTest::mask btest = bol->as_Bool()->_test._test;
2188 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2189 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2190 // At this point, we know that 'x btest y' is true.
2191 switch (btest) {
2192 case BoolTest::eq:
2193 // They are proven equal, so we can collapse the min/max.
2194 // Either value is the answer. Choose the simpler.
2195 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2196 return yvalue;
2197 return xvalue;
2198 case BoolTest::lt: // x < y
2199 case BoolTest::le: // x <= y
2200 return (want_max ? yvalue : xvalue);
2201 case BoolTest::gt: // x > y
2202 case BoolTest::ge: // x >= y
2203 return (want_max ? xvalue : yvalue);
2204 }
2205 }
2206 }
2208 // We failed to find a dominating test.
2209 // Let's pick a test that might GVN with prior tests.
2210 Node* best_bol = NULL;
2211 BoolTest::mask best_btest = BoolTest::illegal;
2212 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2213 Node* cmp = cmps[cmpn];
2214 if (cmp == NULL) continue;
2215 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2216 Node* bol = cmp->fast_out(j);
2217 if (!bol->is_Bool()) continue;
2218 BoolTest::mask btest = bol->as_Bool()->_test._test;
2219 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2220 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2221 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2222 best_bol = bol->as_Bool();
2223 best_btest = btest;
2224 }
2225 }
2226 }
2228 Node* answer_if_true = NULL;
2229 Node* answer_if_false = NULL;
2230 switch (best_btest) {
2231 default:
2232 if (cmpxy == NULL)
2233 cmpxy = ideal_cmpxy;
2234 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2235 // and fall through:
2236 case BoolTest::lt: // x < y
2237 case BoolTest::le: // x <= y
2238 answer_if_true = (want_max ? yvalue : xvalue);
2239 answer_if_false = (want_max ? xvalue : yvalue);
2240 break;
2241 case BoolTest::gt: // x > y
2242 case BoolTest::ge: // x >= y
2243 answer_if_true = (want_max ? xvalue : yvalue);
2244 answer_if_false = (want_max ? yvalue : xvalue);
2245 break;
2246 }
2248 jint hi, lo;
2249 if (want_max) {
2250 // We can sharpen the minimum.
2251 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2252 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2253 } else {
2254 // We can sharpen the maximum.
2255 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2256 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2257 }
2259 // Use a flow-free graph structure, to avoid creating excess control edges
2260 // which could hinder other optimizations.
2261 // Since Math.min/max is often used with arraycopy, we want
2262 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2263 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2264 answer_if_false, answer_if_true,
2265 TypeInt::make(lo, hi, widen));
2267 return _gvn.transform(cmov);
2269 /*
2270 // This is not as desirable as it may seem, since Min and Max
2271 // nodes do not have a full set of optimizations.
2272 // And they would interfere, anyway, with 'if' optimizations
2273 // and with CMoveI canonical forms.
2274 switch (id) {
2275 case vmIntrinsics::_min:
2276 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2277 case vmIntrinsics::_max:
2278 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2279 default:
2280 ShouldNotReachHere();
2281 }
2282 */
2283 }
2285 inline int
2286 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2287 const TypePtr* base_type = TypePtr::NULL_PTR;
2288 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2289 if (base_type == NULL) {
2290 // Unknown type.
2291 return Type::AnyPtr;
2292 } else if (base_type == TypePtr::NULL_PTR) {
2293 // Since this is a NULL+long form, we have to switch to a rawptr.
2294 base = _gvn.transform(new (C) CastX2PNode(offset));
2295 offset = MakeConX(0);
2296 return Type::RawPtr;
2297 } else if (base_type->base() == Type::RawPtr) {
2298 return Type::RawPtr;
2299 } else if (base_type->isa_oopptr()) {
2300 // Base is never null => always a heap address.
2301 if (base_type->ptr() == TypePtr::NotNull) {
2302 return Type::OopPtr;
2303 }
2304 // Offset is small => always a heap address.
2305 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2306 if (offset_type != NULL &&
2307 base_type->offset() == 0 && // (should always be?)
2308 offset_type->_lo >= 0 &&
2309 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2310 return Type::OopPtr;
2311 }
2312 // Otherwise, it might either be oop+off or NULL+addr.
2313 return Type::AnyPtr;
2314 } else {
2315 // No information:
2316 return Type::AnyPtr;
2317 }
2318 }
2320 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2321 int kind = classify_unsafe_addr(base, offset);
2322 if (kind == Type::RawPtr) {
2323 return basic_plus_adr(top(), base, offset);
2324 } else {
2325 return basic_plus_adr(base, offset);
2326 }
2327 }
2329 //--------------------------inline_number_methods-----------------------------
2330 // inline int Integer.numberOfLeadingZeros(int)
2331 // inline int Long.numberOfLeadingZeros(long)
2332 //
2333 // inline int Integer.numberOfTrailingZeros(int)
2334 // inline int Long.numberOfTrailingZeros(long)
2335 //
2336 // inline int Integer.bitCount(int)
2337 // inline int Long.bitCount(long)
2338 //
2339 // inline char Character.reverseBytes(char)
2340 // inline short Short.reverseBytes(short)
2341 // inline int Integer.reverseBytes(int)
2342 // inline long Long.reverseBytes(long)
2343 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2344 Node* arg = argument(0);
2345 Node* n;
2346 switch (id) {
2347 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2348 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2349 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2350 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2351 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2352 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2353 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2354 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2355 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2356 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2357 default: fatal_unexpected_iid(id); break;
2358 }
2359 set_result(_gvn.transform(n));
2360 return true;
2361 }
2363 //----------------------------inline_unsafe_access----------------------------
2365 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2367 // Helper that guards and inserts a pre-barrier.
2368 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2369 Node* pre_val, bool need_mem_bar) {
2370 // We could be accessing the referent field of a reference object. If so, when G1
2371 // is enabled, we need to log the value in the referent field in an SATB buffer.
2372 // This routine performs some compile time filters and generates suitable
2373 // runtime filters that guard the pre-barrier code.
2374 // Also add memory barrier for non volatile load from the referent field
2375 // to prevent commoning of loads across safepoint.
2376 if (!UseG1GC && !need_mem_bar)
2377 return;
2379 // Some compile time checks.
2381 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2382 const TypeX* otype = offset->find_intptr_t_type();
2383 if (otype != NULL && otype->is_con() &&
2384 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2385 // Constant offset but not the reference_offset so just return
2386 return;
2387 }
2389 // We only need to generate the runtime guards for instances.
2390 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2391 if (btype != NULL) {
2392 if (btype->isa_aryptr()) {
2393 // Array type so nothing to do
2394 return;
2395 }
2397 const TypeInstPtr* itype = btype->isa_instptr();
2398 if (itype != NULL) {
2399 // Can the klass of base_oop be statically determined to be
2400 // _not_ a sub-class of Reference and _not_ Object?
2401 ciKlass* klass = itype->klass();
2402 if ( klass->is_loaded() &&
2403 !klass->is_subtype_of(env()->Reference_klass()) &&
2404 !env()->Object_klass()->is_subtype_of(klass)) {
2405 return;
2406 }
2407 }
2408 }
2410 // The compile time filters did not reject base_oop/offset so
2411 // we need to generate the following runtime filters
2412 //
2413 // if (offset == java_lang_ref_Reference::_reference_offset) {
2414 // if (instance_of(base, java.lang.ref.Reference)) {
2415 // pre_barrier(_, pre_val, ...);
2416 // }
2417 // }
2419 float likely = PROB_LIKELY( 0.999);
2420 float unlikely = PROB_UNLIKELY(0.999);
2422 IdealKit ideal(this);
2423 #define __ ideal.
2425 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2427 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2428 // Update graphKit memory and control from IdealKit.
2429 sync_kit(ideal);
2431 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2432 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2434 // Update IdealKit memory and control from graphKit.
2435 __ sync_kit(this);
2437 Node* one = __ ConI(1);
2438 // is_instof == 0 if base_oop == NULL
2439 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2441 // Update graphKit from IdeakKit.
2442 sync_kit(ideal);
2444 // Use the pre-barrier to record the value in the referent field
2445 pre_barrier(false /* do_load */,
2446 __ ctrl(),
2447 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2448 pre_val /* pre_val */,
2449 T_OBJECT);
2450 if (need_mem_bar) {
2451 // Add memory barrier to prevent commoning reads from this field
2452 // across safepoint since GC can change its value.
2453 insert_mem_bar(Op_MemBarCPUOrder);
2454 }
2455 // Update IdealKit from graphKit.
2456 __ sync_kit(this);
2458 } __ end_if(); // _ref_type != ref_none
2459 } __ end_if(); // offset == referent_offset
2461 // Final sync IdealKit and GraphKit.
2462 final_sync(ideal);
2463 #undef __
2464 }
2467 // Interpret Unsafe.fieldOffset cookies correctly:
2468 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2470 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2471 // Attempt to infer a sharper value type from the offset and base type.
2472 ciKlass* sharpened_klass = NULL;
2474 // See if it is an instance field, with an object type.
2475 if (alias_type->field() != NULL) {
2476 assert(!is_native_ptr, "native pointer op cannot use a java address");
2477 if (alias_type->field()->type()->is_klass()) {
2478 sharpened_klass = alias_type->field()->type()->as_klass();
2479 }
2480 }
2482 // See if it is a narrow oop array.
2483 if (adr_type->isa_aryptr()) {
2484 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2485 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2486 if (elem_type != NULL) {
2487 sharpened_klass = elem_type->klass();
2488 }
2489 }
2490 }
2492 // The sharpened class might be unloaded if there is no class loader
2493 // contraint in place.
2494 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2495 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2497 #ifndef PRODUCT
2498 if (C->print_intrinsics() || C->print_inlining()) {
2499 tty->print(" from base type: "); adr_type->dump();
2500 tty->print(" sharpened value: "); tjp->dump();
2501 }
2502 #endif
2503 // Sharpen the value type.
2504 return tjp;
2505 }
2506 return NULL;
2507 }
2509 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2510 if (callee()->is_static()) return false; // caller must have the capability!
2512 #ifndef PRODUCT
2513 {
2514 ResourceMark rm;
2515 // Check the signatures.
2516 ciSignature* sig = callee()->signature();
2517 #ifdef ASSERT
2518 if (!is_store) {
2519 // Object getObject(Object base, int/long offset), etc.
2520 BasicType rtype = sig->return_type()->basic_type();
2521 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2522 rtype = T_ADDRESS; // it is really a C void*
2523 assert(rtype == type, "getter must return the expected value");
2524 if (!is_native_ptr) {
2525 assert(sig->count() == 2, "oop getter has 2 arguments");
2526 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2527 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2528 } else {
2529 assert(sig->count() == 1, "native getter has 1 argument");
2530 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2531 }
2532 } else {
2533 // void putObject(Object base, int/long offset, Object x), etc.
2534 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2535 if (!is_native_ptr) {
2536 assert(sig->count() == 3, "oop putter has 3 arguments");
2537 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2538 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2539 } else {
2540 assert(sig->count() == 2, "native putter has 2 arguments");
2541 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2542 }
2543 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2544 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2545 vtype = T_ADDRESS; // it is really a C void*
2546 assert(vtype == type, "putter must accept the expected value");
2547 }
2548 #endif // ASSERT
2549 }
2550 #endif //PRODUCT
2552 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2554 Node* receiver = argument(0); // type: oop
2556 // Build address expression. See the code in inline_unsafe_prefetch.
2557 Node* adr;
2558 Node* heap_base_oop = top();
2559 Node* offset = top();
2560 Node* val;
2562 if (!is_native_ptr) {
2563 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2564 Node* base = argument(1); // type: oop
2565 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2566 offset = argument(2); // type: long
2567 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2568 // to be plain byte offsets, which are also the same as those accepted
2569 // by oopDesc::field_base.
2570 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2571 "fieldOffset must be byte-scaled");
2572 // 32-bit machines ignore the high half!
2573 offset = ConvL2X(offset);
2574 adr = make_unsafe_address(base, offset);
2575 heap_base_oop = base;
2576 val = is_store ? argument(4) : NULL;
2577 } else {
2578 Node* ptr = argument(1); // type: long
2579 ptr = ConvL2X(ptr); // adjust Java long to machine word
2580 adr = make_unsafe_address(NULL, ptr);
2581 val = is_store ? argument(3) : NULL;
2582 }
2584 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2586 // First guess at the value type.
2587 const Type *value_type = Type::get_const_basic_type(type);
2589 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2590 // there was not enough information to nail it down.
2591 Compile::AliasType* alias_type = C->alias_type(adr_type);
2592 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2594 // We will need memory barriers unless we can determine a unique
2595 // alias category for this reference. (Note: If for some reason
2596 // the barriers get omitted and the unsafe reference begins to "pollute"
2597 // the alias analysis of the rest of the graph, either Compile::can_alias
2598 // or Compile::must_alias will throw a diagnostic assert.)
2599 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2601 // If we are reading the value of the referent field of a Reference
2602 // object (either by using Unsafe directly or through reflection)
2603 // then, if G1 is enabled, we need to record the referent in an
2604 // SATB log buffer using the pre-barrier mechanism.
2605 // Also we need to add memory barrier to prevent commoning reads
2606 // from this field across safepoint since GC can change its value.
2607 bool need_read_barrier = !is_native_ptr && !is_store &&
2608 offset != top() && heap_base_oop != top();
2610 if (!is_store && type == T_OBJECT) {
2611 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2612 if (tjp != NULL) {
2613 value_type = tjp;
2614 }
2615 }
2617 receiver = null_check(receiver);
2618 if (stopped()) {
2619 return true;
2620 }
2621 // Heap pointers get a null-check from the interpreter,
2622 // as a courtesy. However, this is not guaranteed by Unsafe,
2623 // and it is not possible to fully distinguish unintended nulls
2624 // from intended ones in this API.
2626 if (is_volatile) {
2627 // We need to emit leading and trailing CPU membars (see below) in
2628 // addition to memory membars when is_volatile. This is a little
2629 // too strong, but avoids the need to insert per-alias-type
2630 // volatile membars (for stores; compare Parse::do_put_xxx), which
2631 // we cannot do effectively here because we probably only have a
2632 // rough approximation of type.
2633 need_mem_bar = true;
2634 // For Stores, place a memory ordering barrier now.
2635 if (is_store) {
2636 insert_mem_bar(Op_MemBarRelease);
2637 } else {
2638 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2639 insert_mem_bar(Op_MemBarVolatile);
2640 }
2641 }
2642 }
2644 // Memory barrier to prevent normal and 'unsafe' accesses from
2645 // bypassing each other. Happens after null checks, so the
2646 // exception paths do not take memory state from the memory barrier,
2647 // so there's no problems making a strong assert about mixing users
2648 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2649 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2650 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2652 if (!is_store) {
2653 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2654 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2655 // load value
2656 switch (type) {
2657 case T_BOOLEAN:
2658 case T_CHAR:
2659 case T_BYTE:
2660 case T_SHORT:
2661 case T_INT:
2662 case T_LONG:
2663 case T_FLOAT:
2664 case T_DOUBLE:
2665 break;
2666 case T_OBJECT:
2667 if (need_read_barrier) {
2668 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2669 }
2670 break;
2671 case T_ADDRESS:
2672 // Cast to an int type.
2673 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2674 p = ConvX2UL(p);
2675 break;
2676 default:
2677 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2678 break;
2679 }
2680 // The load node has the control of the preceding MemBarCPUOrder. All
2681 // following nodes will have the control of the MemBarCPUOrder inserted at
2682 // the end of this method. So, pushing the load onto the stack at a later
2683 // point is fine.
2684 set_result(p);
2685 } else {
2686 // place effect of store into memory
2687 switch (type) {
2688 case T_DOUBLE:
2689 val = dstore_rounding(val);
2690 break;
2691 case T_ADDRESS:
2692 // Repackage the long as a pointer.
2693 val = ConvL2X(val);
2694 val = _gvn.transform(new (C) CastX2PNode(val));
2695 break;
2696 }
2698 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2699 if (type != T_OBJECT ) {
2700 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2701 } else {
2702 // Possibly an oop being stored to Java heap or native memory
2703 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2704 // oop to Java heap.
2705 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2706 } else {
2707 // We can't tell at compile time if we are storing in the Java heap or outside
2708 // of it. So we need to emit code to conditionally do the proper type of
2709 // store.
2711 IdealKit ideal(this);
2712 #define __ ideal.
2713 // QQQ who knows what probability is here??
2714 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2715 // Sync IdealKit and graphKit.
2716 sync_kit(ideal);
2717 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2718 // Update IdealKit memory.
2719 __ sync_kit(this);
2720 } __ else_(); {
2721 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2722 } __ end_if();
2723 // Final sync IdealKit and GraphKit.
2724 final_sync(ideal);
2725 #undef __
2726 }
2727 }
2728 }
2730 if (is_volatile) {
2731 if (!is_store) {
2732 insert_mem_bar(Op_MemBarAcquire);
2733 } else {
2734 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2735 insert_mem_bar(Op_MemBarVolatile);
2736 }
2737 }
2738 }
2740 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2742 return true;
2743 }
2745 //----------------------------inline_unsafe_prefetch----------------------------
2747 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2748 #ifndef PRODUCT
2749 {
2750 ResourceMark rm;
2751 // Check the signatures.
2752 ciSignature* sig = callee()->signature();
2753 #ifdef ASSERT
2754 // Object getObject(Object base, int/long offset), etc.
2755 BasicType rtype = sig->return_type()->basic_type();
2756 if (!is_native_ptr) {
2757 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2758 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2759 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2760 } else {
2761 assert(sig->count() == 1, "native prefetch has 1 argument");
2762 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2763 }
2764 #endif // ASSERT
2765 }
2766 #endif // !PRODUCT
2768 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2770 const int idx = is_static ? 0 : 1;
2771 if (!is_static) {
2772 null_check_receiver();
2773 if (stopped()) {
2774 return true;
2775 }
2776 }
2778 // Build address expression. See the code in inline_unsafe_access.
2779 Node *adr;
2780 if (!is_native_ptr) {
2781 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2782 Node* base = argument(idx + 0); // type: oop
2783 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2784 Node* offset = argument(idx + 1); // type: long
2785 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2786 // to be plain byte offsets, which are also the same as those accepted
2787 // by oopDesc::field_base.
2788 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2789 "fieldOffset must be byte-scaled");
2790 // 32-bit machines ignore the high half!
2791 offset = ConvL2X(offset);
2792 adr = make_unsafe_address(base, offset);
2793 } else {
2794 Node* ptr = argument(idx + 0); // type: long
2795 ptr = ConvL2X(ptr); // adjust Java long to machine word
2796 adr = make_unsafe_address(NULL, ptr);
2797 }
2799 // Generate the read or write prefetch
2800 Node *prefetch;
2801 if (is_store) {
2802 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2803 } else {
2804 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2805 }
2806 prefetch->init_req(0, control());
2807 set_i_o(_gvn.transform(prefetch));
2809 return true;
2810 }
2812 //----------------------------inline_unsafe_load_store----------------------------
2813 // This method serves a couple of different customers (depending on LoadStoreKind):
2814 //
2815 // LS_cmpxchg:
2816 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2817 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2818 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2819 //
2820 // LS_xadd:
2821 // public int getAndAddInt( Object o, long offset, int delta)
2822 // public long getAndAddLong(Object o, long offset, long delta)
2823 //
2824 // LS_xchg:
2825 // int getAndSet(Object o, long offset, int newValue)
2826 // long getAndSet(Object o, long offset, long newValue)
2827 // Object getAndSet(Object o, long offset, Object newValue)
2828 //
2829 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2830 // This basic scheme here is the same as inline_unsafe_access, but
2831 // differs in enough details that combining them would make the code
2832 // overly confusing. (This is a true fact! I originally combined
2833 // them, but even I was confused by it!) As much code/comments as
2834 // possible are retained from inline_unsafe_access though to make
2835 // the correspondences clearer. - dl
2837 if (callee()->is_static()) return false; // caller must have the capability!
2839 #ifndef PRODUCT
2840 BasicType rtype;
2841 {
2842 ResourceMark rm;
2843 // Check the signatures.
2844 ciSignature* sig = callee()->signature();
2845 rtype = sig->return_type()->basic_type();
2846 if (kind == LS_xadd || kind == LS_xchg) {
2847 // Check the signatures.
2848 #ifdef ASSERT
2849 assert(rtype == type, "get and set must return the expected type");
2850 assert(sig->count() == 3, "get and set has 3 arguments");
2851 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2852 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2853 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2854 #endif // ASSERT
2855 } else if (kind == LS_cmpxchg) {
2856 // Check the signatures.
2857 #ifdef ASSERT
2858 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2859 assert(sig->count() == 4, "CAS has 4 arguments");
2860 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2861 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2862 #endif // ASSERT
2863 } else {
2864 ShouldNotReachHere();
2865 }
2866 }
2867 #endif //PRODUCT
2869 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2871 // Get arguments:
2872 Node* receiver = NULL;
2873 Node* base = NULL;
2874 Node* offset = NULL;
2875 Node* oldval = NULL;
2876 Node* newval = NULL;
2877 if (kind == LS_cmpxchg) {
2878 const bool two_slot_type = type2size[type] == 2;
2879 receiver = argument(0); // type: oop
2880 base = argument(1); // type: oop
2881 offset = argument(2); // type: long
2882 oldval = argument(4); // type: oop, int, or long
2883 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2884 } else if (kind == LS_xadd || kind == LS_xchg){
2885 receiver = argument(0); // type: oop
2886 base = argument(1); // type: oop
2887 offset = argument(2); // type: long
2888 oldval = NULL;
2889 newval = argument(4); // type: oop, int, or long
2890 }
2892 // Null check receiver.
2893 receiver = null_check(receiver);
2894 if (stopped()) {
2895 return true;
2896 }
2898 // Build field offset expression.
2899 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2900 // to be plain byte offsets, which are also the same as those accepted
2901 // by oopDesc::field_base.
2902 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2903 // 32-bit machines ignore the high half of long offsets
2904 offset = ConvL2X(offset);
2905 Node* adr = make_unsafe_address(base, offset);
2906 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2908 // For CAS, unlike inline_unsafe_access, there seems no point in
2909 // trying to refine types. Just use the coarse types here.
2910 const Type *value_type = Type::get_const_basic_type(type);
2911 Compile::AliasType* alias_type = C->alias_type(adr_type);
2912 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2914 if (kind == LS_xchg && type == T_OBJECT) {
2915 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2916 if (tjp != NULL) {
2917 value_type = tjp;
2918 }
2919 }
2921 int alias_idx = C->get_alias_index(adr_type);
2923 // Memory-model-wise, a LoadStore acts like a little synchronized
2924 // block, so needs barriers on each side. These don't translate
2925 // into actual barriers on most machines, but we still need rest of
2926 // compiler to respect ordering.
2928 insert_mem_bar(Op_MemBarRelease);
2929 insert_mem_bar(Op_MemBarCPUOrder);
2931 // 4984716: MemBars must be inserted before this
2932 // memory node in order to avoid a false
2933 // dependency which will confuse the scheduler.
2934 Node *mem = memory(alias_idx);
2936 // For now, we handle only those cases that actually exist: ints,
2937 // longs, and Object. Adding others should be straightforward.
2938 Node* load_store;
2939 switch(type) {
2940 case T_INT:
2941 if (kind == LS_xadd) {
2942 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2943 } else if (kind == LS_xchg) {
2944 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2945 } else if (kind == LS_cmpxchg) {
2946 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2947 } else {
2948 ShouldNotReachHere();
2949 }
2950 break;
2951 case T_LONG:
2952 if (kind == LS_xadd) {
2953 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2954 } else if (kind == LS_xchg) {
2955 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2956 } else if (kind == LS_cmpxchg) {
2957 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2958 } else {
2959 ShouldNotReachHere();
2960 }
2961 break;
2962 case T_OBJECT:
2963 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2964 // could be delayed during Parse (for example, in adjust_map_after_if()).
2965 // Execute transformation here to avoid barrier generation in such case.
2966 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2967 newval = _gvn.makecon(TypePtr::NULL_PTR);
2969 // Reference stores need a store barrier.
2970 if (kind == LS_xchg) {
2971 // If pre-barrier must execute before the oop store, old value will require do_load here.
2972 if (!can_move_pre_barrier()) {
2973 pre_barrier(true /* do_load*/,
2974 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2975 NULL /* pre_val*/,
2976 T_OBJECT);
2977 } // Else move pre_barrier to use load_store value, see below.
2978 } else if (kind == LS_cmpxchg) {
2979 // Same as for newval above:
2980 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2981 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2982 }
2983 // The only known value which might get overwritten is oldval.
2984 pre_barrier(false /* do_load */,
2985 control(), NULL, NULL, max_juint, NULL, NULL,
2986 oldval /* pre_val */,
2987 T_OBJECT);
2988 } else {
2989 ShouldNotReachHere();
2990 }
2992 #ifdef _LP64
2993 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2994 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2995 if (kind == LS_xchg) {
2996 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2997 newval_enc, adr_type, value_type->make_narrowoop()));
2998 } else {
2999 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3000 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3001 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3002 newval_enc, oldval_enc));
3003 }
3004 } else
3005 #endif
3006 {
3007 if (kind == LS_xchg) {
3008 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3009 } else {
3010 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3011 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3012 }
3013 }
3014 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3015 break;
3016 default:
3017 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3018 break;
3019 }
3021 // SCMemProjNodes represent the memory state of a LoadStore. Their
3022 // main role is to prevent LoadStore nodes from being optimized away
3023 // when their results aren't used.
3024 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3025 set_memory(proj, alias_idx);
3027 if (type == T_OBJECT && kind == LS_xchg) {
3028 #ifdef _LP64
3029 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3030 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3031 }
3032 #endif
3033 if (can_move_pre_barrier()) {
3034 // Don't need to load pre_val. The old value is returned by load_store.
3035 // The pre_barrier can execute after the xchg as long as no safepoint
3036 // gets inserted between them.
3037 pre_barrier(false /* do_load */,
3038 control(), NULL, NULL, max_juint, NULL, NULL,
3039 load_store /* pre_val */,
3040 T_OBJECT);
3041 }
3042 }
3044 // Add the trailing membar surrounding the access
3045 insert_mem_bar(Op_MemBarCPUOrder);
3046 insert_mem_bar(Op_MemBarAcquire);
3048 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3049 set_result(load_store);
3050 return true;
3051 }
3053 //----------------------------inline_unsafe_ordered_store----------------------
3054 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3055 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3056 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3057 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3058 // This is another variant of inline_unsafe_access, differing in
3059 // that it always issues store-store ("release") barrier and ensures
3060 // store-atomicity (which only matters for "long").
3062 if (callee()->is_static()) return false; // caller must have the capability!
3064 #ifndef PRODUCT
3065 {
3066 ResourceMark rm;
3067 // Check the signatures.
3068 ciSignature* sig = callee()->signature();
3069 #ifdef ASSERT
3070 BasicType rtype = sig->return_type()->basic_type();
3071 assert(rtype == T_VOID, "must return void");
3072 assert(sig->count() == 3, "has 3 arguments");
3073 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3074 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3075 #endif // ASSERT
3076 }
3077 #endif //PRODUCT
3079 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3081 // Get arguments:
3082 Node* receiver = argument(0); // type: oop
3083 Node* base = argument(1); // type: oop
3084 Node* offset = argument(2); // type: long
3085 Node* val = argument(4); // type: oop, int, or long
3087 // Null check receiver.
3088 receiver = null_check(receiver);
3089 if (stopped()) {
3090 return true;
3091 }
3093 // Build field offset expression.
3094 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3095 // 32-bit machines ignore the high half of long offsets
3096 offset = ConvL2X(offset);
3097 Node* adr = make_unsafe_address(base, offset);
3098 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3099 const Type *value_type = Type::get_const_basic_type(type);
3100 Compile::AliasType* alias_type = C->alias_type(adr_type);
3102 insert_mem_bar(Op_MemBarRelease);
3103 insert_mem_bar(Op_MemBarCPUOrder);
3104 // Ensure that the store is atomic for longs:
3105 const bool require_atomic_access = true;
3106 Node* store;
3107 if (type == T_OBJECT) // reference stores need a store barrier.
3108 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3109 else {
3110 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3111 }
3112 insert_mem_bar(Op_MemBarCPUOrder);
3113 return true;
3114 }
3116 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3117 // Regardless of form, don't allow previous ld/st to move down,
3118 // then issue acquire, release, or volatile mem_bar.
3119 insert_mem_bar(Op_MemBarCPUOrder);
3120 switch(id) {
3121 case vmIntrinsics::_loadFence:
3122 insert_mem_bar(Op_LoadFence);
3123 return true;
3124 case vmIntrinsics::_storeFence:
3125 insert_mem_bar(Op_StoreFence);
3126 return true;
3127 case vmIntrinsics::_fullFence:
3128 insert_mem_bar(Op_MemBarVolatile);
3129 return true;
3130 default:
3131 fatal_unexpected_iid(id);
3132 return false;
3133 }
3134 }
3136 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3137 if (!kls->is_Con()) {
3138 return true;
3139 }
3140 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3141 if (klsptr == NULL) {
3142 return true;
3143 }
3144 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3145 // don't need a guard for a klass that is already initialized
3146 return !ik->is_initialized();
3147 }
3149 //----------------------------inline_unsafe_allocate---------------------------
3150 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3151 bool LibraryCallKit::inline_unsafe_allocate() {
3152 if (callee()->is_static()) return false; // caller must have the capability!
3154 null_check_receiver(); // null-check, then ignore
3155 Node* cls = null_check(argument(1));
3156 if (stopped()) return true;
3158 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3159 kls = null_check(kls);
3160 if (stopped()) return true; // argument was like int.class
3162 Node* test = NULL;
3163 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3164 // Note: The argument might still be an illegal value like
3165 // Serializable.class or Object[].class. The runtime will handle it.
3166 // But we must make an explicit check for initialization.
3167 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3168 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3169 // can generate code to load it as unsigned byte.
3170 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3171 Node* bits = intcon(InstanceKlass::fully_initialized);
3172 test = _gvn.transform(new (C) SubINode(inst, bits));
3173 // The 'test' is non-zero if we need to take a slow path.
3174 }
3176 Node* obj = new_instance(kls, test);
3177 set_result(obj);
3178 return true;
3179 }
3181 #ifdef TRACE_HAVE_INTRINSICS
3182 /*
3183 * oop -> myklass
3184 * myklass->trace_id |= USED
3185 * return myklass->trace_id & ~0x3
3186 */
3187 bool LibraryCallKit::inline_native_classID() {
3188 null_check_receiver(); // null-check, then ignore
3189 Node* cls = null_check(argument(1), T_OBJECT);
3190 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3191 kls = null_check(kls, T_OBJECT);
3192 ByteSize offset = TRACE_ID_OFFSET;
3193 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3194 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3195 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3196 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3197 Node* clsused = longcon(0x01l); // set the class bit
3198 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3200 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3201 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3202 set_result(andl);
3203 return true;
3204 }
3206 bool LibraryCallKit::inline_native_threadID() {
3207 Node* tls_ptr = NULL;
3208 Node* cur_thr = generate_current_thread(tls_ptr);
3209 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3210 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3211 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3213 Node* threadid = NULL;
3214 size_t thread_id_size = OSThread::thread_id_size();
3215 if (thread_id_size == (size_t) BytesPerLong) {
3216 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3217 } else if (thread_id_size == (size_t) BytesPerInt) {
3218 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3219 } else {
3220 ShouldNotReachHere();
3221 }
3222 set_result(threadid);
3223 return true;
3224 }
3225 #endif
3227 //------------------------inline_native_time_funcs--------------
3228 // inline code for System.currentTimeMillis() and System.nanoTime()
3229 // these have the same type and signature
3230 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3231 const TypeFunc* tf = OptoRuntime::void_long_Type();
3232 const TypePtr* no_memory_effects = NULL;
3233 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3234 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3235 #ifdef ASSERT
3236 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3237 assert(value_top == top(), "second value must be top");
3238 #endif
3239 set_result(value);
3240 return true;
3241 }
3243 //------------------------inline_native_currentThread------------------
3244 bool LibraryCallKit::inline_native_currentThread() {
3245 Node* junk = NULL;
3246 set_result(generate_current_thread(junk));
3247 return true;
3248 }
3250 //------------------------inline_native_isInterrupted------------------
3251 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3252 bool LibraryCallKit::inline_native_isInterrupted() {
3253 // Add a fast path to t.isInterrupted(clear_int):
3254 // (t == Thread.current() &&
3255 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3256 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3257 // So, in the common case that the interrupt bit is false,
3258 // we avoid making a call into the VM. Even if the interrupt bit
3259 // is true, if the clear_int argument is false, we avoid the VM call.
3260 // However, if the receiver is not currentThread, we must call the VM,
3261 // because there must be some locking done around the operation.
3263 // We only go to the fast case code if we pass two guards.
3264 // Paths which do not pass are accumulated in the slow_region.
3266 enum {
3267 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3268 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3269 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3270 PATH_LIMIT
3271 };
3273 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3274 // out of the function.
3275 insert_mem_bar(Op_MemBarCPUOrder);
3277 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3278 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3280 RegionNode* slow_region = new (C) RegionNode(1);
3281 record_for_igvn(slow_region);
3283 // (a) Receiving thread must be the current thread.
3284 Node* rec_thr = argument(0);
3285 Node* tls_ptr = NULL;
3286 Node* cur_thr = generate_current_thread(tls_ptr);
3287 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3288 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3290 generate_slow_guard(bol_thr, slow_region);
3292 // (b) Interrupt bit on TLS must be false.
3293 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3294 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3295 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3297 // Set the control input on the field _interrupted read to prevent it floating up.
3298 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3299 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3300 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3302 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3304 // First fast path: if (!TLS._interrupted) return false;
3305 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3306 result_rgn->init_req(no_int_result_path, false_bit);
3307 result_val->init_req(no_int_result_path, intcon(0));
3309 // drop through to next case
3310 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3312 #ifndef TARGET_OS_FAMILY_windows
3313 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3314 Node* clr_arg = argument(1);
3315 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3316 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3317 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3319 // Second fast path: ... else if (!clear_int) return true;
3320 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3321 result_rgn->init_req(no_clear_result_path, false_arg);
3322 result_val->init_req(no_clear_result_path, intcon(1));
3324 // drop through to next case
3325 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3326 #else
3327 // To return true on Windows you must read the _interrupted field
3328 // and check the the event state i.e. take the slow path.
3329 #endif // TARGET_OS_FAMILY_windows
3331 // (d) Otherwise, go to the slow path.
3332 slow_region->add_req(control());
3333 set_control( _gvn.transform(slow_region));
3335 if (stopped()) {
3336 // There is no slow path.
3337 result_rgn->init_req(slow_result_path, top());
3338 result_val->init_req(slow_result_path, top());
3339 } else {
3340 // non-virtual because it is a private non-static
3341 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3343 Node* slow_val = set_results_for_java_call(slow_call);
3344 // this->control() comes from set_results_for_java_call
3346 Node* fast_io = slow_call->in(TypeFunc::I_O);
3347 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3349 // These two phis are pre-filled with copies of of the fast IO and Memory
3350 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3351 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3353 result_rgn->init_req(slow_result_path, control());
3354 result_io ->init_req(slow_result_path, i_o());
3355 result_mem->init_req(slow_result_path, reset_memory());
3356 result_val->init_req(slow_result_path, slow_val);
3358 set_all_memory(_gvn.transform(result_mem));
3359 set_i_o( _gvn.transform(result_io));
3360 }
3362 C->set_has_split_ifs(true); // Has chance for split-if optimization
3363 set_result(result_rgn, result_val);
3364 return true;
3365 }
3367 //---------------------------load_mirror_from_klass----------------------------
3368 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3369 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3370 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3371 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3372 }
3374 //-----------------------load_klass_from_mirror_common-------------------------
3375 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3376 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3377 // and branch to the given path on the region.
3378 // If never_see_null, take an uncommon trap on null, so we can optimistically
3379 // compile for the non-null case.
3380 // If the region is NULL, force never_see_null = true.
3381 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3382 bool never_see_null,
3383 RegionNode* region,
3384 int null_path,
3385 int offset) {
3386 if (region == NULL) never_see_null = true;
3387 Node* p = basic_plus_adr(mirror, offset);
3388 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3389 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3390 Node* null_ctl = top();
3391 kls = null_check_oop(kls, &null_ctl, never_see_null);
3392 if (region != NULL) {
3393 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3394 region->init_req(null_path, null_ctl);
3395 } else {
3396 assert(null_ctl == top(), "no loose ends");
3397 }
3398 return kls;
3399 }
3401 //--------------------(inline_native_Class_query helpers)---------------------
3402 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3403 // Fall through if (mods & mask) == bits, take the guard otherwise.
3404 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3405 // Branch around if the given klass has the given modifier bit set.
3406 // Like generate_guard, adds a new path onto the region.
3407 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3408 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3409 Node* mask = intcon(modifier_mask);
3410 Node* bits = intcon(modifier_bits);
3411 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3412 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3413 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3414 return generate_fair_guard(bol, region);
3415 }
3416 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3417 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3418 }
3420 //-------------------------inline_native_Class_query-------------------
3421 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3422 const Type* return_type = TypeInt::BOOL;
3423 Node* prim_return_value = top(); // what happens if it's a primitive class?
3424 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3425 bool expect_prim = false; // most of these guys expect to work on refs
3427 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3429 Node* mirror = argument(0);
3430 Node* obj = top();
3432 switch (id) {
3433 case vmIntrinsics::_isInstance:
3434 // nothing is an instance of a primitive type
3435 prim_return_value = intcon(0);
3436 obj = argument(1);
3437 break;
3438 case vmIntrinsics::_getModifiers:
3439 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3440 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3441 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3442 break;
3443 case vmIntrinsics::_isInterface:
3444 prim_return_value = intcon(0);
3445 break;
3446 case vmIntrinsics::_isArray:
3447 prim_return_value = intcon(0);
3448 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3449 break;
3450 case vmIntrinsics::_isPrimitive:
3451 prim_return_value = intcon(1);
3452 expect_prim = true; // obviously
3453 break;
3454 case vmIntrinsics::_getSuperclass:
3455 prim_return_value = null();
3456 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3457 break;
3458 case vmIntrinsics::_getComponentType:
3459 prim_return_value = null();
3460 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3461 break;
3462 case vmIntrinsics::_getClassAccessFlags:
3463 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3464 return_type = TypeInt::INT; // not bool! 6297094
3465 break;
3466 default:
3467 fatal_unexpected_iid(id);
3468 break;
3469 }
3471 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3472 if (mirror_con == NULL) return false; // cannot happen?
3474 #ifndef PRODUCT
3475 if (C->print_intrinsics() || C->print_inlining()) {
3476 ciType* k = mirror_con->java_mirror_type();
3477 if (k) {
3478 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3479 k->print_name();
3480 tty->cr();
3481 }
3482 }
3483 #endif
3485 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3486 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3487 record_for_igvn(region);
3488 PhiNode* phi = new (C) PhiNode(region, return_type);
3490 // The mirror will never be null of Reflection.getClassAccessFlags, however
3491 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3492 // if it is. See bug 4774291.
3494 // For Reflection.getClassAccessFlags(), the null check occurs in
3495 // the wrong place; see inline_unsafe_access(), above, for a similar
3496 // situation.
3497 mirror = null_check(mirror);
3498 // If mirror or obj is dead, only null-path is taken.
3499 if (stopped()) return true;
3501 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3503 // Now load the mirror's klass metaobject, and null-check it.
3504 // Side-effects region with the control path if the klass is null.
3505 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3506 // If kls is null, we have a primitive mirror.
3507 phi->init_req(_prim_path, prim_return_value);
3508 if (stopped()) { set_result(region, phi); return true; }
3509 bool safe_for_replace = (region->in(_prim_path) == top());
3511 Node* p; // handy temp
3512 Node* null_ctl;
3514 // Now that we have the non-null klass, we can perform the real query.
3515 // For constant classes, the query will constant-fold in LoadNode::Value.
3516 Node* query_value = top();
3517 switch (id) {
3518 case vmIntrinsics::_isInstance:
3519 // nothing is an instance of a primitive type
3520 query_value = gen_instanceof(obj, kls, safe_for_replace);
3521 break;
3523 case vmIntrinsics::_getModifiers:
3524 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3525 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3526 break;
3528 case vmIntrinsics::_isInterface:
3529 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3530 if (generate_interface_guard(kls, region) != NULL)
3531 // A guard was added. If the guard is taken, it was an interface.
3532 phi->add_req(intcon(1));
3533 // If we fall through, it's a plain class.
3534 query_value = intcon(0);
3535 break;
3537 case vmIntrinsics::_isArray:
3538 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3539 if (generate_array_guard(kls, region) != NULL)
3540 // A guard was added. If the guard is taken, it was an array.
3541 phi->add_req(intcon(1));
3542 // If we fall through, it's a plain class.
3543 query_value = intcon(0);
3544 break;
3546 case vmIntrinsics::_isPrimitive:
3547 query_value = intcon(0); // "normal" path produces false
3548 break;
3550 case vmIntrinsics::_getSuperclass:
3551 // The rules here are somewhat unfortunate, but we can still do better
3552 // with random logic than with a JNI call.
3553 // Interfaces store null or Object as _super, but must report null.
3554 // Arrays store an intermediate super as _super, but must report Object.
3555 // Other types can report the actual _super.
3556 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3557 if (generate_interface_guard(kls, region) != NULL)
3558 // A guard was added. If the guard is taken, it was an interface.
3559 phi->add_req(null());
3560 if (generate_array_guard(kls, region) != NULL)
3561 // A guard was added. If the guard is taken, it was an array.
3562 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3563 // If we fall through, it's a plain class. Get its _super.
3564 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3565 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3566 null_ctl = top();
3567 kls = null_check_oop(kls, &null_ctl);
3568 if (null_ctl != top()) {
3569 // If the guard is taken, Object.superClass is null (both klass and mirror).
3570 region->add_req(null_ctl);
3571 phi ->add_req(null());
3572 }
3573 if (!stopped()) {
3574 query_value = load_mirror_from_klass(kls);
3575 }
3576 break;
3578 case vmIntrinsics::_getComponentType:
3579 if (generate_array_guard(kls, region) != NULL) {
3580 // Be sure to pin the oop load to the guard edge just created:
3581 Node* is_array_ctrl = region->in(region->req()-1);
3582 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3583 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3584 phi->add_req(cmo);
3585 }
3586 query_value = null(); // non-array case is null
3587 break;
3589 case vmIntrinsics::_getClassAccessFlags:
3590 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3591 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3592 break;
3594 default:
3595 fatal_unexpected_iid(id);
3596 break;
3597 }
3599 // Fall-through is the normal case of a query to a real class.
3600 phi->init_req(1, query_value);
3601 region->init_req(1, control());
3603 C->set_has_split_ifs(true); // Has chance for split-if optimization
3604 set_result(region, phi);
3605 return true;
3606 }
3608 //--------------------------inline_native_subtype_check------------------------
3609 // This intrinsic takes the JNI calls out of the heart of
3610 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3611 bool LibraryCallKit::inline_native_subtype_check() {
3612 // Pull both arguments off the stack.
3613 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3614 args[0] = argument(0);
3615 args[1] = argument(1);
3616 Node* klasses[2]; // corresponding Klasses: superk, subk
3617 klasses[0] = klasses[1] = top();
3619 enum {
3620 // A full decision tree on {superc is prim, subc is prim}:
3621 _prim_0_path = 1, // {P,N} => false
3622 // {P,P} & superc!=subc => false
3623 _prim_same_path, // {P,P} & superc==subc => true
3624 _prim_1_path, // {N,P} => false
3625 _ref_subtype_path, // {N,N} & subtype check wins => true
3626 _both_ref_path, // {N,N} & subtype check loses => false
3627 PATH_LIMIT
3628 };
3630 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3631 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3632 record_for_igvn(region);
3634 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3635 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3636 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3638 // First null-check both mirrors and load each mirror's klass metaobject.
3639 int which_arg;
3640 for (which_arg = 0; which_arg <= 1; which_arg++) {
3641 Node* arg = args[which_arg];
3642 arg = null_check(arg);
3643 if (stopped()) break;
3644 args[which_arg] = arg;
3646 Node* p = basic_plus_adr(arg, class_klass_offset);
3647 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3648 klasses[which_arg] = _gvn.transform(kls);
3649 }
3651 // Having loaded both klasses, test each for null.
3652 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3653 for (which_arg = 0; which_arg <= 1; which_arg++) {
3654 Node* kls = klasses[which_arg];
3655 Node* null_ctl = top();
3656 kls = null_check_oop(kls, &null_ctl, never_see_null);
3657 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3658 region->init_req(prim_path, null_ctl);
3659 if (stopped()) break;
3660 klasses[which_arg] = kls;
3661 }
3663 if (!stopped()) {
3664 // now we have two reference types, in klasses[0..1]
3665 Node* subk = klasses[1]; // the argument to isAssignableFrom
3666 Node* superk = klasses[0]; // the receiver
3667 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3668 // now we have a successful reference subtype check
3669 region->set_req(_ref_subtype_path, control());
3670 }
3672 // If both operands are primitive (both klasses null), then
3673 // we must return true when they are identical primitives.
3674 // It is convenient to test this after the first null klass check.
3675 set_control(region->in(_prim_0_path)); // go back to first null check
3676 if (!stopped()) {
3677 // Since superc is primitive, make a guard for the superc==subc case.
3678 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3679 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3680 generate_guard(bol_eq, region, PROB_FAIR);
3681 if (region->req() == PATH_LIMIT+1) {
3682 // A guard was added. If the added guard is taken, superc==subc.
3683 region->swap_edges(PATH_LIMIT, _prim_same_path);
3684 region->del_req(PATH_LIMIT);
3685 }
3686 region->set_req(_prim_0_path, control()); // Not equal after all.
3687 }
3689 // these are the only paths that produce 'true':
3690 phi->set_req(_prim_same_path, intcon(1));
3691 phi->set_req(_ref_subtype_path, intcon(1));
3693 // pull together the cases:
3694 assert(region->req() == PATH_LIMIT, "sane region");
3695 for (uint i = 1; i < region->req(); i++) {
3696 Node* ctl = region->in(i);
3697 if (ctl == NULL || ctl == top()) {
3698 region->set_req(i, top());
3699 phi ->set_req(i, top());
3700 } else if (phi->in(i) == NULL) {
3701 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3702 }
3703 }
3705 set_control(_gvn.transform(region));
3706 set_result(_gvn.transform(phi));
3707 return true;
3708 }
3710 //---------------------generate_array_guard_common------------------------
3711 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3712 bool obj_array, bool not_array) {
3713 // If obj_array/non_array==false/false:
3714 // Branch around if the given klass is in fact an array (either obj or prim).
3715 // If obj_array/non_array==false/true:
3716 // Branch around if the given klass is not an array klass of any kind.
3717 // If obj_array/non_array==true/true:
3718 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3719 // If obj_array/non_array==true/false:
3720 // Branch around if the kls is an oop array (Object[] or subtype)
3721 //
3722 // Like generate_guard, adds a new path onto the region.
3723 jint layout_con = 0;
3724 Node* layout_val = get_layout_helper(kls, layout_con);
3725 if (layout_val == NULL) {
3726 bool query = (obj_array
3727 ? Klass::layout_helper_is_objArray(layout_con)
3728 : Klass::layout_helper_is_array(layout_con));
3729 if (query == not_array) {
3730 return NULL; // never a branch
3731 } else { // always a branch
3732 Node* always_branch = control();
3733 if (region != NULL)
3734 region->add_req(always_branch);
3735 set_control(top());
3736 return always_branch;
3737 }
3738 }
3739 // Now test the correct condition.
3740 jint nval = (obj_array
3741 ? ((jint)Klass::_lh_array_tag_type_value
3742 << Klass::_lh_array_tag_shift)
3743 : Klass::_lh_neutral_value);
3744 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3745 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3746 // invert the test if we are looking for a non-array
3747 if (not_array) btest = BoolTest(btest).negate();
3748 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3749 return generate_fair_guard(bol, region);
3750 }
3753 //-----------------------inline_native_newArray--------------------------
3754 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3755 bool LibraryCallKit::inline_native_newArray() {
3756 Node* mirror = argument(0);
3757 Node* count_val = argument(1);
3759 mirror = null_check(mirror);
3760 // If mirror or obj is dead, only null-path is taken.
3761 if (stopped()) return true;
3763 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3764 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3765 PhiNode* result_val = new(C) PhiNode(result_reg,
3766 TypeInstPtr::NOTNULL);
3767 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3768 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3769 TypePtr::BOTTOM);
3771 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3772 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3773 result_reg, _slow_path);
3774 Node* normal_ctl = control();
3775 Node* no_array_ctl = result_reg->in(_slow_path);
3777 // Generate code for the slow case. We make a call to newArray().
3778 set_control(no_array_ctl);
3779 if (!stopped()) {
3780 // Either the input type is void.class, or else the
3781 // array klass has not yet been cached. Either the
3782 // ensuing call will throw an exception, or else it
3783 // will cache the array klass for next time.
3784 PreserveJVMState pjvms(this);
3785 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3786 Node* slow_result = set_results_for_java_call(slow_call);
3787 // this->control() comes from set_results_for_java_call
3788 result_reg->set_req(_slow_path, control());
3789 result_val->set_req(_slow_path, slow_result);
3790 result_io ->set_req(_slow_path, i_o());
3791 result_mem->set_req(_slow_path, reset_memory());
3792 }
3794 set_control(normal_ctl);
3795 if (!stopped()) {
3796 // Normal case: The array type has been cached in the java.lang.Class.
3797 // The following call works fine even if the array type is polymorphic.
3798 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3799 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3800 result_reg->init_req(_normal_path, control());
3801 result_val->init_req(_normal_path, obj);
3802 result_io ->init_req(_normal_path, i_o());
3803 result_mem->init_req(_normal_path, reset_memory());
3804 }
3806 // Return the combined state.
3807 set_i_o( _gvn.transform(result_io) );
3808 set_all_memory( _gvn.transform(result_mem));
3810 C->set_has_split_ifs(true); // Has chance for split-if optimization
3811 set_result(result_reg, result_val);
3812 return true;
3813 }
3815 //----------------------inline_native_getLength--------------------------
3816 // public static native int java.lang.reflect.Array.getLength(Object array);
3817 bool LibraryCallKit::inline_native_getLength() {
3818 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3820 Node* array = null_check(argument(0));
3821 // If array is dead, only null-path is taken.
3822 if (stopped()) return true;
3824 // Deoptimize if it is a non-array.
3825 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3827 if (non_array != NULL) {
3828 PreserveJVMState pjvms(this);
3829 set_control(non_array);
3830 uncommon_trap(Deoptimization::Reason_intrinsic,
3831 Deoptimization::Action_maybe_recompile);
3832 }
3834 // If control is dead, only non-array-path is taken.
3835 if (stopped()) return true;
3837 // The works fine even if the array type is polymorphic.
3838 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3839 Node* result = load_array_length(array);
3841 C->set_has_split_ifs(true); // Has chance for split-if optimization
3842 set_result(result);
3843 return true;
3844 }
3846 //------------------------inline_array_copyOf----------------------------
3847 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3848 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3849 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3850 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3852 // Get the arguments.
3853 Node* original = argument(0);
3854 Node* start = is_copyOfRange? argument(1): intcon(0);
3855 Node* end = is_copyOfRange? argument(2): argument(1);
3856 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3858 Node* newcopy;
3860 // Set the original stack and the reexecute bit for the interpreter to reexecute
3861 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3862 { PreserveReexecuteState preexecs(this);
3863 jvms()->set_should_reexecute(true);
3865 array_type_mirror = null_check(array_type_mirror);
3866 original = null_check(original);
3868 // Check if a null path was taken unconditionally.
3869 if (stopped()) return true;
3871 Node* orig_length = load_array_length(original);
3873 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3874 klass_node = null_check(klass_node);
3876 RegionNode* bailout = new (C) RegionNode(1);
3877 record_for_igvn(bailout);
3879 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3880 // Bail out if that is so.
3881 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3882 if (not_objArray != NULL) {
3883 // Improve the klass node's type from the new optimistic assumption:
3884 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3885 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3886 Node* cast = new (C) CastPPNode(klass_node, akls);
3887 cast->init_req(0, control());
3888 klass_node = _gvn.transform(cast);
3889 }
3891 // Bail out if either start or end is negative.
3892 generate_negative_guard(start, bailout, &start);
3893 generate_negative_guard(end, bailout, &end);
3895 Node* length = end;
3896 if (_gvn.type(start) != TypeInt::ZERO) {
3897 length = _gvn.transform(new (C) SubINode(end, start));
3898 }
3900 // Bail out if length is negative.
3901 // Without this the new_array would throw
3902 // NegativeArraySizeException but IllegalArgumentException is what
3903 // should be thrown
3904 generate_negative_guard(length, bailout, &length);
3906 if (bailout->req() > 1) {
3907 PreserveJVMState pjvms(this);
3908 set_control(_gvn.transform(bailout));
3909 uncommon_trap(Deoptimization::Reason_intrinsic,
3910 Deoptimization::Action_maybe_recompile);
3911 }
3913 if (!stopped()) {
3914 // How many elements will we copy from the original?
3915 // The answer is MinI(orig_length - start, length).
3916 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3917 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3919 newcopy = new_array(klass_node, length, 0); // no argments to push
3921 // Generate a direct call to the right arraycopy function(s).
3922 // We know the copy is disjoint but we might not know if the
3923 // oop stores need checking.
3924 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3925 // This will fail a store-check if x contains any non-nulls.
3926 bool disjoint_bases = true;
3927 // if start > orig_length then the length of the copy may be
3928 // negative.
3929 bool length_never_negative = !is_copyOfRange;
3930 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3931 original, start, newcopy, intcon(0), moved,
3932 disjoint_bases, length_never_negative);
3933 }
3934 } // original reexecute is set back here
3936 C->set_has_split_ifs(true); // Has chance for split-if optimization
3937 if (!stopped()) {
3938 set_result(newcopy);
3939 }
3940 return true;
3941 }
3944 //----------------------generate_virtual_guard---------------------------
3945 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3946 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3947 RegionNode* slow_region) {
3948 ciMethod* method = callee();
3949 int vtable_index = method->vtable_index();
3950 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3951 err_msg_res("bad index %d", vtable_index));
3952 // Get the Method* out of the appropriate vtable entry.
3953 int entry_offset = (InstanceKlass::vtable_start_offset() +
3954 vtable_index*vtableEntry::size()) * wordSize +
3955 vtableEntry::method_offset_in_bytes();
3956 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3957 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3959 // Compare the target method with the expected method (e.g., Object.hashCode).
3960 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3962 Node* native_call = makecon(native_call_addr);
3963 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3964 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3966 return generate_slow_guard(test_native, slow_region);
3967 }
3969 //-----------------------generate_method_call----------------------------
3970 // Use generate_method_call to make a slow-call to the real
3971 // method if the fast path fails. An alternative would be to
3972 // use a stub like OptoRuntime::slow_arraycopy_Java.
3973 // This only works for expanding the current library call,
3974 // not another intrinsic. (E.g., don't use this for making an
3975 // arraycopy call inside of the copyOf intrinsic.)
3976 CallJavaNode*
3977 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3978 // When compiling the intrinsic method itself, do not use this technique.
3979 guarantee(callee() != C->method(), "cannot make slow-call to self");
3981 ciMethod* method = callee();
3982 // ensure the JVMS we have will be correct for this call
3983 guarantee(method_id == method->intrinsic_id(), "must match");
3985 const TypeFunc* tf = TypeFunc::make(method);
3986 CallJavaNode* slow_call;
3987 if (is_static) {
3988 assert(!is_virtual, "");
3989 slow_call = new(C) CallStaticJavaNode(C, tf,
3990 SharedRuntime::get_resolve_static_call_stub(),
3991 method, bci());
3992 } else if (is_virtual) {
3993 null_check_receiver();
3994 int vtable_index = Method::invalid_vtable_index;
3995 if (UseInlineCaches) {
3996 // Suppress the vtable call
3997 } else {
3998 // hashCode and clone are not a miranda methods,
3999 // so the vtable index is fixed.
4000 // No need to use the linkResolver to get it.
4001 vtable_index = method->vtable_index();
4002 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4003 err_msg_res("bad index %d", vtable_index));
4004 }
4005 slow_call = new(C) CallDynamicJavaNode(tf,
4006 SharedRuntime::get_resolve_virtual_call_stub(),
4007 method, vtable_index, bci());
4008 } else { // neither virtual nor static: opt_virtual
4009 null_check_receiver();
4010 slow_call = new(C) CallStaticJavaNode(C, tf,
4011 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4012 method, bci());
4013 slow_call->set_optimized_virtual(true);
4014 }
4015 set_arguments_for_java_call(slow_call);
4016 set_edges_for_java_call(slow_call);
4017 return slow_call;
4018 }
4021 /**
4022 * Build special case code for calls to hashCode on an object. This call may
4023 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4024 * slightly different code.
4025 */
4026 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4027 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4028 assert(!(is_virtual && is_static), "either virtual, special, or static");
4030 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4032 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4033 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4034 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4035 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4036 Node* obj = NULL;
4037 if (!is_static) {
4038 // Check for hashing null object
4039 obj = null_check_receiver();
4040 if (stopped()) return true; // unconditionally null
4041 result_reg->init_req(_null_path, top());
4042 result_val->init_req(_null_path, top());
4043 } else {
4044 // Do a null check, and return zero if null.
4045 // System.identityHashCode(null) == 0
4046 obj = argument(0);
4047 Node* null_ctl = top();
4048 obj = null_check_oop(obj, &null_ctl);
4049 result_reg->init_req(_null_path, null_ctl);
4050 result_val->init_req(_null_path, _gvn.intcon(0));
4051 }
4053 // Unconditionally null? Then return right away.
4054 if (stopped()) {
4055 set_control( result_reg->in(_null_path));
4056 if (!stopped())
4057 set_result(result_val->in(_null_path));
4058 return true;
4059 }
4061 // We only go to the fast case code if we pass a number of guards. The
4062 // paths which do not pass are accumulated in the slow_region.
4063 RegionNode* slow_region = new (C) RegionNode(1);
4064 record_for_igvn(slow_region);
4066 // If this is a virtual call, we generate a funny guard. We pull out
4067 // the vtable entry corresponding to hashCode() from the target object.
4068 // If the target method which we are calling happens to be the native
4069 // Object hashCode() method, we pass the guard. We do not need this
4070 // guard for non-virtual calls -- the caller is known to be the native
4071 // Object hashCode().
4072 if (is_virtual) {
4073 // After null check, get the object's klass.
4074 Node* obj_klass = load_object_klass(obj);
4075 generate_virtual_guard(obj_klass, slow_region);
4076 }
4078 // Get the header out of the object, use LoadMarkNode when available
4079 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4080 // The control of the load must be NULL. Otherwise, the load can move before
4081 // the null check after castPP removal.
4082 Node* no_ctrl = NULL;
4083 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4085 // Test the header to see if it is unlocked.
4086 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4087 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4088 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4089 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4090 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4092 generate_slow_guard(test_unlocked, slow_region);
4094 // Get the hash value and check to see that it has been properly assigned.
4095 // We depend on hash_mask being at most 32 bits and avoid the use of
4096 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4097 // vm: see markOop.hpp.
4098 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4099 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4100 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4101 // This hack lets the hash bits live anywhere in the mark object now, as long
4102 // as the shift drops the relevant bits into the low 32 bits. Note that
4103 // Java spec says that HashCode is an int so there's no point in capturing
4104 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4105 hshifted_header = ConvX2I(hshifted_header);
4106 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4108 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4109 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4110 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4112 generate_slow_guard(test_assigned, slow_region);
4114 Node* init_mem = reset_memory();
4115 // fill in the rest of the null path:
4116 result_io ->init_req(_null_path, i_o());
4117 result_mem->init_req(_null_path, init_mem);
4119 result_val->init_req(_fast_path, hash_val);
4120 result_reg->init_req(_fast_path, control());
4121 result_io ->init_req(_fast_path, i_o());
4122 result_mem->init_req(_fast_path, init_mem);
4124 // Generate code for the slow case. We make a call to hashCode().
4125 set_control(_gvn.transform(slow_region));
4126 if (!stopped()) {
4127 // No need for PreserveJVMState, because we're using up the present state.
4128 set_all_memory(init_mem);
4129 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4130 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4131 Node* slow_result = set_results_for_java_call(slow_call);
4132 // this->control() comes from set_results_for_java_call
4133 result_reg->init_req(_slow_path, control());
4134 result_val->init_req(_slow_path, slow_result);
4135 result_io ->set_req(_slow_path, i_o());
4136 result_mem ->set_req(_slow_path, reset_memory());
4137 }
4139 // Return the combined state.
4140 set_i_o( _gvn.transform(result_io) );
4141 set_all_memory( _gvn.transform(result_mem));
4143 set_result(result_reg, result_val);
4144 return true;
4145 }
4147 //---------------------------inline_native_getClass----------------------------
4148 // public final native Class<?> java.lang.Object.getClass();
4149 //
4150 // Build special case code for calls to getClass on an object.
4151 bool LibraryCallKit::inline_native_getClass() {
4152 Node* obj = null_check_receiver();
4153 if (stopped()) return true;
4154 set_result(load_mirror_from_klass(load_object_klass(obj)));
4155 return true;
4156 }
4158 //-----------------inline_native_Reflection_getCallerClass---------------------
4159 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4160 //
4161 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4162 //
4163 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4164 // in that it must skip particular security frames and checks for
4165 // caller sensitive methods.
4166 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4167 #ifndef PRODUCT
4168 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4169 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4170 }
4171 #endif
4173 if (!jvms()->has_method()) {
4174 #ifndef PRODUCT
4175 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4176 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4177 }
4178 #endif
4179 return false;
4180 }
4182 // Walk back up the JVM state to find the caller at the required
4183 // depth.
4184 JVMState* caller_jvms = jvms();
4186 // Cf. JVM_GetCallerClass
4187 // NOTE: Start the loop at depth 1 because the current JVM state does
4188 // not include the Reflection.getCallerClass() frame.
4189 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4190 ciMethod* m = caller_jvms->method();
4191 switch (n) {
4192 case 0:
4193 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4194 break;
4195 case 1:
4196 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4197 if (!m->caller_sensitive()) {
4198 #ifndef PRODUCT
4199 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4200 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4201 }
4202 #endif
4203 return false; // bail-out; let JVM_GetCallerClass do the work
4204 }
4205 break;
4206 default:
4207 if (!m->is_ignored_by_security_stack_walk()) {
4208 // We have reached the desired frame; return the holder class.
4209 // Acquire method holder as java.lang.Class and push as constant.
4210 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4211 ciInstance* caller_mirror = caller_klass->java_mirror();
4212 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4214 #ifndef PRODUCT
4215 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4216 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());
4217 tty->print_cr(" JVM state at this point:");
4218 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4219 ciMethod* m = jvms()->of_depth(i)->method();
4220 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4221 }
4222 }
4223 #endif
4224 return true;
4225 }
4226 break;
4227 }
4228 }
4230 #ifndef PRODUCT
4231 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4232 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4233 tty->print_cr(" JVM state at this point:");
4234 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4235 ciMethod* m = jvms()->of_depth(i)->method();
4236 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4237 }
4238 }
4239 #endif
4241 return false; // bail-out; let JVM_GetCallerClass do the work
4242 }
4244 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4245 Node* arg = argument(0);
4246 Node* result;
4248 switch (id) {
4249 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4250 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4251 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4252 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4254 case vmIntrinsics::_doubleToLongBits: {
4255 // two paths (plus control) merge in a wood
4256 RegionNode *r = new (C) RegionNode(3);
4257 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4259 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4260 // Build the boolean node
4261 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4263 // Branch either way.
4264 // NaN case is less traveled, which makes all the difference.
4265 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4266 Node *opt_isnan = _gvn.transform(ifisnan);
4267 assert( opt_isnan->is_If(), "Expect an IfNode");
4268 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4269 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4271 set_control(iftrue);
4273 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4274 Node *slow_result = longcon(nan_bits); // return NaN
4275 phi->init_req(1, _gvn.transform( slow_result ));
4276 r->init_req(1, iftrue);
4278 // Else fall through
4279 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4280 set_control(iffalse);
4282 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4283 r->init_req(2, iffalse);
4285 // Post merge
4286 set_control(_gvn.transform(r));
4287 record_for_igvn(r);
4289 C->set_has_split_ifs(true); // Has chance for split-if optimization
4290 result = phi;
4291 assert(result->bottom_type()->isa_long(), "must be");
4292 break;
4293 }
4295 case vmIntrinsics::_floatToIntBits: {
4296 // two paths (plus control) merge in a wood
4297 RegionNode *r = new (C) RegionNode(3);
4298 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4300 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4301 // Build the boolean node
4302 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4304 // Branch either way.
4305 // NaN case is less traveled, which makes all the difference.
4306 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4307 Node *opt_isnan = _gvn.transform(ifisnan);
4308 assert( opt_isnan->is_If(), "Expect an IfNode");
4309 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4310 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4312 set_control(iftrue);
4314 static const jint nan_bits = 0x7fc00000;
4315 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4316 phi->init_req(1, _gvn.transform( slow_result ));
4317 r->init_req(1, iftrue);
4319 // Else fall through
4320 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4321 set_control(iffalse);
4323 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4324 r->init_req(2, iffalse);
4326 // Post merge
4327 set_control(_gvn.transform(r));
4328 record_for_igvn(r);
4330 C->set_has_split_ifs(true); // Has chance for split-if optimization
4331 result = phi;
4332 assert(result->bottom_type()->isa_int(), "must be");
4333 break;
4334 }
4336 default:
4337 fatal_unexpected_iid(id);
4338 break;
4339 }
4340 set_result(_gvn.transform(result));
4341 return true;
4342 }
4344 #ifdef _LP64
4345 #define XTOP ,top() /*additional argument*/
4346 #else //_LP64
4347 #define XTOP /*no additional argument*/
4348 #endif //_LP64
4350 //----------------------inline_unsafe_copyMemory-------------------------
4351 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4352 bool LibraryCallKit::inline_unsafe_copyMemory() {
4353 if (callee()->is_static()) return false; // caller must have the capability!
4354 null_check_receiver(); // null-check receiver
4355 if (stopped()) return true;
4357 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4359 Node* src_ptr = argument(1); // type: oop
4360 Node* src_off = ConvL2X(argument(2)); // type: long
4361 Node* dst_ptr = argument(4); // type: oop
4362 Node* dst_off = ConvL2X(argument(5)); // type: long
4363 Node* size = ConvL2X(argument(7)); // type: long
4365 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4366 "fieldOffset must be byte-scaled");
4368 Node* src = make_unsafe_address(src_ptr, src_off);
4369 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4371 // Conservatively insert a memory barrier on all memory slices.
4372 // Do not let writes of the copy source or destination float below the copy.
4373 insert_mem_bar(Op_MemBarCPUOrder);
4375 // Call it. Note that the length argument is not scaled.
4376 make_runtime_call(RC_LEAF|RC_NO_FP,
4377 OptoRuntime::fast_arraycopy_Type(),
4378 StubRoutines::unsafe_arraycopy(),
4379 "unsafe_arraycopy",
4380 TypeRawPtr::BOTTOM,
4381 src, dst, size XTOP);
4383 // Do not let reads of the copy destination float above the copy.
4384 insert_mem_bar(Op_MemBarCPUOrder);
4386 return true;
4387 }
4389 //------------------------clone_coping-----------------------------------
4390 // Helper function for inline_native_clone.
4391 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4392 assert(obj_size != NULL, "");
4393 Node* raw_obj = alloc_obj->in(1);
4394 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4396 AllocateNode* alloc = NULL;
4397 if (ReduceBulkZeroing) {
4398 // We will be completely responsible for initializing this object -
4399 // mark Initialize node as complete.
4400 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4401 // The object was just allocated - there should be no any stores!
4402 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4403 // Mark as complete_with_arraycopy so that on AllocateNode
4404 // expansion, we know this AllocateNode is initialized by an array
4405 // copy and a StoreStore barrier exists after the array copy.
4406 alloc->initialization()->set_complete_with_arraycopy();
4407 }
4409 // Copy the fastest available way.
4410 // TODO: generate fields copies for small objects instead.
4411 Node* src = obj;
4412 Node* dest = alloc_obj;
4413 Node* size = _gvn.transform(obj_size);
4415 // Exclude the header but include array length to copy by 8 bytes words.
4416 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4417 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4418 instanceOopDesc::base_offset_in_bytes();
4419 // base_off:
4420 // 8 - 32-bit VM
4421 // 12 - 64-bit VM, compressed klass
4422 // 16 - 64-bit VM, normal klass
4423 if (base_off % BytesPerLong != 0) {
4424 assert(UseCompressedClassPointers, "");
4425 if (is_array) {
4426 // Exclude length to copy by 8 bytes words.
4427 base_off += sizeof(int);
4428 } else {
4429 // Include klass to copy by 8 bytes words.
4430 base_off = instanceOopDesc::klass_offset_in_bytes();
4431 }
4432 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4433 }
4434 src = basic_plus_adr(src, base_off);
4435 dest = basic_plus_adr(dest, base_off);
4437 // Compute the length also, if needed:
4438 Node* countx = size;
4439 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4440 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4442 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4443 bool disjoint_bases = true;
4444 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4445 src, NULL, dest, NULL, countx,
4446 /*dest_uninitialized*/true);
4448 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4449 if (card_mark) {
4450 assert(!is_array, "");
4451 // Put in store barrier for any and all oops we are sticking
4452 // into this object. (We could avoid this if we could prove
4453 // that the object type contains no oop fields at all.)
4454 Node* no_particular_value = NULL;
4455 Node* no_particular_field = NULL;
4456 int raw_adr_idx = Compile::AliasIdxRaw;
4457 post_barrier(control(),
4458 memory(raw_adr_type),
4459 alloc_obj,
4460 no_particular_field,
4461 raw_adr_idx,
4462 no_particular_value,
4463 T_OBJECT,
4464 false);
4465 }
4467 // Do not let reads from the cloned object float above the arraycopy.
4468 if (alloc != NULL) {
4469 // Do not let stores that initialize this object be reordered with
4470 // a subsequent store that would make this object accessible by
4471 // other threads.
4472 // Record what AllocateNode this StoreStore protects so that
4473 // escape analysis can go from the MemBarStoreStoreNode to the
4474 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4475 // based on the escape status of the AllocateNode.
4476 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4477 } else {
4478 insert_mem_bar(Op_MemBarCPUOrder);
4479 }
4480 }
4482 //------------------------inline_native_clone----------------------------
4483 // protected native Object java.lang.Object.clone();
4484 //
4485 // Here are the simple edge cases:
4486 // null receiver => normal trap
4487 // virtual and clone was overridden => slow path to out-of-line clone
4488 // not cloneable or finalizer => slow path to out-of-line Object.clone
4489 //
4490 // The general case has two steps, allocation and copying.
4491 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4492 //
4493 // Copying also has two cases, oop arrays and everything else.
4494 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4495 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4496 //
4497 // These steps fold up nicely if and when the cloned object's klass
4498 // can be sharply typed as an object array, a type array, or an instance.
4499 //
4500 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4501 PhiNode* result_val;
4503 // Set the reexecute bit for the interpreter to reexecute
4504 // the bytecode that invokes Object.clone if deoptimization happens.
4505 { PreserveReexecuteState preexecs(this);
4506 jvms()->set_should_reexecute(true);
4508 Node* obj = null_check_receiver();
4509 if (stopped()) return true;
4511 Node* obj_klass = load_object_klass(obj);
4512 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4513 const TypeOopPtr* toop = ((tklass != NULL)
4514 ? tklass->as_instance_type()
4515 : TypeInstPtr::NOTNULL);
4517 // Conservatively insert a memory barrier on all memory slices.
4518 // Do not let writes into the original float below the clone.
4519 insert_mem_bar(Op_MemBarCPUOrder);
4521 // paths into result_reg:
4522 enum {
4523 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4524 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4525 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4526 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4527 PATH_LIMIT
4528 };
4529 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4530 result_val = new(C) PhiNode(result_reg,
4531 TypeInstPtr::NOTNULL);
4532 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4533 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4534 TypePtr::BOTTOM);
4535 record_for_igvn(result_reg);
4537 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4538 int raw_adr_idx = Compile::AliasIdxRaw;
4540 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4541 if (array_ctl != NULL) {
4542 // It's an array.
4543 PreserveJVMState pjvms(this);
4544 set_control(array_ctl);
4545 Node* obj_length = load_array_length(obj);
4546 Node* obj_size = NULL;
4547 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4549 if (!use_ReduceInitialCardMarks()) {
4550 // If it is an oop array, it requires very special treatment,
4551 // because card marking is required on each card of the array.
4552 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4553 if (is_obja != NULL) {
4554 PreserveJVMState pjvms2(this);
4555 set_control(is_obja);
4556 // Generate a direct call to the right arraycopy function(s).
4557 bool disjoint_bases = true;
4558 bool length_never_negative = true;
4559 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4560 obj, intcon(0), alloc_obj, intcon(0),
4561 obj_length,
4562 disjoint_bases, length_never_negative);
4563 result_reg->init_req(_objArray_path, control());
4564 result_val->init_req(_objArray_path, alloc_obj);
4565 result_i_o ->set_req(_objArray_path, i_o());
4566 result_mem ->set_req(_objArray_path, reset_memory());
4567 }
4568 }
4569 // Otherwise, there are no card marks to worry about.
4570 // (We can dispense with card marks if we know the allocation
4571 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4572 // causes the non-eden paths to take compensating steps to
4573 // simulate a fresh allocation, so that no further
4574 // card marks are required in compiled code to initialize
4575 // the object.)
4577 if (!stopped()) {
4578 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4580 // Present the results of the copy.
4581 result_reg->init_req(_array_path, control());
4582 result_val->init_req(_array_path, alloc_obj);
4583 result_i_o ->set_req(_array_path, i_o());
4584 result_mem ->set_req(_array_path, reset_memory());
4585 }
4586 }
4588 // We only go to the instance fast case code if we pass a number of guards.
4589 // The paths which do not pass are accumulated in the slow_region.
4590 RegionNode* slow_region = new (C) RegionNode(1);
4591 record_for_igvn(slow_region);
4592 if (!stopped()) {
4593 // It's an instance (we did array above). Make the slow-path tests.
4594 // If this is a virtual call, we generate a funny guard. We grab
4595 // the vtable entry corresponding to clone() from the target object.
4596 // If the target method which we are calling happens to be the
4597 // Object clone() method, we pass the guard. We do not need this
4598 // guard for non-virtual calls; the caller is known to be the native
4599 // Object clone().
4600 if (is_virtual) {
4601 generate_virtual_guard(obj_klass, slow_region);
4602 }
4604 // The object must be cloneable and must not have a finalizer.
4605 // Both of these conditions may be checked in a single test.
4606 // We could optimize the cloneable test further, but we don't care.
4607 generate_access_flags_guard(obj_klass,
4608 // Test both conditions:
4609 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4610 // Must be cloneable but not finalizer:
4611 JVM_ACC_IS_CLONEABLE,
4612 slow_region);
4613 }
4615 if (!stopped()) {
4616 // It's an instance, and it passed the slow-path tests.
4617 PreserveJVMState pjvms(this);
4618 Node* obj_size = NULL;
4619 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4620 // is reexecuted if deoptimization occurs and there could be problems when merging
4621 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4622 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4624 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4626 // Present the results of the slow call.
4627 result_reg->init_req(_instance_path, control());
4628 result_val->init_req(_instance_path, alloc_obj);
4629 result_i_o ->set_req(_instance_path, i_o());
4630 result_mem ->set_req(_instance_path, reset_memory());
4631 }
4633 // Generate code for the slow case. We make a call to clone().
4634 set_control(_gvn.transform(slow_region));
4635 if (!stopped()) {
4636 PreserveJVMState pjvms(this);
4637 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4638 Node* slow_result = set_results_for_java_call(slow_call);
4639 // this->control() comes from set_results_for_java_call
4640 result_reg->init_req(_slow_path, control());
4641 result_val->init_req(_slow_path, slow_result);
4642 result_i_o ->set_req(_slow_path, i_o());
4643 result_mem ->set_req(_slow_path, reset_memory());
4644 }
4646 // Return the combined state.
4647 set_control( _gvn.transform(result_reg));
4648 set_i_o( _gvn.transform(result_i_o));
4649 set_all_memory( _gvn.transform(result_mem));
4650 } // original reexecute is set back here
4652 set_result(_gvn.transform(result_val));
4653 return true;
4654 }
4656 //------------------------------basictype2arraycopy----------------------------
4657 address LibraryCallKit::basictype2arraycopy(BasicType t,
4658 Node* src_offset,
4659 Node* dest_offset,
4660 bool disjoint_bases,
4661 const char* &name,
4662 bool dest_uninitialized) {
4663 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4664 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4666 bool aligned = false;
4667 bool disjoint = disjoint_bases;
4669 // if the offsets are the same, we can treat the memory regions as
4670 // disjoint, because either the memory regions are in different arrays,
4671 // or they are identical (which we can treat as disjoint.) We can also
4672 // treat a copy with a destination index less that the source index
4673 // as disjoint since a low->high copy will work correctly in this case.
4674 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4675 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4676 // both indices are constants
4677 int s_offs = src_offset_inttype->get_con();
4678 int d_offs = dest_offset_inttype->get_con();
4679 int element_size = type2aelembytes(t);
4680 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4681 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4682 if (s_offs >= d_offs) disjoint = true;
4683 } else if (src_offset == dest_offset && src_offset != NULL) {
4684 // This can occur if the offsets are identical non-constants.
4685 disjoint = true;
4686 }
4688 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4689 }
4692 //------------------------------inline_arraycopy-----------------------
4693 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4694 // Object dest, int destPos,
4695 // int length);
4696 bool LibraryCallKit::inline_arraycopy() {
4697 // Get the arguments.
4698 Node* src = argument(0); // type: oop
4699 Node* src_offset = argument(1); // type: int
4700 Node* dest = argument(2); // type: oop
4701 Node* dest_offset = argument(3); // type: int
4702 Node* length = argument(4); // type: int
4704 // Compile time checks. If any of these checks cannot be verified at compile time,
4705 // we do not make a fast path for this call. Instead, we let the call remain as it
4706 // is. The checks we choose to mandate at compile time are:
4707 //
4708 // (1) src and dest are arrays.
4709 const Type* src_type = src->Value(&_gvn);
4710 const Type* dest_type = dest->Value(&_gvn);
4711 const TypeAryPtr* top_src = src_type->isa_aryptr();
4712 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4714 // Do we have the type of src?
4715 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4716 // Do we have the type of dest?
4717 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4718 // Is the type for src from speculation?
4719 bool src_spec = false;
4720 // Is the type for dest from speculation?
4721 bool dest_spec = false;
4723 if (!has_src || !has_dest) {
4724 // We don't have sufficient type information, let's see if
4725 // speculative types can help. We need to have types for both src
4726 // and dest so that it pays off.
4728 // Do we already have or could we have type information for src
4729 bool could_have_src = has_src;
4730 // Do we already have or could we have type information for dest
4731 bool could_have_dest = has_dest;
4733 ciKlass* src_k = NULL;
4734 if (!has_src) {
4735 src_k = src_type->speculative_type();
4736 if (src_k != NULL && src_k->is_array_klass()) {
4737 could_have_src = true;
4738 }
4739 }
4741 ciKlass* dest_k = NULL;
4742 if (!has_dest) {
4743 dest_k = dest_type->speculative_type();
4744 if (dest_k != NULL && dest_k->is_array_klass()) {
4745 could_have_dest = true;
4746 }
4747 }
4749 if (could_have_src && could_have_dest) {
4750 // This is going to pay off so emit the required guards
4751 if (!has_src) {
4752 src = maybe_cast_profiled_obj(src, src_k);
4753 src_type = _gvn.type(src);
4754 top_src = src_type->isa_aryptr();
4755 has_src = (top_src != NULL && top_src->klass() != NULL);
4756 src_spec = true;
4757 }
4758 if (!has_dest) {
4759 dest = maybe_cast_profiled_obj(dest, dest_k);
4760 dest_type = _gvn.type(dest);
4761 top_dest = dest_type->isa_aryptr();
4762 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4763 dest_spec = true;
4764 }
4765 }
4766 }
4768 if (!has_src || !has_dest) {
4769 // Conservatively insert a memory barrier on all memory slices.
4770 // Do not let writes into the source float below the arraycopy.
4771 insert_mem_bar(Op_MemBarCPUOrder);
4773 // Call StubRoutines::generic_arraycopy stub.
4774 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4775 src, src_offset, dest, dest_offset, length);
4777 // Do not let reads from the destination float above the arraycopy.
4778 // Since we cannot type the arrays, we don't know which slices
4779 // might be affected. We could restrict this barrier only to those
4780 // memory slices which pertain to array elements--but don't bother.
4781 if (!InsertMemBarAfterArraycopy)
4782 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4783 insert_mem_bar(Op_MemBarCPUOrder);
4784 return true;
4785 }
4787 // (2) src and dest arrays must have elements of the same BasicType
4788 // Figure out the size and type of the elements we will be copying.
4789 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4790 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4791 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4792 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4794 if (src_elem != dest_elem || dest_elem == T_VOID) {
4795 // The component types are not the same or are not recognized. Punt.
4796 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4797 generate_slow_arraycopy(TypePtr::BOTTOM,
4798 src, src_offset, dest, dest_offset, length,
4799 /*dest_uninitialized*/false);
4800 return true;
4801 }
4803 if (src_elem == T_OBJECT) {
4804 // If both arrays are object arrays then having the exact types
4805 // for both will remove the need for a subtype check at runtime
4806 // before the call and may make it possible to pick a faster copy
4807 // routine (without a subtype check on every element)
4808 // Do we have the exact type of src?
4809 bool could_have_src = src_spec;
4810 // Do we have the exact type of dest?
4811 bool could_have_dest = dest_spec;
4812 ciKlass* src_k = top_src->klass();
4813 ciKlass* dest_k = top_dest->klass();
4814 if (!src_spec) {
4815 src_k = src_type->speculative_type();
4816 if (src_k != NULL && src_k->is_array_klass()) {
4817 could_have_src = true;
4818 }
4819 }
4820 if (!dest_spec) {
4821 dest_k = dest_type->speculative_type();
4822 if (dest_k != NULL && dest_k->is_array_klass()) {
4823 could_have_dest = true;
4824 }
4825 }
4826 if (could_have_src && could_have_dest) {
4827 // If we can have both exact types, emit the missing guards
4828 if (could_have_src && !src_spec) {
4829 src = maybe_cast_profiled_obj(src, src_k);
4830 }
4831 if (could_have_dest && !dest_spec) {
4832 dest = maybe_cast_profiled_obj(dest, dest_k);
4833 }
4834 }
4835 }
4837 //---------------------------------------------------------------------------
4838 // We will make a fast path for this call to arraycopy.
4840 // We have the following tests left to perform:
4841 //
4842 // (3) src and dest must not be null.
4843 // (4) src_offset must not be negative.
4844 // (5) dest_offset must not be negative.
4845 // (6) length must not be negative.
4846 // (7) src_offset + length must not exceed length of src.
4847 // (8) dest_offset + length must not exceed length of dest.
4848 // (9) each element of an oop array must be assignable
4850 RegionNode* slow_region = new (C) RegionNode(1);
4851 record_for_igvn(slow_region);
4853 // (3) operands must not be null
4854 // We currently perform our null checks with the null_check routine.
4855 // This means that the null exceptions will be reported in the caller
4856 // rather than (correctly) reported inside of the native arraycopy call.
4857 // This should be corrected, given time. We do our null check with the
4858 // stack pointer restored.
4859 src = null_check(src, T_ARRAY);
4860 dest = null_check(dest, T_ARRAY);
4862 // (4) src_offset must not be negative.
4863 generate_negative_guard(src_offset, slow_region);
4865 // (5) dest_offset must not be negative.
4866 generate_negative_guard(dest_offset, slow_region);
4868 // (6) length must not be negative (moved to generate_arraycopy()).
4869 // generate_negative_guard(length, slow_region);
4871 // (7) src_offset + length must not exceed length of src.
4872 generate_limit_guard(src_offset, length,
4873 load_array_length(src),
4874 slow_region);
4876 // (8) dest_offset + length must not exceed length of dest.
4877 generate_limit_guard(dest_offset, length,
4878 load_array_length(dest),
4879 slow_region);
4881 // (9) each element of an oop array must be assignable
4882 // The generate_arraycopy subroutine checks this.
4884 // This is where the memory effects are placed:
4885 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4886 generate_arraycopy(adr_type, dest_elem,
4887 src, src_offset, dest, dest_offset, length,
4888 false, false, slow_region);
4890 return true;
4891 }
4893 //-----------------------------generate_arraycopy----------------------
4894 // Generate an optimized call to arraycopy.
4895 // Caller must guard against non-arrays.
4896 // Caller must determine a common array basic-type for both arrays.
4897 // Caller must validate offsets against array bounds.
4898 // The slow_region has already collected guard failure paths
4899 // (such as out of bounds length or non-conformable array types).
4900 // The generated code has this shape, in general:
4901 //
4902 // if (length == 0) return // via zero_path
4903 // slowval = -1
4904 // if (types unknown) {
4905 // slowval = call generic copy loop
4906 // if (slowval == 0) return // via checked_path
4907 // } else if (indexes in bounds) {
4908 // if ((is object array) && !(array type check)) {
4909 // slowval = call checked copy loop
4910 // if (slowval == 0) return // via checked_path
4911 // } else {
4912 // call bulk copy loop
4913 // return // via fast_path
4914 // }
4915 // }
4916 // // adjust params for remaining work:
4917 // if (slowval != -1) {
4918 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4919 // }
4920 // slow_region:
4921 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4922 // return // via slow_call_path
4923 //
4924 // This routine is used from several intrinsics: System.arraycopy,
4925 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4926 //
4927 void
4928 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4929 BasicType basic_elem_type,
4930 Node* src, Node* src_offset,
4931 Node* dest, Node* dest_offset,
4932 Node* copy_length,
4933 bool disjoint_bases,
4934 bool length_never_negative,
4935 RegionNode* slow_region) {
4937 if (slow_region == NULL) {
4938 slow_region = new(C) RegionNode(1);
4939 record_for_igvn(slow_region);
4940 }
4942 Node* original_dest = dest;
4943 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4944 bool dest_uninitialized = false;
4946 // See if this is the initialization of a newly-allocated array.
4947 // If so, we will take responsibility here for initializing it to zero.
4948 // (Note: Because tightly_coupled_allocation performs checks on the
4949 // out-edges of the dest, we need to avoid making derived pointers
4950 // from it until we have checked its uses.)
4951 if (ReduceBulkZeroing
4952 && !ZeroTLAB // pointless if already zeroed
4953 && basic_elem_type != T_CONFLICT // avoid corner case
4954 && !src->eqv_uncast(dest)
4955 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4956 != NULL)
4957 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4958 && alloc->maybe_set_complete(&_gvn)) {
4959 // "You break it, you buy it."
4960 InitializeNode* init = alloc->initialization();
4961 assert(init->is_complete(), "we just did this");
4962 init->set_complete_with_arraycopy();
4963 assert(dest->is_CheckCastPP(), "sanity");
4964 assert(dest->in(0)->in(0) == init, "dest pinned");
4965 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4966 // From this point on, every exit path is responsible for
4967 // initializing any non-copied parts of the object to zero.
4968 // Also, if this flag is set we make sure that arraycopy interacts properly
4969 // with G1, eliding pre-barriers. See CR 6627983.
4970 dest_uninitialized = true;
4971 } else {
4972 // No zeroing elimination here.
4973 alloc = NULL;
4974 //original_dest = dest;
4975 //dest_uninitialized = false;
4976 }
4978 // Results are placed here:
4979 enum { fast_path = 1, // normal void-returning assembly stub
4980 checked_path = 2, // special assembly stub with cleanup
4981 slow_call_path = 3, // something went wrong; call the VM
4982 zero_path = 4, // bypass when length of copy is zero
4983 bcopy_path = 5, // copy primitive array by 64-bit blocks
4984 PATH_LIMIT = 6
4985 };
4986 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4987 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4988 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4989 record_for_igvn(result_region);
4990 _gvn.set_type_bottom(result_i_o);
4991 _gvn.set_type_bottom(result_memory);
4992 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4994 // The slow_control path:
4995 Node* slow_control;
4996 Node* slow_i_o = i_o();
4997 Node* slow_mem = memory(adr_type);
4998 debug_only(slow_control = (Node*) badAddress);
5000 // Checked control path:
5001 Node* checked_control = top();
5002 Node* checked_mem = NULL;
5003 Node* checked_i_o = NULL;
5004 Node* checked_value = NULL;
5006 if (basic_elem_type == T_CONFLICT) {
5007 assert(!dest_uninitialized, "");
5008 Node* cv = generate_generic_arraycopy(adr_type,
5009 src, src_offset, dest, dest_offset,
5010 copy_length, dest_uninitialized);
5011 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5012 checked_control = control();
5013 checked_i_o = i_o();
5014 checked_mem = memory(adr_type);
5015 checked_value = cv;
5016 set_control(top()); // no fast path
5017 }
5019 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5020 if (not_pos != NULL) {
5021 PreserveJVMState pjvms(this);
5022 set_control(not_pos);
5024 // (6) length must not be negative.
5025 if (!length_never_negative) {
5026 generate_negative_guard(copy_length, slow_region);
5027 }
5029 // copy_length is 0.
5030 if (!stopped() && dest_uninitialized) {
5031 Node* dest_length = alloc->in(AllocateNode::ALength);
5032 if (copy_length->eqv_uncast(dest_length)
5033 || _gvn.find_int_con(dest_length, 1) <= 0) {
5034 // There is no zeroing to do. No need for a secondary raw memory barrier.
5035 } else {
5036 // Clear the whole thing since there are no source elements to copy.
5037 generate_clear_array(adr_type, dest, basic_elem_type,
5038 intcon(0), NULL,
5039 alloc->in(AllocateNode::AllocSize));
5040 // Use a secondary InitializeNode as raw memory barrier.
5041 // Currently it is needed only on this path since other
5042 // paths have stub or runtime calls as raw memory barriers.
5043 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5044 Compile::AliasIdxRaw,
5045 top())->as_Initialize();
5046 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5047 }
5048 }
5050 // Present the results of the fast call.
5051 result_region->init_req(zero_path, control());
5052 result_i_o ->init_req(zero_path, i_o());
5053 result_memory->init_req(zero_path, memory(adr_type));
5054 }
5056 if (!stopped() && dest_uninitialized) {
5057 // We have to initialize the *uncopied* part of the array to zero.
5058 // The copy destination is the slice dest[off..off+len]. The other slices
5059 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5060 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5061 Node* dest_length = alloc->in(AllocateNode::ALength);
5062 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5063 copy_length));
5065 // If there is a head section that needs zeroing, do it now.
5066 if (find_int_con(dest_offset, -1) != 0) {
5067 generate_clear_array(adr_type, dest, basic_elem_type,
5068 intcon(0), dest_offset,
5069 NULL);
5070 }
5072 // Next, perform a dynamic check on the tail length.
5073 // It is often zero, and we can win big if we prove this.
5074 // There are two wins: Avoid generating the ClearArray
5075 // with its attendant messy index arithmetic, and upgrade
5076 // the copy to a more hardware-friendly word size of 64 bits.
5077 Node* tail_ctl = NULL;
5078 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5079 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5080 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5081 tail_ctl = generate_slow_guard(bol_lt, NULL);
5082 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5083 }
5085 // At this point, let's assume there is no tail.
5086 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5087 // There is no tail. Try an upgrade to a 64-bit copy.
5088 bool didit = false;
5089 { PreserveJVMState pjvms(this);
5090 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5091 src, src_offset, dest, dest_offset,
5092 dest_size, dest_uninitialized);
5093 if (didit) {
5094 // Present the results of the block-copying fast call.
5095 result_region->init_req(bcopy_path, control());
5096 result_i_o ->init_req(bcopy_path, i_o());
5097 result_memory->init_req(bcopy_path, memory(adr_type));
5098 }
5099 }
5100 if (didit)
5101 set_control(top()); // no regular fast path
5102 }
5104 // Clear the tail, if any.
5105 if (tail_ctl != NULL) {
5106 Node* notail_ctl = stopped() ? NULL : control();
5107 set_control(tail_ctl);
5108 if (notail_ctl == NULL) {
5109 generate_clear_array(adr_type, dest, basic_elem_type,
5110 dest_tail, NULL,
5111 dest_size);
5112 } else {
5113 // Make a local merge.
5114 Node* done_ctl = new(C) RegionNode(3);
5115 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5116 done_ctl->init_req(1, notail_ctl);
5117 done_mem->init_req(1, memory(adr_type));
5118 generate_clear_array(adr_type, dest, basic_elem_type,
5119 dest_tail, NULL,
5120 dest_size);
5121 done_ctl->init_req(2, control());
5122 done_mem->init_req(2, memory(adr_type));
5123 set_control( _gvn.transform(done_ctl));
5124 set_memory( _gvn.transform(done_mem), adr_type );
5125 }
5126 }
5127 }
5129 BasicType copy_type = basic_elem_type;
5130 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5131 if (!stopped() && copy_type == T_OBJECT) {
5132 // If src and dest have compatible element types, we can copy bits.
5133 // Types S[] and D[] are compatible if D is a supertype of S.
5134 //
5135 // If they are not, we will use checked_oop_disjoint_arraycopy,
5136 // which performs a fast optimistic per-oop check, and backs off
5137 // further to JVM_ArrayCopy on the first per-oop check that fails.
5138 // (Actually, we don't move raw bits only; the GC requires card marks.)
5140 // Get the Klass* for both src and dest
5141 Node* src_klass = load_object_klass(src);
5142 Node* dest_klass = load_object_klass(dest);
5144 // Generate the subtype check.
5145 // This might fold up statically, or then again it might not.
5146 //
5147 // Non-static example: Copying List<String>.elements to a new String[].
5148 // The backing store for a List<String> is always an Object[],
5149 // but its elements are always type String, if the generic types
5150 // are correct at the source level.
5151 //
5152 // Test S[] against D[], not S against D, because (probably)
5153 // the secondary supertype cache is less busy for S[] than S.
5154 // This usually only matters when D is an interface.
5155 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5156 // Plug failing path into checked_oop_disjoint_arraycopy
5157 if (not_subtype_ctrl != top()) {
5158 PreserveJVMState pjvms(this);
5159 set_control(not_subtype_ctrl);
5160 // (At this point we can assume disjoint_bases, since types differ.)
5161 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5162 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5163 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5164 Node* dest_elem_klass = _gvn.transform(n1);
5165 Node* cv = generate_checkcast_arraycopy(adr_type,
5166 dest_elem_klass,
5167 src, src_offset, dest, dest_offset,
5168 ConvI2X(copy_length), dest_uninitialized);
5169 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5170 checked_control = control();
5171 checked_i_o = i_o();
5172 checked_mem = memory(adr_type);
5173 checked_value = cv;
5174 }
5175 // At this point we know we do not need type checks on oop stores.
5177 // Let's see if we need card marks:
5178 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5179 // If we do not need card marks, copy using the jint or jlong stub.
5180 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5181 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5182 "sizes agree");
5183 }
5184 }
5186 if (!stopped()) {
5187 // Generate the fast path, if possible.
5188 PreserveJVMState pjvms(this);
5189 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5190 src, src_offset, dest, dest_offset,
5191 ConvI2X(copy_length), dest_uninitialized);
5193 // Present the results of the fast call.
5194 result_region->init_req(fast_path, control());
5195 result_i_o ->init_req(fast_path, i_o());
5196 result_memory->init_req(fast_path, memory(adr_type));
5197 }
5199 // Here are all the slow paths up to this point, in one bundle:
5200 slow_control = top();
5201 if (slow_region != NULL)
5202 slow_control = _gvn.transform(slow_region);
5203 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5205 set_control(checked_control);
5206 if (!stopped()) {
5207 // Clean up after the checked call.
5208 // The returned value is either 0 or -1^K,
5209 // where K = number of partially transferred array elements.
5210 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5211 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5212 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5214 // If it is 0, we are done, so transfer to the end.
5215 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5216 result_region->init_req(checked_path, checks_done);
5217 result_i_o ->init_req(checked_path, checked_i_o);
5218 result_memory->init_req(checked_path, checked_mem);
5220 // If it is not zero, merge into the slow call.
5221 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5222 RegionNode* slow_reg2 = new(C) RegionNode(3);
5223 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5224 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5225 record_for_igvn(slow_reg2);
5226 slow_reg2 ->init_req(1, slow_control);
5227 slow_i_o2 ->init_req(1, slow_i_o);
5228 slow_mem2 ->init_req(1, slow_mem);
5229 slow_reg2 ->init_req(2, control());
5230 slow_i_o2 ->init_req(2, checked_i_o);
5231 slow_mem2 ->init_req(2, checked_mem);
5233 slow_control = _gvn.transform(slow_reg2);
5234 slow_i_o = _gvn.transform(slow_i_o2);
5235 slow_mem = _gvn.transform(slow_mem2);
5237 if (alloc != NULL) {
5238 // We'll restart from the very beginning, after zeroing the whole thing.
5239 // This can cause double writes, but that's OK since dest is brand new.
5240 // So we ignore the low 31 bits of the value returned from the stub.
5241 } else {
5242 // We must continue the copy exactly where it failed, or else
5243 // another thread might see the wrong number of writes to dest.
5244 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5245 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5246 slow_offset->init_req(1, intcon(0));
5247 slow_offset->init_req(2, checked_offset);
5248 slow_offset = _gvn.transform(slow_offset);
5250 // Adjust the arguments by the conditionally incoming offset.
5251 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5252 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5253 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5255 // Tweak the node variables to adjust the code produced below:
5256 src_offset = src_off_plus;
5257 dest_offset = dest_off_plus;
5258 copy_length = length_minus;
5259 }
5260 }
5262 set_control(slow_control);
5263 if (!stopped()) {
5264 // Generate the slow path, if needed.
5265 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5267 set_memory(slow_mem, adr_type);
5268 set_i_o(slow_i_o);
5270 if (dest_uninitialized) {
5271 generate_clear_array(adr_type, dest, basic_elem_type,
5272 intcon(0), NULL,
5273 alloc->in(AllocateNode::AllocSize));
5274 }
5276 generate_slow_arraycopy(adr_type,
5277 src, src_offset, dest, dest_offset,
5278 copy_length, /*dest_uninitialized*/false);
5280 result_region->init_req(slow_call_path, control());
5281 result_i_o ->init_req(slow_call_path, i_o());
5282 result_memory->init_req(slow_call_path, memory(adr_type));
5283 }
5285 // Remove unused edges.
5286 for (uint i = 1; i < result_region->req(); i++) {
5287 if (result_region->in(i) == NULL)
5288 result_region->init_req(i, top());
5289 }
5291 // Finished; return the combined state.
5292 set_control( _gvn.transform(result_region));
5293 set_i_o( _gvn.transform(result_i_o) );
5294 set_memory( _gvn.transform(result_memory), adr_type );
5296 // The memory edges above are precise in order to model effects around
5297 // array copies accurately to allow value numbering of field loads around
5298 // arraycopy. Such field loads, both before and after, are common in Java
5299 // collections and similar classes involving header/array data structures.
5300 //
5301 // But with low number of register or when some registers are used or killed
5302 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5303 // The next memory barrier is added to avoid it. If the arraycopy can be
5304 // optimized away (which it can, sometimes) then we can manually remove
5305 // the membar also.
5306 //
5307 // Do not let reads from the cloned object float above the arraycopy.
5308 if (alloc != NULL) {
5309 // Do not let stores that initialize this object be reordered with
5310 // a subsequent store that would make this object accessible by
5311 // other threads.
5312 // Record what AllocateNode this StoreStore protects so that
5313 // escape analysis can go from the MemBarStoreStoreNode to the
5314 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5315 // based on the escape status of the AllocateNode.
5316 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5317 } else if (InsertMemBarAfterArraycopy)
5318 insert_mem_bar(Op_MemBarCPUOrder);
5319 }
5322 // Helper function which determines if an arraycopy immediately follows
5323 // an allocation, with no intervening tests or other escapes for the object.
5324 AllocateArrayNode*
5325 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5326 RegionNode* slow_region) {
5327 if (stopped()) return NULL; // no fast path
5328 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5330 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5331 if (alloc == NULL) return NULL;
5333 Node* rawmem = memory(Compile::AliasIdxRaw);
5334 // Is the allocation's memory state untouched?
5335 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5336 // Bail out if there have been raw-memory effects since the allocation.
5337 // (Example: There might have been a call or safepoint.)
5338 return NULL;
5339 }
5340 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5341 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5342 return NULL;
5343 }
5345 // There must be no unexpected observers of this allocation.
5346 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5347 Node* obs = ptr->fast_out(i);
5348 if (obs != this->map()) {
5349 return NULL;
5350 }
5351 }
5353 // This arraycopy must unconditionally follow the allocation of the ptr.
5354 Node* alloc_ctl = ptr->in(0);
5355 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5357 Node* ctl = control();
5358 while (ctl != alloc_ctl) {
5359 // There may be guards which feed into the slow_region.
5360 // Any other control flow means that we might not get a chance
5361 // to finish initializing the allocated object.
5362 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5363 IfNode* iff = ctl->in(0)->as_If();
5364 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5365 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5366 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5367 ctl = iff->in(0); // This test feeds the known slow_region.
5368 continue;
5369 }
5370 // One more try: Various low-level checks bottom out in
5371 // uncommon traps. If the debug-info of the trap omits
5372 // any reference to the allocation, as we've already
5373 // observed, then there can be no objection to the trap.
5374 bool found_trap = false;
5375 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5376 Node* obs = not_ctl->fast_out(j);
5377 if (obs->in(0) == not_ctl && obs->is_Call() &&
5378 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5379 found_trap = true; break;
5380 }
5381 }
5382 if (found_trap) {
5383 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5384 continue;
5385 }
5386 }
5387 return NULL;
5388 }
5390 // If we get this far, we have an allocation which immediately
5391 // precedes the arraycopy, and we can take over zeroing the new object.
5392 // The arraycopy will finish the initialization, and provide
5393 // a new control state to which we will anchor the destination pointer.
5395 return alloc;
5396 }
5398 // Helper for initialization of arrays, creating a ClearArray.
5399 // It writes zero bits in [start..end), within the body of an array object.
5400 // The memory effects are all chained onto the 'adr_type' alias category.
5401 //
5402 // Since the object is otherwise uninitialized, we are free
5403 // to put a little "slop" around the edges of the cleared area,
5404 // as long as it does not go back into the array's header,
5405 // or beyond the array end within the heap.
5406 //
5407 // The lower edge can be rounded down to the nearest jint and the
5408 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5409 //
5410 // Arguments:
5411 // adr_type memory slice where writes are generated
5412 // dest oop of the destination array
5413 // basic_elem_type element type of the destination
5414 // slice_idx array index of first element to store
5415 // slice_len number of elements to store (or NULL)
5416 // dest_size total size in bytes of the array object
5417 //
5418 // Exactly one of slice_len or dest_size must be non-NULL.
5419 // If dest_size is non-NULL, zeroing extends to the end of the object.
5420 // If slice_len is non-NULL, the slice_idx value must be a constant.
5421 void
5422 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5423 Node* dest,
5424 BasicType basic_elem_type,
5425 Node* slice_idx,
5426 Node* slice_len,
5427 Node* dest_size) {
5428 // one or the other but not both of slice_len and dest_size:
5429 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5430 if (slice_len == NULL) slice_len = top();
5431 if (dest_size == NULL) dest_size = top();
5433 // operate on this memory slice:
5434 Node* mem = memory(adr_type); // memory slice to operate on
5436 // scaling and rounding of indexes:
5437 int scale = exact_log2(type2aelembytes(basic_elem_type));
5438 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5439 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5440 int bump_bit = (-1 << scale) & BytesPerInt;
5442 // determine constant starts and ends
5443 const intptr_t BIG_NEG = -128;
5444 assert(BIG_NEG + 2*abase < 0, "neg enough");
5445 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5446 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5447 if (slice_len_con == 0) {
5448 return; // nothing to do here
5449 }
5450 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5451 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5452 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5453 assert(end_con < 0, "not two cons");
5454 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5455 BytesPerLong);
5456 }
5458 if (start_con >= 0 && end_con >= 0) {
5459 // Constant start and end. Simple.
5460 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5461 start_con, end_con, &_gvn);
5462 } else if (start_con >= 0 && dest_size != top()) {
5463 // Constant start, pre-rounded end after the tail of the array.
5464 Node* end = dest_size;
5465 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5466 start_con, end, &_gvn);
5467 } else if (start_con >= 0 && slice_len != top()) {
5468 // Constant start, non-constant end. End needs rounding up.
5469 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5470 intptr_t end_base = abase + (slice_idx_con << scale);
5471 int end_round = (-1 << scale) & (BytesPerLong - 1);
5472 Node* end = ConvI2X(slice_len);
5473 if (scale != 0)
5474 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5475 end_base += end_round;
5476 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5477 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5478 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5479 start_con, end, &_gvn);
5480 } else if (start_con < 0 && dest_size != top()) {
5481 // Non-constant start, pre-rounded end after the tail of the array.
5482 // This is almost certainly a "round-to-end" operation.
5483 Node* start = slice_idx;
5484 start = ConvI2X(start);
5485 if (scale != 0)
5486 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5487 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5488 if ((bump_bit | clear_low) != 0) {
5489 int to_clear = (bump_bit | clear_low);
5490 // Align up mod 8, then store a jint zero unconditionally
5491 // just before the mod-8 boundary.
5492 if (((abase + bump_bit) & ~to_clear) - bump_bit
5493 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5494 bump_bit = 0;
5495 assert((abase & to_clear) == 0, "array base must be long-aligned");
5496 } else {
5497 // Bump 'start' up to (or past) the next jint boundary:
5498 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5499 assert((abase & clear_low) == 0, "array base must be int-aligned");
5500 }
5501 // Round bumped 'start' down to jlong boundary in body of array.
5502 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5503 if (bump_bit != 0) {
5504 // Store a zero to the immediately preceding jint:
5505 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5506 Node* p1 = basic_plus_adr(dest, x1);
5507 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5508 mem = _gvn.transform(mem);
5509 }
5510 }
5511 Node* end = dest_size; // pre-rounded
5512 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5513 start, end, &_gvn);
5514 } else {
5515 // Non-constant start, unrounded non-constant end.
5516 // (Nobody zeroes a random midsection of an array using this routine.)
5517 ShouldNotReachHere(); // fix caller
5518 }
5520 // Done.
5521 set_memory(mem, adr_type);
5522 }
5525 bool
5526 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5527 BasicType basic_elem_type,
5528 AllocateNode* alloc,
5529 Node* src, Node* src_offset,
5530 Node* dest, Node* dest_offset,
5531 Node* dest_size, bool dest_uninitialized) {
5532 // See if there is an advantage from block transfer.
5533 int scale = exact_log2(type2aelembytes(basic_elem_type));
5534 if (scale >= LogBytesPerLong)
5535 return false; // it is already a block transfer
5537 // Look at the alignment of the starting offsets.
5538 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5540 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5541 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5542 if (src_off_con < 0 || dest_off_con < 0)
5543 // At present, we can only understand constants.
5544 return false;
5546 intptr_t src_off = abase + (src_off_con << scale);
5547 intptr_t dest_off = abase + (dest_off_con << scale);
5549 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5550 // Non-aligned; too bad.
5551 // One more chance: Pick off an initial 32-bit word.
5552 // This is a common case, since abase can be odd mod 8.
5553 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5554 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5555 Node* sptr = basic_plus_adr(src, src_off);
5556 Node* dptr = basic_plus_adr(dest, dest_off);
5557 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5558 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5559 src_off += BytesPerInt;
5560 dest_off += BytesPerInt;
5561 } else {
5562 return false;
5563 }
5564 }
5565 assert(src_off % BytesPerLong == 0, "");
5566 assert(dest_off % BytesPerLong == 0, "");
5568 // Do this copy by giant steps.
5569 Node* sptr = basic_plus_adr(src, src_off);
5570 Node* dptr = basic_plus_adr(dest, dest_off);
5571 Node* countx = dest_size;
5572 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5573 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5575 bool disjoint_bases = true; // since alloc != NULL
5576 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5577 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5579 return true;
5580 }
5583 // Helper function; generates code for the slow case.
5584 // We make a call to a runtime method which emulates the native method,
5585 // but without the native wrapper overhead.
5586 void
5587 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5588 Node* src, Node* src_offset,
5589 Node* dest, Node* dest_offset,
5590 Node* copy_length, bool dest_uninitialized) {
5591 assert(!dest_uninitialized, "Invariant");
5592 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5593 OptoRuntime::slow_arraycopy_Type(),
5594 OptoRuntime::slow_arraycopy_Java(),
5595 "slow_arraycopy", adr_type,
5596 src, src_offset, dest, dest_offset,
5597 copy_length);
5599 // Handle exceptions thrown by this fellow:
5600 make_slow_call_ex(call, env()->Throwable_klass(), false);
5601 }
5603 // Helper function; generates code for cases requiring runtime checks.
5604 Node*
5605 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5606 Node* dest_elem_klass,
5607 Node* src, Node* src_offset,
5608 Node* dest, Node* dest_offset,
5609 Node* copy_length, bool dest_uninitialized) {
5610 if (stopped()) return NULL;
5612 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5613 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5614 return NULL;
5615 }
5617 // Pick out the parameters required to perform a store-check
5618 // for the target array. This is an optimistic check. It will
5619 // look in each non-null element's class, at the desired klass's
5620 // super_check_offset, for the desired klass.
5621 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5622 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5623 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5624 Node* check_offset = ConvI2X(_gvn.transform(n3));
5625 Node* check_value = dest_elem_klass;
5627 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5628 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5630 // (We know the arrays are never conjoint, because their types differ.)
5631 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5632 OptoRuntime::checkcast_arraycopy_Type(),
5633 copyfunc_addr, "checkcast_arraycopy", adr_type,
5634 // five arguments, of which two are
5635 // intptr_t (jlong in LP64)
5636 src_start, dest_start,
5637 copy_length XTOP,
5638 check_offset XTOP,
5639 check_value);
5641 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5642 }
5645 // Helper function; generates code for cases requiring runtime checks.
5646 Node*
5647 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5648 Node* src, Node* src_offset,
5649 Node* dest, Node* dest_offset,
5650 Node* copy_length, bool dest_uninitialized) {
5651 assert(!dest_uninitialized, "Invariant");
5652 if (stopped()) return NULL;
5653 address copyfunc_addr = StubRoutines::generic_arraycopy();
5654 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5655 return NULL;
5656 }
5658 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5659 OptoRuntime::generic_arraycopy_Type(),
5660 copyfunc_addr, "generic_arraycopy", adr_type,
5661 src, src_offset, dest, dest_offset, copy_length);
5663 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5664 }
5666 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5667 void
5668 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5669 BasicType basic_elem_type,
5670 bool disjoint_bases,
5671 Node* src, Node* src_offset,
5672 Node* dest, Node* dest_offset,
5673 Node* copy_length, bool dest_uninitialized) {
5674 if (stopped()) return; // nothing to do
5676 Node* src_start = src;
5677 Node* dest_start = dest;
5678 if (src_offset != NULL || dest_offset != NULL) {
5679 assert(src_offset != NULL && dest_offset != NULL, "");
5680 src_start = array_element_address(src, src_offset, basic_elem_type);
5681 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5682 }
5684 // Figure out which arraycopy runtime method to call.
5685 const char* copyfunc_name = "arraycopy";
5686 address copyfunc_addr =
5687 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5688 disjoint_bases, copyfunc_name, dest_uninitialized);
5690 // Call it. Note that the count_ix value is not scaled to a byte-size.
5691 make_runtime_call(RC_LEAF|RC_NO_FP,
5692 OptoRuntime::fast_arraycopy_Type(),
5693 copyfunc_addr, copyfunc_name, adr_type,
5694 src_start, dest_start, copy_length XTOP);
5695 }
5697 //-------------inline_encodeISOArray-----------------------------------
5698 // encode char[] to byte[] in ISO_8859_1
5699 bool LibraryCallKit::inline_encodeISOArray() {
5700 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5701 // no receiver since it is static method
5702 Node *src = argument(0);
5703 Node *src_offset = argument(1);
5704 Node *dst = argument(2);
5705 Node *dst_offset = argument(3);
5706 Node *length = argument(4);
5708 const Type* src_type = src->Value(&_gvn);
5709 const Type* dst_type = dst->Value(&_gvn);
5710 const TypeAryPtr* top_src = src_type->isa_aryptr();
5711 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5712 if (top_src == NULL || top_src->klass() == NULL ||
5713 top_dest == NULL || top_dest->klass() == NULL) {
5714 // failed array check
5715 return false;
5716 }
5718 // Figure out the size and type of the elements we will be copying.
5719 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5720 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5721 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5722 return false;
5723 }
5724 Node* src_start = array_element_address(src, src_offset, src_elem);
5725 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5726 // 'src_start' points to src array + scaled offset
5727 // 'dst_start' points to dst array + scaled offset
5729 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5730 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5731 enc = _gvn.transform(enc);
5732 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5733 set_memory(res_mem, mtype);
5734 set_result(enc);
5735 return true;
5736 }
5738 /**
5739 * Calculate CRC32 for byte.
5740 * int java.util.zip.CRC32.update(int crc, int b)
5741 */
5742 bool LibraryCallKit::inline_updateCRC32() {
5743 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5744 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5745 // no receiver since it is static method
5746 Node* crc = argument(0); // type: int
5747 Node* b = argument(1); // type: int
5749 /*
5750 * int c = ~ crc;
5751 * b = timesXtoThe32[(b ^ c) & 0xFF];
5752 * b = b ^ (c >>> 8);
5753 * crc = ~b;
5754 */
5756 Node* M1 = intcon(-1);
5757 crc = _gvn.transform(new (C) XorINode(crc, M1));
5758 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5759 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5761 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5762 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5763 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5764 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5766 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5767 result = _gvn.transform(new (C) XorINode(crc, result));
5768 result = _gvn.transform(new (C) XorINode(result, M1));
5769 set_result(result);
5770 return true;
5771 }
5773 /**
5774 * Calculate CRC32 for byte[] array.
5775 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5776 */
5777 bool LibraryCallKit::inline_updateBytesCRC32() {
5778 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5779 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5780 // no receiver since it is static method
5781 Node* crc = argument(0); // type: int
5782 Node* src = argument(1); // type: oop
5783 Node* offset = argument(2); // type: int
5784 Node* length = argument(3); // type: int
5786 const Type* src_type = src->Value(&_gvn);
5787 const TypeAryPtr* top_src = src_type->isa_aryptr();
5788 if (top_src == NULL || top_src->klass() == NULL) {
5789 // failed array check
5790 return false;
5791 }
5793 // Figure out the size and type of the elements we will be copying.
5794 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5795 if (src_elem != T_BYTE) {
5796 return false;
5797 }
5799 // 'src_start' points to src array + scaled offset
5800 Node* src_start = array_element_address(src, offset, src_elem);
5802 // We assume that range check is done by caller.
5803 // TODO: generate range check (offset+length < src.length) in debug VM.
5805 // Call the stub.
5806 address stubAddr = StubRoutines::updateBytesCRC32();
5807 const char *stubName = "updateBytesCRC32";
5809 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5810 stubAddr, stubName, TypePtr::BOTTOM,
5811 crc, src_start, length);
5812 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5813 set_result(result);
5814 return true;
5815 }
5817 /**
5818 * Calculate CRC32 for ByteBuffer.
5819 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5820 */
5821 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5822 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5823 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5824 // no receiver since it is static method
5825 Node* crc = argument(0); // type: int
5826 Node* src = argument(1); // type: long
5827 Node* offset = argument(3); // type: int
5828 Node* length = argument(4); // type: int
5830 src = ConvL2X(src); // adjust Java long to machine word
5831 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5832 offset = ConvI2X(offset);
5834 // 'src_start' points to src array + scaled offset
5835 Node* src_start = basic_plus_adr(top(), base, offset);
5837 // Call the stub.
5838 address stubAddr = StubRoutines::updateBytesCRC32();
5839 const char *stubName = "updateBytesCRC32";
5841 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5842 stubAddr, stubName, TypePtr::BOTTOM,
5843 crc, src_start, length);
5844 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5845 set_result(result);
5846 return true;
5847 }
5849 //----------------------------inline_reference_get----------------------------
5850 // public T java.lang.ref.Reference.get();
5851 bool LibraryCallKit::inline_reference_get() {
5852 const int referent_offset = java_lang_ref_Reference::referent_offset;
5853 guarantee(referent_offset > 0, "should have already been set");
5855 // Get the argument:
5856 Node* reference_obj = null_check_receiver();
5857 if (stopped()) return true;
5859 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5861 ciInstanceKlass* klass = env()->Object_klass();
5862 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5864 Node* no_ctrl = NULL;
5865 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5867 // Use the pre-barrier to record the value in the referent field
5868 pre_barrier(false /* do_load */,
5869 control(),
5870 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5871 result /* pre_val */,
5872 T_OBJECT);
5874 // Add memory barrier to prevent commoning reads from this field
5875 // across safepoint since GC can change its value.
5876 insert_mem_bar(Op_MemBarCPUOrder);
5878 set_result(result);
5879 return true;
5880 }
5883 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5884 bool is_exact=true, bool is_static=false) {
5886 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5887 assert(tinst != NULL, "obj is null");
5888 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5889 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5891 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5892 ciSymbol::make(fieldTypeString),
5893 is_static);
5894 if (field == NULL) return (Node *) NULL;
5895 assert (field != NULL, "undefined field");
5897 // Next code copied from Parse::do_get_xxx():
5899 // Compute address and memory type.
5900 int offset = field->offset_in_bytes();
5901 bool is_vol = field->is_volatile();
5902 ciType* field_klass = field->type();
5903 assert(field_klass->is_loaded(), "should be loaded");
5904 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5905 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5906 BasicType bt = field->layout_type();
5908 // Build the resultant type of the load
5909 const Type *type;
5910 if (bt == T_OBJECT) {
5911 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5912 } else {
5913 type = Type::get_const_basic_type(bt);
5914 }
5916 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5917 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5918 }
5919 // Build the load.
5920 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5921 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
5922 // If reference is volatile, prevent following memory ops from
5923 // floating up past the volatile read. Also prevents commoning
5924 // another volatile read.
5925 if (is_vol) {
5926 // Memory barrier includes bogus read of value to force load BEFORE membar
5927 insert_mem_bar(Op_MemBarAcquire, loadedField);
5928 }
5929 return loadedField;
5930 }
5933 //------------------------------inline_aescrypt_Block-----------------------
5934 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5935 address stubAddr;
5936 const char *stubName;
5937 assert(UseAES, "need AES instruction support");
5939 switch(id) {
5940 case vmIntrinsics::_aescrypt_encryptBlock:
5941 stubAddr = StubRoutines::aescrypt_encryptBlock();
5942 stubName = "aescrypt_encryptBlock";
5943 break;
5944 case vmIntrinsics::_aescrypt_decryptBlock:
5945 stubAddr = StubRoutines::aescrypt_decryptBlock();
5946 stubName = "aescrypt_decryptBlock";
5947 break;
5948 }
5949 if (stubAddr == NULL) return false;
5951 Node* aescrypt_object = argument(0);
5952 Node* src = argument(1);
5953 Node* src_offset = argument(2);
5954 Node* dest = argument(3);
5955 Node* dest_offset = argument(4);
5957 // (1) src and dest are arrays.
5958 const Type* src_type = src->Value(&_gvn);
5959 const Type* dest_type = dest->Value(&_gvn);
5960 const TypeAryPtr* top_src = src_type->isa_aryptr();
5961 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5962 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5964 // for the quick and dirty code we will skip all the checks.
5965 // we are just trying to get the call to be generated.
5966 Node* src_start = src;
5967 Node* dest_start = dest;
5968 if (src_offset != NULL || dest_offset != NULL) {
5969 assert(src_offset != NULL && dest_offset != NULL, "");
5970 src_start = array_element_address(src, src_offset, T_BYTE);
5971 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5972 }
5974 // now need to get the start of its expanded key array
5975 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5976 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5977 if (k_start == NULL) return false;
5979 if (Matcher::pass_original_key_for_aes()) {
5980 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5981 // compatibility issues between Java key expansion and SPARC crypto instructions
5982 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5983 if (original_k_start == NULL) return false;
5985 // Call the stub.
5986 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5987 stubAddr, stubName, TypePtr::BOTTOM,
5988 src_start, dest_start, k_start, original_k_start);
5989 } else {
5990 // Call the stub.
5991 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5992 stubAddr, stubName, TypePtr::BOTTOM,
5993 src_start, dest_start, k_start);
5994 }
5996 return true;
5997 }
5999 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6000 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6001 address stubAddr;
6002 const char *stubName;
6004 assert(UseAES, "need AES instruction support");
6006 switch(id) {
6007 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6008 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6009 stubName = "cipherBlockChaining_encryptAESCrypt";
6010 break;
6011 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6012 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6013 stubName = "cipherBlockChaining_decryptAESCrypt";
6014 break;
6015 }
6016 if (stubAddr == NULL) return false;
6018 Node* cipherBlockChaining_object = argument(0);
6019 Node* src = argument(1);
6020 Node* src_offset = argument(2);
6021 Node* len = argument(3);
6022 Node* dest = argument(4);
6023 Node* dest_offset = argument(5);
6025 // (1) src and dest are arrays.
6026 const Type* src_type = src->Value(&_gvn);
6027 const Type* dest_type = dest->Value(&_gvn);
6028 const TypeAryPtr* top_src = src_type->isa_aryptr();
6029 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6030 assert (top_src != NULL && top_src->klass() != NULL
6031 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6033 // checks are the responsibility of the caller
6034 Node* src_start = src;
6035 Node* dest_start = dest;
6036 if (src_offset != NULL || dest_offset != NULL) {
6037 assert(src_offset != NULL && dest_offset != NULL, "");
6038 src_start = array_element_address(src, src_offset, T_BYTE);
6039 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6040 }
6042 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6043 // (because of the predicated logic executed earlier).
6044 // so we cast it here safely.
6045 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6047 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6048 if (embeddedCipherObj == NULL) return false;
6050 // cast it to what we know it will be at runtime
6051 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6052 assert(tinst != NULL, "CBC obj is null");
6053 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6054 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6055 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6057 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6058 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6059 const TypeOopPtr* xtype = aklass->as_instance_type();
6060 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6061 aescrypt_object = _gvn.transform(aescrypt_object);
6063 // we need to get the start of the aescrypt_object's expanded key array
6064 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6065 if (k_start == NULL) return false;
6067 // similarly, get the start address of the r vector
6068 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6069 if (objRvec == NULL) return false;
6070 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6072 Node* cbcCrypt;
6073 if (Matcher::pass_original_key_for_aes()) {
6074 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6075 // compatibility issues between Java key expansion and SPARC crypto instructions
6076 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6077 if (original_k_start == NULL) return false;
6079 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6080 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6081 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6082 stubAddr, stubName, TypePtr::BOTTOM,
6083 src_start, dest_start, k_start, r_start, len, original_k_start);
6084 } else {
6085 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6086 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6087 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6088 stubAddr, stubName, TypePtr::BOTTOM,
6089 src_start, dest_start, k_start, r_start, len);
6090 }
6092 // return cipher length (int)
6093 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6094 set_result(retvalue);
6095 return true;
6096 }
6098 //------------------------------get_key_start_from_aescrypt_object-----------------------
6099 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6100 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6101 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6102 if (objAESCryptKey == NULL) return (Node *) NULL;
6104 // now have the array, need to get the start address of the K array
6105 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6106 return k_start;
6107 }
6109 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6110 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6111 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6112 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6113 if (objAESCryptKey == NULL) return (Node *) NULL;
6115 // now have the array, need to get the start address of the lastKey array
6116 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6117 return original_k_start;
6118 }
6120 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6121 // Return node representing slow path of predicate check.
6122 // the pseudo code we want to emulate with this predicate is:
6123 // for encryption:
6124 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6125 // for decryption:
6126 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6127 // note cipher==plain is more conservative than the original java code but that's OK
6128 //
6129 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6130 // The receiver was checked for NULL already.
6131 Node* objCBC = argument(0);
6133 // Load embeddedCipher field of CipherBlockChaining object.
6134 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6136 // get AESCrypt klass for instanceOf check
6137 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6138 // will have same classloader as CipherBlockChaining object
6139 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6140 assert(tinst != NULL, "CBCobj is null");
6141 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6143 // we want to do an instanceof comparison against the AESCrypt class
6144 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6145 if (!klass_AESCrypt->is_loaded()) {
6146 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6147 Node* ctrl = control();
6148 set_control(top()); // no regular fast path
6149 return ctrl;
6150 }
6151 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6153 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6154 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6155 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6157 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6159 // for encryption, we are done
6160 if (!decrypting)
6161 return instof_false; // even if it is NULL
6163 // for decryption, we need to add a further check to avoid
6164 // taking the intrinsic path when cipher and plain are the same
6165 // see the original java code for why.
6166 RegionNode* region = new(C) RegionNode(3);
6167 region->init_req(1, instof_false);
6168 Node* src = argument(1);
6169 Node* dest = argument(4);
6170 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6171 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6172 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6173 region->init_req(2, src_dest_conjoint);
6175 record_for_igvn(region);
6176 return _gvn.transform(region);
6177 }
6179 //------------------------------inline_sha_implCompress-----------------------
6180 //
6181 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6182 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6183 //
6184 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6185 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6186 //
6187 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6188 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6189 //
6190 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6191 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6193 Node* sha_obj = argument(0);
6194 Node* src = argument(1); // type oop
6195 Node* ofs = argument(2); // type int
6197 const Type* src_type = src->Value(&_gvn);
6198 const TypeAryPtr* top_src = src_type->isa_aryptr();
6199 if (top_src == NULL || top_src->klass() == NULL) {
6200 // failed array check
6201 return false;
6202 }
6203 // Figure out the size and type of the elements we will be copying.
6204 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6205 if (src_elem != T_BYTE) {
6206 return false;
6207 }
6208 // 'src_start' points to src array + offset
6209 Node* src_start = array_element_address(src, ofs, src_elem);
6210 Node* state = NULL;
6211 address stubAddr;
6212 const char *stubName;
6214 switch(id) {
6215 case vmIntrinsics::_sha_implCompress:
6216 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6217 state = get_state_from_sha_object(sha_obj);
6218 stubAddr = StubRoutines::sha1_implCompress();
6219 stubName = "sha1_implCompress";
6220 break;
6221 case vmIntrinsics::_sha2_implCompress:
6222 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6223 state = get_state_from_sha_object(sha_obj);
6224 stubAddr = StubRoutines::sha256_implCompress();
6225 stubName = "sha256_implCompress";
6226 break;
6227 case vmIntrinsics::_sha5_implCompress:
6228 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6229 state = get_state_from_sha5_object(sha_obj);
6230 stubAddr = StubRoutines::sha512_implCompress();
6231 stubName = "sha512_implCompress";
6232 break;
6233 default:
6234 fatal_unexpected_iid(id);
6235 return false;
6236 }
6237 if (state == NULL) return false;
6239 // Call the stub.
6240 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6241 stubAddr, stubName, TypePtr::BOTTOM,
6242 src_start, state);
6244 return true;
6245 }
6247 //------------------------------inline_digestBase_implCompressMB-----------------------
6248 //
6249 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6250 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6251 //
6252 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6253 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6254 "need SHA1/SHA256/SHA512 instruction support");
6255 assert((uint)predicate < 3, "sanity");
6256 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6258 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6259 Node* src = argument(1); // byte[] array
6260 Node* ofs = argument(2); // type int
6261 Node* limit = argument(3); // type int
6263 const Type* src_type = src->Value(&_gvn);
6264 const TypeAryPtr* top_src = src_type->isa_aryptr();
6265 if (top_src == NULL || top_src->klass() == NULL) {
6266 // failed array check
6267 return false;
6268 }
6269 // Figure out the size and type of the elements we will be copying.
6270 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6271 if (src_elem != T_BYTE) {
6272 return false;
6273 }
6274 // 'src_start' points to src array + offset
6275 Node* src_start = array_element_address(src, ofs, src_elem);
6277 const char* klass_SHA_name = NULL;
6278 const char* stub_name = NULL;
6279 address stub_addr = NULL;
6280 bool long_state = false;
6282 switch (predicate) {
6283 case 0:
6284 if (UseSHA1Intrinsics) {
6285 klass_SHA_name = "sun/security/provider/SHA";
6286 stub_name = "sha1_implCompressMB";
6287 stub_addr = StubRoutines::sha1_implCompressMB();
6288 }
6289 break;
6290 case 1:
6291 if (UseSHA256Intrinsics) {
6292 klass_SHA_name = "sun/security/provider/SHA2";
6293 stub_name = "sha256_implCompressMB";
6294 stub_addr = StubRoutines::sha256_implCompressMB();
6295 }
6296 break;
6297 case 2:
6298 if (UseSHA512Intrinsics) {
6299 klass_SHA_name = "sun/security/provider/SHA5";
6300 stub_name = "sha512_implCompressMB";
6301 stub_addr = StubRoutines::sha512_implCompressMB();
6302 long_state = true;
6303 }
6304 break;
6305 default:
6306 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6307 }
6308 if (klass_SHA_name != NULL) {
6309 // get DigestBase klass to lookup for SHA klass
6310 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6311 assert(tinst != NULL, "digestBase_obj is not instance???");
6312 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6314 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6315 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6316 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6317 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6318 }
6319 return false;
6320 }
6321 //------------------------------inline_sha_implCompressMB-----------------------
6322 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6323 bool long_state, address stubAddr, const char *stubName,
6324 Node* src_start, Node* ofs, Node* limit) {
6325 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6326 const TypeOopPtr* xtype = aklass->as_instance_type();
6327 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6328 sha_obj = _gvn.transform(sha_obj);
6330 Node* state;
6331 if (long_state) {
6332 state = get_state_from_sha5_object(sha_obj);
6333 } else {
6334 state = get_state_from_sha_object(sha_obj);
6335 }
6336 if (state == NULL) return false;
6338 // Call the stub.
6339 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6340 OptoRuntime::digestBase_implCompressMB_Type(),
6341 stubAddr, stubName, TypePtr::BOTTOM,
6342 src_start, state, ofs, limit);
6343 // return ofs (int)
6344 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6345 set_result(result);
6347 return true;
6348 }
6350 //------------------------------get_state_from_sha_object-----------------------
6351 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6352 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6353 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6354 if (sha_state == NULL) return (Node *) NULL;
6356 // now have the array, need to get the start address of the state array
6357 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6358 return state;
6359 }
6361 //------------------------------get_state_from_sha5_object-----------------------
6362 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6363 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6364 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6365 if (sha_state == NULL) return (Node *) NULL;
6367 // now have the array, need to get the start address of the state array
6368 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6369 return state;
6370 }
6372 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6373 // Return node representing slow path of predicate check.
6374 // the pseudo code we want to emulate with this predicate is:
6375 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6376 //
6377 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6378 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6379 "need SHA1/SHA256/SHA512 instruction support");
6380 assert((uint)predicate < 3, "sanity");
6382 // The receiver was checked for NULL already.
6383 Node* digestBaseObj = argument(0);
6385 // get DigestBase klass for instanceOf check
6386 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6387 assert(tinst != NULL, "digestBaseObj is null");
6388 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6390 const char* klass_SHA_name = NULL;
6391 switch (predicate) {
6392 case 0:
6393 if (UseSHA1Intrinsics) {
6394 // we want to do an instanceof comparison against the SHA class
6395 klass_SHA_name = "sun/security/provider/SHA";
6396 }
6397 break;
6398 case 1:
6399 if (UseSHA256Intrinsics) {
6400 // we want to do an instanceof comparison against the SHA2 class
6401 klass_SHA_name = "sun/security/provider/SHA2";
6402 }
6403 break;
6404 case 2:
6405 if (UseSHA512Intrinsics) {
6406 // we want to do an instanceof comparison against the SHA5 class
6407 klass_SHA_name = "sun/security/provider/SHA5";
6408 }
6409 break;
6410 default:
6411 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6412 }
6414 ciKlass* klass_SHA = NULL;
6415 if (klass_SHA_name != NULL) {
6416 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6417 }
6418 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6419 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6420 Node* ctrl = control();
6421 set_control(top()); // no intrinsic path
6422 return ctrl;
6423 }
6424 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6426 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6427 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6428 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6429 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6431 return instof_false; // even if it is NULL
6432 }