1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/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"
43
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;
53
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 };
73
74
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
81
82 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
83
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 }
104
105 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
106
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(); }
112
113 bool try_to_inline(int predicate);
114 Node* try_to_predicate(int predicate);
115
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 }
123
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 }
128
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; }
132
133 virtual int reexecute_sp() { return _reexecute_sp; }
134
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);
197
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();
246
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();
257
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 bool inline_multiplyToLen();
326 };
327
328
329 //---------------------------make_vm_intrinsic----------------------------
330 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
331 vmIntrinsics::ID id = m->intrinsic_id();
332 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
333
334 ccstr disable_intr = NULL;
335
336 if ((DisableIntrinsic[0] != '\0'
337 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
338 (method_has_option_value("DisableIntrinsic", disable_intr)
339 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
340 // disabled by a user request on the command line:
341 // example: -XX:DisableIntrinsic=_hashCode,_getClass
342 return NULL;
343 }
344
345 if (!m->is_loaded()) {
346 // do not attempt to inline unloaded methods
347 return NULL;
348 }
349
350 // Only a few intrinsics implement a virtual dispatch.
351 // They are expensive calls which are also frequently overridden.
352 if (is_virtual) {
353 switch (id) {
354 case vmIntrinsics::_hashCode:
355 case vmIntrinsics::_clone:
356 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
357 break;
358 default:
359 return NULL;
360 }
361 }
362
363 // -XX:-InlineNatives disables nearly all intrinsics:
364 if (!InlineNatives) {
365 switch (id) {
366 case vmIntrinsics::_indexOf:
367 case vmIntrinsics::_compareTo:
368 case vmIntrinsics::_equals:
369 case vmIntrinsics::_equalsC:
370 case vmIntrinsics::_getAndAddInt:
371 case vmIntrinsics::_getAndAddLong:
372 case vmIntrinsics::_getAndSetInt:
373 case vmIntrinsics::_getAndSetLong:
374 case vmIntrinsics::_getAndSetObject:
375 case vmIntrinsics::_loadFence:
376 case vmIntrinsics::_storeFence:
377 case vmIntrinsics::_fullFence:
378 break; // InlineNatives does not control String.compareTo
379 case vmIntrinsics::_Reference_get:
380 break; // InlineNatives does not control Reference.get
381 default:
382 return NULL;
383 }
384 }
385
386 int predicates = 0;
387 bool does_virtual_dispatch = false;
388
389 switch (id) {
390 case vmIntrinsics::_compareTo:
391 if (!SpecialStringCompareTo) return NULL;
392 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
393 break;
394 case vmIntrinsics::_indexOf:
395 if (!SpecialStringIndexOf) return NULL;
396 break;
397 case vmIntrinsics::_equals:
398 if (!SpecialStringEquals) return NULL;
399 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
400 break;
401 case vmIntrinsics::_equalsC:
402 if (!SpecialArraysEquals) return NULL;
403 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
404 break;
405 case vmIntrinsics::_arraycopy:
406 if (!InlineArrayCopy) return NULL;
407 break;
408 case vmIntrinsics::_copyMemory:
409 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
410 if (!InlineArrayCopy) return NULL;
411 break;
412 case vmIntrinsics::_hashCode:
413 if (!InlineObjectHash) return NULL;
414 does_virtual_dispatch = true;
415 break;
416 case vmIntrinsics::_clone:
417 does_virtual_dispatch = true;
418 case vmIntrinsics::_copyOf:
419 case vmIntrinsics::_copyOfRange:
420 if (!InlineObjectCopy) return NULL;
421 // These also use the arraycopy intrinsic mechanism:
422 if (!InlineArrayCopy) return NULL;
423 break;
424 case vmIntrinsics::_encodeISOArray:
425 if (!SpecialEncodeISOArray) return NULL;
426 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
427 break;
428 case vmIntrinsics::_checkIndex:
429 // We do not intrinsify this. The optimizer does fine with it.
430 return NULL;
431
432 case vmIntrinsics::_getCallerClass:
433 if (!UseNewReflection) return NULL;
434 if (!InlineReflectionGetCallerClass) return NULL;
435 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
436 break;
437
438 case vmIntrinsics::_bitCount_i:
439 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
440 break;
441
442 case vmIntrinsics::_bitCount_l:
443 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
444 break;
445
446 case vmIntrinsics::_numberOfLeadingZeros_i:
447 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
448 break;
449
450 case vmIntrinsics::_numberOfLeadingZeros_l:
451 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
452 break;
453
454 case vmIntrinsics::_numberOfTrailingZeros_i:
455 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
456 break;
457
458 case vmIntrinsics::_numberOfTrailingZeros_l:
459 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
460 break;
461
462 case vmIntrinsics::_reverseBytes_c:
463 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
464 break;
465 case vmIntrinsics::_reverseBytes_s:
466 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
467 break;
468 case vmIntrinsics::_reverseBytes_i:
469 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
470 break;
471 case vmIntrinsics::_reverseBytes_l:
472 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
473 break;
474
475 case vmIntrinsics::_Reference_get:
476 // Use the intrinsic version of Reference.get() so that the value in
477 // the referent field can be registered by the G1 pre-barrier code.
478 // Also add memory barrier to prevent commoning reads from this field
479 // across safepoint since GC can change it value.
480 break;
481
482 case vmIntrinsics::_compareAndSwapObject:
483 #ifdef _LP64
484 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
485 #endif
486 break;
487
488 case vmIntrinsics::_compareAndSwapLong:
489 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
490 break;
491
492 case vmIntrinsics::_getAndAddInt:
493 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
494 break;
495
496 case vmIntrinsics::_getAndAddLong:
497 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
498 break;
499
500 case vmIntrinsics::_getAndSetInt:
501 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
502 break;
503
504 case vmIntrinsics::_getAndSetLong:
505 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
506 break;
507
508 case vmIntrinsics::_getAndSetObject:
509 #ifdef _LP64
510 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
511 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
512 break;
513 #else
514 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
515 break;
516 #endif
517
518 case vmIntrinsics::_aescrypt_encryptBlock:
519 case vmIntrinsics::_aescrypt_decryptBlock:
520 if (!UseAESIntrinsics) return NULL;
521 break;
522
523 case vmIntrinsics::_multiplyToLen:
524 if (!UseMultiplyToLenIntrinsic) return NULL;
525 break;
526
527 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
528 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
529 if (!UseAESIntrinsics) return NULL;
530 // these two require the predicated logic
531 predicates = 1;
532 break;
533
534 case vmIntrinsics::_sha_implCompress:
535 if (!UseSHA1Intrinsics) return NULL;
536 break;
537
538 case vmIntrinsics::_sha2_implCompress:
539 if (!UseSHA256Intrinsics) return NULL;
540 break;
541
542 case vmIntrinsics::_sha5_implCompress:
543 if (!UseSHA512Intrinsics) return NULL;
544 break;
545
546 case vmIntrinsics::_digestBase_implCompressMB:
547 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
548 predicates = 3;
549 break;
550
551 case vmIntrinsics::_updateCRC32:
552 case vmIntrinsics::_updateBytesCRC32:
553 case vmIntrinsics::_updateByteBufferCRC32:
554 if (!UseCRC32Intrinsics) return NULL;
555 break;
556
557 case vmIntrinsics::_incrementExactI:
558 case vmIntrinsics::_addExactI:
559 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
560 break;
561 case vmIntrinsics::_incrementExactL:
562 case vmIntrinsics::_addExactL:
563 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
564 break;
565 case vmIntrinsics::_decrementExactI:
566 case vmIntrinsics::_subtractExactI:
567 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
568 break;
569 case vmIntrinsics::_decrementExactL:
570 case vmIntrinsics::_subtractExactL:
571 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
572 break;
573 case vmIntrinsics::_negateExactI:
574 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
575 break;
576 case vmIntrinsics::_negateExactL:
577 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
578 break;
579 case vmIntrinsics::_multiplyExactI:
580 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
581 break;
582 case vmIntrinsics::_multiplyExactL:
583 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
584 break;
585
586 default:
587 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
588 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
589 break;
590 }
591
592 // -XX:-InlineClassNatives disables natives from the Class class.
593 // The flag applies to all reflective calls, notably Array.newArray
594 // (visible to Java programmers as Array.newInstance).
595 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
596 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
597 if (!InlineClassNatives) return NULL;
598 }
599
600 // -XX:-InlineThreadNatives disables natives from the Thread class.
601 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
602 if (!InlineThreadNatives) return NULL;
603 }
604
605 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
606 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
607 m->holder()->name() == ciSymbol::java_lang_Float() ||
608 m->holder()->name() == ciSymbol::java_lang_Double()) {
609 if (!InlineMathNatives) return NULL;
610 }
611
612 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
613 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
614 if (!InlineUnsafeOps) return NULL;
615 }
616
617 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
618 }
619
620 //----------------------register_library_intrinsics-----------------------
621 // Initialize this file's data structures, for each Compile instance.
622 void Compile::register_library_intrinsics() {
623 // Nothing to do here.
624 }
625
626 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
627 LibraryCallKit kit(jvms, this);
628 Compile* C = kit.C;
629 int nodes = C->unique();
630 #ifndef PRODUCT
631 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
632 char buf[1000];
633 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
634 tty->print_cr("Intrinsic %s", str);
635 }
636 #endif
637 ciMethod* callee = kit.callee();
638 const int bci = kit.bci();
639
640 // Try to inline the intrinsic.
641 if (kit.try_to_inline(_last_predicate)) {
642 if (C->print_intrinsics() || C->print_inlining()) {
643 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
644 }
645 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
646 if (C->log()) {
647 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
648 vmIntrinsics::name_at(intrinsic_id()),
649 (is_virtual() ? " virtual='1'" : ""),
650 C->unique() - nodes);
651 }
652 // Push the result from the inlined method onto the stack.
653 kit.push_result();
654 return kit.transfer_exceptions_into_jvms();
655 }
656
657 // The intrinsic bailed out
658 if (C->print_intrinsics() || C->print_inlining()) {
659 if (jvms->has_method()) {
660 // Not a root compile.
661 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
662 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
663 } else {
664 // Root compile
665 tty->print("Did not generate intrinsic %s%s at bci:%d in",
666 vmIntrinsics::name_at(intrinsic_id()),
667 (is_virtual() ? " (virtual)" : ""), bci);
668 }
669 }
670 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
671 return NULL;
672 }
673
674 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
675 LibraryCallKit kit(jvms, this);
676 Compile* C = kit.C;
677 int nodes = C->unique();
678 _last_predicate = predicate;
679 #ifndef PRODUCT
680 assert(is_predicated() && predicate < predicates_count(), "sanity");
681 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
682 char buf[1000];
683 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
684 tty->print_cr("Predicate for intrinsic %s", str);
685 }
686 #endif
687 ciMethod* callee = kit.callee();
688 const int bci = kit.bci();
689
690 Node* slow_ctl = kit.try_to_predicate(predicate);
691 if (!kit.failing()) {
692 if (C->print_intrinsics() || C->print_inlining()) {
693 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
694 }
695 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
696 if (C->log()) {
697 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
698 vmIntrinsics::name_at(intrinsic_id()),
699 (is_virtual() ? " virtual='1'" : ""),
700 C->unique() - nodes);
701 }
702 return slow_ctl; // Could be NULL if the check folds.
703 }
704
705 // The intrinsic bailed out
706 if (C->print_intrinsics() || C->print_inlining()) {
707 if (jvms->has_method()) {
708 // Not a root compile.
709 const char* msg = "failed to generate predicate for intrinsic";
710 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
711 } else {
712 // Root compile
713 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
714 vmIntrinsics::name_at(intrinsic_id()),
715 (is_virtual() ? " (virtual)" : ""), bci);
716 }
717 }
718 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
719 return NULL;
720 }
721
722 bool LibraryCallKit::try_to_inline(int predicate) {
723 // Handle symbolic names for otherwise undistinguished boolean switches:
724 const bool is_store = true;
725 const bool is_native_ptr = true;
726 const bool is_static = true;
727 const bool is_volatile = true;
728
729 if (!jvms()->has_method()) {
730 // Root JVMState has a null method.
731 assert(map()->memory()->Opcode() == Op_Parm, "");
732 // Insert the memory aliasing node
733 set_all_memory(reset_memory());
734 }
735 assert(merged_memory(), "");
736
737
738 switch (intrinsic_id()) {
739 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
740 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
741 case vmIntrinsics::_getClass: return inline_native_getClass();
742
743 case vmIntrinsics::_dsin:
744 case vmIntrinsics::_dcos:
745 case vmIntrinsics::_dtan:
746 case vmIntrinsics::_dabs:
747 case vmIntrinsics::_datan2:
748 case vmIntrinsics::_dsqrt:
749 case vmIntrinsics::_dexp:
750 case vmIntrinsics::_dlog:
751 case vmIntrinsics::_dlog10:
752 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
753
754 case vmIntrinsics::_min:
755 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
756
757 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
758 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
759 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
760 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
761 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
762 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
763 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
764 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
765 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
766 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
767 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
768 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
769
770 case vmIntrinsics::_arraycopy: return inline_arraycopy();
771
772 case vmIntrinsics::_compareTo: return inline_string_compareTo();
773 case vmIntrinsics::_indexOf: return inline_string_indexOf();
774 case vmIntrinsics::_equals: return inline_string_equals();
775
776 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
777 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
778 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
779 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
780 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
781 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
782 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
783 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
784 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
785
786 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
787 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
788 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
789 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
790 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
791 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
792 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
793 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
794 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
795
796 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
797 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
798 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
799 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
800 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
801 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
802 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
803 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
804
805 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
806 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
807 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
808 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
809 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
810 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
811 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
812 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
813
814 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
815 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
816 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
817 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
818 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
819 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
820 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
821 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
822 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
823
824 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
825 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
826 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
827 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
828 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
829 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
830 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
831 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
832 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
833
834 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
835 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
836 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
837 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
838
839 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
840 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
841 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
842
843 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
844 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
845 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
846
847 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
848 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
849 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
850 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
851 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
852
853 case vmIntrinsics::_loadFence:
854 case vmIntrinsics::_storeFence:
855 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
856
857 case vmIntrinsics::_currentThread: return inline_native_currentThread();
858 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
859
860 #ifdef TRACE_HAVE_INTRINSICS
861 case vmIntrinsics::_classID: return inline_native_classID();
862 case vmIntrinsics::_threadID: return inline_native_threadID();
863 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
864 #endif
865 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
866 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
867 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
868 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
869 case vmIntrinsics::_newArray: return inline_native_newArray();
870 case vmIntrinsics::_getLength: return inline_native_getLength();
871 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
872 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
873 case vmIntrinsics::_equalsC: return inline_array_equals();
874 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
875
876 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
877
878 case vmIntrinsics::_isInstance:
879 case vmIntrinsics::_getModifiers:
880 case vmIntrinsics::_isInterface:
881 case vmIntrinsics::_isArray:
882 case vmIntrinsics::_isPrimitive:
883 case vmIntrinsics::_getSuperclass:
884 case vmIntrinsics::_getComponentType:
885 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
886
887 case vmIntrinsics::_floatToRawIntBits:
888 case vmIntrinsics::_floatToIntBits:
889 case vmIntrinsics::_intBitsToFloat:
890 case vmIntrinsics::_doubleToRawLongBits:
891 case vmIntrinsics::_doubleToLongBits:
892 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
893
894 case vmIntrinsics::_numberOfLeadingZeros_i:
895 case vmIntrinsics::_numberOfLeadingZeros_l:
896 case vmIntrinsics::_numberOfTrailingZeros_i:
897 case vmIntrinsics::_numberOfTrailingZeros_l:
898 case vmIntrinsics::_bitCount_i:
899 case vmIntrinsics::_bitCount_l:
900 case vmIntrinsics::_reverseBytes_i:
901 case vmIntrinsics::_reverseBytes_l:
902 case vmIntrinsics::_reverseBytes_s:
903 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
904
905 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
906
907 case vmIntrinsics::_Reference_get: return inline_reference_get();
908
909 case vmIntrinsics::_aescrypt_encryptBlock:
910 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
911
912 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
913 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
914 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
915
916 case vmIntrinsics::_sha_implCompress:
917 case vmIntrinsics::_sha2_implCompress:
918 case vmIntrinsics::_sha5_implCompress:
919 return inline_sha_implCompress(intrinsic_id());
920
921 case vmIntrinsics::_digestBase_implCompressMB:
922 return inline_digestBase_implCompressMB(predicate);
923
924 case vmIntrinsics::_multiplyToLen:
925 return inline_multiplyToLen();
926
927 case vmIntrinsics::_encodeISOArray:
928 return inline_encodeISOArray();
929
930 case vmIntrinsics::_updateCRC32:
931 return inline_updateCRC32();
932 case vmIntrinsics::_updateBytesCRC32:
933 return inline_updateBytesCRC32();
934 case vmIntrinsics::_updateByteBufferCRC32:
935 return inline_updateByteBufferCRC32();
936
937 default:
938 // If you get here, it may be that someone has added a new intrinsic
939 // to the list in vmSymbols.hpp without implementing it here.
940 #ifndef PRODUCT
941 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
942 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
943 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
944 }
945 #endif
946 return false;
947 }
948 }
949
950 Node* LibraryCallKit::try_to_predicate(int predicate) {
951 if (!jvms()->has_method()) {
952 // Root JVMState has a null method.
953 assert(map()->memory()->Opcode() == Op_Parm, "");
954 // Insert the memory aliasing node
955 set_all_memory(reset_memory());
956 }
957 assert(merged_memory(), "");
958
959 switch (intrinsic_id()) {
960 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
961 return inline_cipherBlockChaining_AESCrypt_predicate(false);
962 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
963 return inline_cipherBlockChaining_AESCrypt_predicate(true);
964 case vmIntrinsics::_digestBase_implCompressMB:
965 return inline_digestBase_implCompressMB_predicate(predicate);
966
967 default:
968 // If you get here, it may be that someone has added a new intrinsic
969 // to the list in vmSymbols.hpp without implementing it here.
970 #ifndef PRODUCT
971 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
972 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
973 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
974 }
975 #endif
976 Node* slow_ctl = control();
977 set_control(top()); // No fast path instrinsic
978 return slow_ctl;
979 }
980 }
981
982 //------------------------------set_result-------------------------------
983 // Helper function for finishing intrinsics.
984 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
985 record_for_igvn(region);
986 set_control(_gvn.transform(region));
987 set_result( _gvn.transform(value));
988 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
989 }
990
991 //------------------------------generate_guard---------------------------
992 // Helper function for generating guarded fast-slow graph structures.
993 // The given 'test', if true, guards a slow path. If the test fails
994 // then a fast path can be taken. (We generally hope it fails.)
995 // In all cases, GraphKit::control() is updated to the fast path.
996 // The returned value represents the control for the slow path.
997 // The return value is never 'top'; it is either a valid control
998 // or NULL if it is obvious that the slow path can never be taken.
999 // Also, if region and the slow control are not NULL, the slow edge
1000 // is appended to the region.
1001 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1002 if (stopped()) {
1003 // Already short circuited.
1004 return NULL;
1005 }
1006
1007 // Build an if node and its projections.
1008 // If test is true we take the slow path, which we assume is uncommon.
1009 if (_gvn.type(test) == TypeInt::ZERO) {
1010 // The slow branch is never taken. No need to build this guard.
1011 return NULL;
1012 }
1013
1014 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1015
1016 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1017 if (if_slow == top()) {
1018 // The slow branch is never taken. No need to build this guard.
1019 return NULL;
1020 }
1021
1022 if (region != NULL)
1023 region->add_req(if_slow);
1024
1025 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1026 set_control(if_fast);
1027
1028 return if_slow;
1029 }
1030
1031 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1032 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1033 }
1034 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1035 return generate_guard(test, region, PROB_FAIR);
1036 }
1037
1038 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1039 Node* *pos_index) {
1040 if (stopped())
1041 return NULL; // already stopped
1042 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1043 return NULL; // index is already adequately typed
1044 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1045 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1046 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1047 if (is_neg != NULL && pos_index != NULL) {
1048 // Emulate effect of Parse::adjust_map_after_if.
1049 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1050 ccast->set_req(0, control());
1051 (*pos_index) = _gvn.transform(ccast);
1052 }
1053 return is_neg;
1054 }
1055
1056 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1057 Node* *pos_index) {
1058 if (stopped())
1059 return NULL; // already stopped
1060 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1061 return NULL; // index is already adequately typed
1062 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1063 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1064 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1065 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1066 if (is_notp != NULL && pos_index != NULL) {
1067 // Emulate effect of Parse::adjust_map_after_if.
1068 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1069 ccast->set_req(0, control());
1070 (*pos_index) = _gvn.transform(ccast);
1071 }
1072 return is_notp;
1073 }
1074
1075 // Make sure that 'position' is a valid limit index, in [0..length].
1076 // There are two equivalent plans for checking this:
1077 // A. (offset + copyLength) unsigned<= arrayLength
1078 // B. offset <= (arrayLength - copyLength)
1079 // We require that all of the values above, except for the sum and
1080 // difference, are already known to be non-negative.
1081 // Plan A is robust in the face of overflow, if offset and copyLength
1082 // are both hugely positive.
1083 //
1084 // Plan B is less direct and intuitive, but it does not overflow at
1085 // all, since the difference of two non-negatives is always
1086 // representable. Whenever Java methods must perform the equivalent
1087 // check they generally use Plan B instead of Plan A.
1088 // For the moment we use Plan A.
1089 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1090 Node* subseq_length,
1091 Node* array_length,
1092 RegionNode* region) {
1093 if (stopped())
1094 return NULL; // already stopped
1095 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1096 if (zero_offset && subseq_length->eqv_uncast(array_length))
1097 return NULL; // common case of whole-array copy
1098 Node* last = subseq_length;
1099 if (!zero_offset) // last += offset
1100 last = _gvn.transform(new (C) AddINode(last, offset));
1101 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1102 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1103 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1104 return is_over;
1105 }
1106
1107
1108 //--------------------------generate_current_thread--------------------
1109 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1110 ciKlass* thread_klass = env()->Thread_klass();
1111 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1112 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1113 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1114 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1115 tls_output = thread;
1116 return threadObj;
1117 }
1118
1119
1120 //------------------------------make_string_method_node------------------------
1121 // Helper method for String intrinsic functions. This version is called
1122 // with str1 and str2 pointing to String object nodes.
1123 //
1124 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1125 Node* no_ctrl = NULL;
1126
1127 // Get start addr of string
1128 Node* str1_value = load_String_value(no_ctrl, str1);
1129 Node* str1_offset = load_String_offset(no_ctrl, str1);
1130 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1131
1132 // Get length of string 1
1133 Node* str1_len = load_String_length(no_ctrl, str1);
1134
1135 Node* str2_value = load_String_value(no_ctrl, str2);
1136 Node* str2_offset = load_String_offset(no_ctrl, str2);
1137 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1138
1139 Node* str2_len = NULL;
1140 Node* result = NULL;
1141
1142 switch (opcode) {
1143 case Op_StrIndexOf:
1144 // Get length of string 2
1145 str2_len = load_String_length(no_ctrl, str2);
1146
1147 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1148 str1_start, str1_len, str2_start, str2_len);
1149 break;
1150 case Op_StrComp:
1151 // Get length of string 2
1152 str2_len = load_String_length(no_ctrl, str2);
1153
1154 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1155 str1_start, str1_len, str2_start, str2_len);
1156 break;
1157 case Op_StrEquals:
1158 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1159 str1_start, str2_start, str1_len);
1160 break;
1161 default:
1162 ShouldNotReachHere();
1163 return NULL;
1164 }
1165
1166 // All these intrinsics have checks.
1167 C->set_has_split_ifs(true); // Has chance for split-if optimization
1168
1169 return _gvn.transform(result);
1170 }
1171
1172 // Helper method for String intrinsic functions. This version is called
1173 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1174 // to Int nodes containing the lenghts of str1 and str2.
1175 //
1176 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1177 Node* result = NULL;
1178 switch (opcode) {
1179 case Op_StrIndexOf:
1180 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1181 str1_start, cnt1, str2_start, cnt2);
1182 break;
1183 case Op_StrComp:
1184 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1185 str1_start, cnt1, str2_start, cnt2);
1186 break;
1187 case Op_StrEquals:
1188 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1189 str1_start, str2_start, cnt1);
1190 break;
1191 default:
1192 ShouldNotReachHere();
1193 return NULL;
1194 }
1195
1196 // All these intrinsics have checks.
1197 C->set_has_split_ifs(true); // Has chance for split-if optimization
1198
1199 return _gvn.transform(result);
1200 }
1201
1202 //------------------------------inline_string_compareTo------------------------
1203 // public int java.lang.String.compareTo(String anotherString);
1204 bool LibraryCallKit::inline_string_compareTo() {
1205 Node* receiver = null_check(argument(0));
1206 Node* arg = null_check(argument(1));
1207 if (stopped()) {
1208 return true;
1209 }
1210 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1211 return true;
1212 }
1213
1214 //------------------------------inline_string_equals------------------------
1215 bool LibraryCallKit::inline_string_equals() {
1216 Node* receiver = null_check_receiver();
1217 // NOTE: Do not null check argument for String.equals() because spec
1218 // allows to specify NULL as argument.
1219 Node* argument = this->argument(1);
1220 if (stopped()) {
1221 return true;
1222 }
1223
1224 // paths (plus control) merge
1225 RegionNode* region = new (C) RegionNode(5);
1226 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1227
1228 // does source == target string?
1229 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1230 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1231
1232 Node* if_eq = generate_slow_guard(bol, NULL);
1233 if (if_eq != NULL) {
1234 // receiver == argument
1235 phi->init_req(2, intcon(1));
1236 region->init_req(2, if_eq);
1237 }
1238
1239 // get String klass for instanceOf
1240 ciInstanceKlass* klass = env()->String_klass();
1241
1242 if (!stopped()) {
1243 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1244 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1245 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1246
1247 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1248 //instanceOf == true, fallthrough
1249
1250 if (inst_false != NULL) {
1251 phi->init_req(3, intcon(0));
1252 region->init_req(3, inst_false);
1253 }
1254 }
1255
1256 if (!stopped()) {
1257 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1258
1259 // Properly cast the argument to String
1260 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1261 // This path is taken only when argument's type is String:NotNull.
1262 argument = cast_not_null(argument, false);
1263
1264 Node* no_ctrl = NULL;
1265
1266 // Get start addr of receiver
1267 Node* receiver_val = load_String_value(no_ctrl, receiver);
1268 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1269 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1270
1271 // Get length of receiver
1272 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1273
1274 // Get start addr of argument
1275 Node* argument_val = load_String_value(no_ctrl, argument);
1276 Node* argument_offset = load_String_offset(no_ctrl, argument);
1277 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1278
1279 // Get length of argument
1280 Node* argument_cnt = load_String_length(no_ctrl, argument);
1281
1282 // Check for receiver count != argument count
1283 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1284 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1285 Node* if_ne = generate_slow_guard(bol, NULL);
1286 if (if_ne != NULL) {
1287 phi->init_req(4, intcon(0));
1288 region->init_req(4, if_ne);
1289 }
1290
1291 // Check for count == 0 is done by assembler code for StrEquals.
1292
1293 if (!stopped()) {
1294 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1295 phi->init_req(1, equals);
1296 region->init_req(1, control());
1297 }
1298 }
1299
1300 // post merge
1301 set_control(_gvn.transform(region));
1302 record_for_igvn(region);
1303
1304 set_result(_gvn.transform(phi));
1305 return true;
1306 }
1307
1308 //------------------------------inline_array_equals----------------------------
1309 bool LibraryCallKit::inline_array_equals() {
1310 Node* arg1 = argument(0);
1311 Node* arg2 = argument(1);
1312 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1313 return true;
1314 }
1315
1316 // Java version of String.indexOf(constant string)
1317 // class StringDecl {
1318 // StringDecl(char[] ca) {
1319 // offset = 0;
1320 // count = ca.length;
1321 // value = ca;
1322 // }
1323 // int offset;
1324 // int count;
1325 // char[] value;
1326 // }
1327 //
1328 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1329 // int targetOffset, int cache_i, int md2) {
1330 // int cache = cache_i;
1331 // int sourceOffset = string_object.offset;
1332 // int sourceCount = string_object.count;
1333 // int targetCount = target_object.length;
1334 //
1335 // int targetCountLess1 = targetCount - 1;
1336 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1337 //
1338 // char[] source = string_object.value;
1339 // char[] target = target_object;
1340 // int lastChar = target[targetCountLess1];
1341 //
1342 // outer_loop:
1343 // for (int i = sourceOffset; i < sourceEnd; ) {
1344 // int src = source[i + targetCountLess1];
1345 // if (src == lastChar) {
1346 // // With random strings and a 4-character alphabet,
1347 // // reverse matching at this point sets up 0.8% fewer
1348 // // frames, but (paradoxically) makes 0.3% more probes.
1349 // // Since those probes are nearer the lastChar probe,
1350 // // there is may be a net D$ win with reverse matching.
1351 // // But, reversing loop inhibits unroll of inner loop
1352 // // for unknown reason. So, does running outer loop from
1353 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1354 // for (int j = 0; j < targetCountLess1; j++) {
1355 // if (target[targetOffset + j] != source[i+j]) {
1356 // if ((cache & (1 << source[i+j])) == 0) {
1357 // if (md2 < j+1) {
1358 // i += j+1;
1359 // continue outer_loop;
1360 // }
1361 // }
1362 // i += md2;
1363 // continue outer_loop;
1364 // }
1365 // }
1366 // return i - sourceOffset;
1367 // }
1368 // if ((cache & (1 << src)) == 0) {
1369 // i += targetCountLess1;
1370 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1371 // i++;
1372 // }
1373 // return -1;
1374 // }
1375
1376 //------------------------------string_indexOf------------------------
1377 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1378 jint cache_i, jint md2_i) {
1379
1380 Node* no_ctrl = NULL;
1381 float likely = PROB_LIKELY(0.9);
1382 float unlikely = PROB_UNLIKELY(0.9);
1383
1384 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1385
1386 Node* source = load_String_value(no_ctrl, string_object);
1387 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1388 Node* sourceCount = load_String_length(no_ctrl, string_object);
1389
1390 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1391 jint target_length = target_array->length();
1392 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1393 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1394
1395 // String.value field is known to be @Stable.
1396 if (UseImplicitStableValues) {
1397 target = cast_array_to_stable(target, target_type);
1398 }
1399
1400 IdealKit kit(this, false, true);
1401 #define __ kit.
1402 Node* zero = __ ConI(0);
1403 Node* one = __ ConI(1);
1404 Node* cache = __ ConI(cache_i);
1405 Node* md2 = __ ConI(md2_i);
1406 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1407 Node* targetCount = __ ConI(target_length);
1408 Node* targetCountLess1 = __ ConI(target_length - 1);
1409 Node* targetOffset = __ ConI(targetOffset_i);
1410 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1411
1412 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1413 Node* outer_loop = __ make_label(2 /* goto */);
1414 Node* return_ = __ make_label(1);
1415
1416 __ set(rtn,__ ConI(-1));
1417 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1418 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1419 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1420 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1421 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1422 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1423 Node* tpj = __ AddI(targetOffset, __ value(j));
1424 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1425 Node* ipj = __ AddI(__ value(i), __ value(j));
1426 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1427 __ if_then(targ, BoolTest::ne, src2); {
1428 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1429 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1430 __ increment(i, __ AddI(__ value(j), one));
1431 __ goto_(outer_loop);
1432 } __ end_if(); __ dead(j);
1433 }__ end_if(); __ dead(j);
1434 __ increment(i, md2);
1435 __ goto_(outer_loop);
1436 }__ end_if();
1437 __ increment(j, one);
1438 }__ end_loop(); __ dead(j);
1439 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1440 __ goto_(return_);
1441 }__ end_if();
1442 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1443 __ increment(i, targetCountLess1);
1444 }__ end_if();
1445 __ increment(i, one);
1446 __ bind(outer_loop);
1447 }__ end_loop(); __ dead(i);
1448 __ bind(return_);
1449
1450 // Final sync IdealKit and GraphKit.
1451 final_sync(kit);
1452 Node* result = __ value(rtn);
1453 #undef __
1454 C->set_has_loops(true);
1455 return result;
1456 }
1457
1458 //------------------------------inline_string_indexOf------------------------
1459 bool LibraryCallKit::inline_string_indexOf() {
1460 Node* receiver = argument(0);
1461 Node* arg = argument(1);
1462
1463 Node* result;
1464 // Disable the use of pcmpestri until it can be guaranteed that
1465 // the load doesn't cross into the uncommited space.
1466 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1467 UseSSE42Intrinsics) {
1468 // Generate SSE4.2 version of indexOf
1469 // We currently only have match rules that use SSE4.2
1470
1471 receiver = null_check(receiver);
1472 arg = null_check(arg);
1473 if (stopped()) {
1474 return true;
1475 }
1476
1477 ciInstanceKlass* str_klass = env()->String_klass();
1478 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1479
1480 // Make the merge point
1481 RegionNode* result_rgn = new (C) RegionNode(4);
1482 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1483 Node* no_ctrl = NULL;
1484
1485 // Get start addr of source string
1486 Node* source = load_String_value(no_ctrl, receiver);
1487 Node* source_offset = load_String_offset(no_ctrl, receiver);
1488 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1489
1490 // Get length of source string
1491 Node* source_cnt = load_String_length(no_ctrl, receiver);
1492
1493 // Get start addr of substring
1494 Node* substr = load_String_value(no_ctrl, arg);
1495 Node* substr_offset = load_String_offset(no_ctrl, arg);
1496 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1497
1498 // Get length of source string
1499 Node* substr_cnt = load_String_length(no_ctrl, arg);
1500
1501 // Check for substr count > string count
1502 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1503 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1504 Node* if_gt = generate_slow_guard(bol, NULL);
1505 if (if_gt != NULL) {
1506 result_phi->init_req(2, intcon(-1));
1507 result_rgn->init_req(2, if_gt);
1508 }
1509
1510 if (!stopped()) {
1511 // Check for substr count == 0
1512 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1513 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1514 Node* if_zero = generate_slow_guard(bol, NULL);
1515 if (if_zero != NULL) {
1516 result_phi->init_req(3, intcon(0));
1517 result_rgn->init_req(3, if_zero);
1518 }
1519 }
1520
1521 if (!stopped()) {
1522 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1523 result_phi->init_req(1, result);
1524 result_rgn->init_req(1, control());
1525 }
1526 set_control(_gvn.transform(result_rgn));
1527 record_for_igvn(result_rgn);
1528 result = _gvn.transform(result_phi);
1529
1530 } else { // Use LibraryCallKit::string_indexOf
1531 // don't intrinsify if argument isn't a constant string.
1532 if (!arg->is_Con()) {
1533 return false;
1534 }
1535 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1536 if (str_type == NULL) {
1537 return false;
1538 }
1539 ciInstanceKlass* klass = env()->String_klass();
1540 ciObject* str_const = str_type->const_oop();
1541 if (str_const == NULL || str_const->klass() != klass) {
1542 return false;
1543 }
1544 ciInstance* str = str_const->as_instance();
1545 assert(str != NULL, "must be instance");
1546
1547 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1548 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1549
1550 int o;
1551 int c;
1552 if (java_lang_String::has_offset_field()) {
1553 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1554 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1555 } else {
1556 o = 0;
1557 c = pat->length();
1558 }
1559
1560 // constant strings have no offset and count == length which
1561 // simplifies the resulting code somewhat so lets optimize for that.
1562 if (o != 0 || c != pat->length()) {
1563 return false;
1564 }
1565
1566 receiver = null_check(receiver, T_OBJECT);
1567 // NOTE: No null check on the argument is needed since it's a constant String oop.
1568 if (stopped()) {
1569 return true;
1570 }
1571
1572 // The null string as a pattern always returns 0 (match at beginning of string)
1573 if (c == 0) {
1574 set_result(intcon(0));
1575 return true;
1576 }
1577
1578 // Generate default indexOf
1579 jchar lastChar = pat->char_at(o + (c - 1));
1580 int cache = 0;
1581 int i;
1582 for (i = 0; i < c - 1; i++) {
1583 assert(i < pat->length(), "out of range");
1584 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1585 }
1586
1587 int md2 = c;
1588 for (i = 0; i < c - 1; i++) {
1589 assert(i < pat->length(), "out of range");
1590 if (pat->char_at(o + i) == lastChar) {
1591 md2 = (c - 1) - i;
1592 }
1593 }
1594
1595 result = string_indexOf(receiver, pat, o, cache, md2);
1596 }
1597 set_result(result);
1598 return true;
1599 }
1600
1601 //--------------------------round_double_node--------------------------------
1602 // Round a double node if necessary.
1603 Node* LibraryCallKit::round_double_node(Node* n) {
1604 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1605 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1606 return n;
1607 }
1608
1609 //------------------------------inline_math-----------------------------------
1610 // public static double Math.abs(double)
1611 // public static double Math.sqrt(double)
1612 // public static double Math.log(double)
1613 // public static double Math.log10(double)
1614 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1615 Node* arg = round_double_node(argument(0));
1616 Node* n;
1617 switch (id) {
1618 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1619 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1620 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1621 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1622 default: fatal_unexpected_iid(id); break;
1623 }
1624 set_result(_gvn.transform(n));
1625 return true;
1626 }
1627
1628 //------------------------------inline_trig----------------------------------
1629 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1630 // argument reduction which will turn into a fast/slow diamond.
1631 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1632 Node* arg = round_double_node(argument(0));
1633 Node* n = NULL;
1634
1635 switch (id) {
1636 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1637 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1638 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1639 default: fatal_unexpected_iid(id); break;
1640 }
1641 n = _gvn.transform(n);
1642
1643 // Rounding required? Check for argument reduction!
1644 if (Matcher::strict_fp_requires_explicit_rounding) {
1645 static const double pi_4 = 0.7853981633974483;
1646 static const double neg_pi_4 = -0.7853981633974483;
1647 // pi/2 in 80-bit extended precision
1648 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1649 // -pi/2 in 80-bit extended precision
1650 // 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};
1651 // Cutoff value for using this argument reduction technique
1652 //static const double pi_2_minus_epsilon = 1.564660403643354;
1653 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1654
1655 // Pseudocode for sin:
1656 // if (x <= Math.PI / 4.0) {
1657 // if (x >= -Math.PI / 4.0) return fsin(x);
1658 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1659 // } else {
1660 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1661 // }
1662 // return StrictMath.sin(x);
1663
1664 // Pseudocode for cos:
1665 // if (x <= Math.PI / 4.0) {
1666 // if (x >= -Math.PI / 4.0) return fcos(x);
1667 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1668 // } else {
1669 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1670 // }
1671 // return StrictMath.cos(x);
1672
1673 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1674 // requires a special machine instruction to load it. Instead we'll try
1675 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1676 // probably do the math inside the SIN encoding.
1677
1678 // Make the merge point
1679 RegionNode* r = new (C) RegionNode(3);
1680 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1681
1682 // Flatten arg so we need only 1 test
1683 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1684 // Node for PI/4 constant
1685 Node *pi4 = makecon(TypeD::make(pi_4));
1686 // Check PI/4 : abs(arg)
1687 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1688 // Check: If PI/4 < abs(arg) then go slow
1689 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1690 // Branch either way
1691 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1692 set_control(opt_iff(r,iff));
1693
1694 // Set fast path result
1695 phi->init_req(2, n);
1696
1697 // Slow path - non-blocking leaf call
1698 Node* call = NULL;
1699 switch (id) {
1700 case vmIntrinsics::_dsin:
1701 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1702 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1703 "Sin", NULL, arg, top());
1704 break;
1705 case vmIntrinsics::_dcos:
1706 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1707 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1708 "Cos", NULL, arg, top());
1709 break;
1710 case vmIntrinsics::_dtan:
1711 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1712 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1713 "Tan", NULL, arg, top());
1714 break;
1715 }
1716 assert(control()->in(0) == call, "");
1717 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1718 r->init_req(1, control());
1719 phi->init_req(1, slow_result);
1720
1721 // Post-merge
1722 set_control(_gvn.transform(r));
1723 record_for_igvn(r);
1724 n = _gvn.transform(phi);
1725
1726 C->set_has_split_ifs(true); // Has chance for split-if optimization
1727 }
1728 set_result(n);
1729 return true;
1730 }
1731
1732 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1733 //-------------------
1734 //result=(result.isNaN())? funcAddr():result;
1735 // Check: If isNaN() by checking result!=result? then either trap
1736 // or go to runtime
1737 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1738 // Build the boolean node
1739 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1740
1741 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1742 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1743 // The pow or exp intrinsic returned a NaN, which requires a call
1744 // to the runtime. Recompile with the runtime call.
1745 uncommon_trap(Deoptimization::Reason_intrinsic,
1746 Deoptimization::Action_make_not_entrant);
1747 }
1748 return result;
1749 } else {
1750 // If this inlining ever returned NaN in the past, we compile a call
1751 // to the runtime to properly handle corner cases
1752
1753 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1754 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1755 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1756
1757 if (!if_slow->is_top()) {
1758 RegionNode* result_region = new (C) RegionNode(3);
1759 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1760
1761 result_region->init_req(1, if_fast);
1762 result_val->init_req(1, result);
1763
1764 set_control(if_slow);
1765
1766 const TypePtr* no_memory_effects = NULL;
1767 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1768 no_memory_effects,
1769 x, top(), y, y ? top() : NULL);
1770 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1771 #ifdef ASSERT
1772 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1773 assert(value_top == top(), "second value must be top");
1774 #endif
1775
1776 result_region->init_req(2, control());
1777 result_val->init_req(2, value);
1778 set_control(_gvn.transform(result_region));
1779 return _gvn.transform(result_val);
1780 } else {
1781 return result;
1782 }
1783 }
1784 }
1785
1786 //------------------------------inline_exp-------------------------------------
1787 // Inline exp instructions, if possible. The Intel hardware only misses
1788 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1789 bool LibraryCallKit::inline_exp() {
1790 Node* arg = round_double_node(argument(0));
1791 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1792
1793 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1794 set_result(n);
1795
1796 C->set_has_split_ifs(true); // Has chance for split-if optimization
1797 return true;
1798 }
1799
1800 //------------------------------inline_pow-------------------------------------
1801 // Inline power instructions, if possible.
1802 bool LibraryCallKit::inline_pow() {
1803 // Pseudocode for pow
1804 // if (y == 2) {
1805 // return x * x;
1806 // } else {
1807 // if (x <= 0.0) {
1808 // long longy = (long)y;
1809 // if ((double)longy == y) { // if y is long
1810 // if (y + 1 == y) longy = 0; // huge number: even
1811 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1812 // } else {
1813 // result = NaN;
1814 // }
1815 // } else {
1816 // result = DPow(x,y);
1817 // }
1818 // if (result != result)? {
1819 // result = uncommon_trap() or runtime_call();
1820 // }
1821 // return result;
1822 // }
1823
1824 Node* x = round_double_node(argument(0));
1825 Node* y = round_double_node(argument(2));
1826
1827 Node* result = NULL;
1828
1829 Node* const_two_node = makecon(TypeD::make(2.0));
1830 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1831 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1832 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1833 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1834 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1835
1836 RegionNode* region_node = new (C) RegionNode(3);
1837 region_node->init_req(1, if_true);
1838
1839 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1840 // special case for x^y where y == 2, we can convert it to x * x
1841 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1842
1843 // set control to if_false since we will now process the false branch
1844 set_control(if_false);
1845
1846 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1847 // Short form: skip the fancy tests and just check for NaN result.
1848 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1849 } else {
1850 // If this inlining ever returned NaN in the past, include all
1851 // checks + call to the runtime.
1852
1853 // Set the merge point for If node with condition of (x <= 0.0)
1854 // There are four possible paths to region node and phi node
1855 RegionNode *r = new (C) RegionNode(4);
1856 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1857
1858 // Build the first if node: if (x <= 0.0)
1859 // Node for 0 constant
1860 Node *zeronode = makecon(TypeD::ZERO);
1861 // Check x:0
1862 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1863 // Check: If (x<=0) then go complex path
1864 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1865 // Branch either way
1866 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1867 // Fast path taken; set region slot 3
1868 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1869 r->init_req(3,fast_taken); // Capture fast-control
1870
1871 // Fast path not-taken, i.e. slow path
1872 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1873
1874 // Set fast path result
1875 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1876 phi->init_req(3, fast_result);
1877
1878 // Complex path
1879 // Build the second if node (if y is long)
1880 // Node for (long)y
1881 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1882 // Node for (double)((long) y)
1883 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1884 // Check (double)((long) y) : y
1885 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1886 // Check if (y isn't long) then go to slow path
1887
1888 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1889 // Branch either way
1890 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1891 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1892
1893 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1894
1895 // Calculate DPow(abs(x), y)*(1 & (long)y)
1896 // Node for constant 1
1897 Node *conone = longcon(1);
1898 // 1& (long)y
1899 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1900
1901 // A huge number is always even. Detect a huge number by checking
1902 // if y + 1 == y and set integer to be tested for parity to 0.
1903 // Required for corner case:
1904 // (long)9.223372036854776E18 = max_jlong
1905 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1906 // max_jlong is odd but 9.223372036854776E18 is even
1907 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1908 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1909 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1910 Node* correctedsign = NULL;
1911 if (ConditionalMoveLimit != 0) {
1912 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1913 } else {
1914 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1915 RegionNode *r = new (C) RegionNode(3);
1916 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1917 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1918 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1919 phi->init_req(1, signnode);
1920 phi->init_req(2, longcon(0));
1921 correctedsign = _gvn.transform(phi);
1922 ylong_path = _gvn.transform(r);
1923 record_for_igvn(r);
1924 }
1925
1926 // zero node
1927 Node *conzero = longcon(0);
1928 // Check (1&(long)y)==0?
1929 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1930 // Check if (1&(long)y)!=0?, if so the result is negative
1931 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1932 // abs(x)
1933 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1934 // abs(x)^y
1935 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1936 // -abs(x)^y
1937 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1938 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1939 Node *signresult = NULL;
1940 if (ConditionalMoveLimit != 0) {
1941 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1942 } else {
1943 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1944 RegionNode *r = new (C) RegionNode(3);
1945 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1946 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1947 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1948 phi->init_req(1, absxpowy);
1949 phi->init_req(2, negabsxpowy);
1950 signresult = _gvn.transform(phi);
1951 ylong_path = _gvn.transform(r);
1952 record_for_igvn(r);
1953 }
1954 // Set complex path fast result
1955 r->init_req(2, ylong_path);
1956 phi->init_req(2, signresult);
1957
1958 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1959 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1960 r->init_req(1,slow_path);
1961 phi->init_req(1,slow_result);
1962
1963 // Post merge
1964 set_control(_gvn.transform(r));
1965 record_for_igvn(r);
1966 result = _gvn.transform(phi);
1967 }
1968
1969 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1970
1971 // control from finish_pow_exp is now input to the region node
1972 region_node->set_req(2, control());
1973 // the result from finish_pow_exp is now input to the phi node
1974 phi_node->init_req(2, result);
1975 set_control(_gvn.transform(region_node));
1976 record_for_igvn(region_node);
1977 set_result(_gvn.transform(phi_node));
1978
1979 C->set_has_split_ifs(true); // Has chance for split-if optimization
1980 return true;
1981 }
1982
1983 //------------------------------runtime_math-----------------------------
1984 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1985 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1986 "must be (DD)D or (D)D type");
1987
1988 // Inputs
1989 Node* a = round_double_node(argument(0));
1990 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1991
1992 const TypePtr* no_memory_effects = NULL;
1993 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1994 no_memory_effects,
1995 a, top(), b, b ? top() : NULL);
1996 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1997 #ifdef ASSERT
1998 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1999 assert(value_top == top(), "second value must be top");
2000 #endif
2001
2002 set_result(value);
2003 return true;
2004 }
2005
2006 //------------------------------inline_math_native-----------------------------
2007 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2008 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2009 switch (id) {
2010 // These intrinsics are not properly supported on all hardware
2011 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2012 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2013 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2014 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2015 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2016 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2017
2018 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2019 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2020 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2021 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2022
2023 // These intrinsics are supported on all hardware
2024 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2025 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2026
2027 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2028 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2029 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2030 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2031 #undef FN_PTR
2032
2033 // These intrinsics are not yet correctly implemented
2034 case vmIntrinsics::_datan2:
2035 return false;
2036
2037 default:
2038 fatal_unexpected_iid(id);
2039 return false;
2040 }
2041 }
2042
2043 static bool is_simple_name(Node* n) {
2044 return (n->req() == 1 // constant
2045 || (n->is_Type() && n->as_Type()->type()->singleton())
2046 || n->is_Proj() // parameter or return value
2047 || n->is_Phi() // local of some sort
2048 );
2049 }
2050
2051 //----------------------------inline_min_max-----------------------------------
2052 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2053 set_result(generate_min_max(id, argument(0), argument(1)));
2054 return true;
2055 }
2056
2057 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2058 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2059 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2060 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2061 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2062
2063 {
2064 PreserveJVMState pjvms(this);
2065 PreserveReexecuteState preexecs(this);
2066 jvms()->set_should_reexecute(true);
2067
2068 set_control(slow_path);
2069 set_i_o(i_o());
2070
2071 uncommon_trap(Deoptimization::Reason_intrinsic,
2072 Deoptimization::Action_none);
2073 }
2074
2075 set_control(fast_path);
2076 set_result(math);
2077 }
2078
2079 template <typename OverflowOp>
2080 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2081 typedef typename OverflowOp::MathOp MathOp;
2082
2083 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2084 Node* operation = _gvn.transform( mathOp );
2085 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2086 inline_math_mathExact(operation, ofcheck);
2087 return true;
2088 }
2089
2090 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2091 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2092 }
2093
2094 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2095 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2096 }
2097
2098 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2099 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2100 }
2101
2102 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2103 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2104 }
2105
2106 bool LibraryCallKit::inline_math_negateExactI() {
2107 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2108 }
2109
2110 bool LibraryCallKit::inline_math_negateExactL() {
2111 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2112 }
2113
2114 bool LibraryCallKit::inline_math_multiplyExactI() {
2115 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2116 }
2117
2118 bool LibraryCallKit::inline_math_multiplyExactL() {
2119 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2120 }
2121
2122 Node*
2123 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2124 // These are the candidate return value:
2125 Node* xvalue = x0;
2126 Node* yvalue = y0;
2127
2128 if (xvalue == yvalue) {
2129 return xvalue;
2130 }
2131
2132 bool want_max = (id == vmIntrinsics::_max);
2133
2134 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2135 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2136 if (txvalue == NULL || tyvalue == NULL) return top();
2137 // This is not really necessary, but it is consistent with a
2138 // hypothetical MaxINode::Value method:
2139 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2140
2141 // %%% This folding logic should (ideally) be in a different place.
2142 // Some should be inside IfNode, and there to be a more reliable
2143 // transformation of ?: style patterns into cmoves. We also want
2144 // more powerful optimizations around cmove and min/max.
2145
2146 // Try to find a dominating comparison of these guys.
2147 // It can simplify the index computation for Arrays.copyOf
2148 // and similar uses of System.arraycopy.
2149 // First, compute the normalized version of CmpI(x, y).
2150 int cmp_op = Op_CmpI;
2151 Node* xkey = xvalue;
2152 Node* ykey = yvalue;
2153 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2154 if (ideal_cmpxy->is_Cmp()) {
2155 // E.g., if we have CmpI(length - offset, count),
2156 // it might idealize to CmpI(length, count + offset)
2157 cmp_op = ideal_cmpxy->Opcode();
2158 xkey = ideal_cmpxy->in(1);
2159 ykey = ideal_cmpxy->in(2);
2160 }
2161
2162 // Start by locating any relevant comparisons.
2163 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2164 Node* cmpxy = NULL;
2165 Node* cmpyx = NULL;
2166 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2167 Node* cmp = start_from->fast_out(k);
2168 if (cmp->outcnt() > 0 && // must have prior uses
2169 cmp->in(0) == NULL && // must be context-independent
2170 cmp->Opcode() == cmp_op) { // right kind of compare
2171 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2172 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2173 }
2174 }
2175
2176 const int NCMPS = 2;
2177 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2178 int cmpn;
2179 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2180 if (cmps[cmpn] != NULL) break; // find a result
2181 }
2182 if (cmpn < NCMPS) {
2183 // Look for a dominating test that tells us the min and max.
2184 int depth = 0; // Limit search depth for speed
2185 Node* dom = control();
2186 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2187 if (++depth >= 100) break;
2188 Node* ifproj = dom;
2189 if (!ifproj->is_Proj()) continue;
2190 Node* iff = ifproj->in(0);
2191 if (!iff->is_If()) continue;
2192 Node* bol = iff->in(1);
2193 if (!bol->is_Bool()) continue;
2194 Node* cmp = bol->in(1);
2195 if (cmp == NULL) continue;
2196 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2197 if (cmps[cmpn] == cmp) break;
2198 if (cmpn == NCMPS) continue;
2199 BoolTest::mask btest = bol->as_Bool()->_test._test;
2200 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2201 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2202 // At this point, we know that 'x btest y' is true.
2203 switch (btest) {
2204 case BoolTest::eq:
2205 // They are proven equal, so we can collapse the min/max.
2206 // Either value is the answer. Choose the simpler.
2207 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2208 return yvalue;
2209 return xvalue;
2210 case BoolTest::lt: // x < y
2211 case BoolTest::le: // x <= y
2212 return (want_max ? yvalue : xvalue);
2213 case BoolTest::gt: // x > y
2214 case BoolTest::ge: // x >= y
2215 return (want_max ? xvalue : yvalue);
2216 }
2217 }
2218 }
2219
2220 // We failed to find a dominating test.
2221 // Let's pick a test that might GVN with prior tests.
2222 Node* best_bol = NULL;
2223 BoolTest::mask best_btest = BoolTest::illegal;
2224 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2225 Node* cmp = cmps[cmpn];
2226 if (cmp == NULL) continue;
2227 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2228 Node* bol = cmp->fast_out(j);
2229 if (!bol->is_Bool()) continue;
2230 BoolTest::mask btest = bol->as_Bool()->_test._test;
2231 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2232 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2233 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2234 best_bol = bol->as_Bool();
2235 best_btest = btest;
2236 }
2237 }
2238 }
2239
2240 Node* answer_if_true = NULL;
2241 Node* answer_if_false = NULL;
2242 switch (best_btest) {
2243 default:
2244 if (cmpxy == NULL)
2245 cmpxy = ideal_cmpxy;
2246 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2247 // and fall through:
2248 case BoolTest::lt: // x < y
2249 case BoolTest::le: // x <= y
2250 answer_if_true = (want_max ? yvalue : xvalue);
2251 answer_if_false = (want_max ? xvalue : yvalue);
2252 break;
2253 case BoolTest::gt: // x > y
2254 case BoolTest::ge: // x >= y
2255 answer_if_true = (want_max ? xvalue : yvalue);
2256 answer_if_false = (want_max ? yvalue : xvalue);
2257 break;
2258 }
2259
2260 jint hi, lo;
2261 if (want_max) {
2262 // We can sharpen the minimum.
2263 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2264 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2265 } else {
2266 // We can sharpen the maximum.
2267 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2268 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2269 }
2270
2271 // Use a flow-free graph structure, to avoid creating excess control edges
2272 // which could hinder other optimizations.
2273 // Since Math.min/max is often used with arraycopy, we want
2274 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2275 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2276 answer_if_false, answer_if_true,
2277 TypeInt::make(lo, hi, widen));
2278
2279 return _gvn.transform(cmov);
2280
2281 /*
2282 // This is not as desirable as it may seem, since Min and Max
2283 // nodes do not have a full set of optimizations.
2284 // And they would interfere, anyway, with 'if' optimizations
2285 // and with CMoveI canonical forms.
2286 switch (id) {
2287 case vmIntrinsics::_min:
2288 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2289 case vmIntrinsics::_max:
2290 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2291 default:
2292 ShouldNotReachHere();
2293 }
2294 */
2295 }
2296
2297 inline int
2298 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2299 const TypePtr* base_type = TypePtr::NULL_PTR;
2300 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2301 if (base_type == NULL) {
2302 // Unknown type.
2303 return Type::AnyPtr;
2304 } else if (base_type == TypePtr::NULL_PTR) {
2305 // Since this is a NULL+long form, we have to switch to a rawptr.
2306 base = _gvn.transform(new (C) CastX2PNode(offset));
2307 offset = MakeConX(0);
2308 return Type::RawPtr;
2309 } else if (base_type->base() == Type::RawPtr) {
2310 return Type::RawPtr;
2311 } else if (base_type->isa_oopptr()) {
2312 // Base is never null => always a heap address.
2313 if (base_type->ptr() == TypePtr::NotNull) {
2314 return Type::OopPtr;
2315 }
2316 // Offset is small => always a heap address.
2317 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2318 if (offset_type != NULL &&
2319 base_type->offset() == 0 && // (should always be?)
2320 offset_type->_lo >= 0 &&
2321 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2322 return Type::OopPtr;
2323 }
2324 // Otherwise, it might either be oop+off or NULL+addr.
2325 return Type::AnyPtr;
2326 } else {
2327 // No information:
2328 return Type::AnyPtr;
2329 }
2330 }
2331
2332 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2333 int kind = classify_unsafe_addr(base, offset);
2334 if (kind == Type::RawPtr) {
2335 return basic_plus_adr(top(), base, offset);
2336 } else {
2337 return basic_plus_adr(base, offset);
2338 }
2339 }
2340
2341 //--------------------------inline_number_methods-----------------------------
2342 // inline int Integer.numberOfLeadingZeros(int)
2343 // inline int Long.numberOfLeadingZeros(long)
2344 //
2345 // inline int Integer.numberOfTrailingZeros(int)
2346 // inline int Long.numberOfTrailingZeros(long)
2347 //
2348 // inline int Integer.bitCount(int)
2349 // inline int Long.bitCount(long)
2350 //
2351 // inline char Character.reverseBytes(char)
2352 // inline short Short.reverseBytes(short)
2353 // inline int Integer.reverseBytes(int)
2354 // inline long Long.reverseBytes(long)
2355 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2356 Node* arg = argument(0);
2357 Node* n;
2358 switch (id) {
2359 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2360 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2361 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2362 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2363 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2364 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2365 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2366 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2367 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2368 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2369 default: fatal_unexpected_iid(id); break;
2370 }
2371 set_result(_gvn.transform(n));
2372 return true;
2373 }
2374
2375 //----------------------------inline_unsafe_access----------------------------
2376
2377 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2378
2379 // Helper that guards and inserts a pre-barrier.
2380 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2381 Node* pre_val, bool need_mem_bar) {
2382 // We could be accessing the referent field of a reference object. If so, when G1
2383 // is enabled, we need to log the value in the referent field in an SATB buffer.
2384 // This routine performs some compile time filters and generates suitable
2385 // runtime filters that guard the pre-barrier code.
2386 // Also add memory barrier for non volatile load from the referent field
2387 // to prevent commoning of loads across safepoint.
2388 if (!UseG1GC && !need_mem_bar)
2389 return;
2390
2391 // Some compile time checks.
2392
2393 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2394 const TypeX* otype = offset->find_intptr_t_type();
2395 if (otype != NULL && otype->is_con() &&
2396 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2397 // Constant offset but not the reference_offset so just return
2398 return;
2399 }
2400
2401 // We only need to generate the runtime guards for instances.
2402 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2403 if (btype != NULL) {
2404 if (btype->isa_aryptr()) {
2405 // Array type so nothing to do
2406 return;
2407 }
2408
2409 const TypeInstPtr* itype = btype->isa_instptr();
2410 if (itype != NULL) {
2411 // Can the klass of base_oop be statically determined to be
2412 // _not_ a sub-class of Reference and _not_ Object?
2413 ciKlass* klass = itype->klass();
2414 if ( klass->is_loaded() &&
2415 !klass->is_subtype_of(env()->Reference_klass()) &&
2416 !env()->Object_klass()->is_subtype_of(klass)) {
2417 return;
2418 }
2419 }
2420 }
2421
2422 // The compile time filters did not reject base_oop/offset so
2423 // we need to generate the following runtime filters
2424 //
2425 // if (offset == java_lang_ref_Reference::_reference_offset) {
2426 // if (instance_of(base, java.lang.ref.Reference)) {
2427 // pre_barrier(_, pre_val, ...);
2428 // }
2429 // }
2430
2431 float likely = PROB_LIKELY( 0.999);
2432 float unlikely = PROB_UNLIKELY(0.999);
2433
2434 IdealKit ideal(this);
2435 #define __ ideal.
2436
2437 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2438
2439 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2440 // Update graphKit memory and control from IdealKit.
2441 sync_kit(ideal);
2442
2443 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2444 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2445
2446 // Update IdealKit memory and control from graphKit.
2447 __ sync_kit(this);
2448
2449 Node* one = __ ConI(1);
2450 // is_instof == 0 if base_oop == NULL
2451 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2452
2453 // Update graphKit from IdeakKit.
2454 sync_kit(ideal);
2455
2456 // Use the pre-barrier to record the value in the referent field
2457 pre_barrier(false /* do_load */,
2458 __ ctrl(),
2459 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2460 pre_val /* pre_val */,
2461 T_OBJECT);
2462 if (need_mem_bar) {
2463 // Add memory barrier to prevent commoning reads from this field
2464 // across safepoint since GC can change its value.
2465 insert_mem_bar(Op_MemBarCPUOrder);
2466 }
2467 // Update IdealKit from graphKit.
2468 __ sync_kit(this);
2469
2470 } __ end_if(); // _ref_type != ref_none
2471 } __ end_if(); // offset == referent_offset
2472
2473 // Final sync IdealKit and GraphKit.
2474 final_sync(ideal);
2475 #undef __
2476 }
2477
2478
2479 // Interpret Unsafe.fieldOffset cookies correctly:
2480 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2481
2482 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2483 // Attempt to infer a sharper value type from the offset and base type.
2484 ciKlass* sharpened_klass = NULL;
2485
2486 // See if it is an instance field, with an object type.
2487 if (alias_type->field() != NULL) {
2488 assert(!is_native_ptr, "native pointer op cannot use a java address");
2489 if (alias_type->field()->type()->is_klass()) {
2490 sharpened_klass = alias_type->field()->type()->as_klass();
2491 }
2492 }
2493
2494 // See if it is a narrow oop array.
2495 if (adr_type->isa_aryptr()) {
2496 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2497 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2498 if (elem_type != NULL) {
2499 sharpened_klass = elem_type->klass();
2500 }
2501 }
2502 }
2503
2504 // The sharpened class might be unloaded if there is no class loader
2505 // contraint in place.
2506 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2507 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2508
2509 #ifndef PRODUCT
2510 if (C->print_intrinsics() || C->print_inlining()) {
2511 tty->print(" from base type: "); adr_type->dump();
2512 tty->print(" sharpened value: "); tjp->dump();
2513 }
2514 #endif
2515 // Sharpen the value type.
2516 return tjp;
2517 }
2518 return NULL;
2519 }
2520
2521 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2522 if (callee()->is_static()) return false; // caller must have the capability!
2523
2524 #ifndef PRODUCT
2525 {
2526 ResourceMark rm;
2527 // Check the signatures.
2528 ciSignature* sig = callee()->signature();
2529 #ifdef ASSERT
2530 if (!is_store) {
2531 // Object getObject(Object base, int/long offset), etc.
2532 BasicType rtype = sig->return_type()->basic_type();
2533 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2534 rtype = T_ADDRESS; // it is really a C void*
2535 assert(rtype == type, "getter must return the expected value");
2536 if (!is_native_ptr) {
2537 assert(sig->count() == 2, "oop getter has 2 arguments");
2538 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2539 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2540 } else {
2541 assert(sig->count() == 1, "native getter has 1 argument");
2542 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2543 }
2544 } else {
2545 // void putObject(Object base, int/long offset, Object x), etc.
2546 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2547 if (!is_native_ptr) {
2548 assert(sig->count() == 3, "oop putter has 3 arguments");
2549 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2550 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2551 } else {
2552 assert(sig->count() == 2, "native putter has 2 arguments");
2553 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2554 }
2555 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2556 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2557 vtype = T_ADDRESS; // it is really a C void*
2558 assert(vtype == type, "putter must accept the expected value");
2559 }
2560 #endif // ASSERT
2561 }
2562 #endif //PRODUCT
2563
2564 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2565
2566 Node* receiver = argument(0); // type: oop
2567
2568 // Build address expression. See the code in inline_unsafe_prefetch.
2569 Node* adr;
2570 Node* heap_base_oop = top();
2571 Node* offset = top();
2572 Node* val;
2573
2574 if (!is_native_ptr) {
2575 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2576 Node* base = argument(1); // type: oop
2577 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2578 offset = argument(2); // type: long
2579 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2580 // to be plain byte offsets, which are also the same as those accepted
2581 // by oopDesc::field_base.
2582 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2583 "fieldOffset must be byte-scaled");
2584 // 32-bit machines ignore the high half!
2585 offset = ConvL2X(offset);
2586 adr = make_unsafe_address(base, offset);
2587 heap_base_oop = base;
2588 val = is_store ? argument(4) : NULL;
2589 } else {
2590 Node* ptr = argument(1); // type: long
2591 ptr = ConvL2X(ptr); // adjust Java long to machine word
2592 adr = make_unsafe_address(NULL, ptr);
2593 val = is_store ? argument(3) : NULL;
2594 }
2595
2596 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2597
2598 // First guess at the value type.
2599 const Type *value_type = Type::get_const_basic_type(type);
2600
2601 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2602 // there was not enough information to nail it down.
2603 Compile::AliasType* alias_type = C->alias_type(adr_type);
2604 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2605
2606 // We will need memory barriers unless we can determine a unique
2607 // alias category for this reference. (Note: If for some reason
2608 // the barriers get omitted and the unsafe reference begins to "pollute"
2609 // the alias analysis of the rest of the graph, either Compile::can_alias
2610 // or Compile::must_alias will throw a diagnostic assert.)
2611 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2612
2613 // If we are reading the value of the referent field of a Reference
2614 // object (either by using Unsafe directly or through reflection)
2615 // then, if G1 is enabled, we need to record the referent in an
2616 // SATB log buffer using the pre-barrier mechanism.
2617 // Also we need to add memory barrier to prevent commoning reads
2618 // from this field across safepoint since GC can change its value.
2619 bool need_read_barrier = !is_native_ptr && !is_store &&
2620 offset != top() && heap_base_oop != top();
2621
2622 if (!is_store && type == T_OBJECT) {
2623 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2624 if (tjp != NULL) {
2625 value_type = tjp;
2626 }
2627 }
2628
2629 receiver = null_check(receiver);
2630 if (stopped()) {
2631 return true;
2632 }
2633 // Heap pointers get a null-check from the interpreter,
2634 // as a courtesy. However, this is not guaranteed by Unsafe,
2635 // and it is not possible to fully distinguish unintended nulls
2636 // from intended ones in this API.
2637
2638 if (is_volatile) {
2639 // We need to emit leading and trailing CPU membars (see below) in
2640 // addition to memory membars when is_volatile. This is a little
2641 // too strong, but avoids the need to insert per-alias-type
2642 // volatile membars (for stores; compare Parse::do_put_xxx), which
2643 // we cannot do effectively here because we probably only have a
2644 // rough approximation of type.
2645 need_mem_bar = true;
2646 // For Stores, place a memory ordering barrier now.
2647 if (is_store) {
2648 insert_mem_bar(Op_MemBarRelease);
2649 } else {
2650 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2651 insert_mem_bar(Op_MemBarVolatile);
2652 }
2653 }
2654 }
2655
2656 // Memory barrier to prevent normal and 'unsafe' accesses from
2657 // bypassing each other. Happens after null checks, so the
2658 // exception paths do not take memory state from the memory barrier,
2659 // so there's no problems making a strong assert about mixing users
2660 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2661 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2662 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2663
2664 if (!is_store) {
2665 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2666 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2667 // load value
2668 switch (type) {
2669 case T_BOOLEAN:
2670 case T_CHAR:
2671 case T_BYTE:
2672 case T_SHORT:
2673 case T_INT:
2674 case T_LONG:
2675 case T_FLOAT:
2676 case T_DOUBLE:
2677 break;
2678 case T_OBJECT:
2679 if (need_read_barrier) {
2680 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2681 }
2682 break;
2683 case T_ADDRESS:
2684 // Cast to an int type.
2685 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2686 p = ConvX2UL(p);
2687 break;
2688 default:
2689 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2690 break;
2691 }
2692 // The load node has the control of the preceding MemBarCPUOrder. All
2693 // following nodes will have the control of the MemBarCPUOrder inserted at
2694 // the end of this method. So, pushing the load onto the stack at a later
2695 // point is fine.
2696 set_result(p);
2697 } else {
2698 // place effect of store into memory
2699 switch (type) {
2700 case T_DOUBLE:
2701 val = dstore_rounding(val);
2702 break;
2703 case T_ADDRESS:
2704 // Repackage the long as a pointer.
2705 val = ConvL2X(val);
2706 val = _gvn.transform(new (C) CastX2PNode(val));
2707 break;
2708 }
2709
2710 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2711 if (type != T_OBJECT ) {
2712 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2713 } else {
2714 // Possibly an oop being stored to Java heap or native memory
2715 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2716 // oop to Java heap.
2717 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2718 } else {
2719 // We can't tell at compile time if we are storing in the Java heap or outside
2720 // of it. So we need to emit code to conditionally do the proper type of
2721 // store.
2722
2723 IdealKit ideal(this);
2724 #define __ ideal.
2725 // QQQ who knows what probability is here??
2726 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2727 // Sync IdealKit and graphKit.
2728 sync_kit(ideal);
2729 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2730 // Update IdealKit memory.
2731 __ sync_kit(this);
2732 } __ else_(); {
2733 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2734 } __ end_if();
2735 // Final sync IdealKit and GraphKit.
2736 final_sync(ideal);
2737 #undef __
2738 }
2739 }
2740 }
2741
2742 if (is_volatile) {
2743 if (!is_store) {
2744 insert_mem_bar(Op_MemBarAcquire);
2745 } else {
2746 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2747 insert_mem_bar(Op_MemBarVolatile);
2748 }
2749 }
2750 }
2751
2752 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2753
2754 return true;
2755 }
2756
2757 //----------------------------inline_unsafe_prefetch----------------------------
2758
2759 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2760 #ifndef PRODUCT
2761 {
2762 ResourceMark rm;
2763 // Check the signatures.
2764 ciSignature* sig = callee()->signature();
2765 #ifdef ASSERT
2766 // Object getObject(Object base, int/long offset), etc.
2767 BasicType rtype = sig->return_type()->basic_type();
2768 if (!is_native_ptr) {
2769 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2770 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2771 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2772 } else {
2773 assert(sig->count() == 1, "native prefetch has 1 argument");
2774 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2775 }
2776 #endif // ASSERT
2777 }
2778 #endif // !PRODUCT
2779
2780 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2781
2782 const int idx = is_static ? 0 : 1;
2783 if (!is_static) {
2784 null_check_receiver();
2785 if (stopped()) {
2786 return true;
2787 }
2788 }
2789
2790 // Build address expression. See the code in inline_unsafe_access.
2791 Node *adr;
2792 if (!is_native_ptr) {
2793 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2794 Node* base = argument(idx + 0); // type: oop
2795 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2796 Node* offset = argument(idx + 1); // type: long
2797 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2798 // to be plain byte offsets, which are also the same as those accepted
2799 // by oopDesc::field_base.
2800 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2801 "fieldOffset must be byte-scaled");
2802 // 32-bit machines ignore the high half!
2803 offset = ConvL2X(offset);
2804 adr = make_unsafe_address(base, offset);
2805 } else {
2806 Node* ptr = argument(idx + 0); // type: long
2807 ptr = ConvL2X(ptr); // adjust Java long to machine word
2808 adr = make_unsafe_address(NULL, ptr);
2809 }
2810
2811 // Generate the read or write prefetch
2812 Node *prefetch;
2813 if (is_store) {
2814 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2815 } else {
2816 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2817 }
2818 prefetch->init_req(0, control());
2819 set_i_o(_gvn.transform(prefetch));
2820
2821 return true;
2822 }
2823
2824 //----------------------------inline_unsafe_load_store----------------------------
2825 // This method serves a couple of different customers (depending on LoadStoreKind):
2826 //
2827 // LS_cmpxchg:
2828 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2829 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2830 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2831 //
2832 // LS_xadd:
2833 // public int getAndAddInt( Object o, long offset, int delta)
2834 // public long getAndAddLong(Object o, long offset, long delta)
2835 //
2836 // LS_xchg:
2837 // int getAndSet(Object o, long offset, int newValue)
2838 // long getAndSet(Object o, long offset, long newValue)
2839 // Object getAndSet(Object o, long offset, Object newValue)
2840 //
2841 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2842 // This basic scheme here is the same as inline_unsafe_access, but
2843 // differs in enough details that combining them would make the code
2844 // overly confusing. (This is a true fact! I originally combined
2845 // them, but even I was confused by it!) As much code/comments as
2846 // possible are retained from inline_unsafe_access though to make
2847 // the correspondences clearer. - dl
2848
2849 if (callee()->is_static()) return false; // caller must have the capability!
2850
2851 #ifndef PRODUCT
2852 BasicType rtype;
2853 {
2854 ResourceMark rm;
2855 // Check the signatures.
2856 ciSignature* sig = callee()->signature();
2857 rtype = sig->return_type()->basic_type();
2858 if (kind == LS_xadd || kind == LS_xchg) {
2859 // Check the signatures.
2860 #ifdef ASSERT
2861 assert(rtype == type, "get and set must return the expected type");
2862 assert(sig->count() == 3, "get and set has 3 arguments");
2863 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2864 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2865 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2866 #endif // ASSERT
2867 } else if (kind == LS_cmpxchg) {
2868 // Check the signatures.
2869 #ifdef ASSERT
2870 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2871 assert(sig->count() == 4, "CAS has 4 arguments");
2872 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2873 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2874 #endif // ASSERT
2875 } else {
2876 ShouldNotReachHere();
2877 }
2878 }
2879 #endif //PRODUCT
2880
2881 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2882
2883 // Get arguments:
2884 Node* receiver = NULL;
2885 Node* base = NULL;
2886 Node* offset = NULL;
2887 Node* oldval = NULL;
2888 Node* newval = NULL;
2889 if (kind == LS_cmpxchg) {
2890 const bool two_slot_type = type2size[type] == 2;
2891 receiver = argument(0); // type: oop
2892 base = argument(1); // type: oop
2893 offset = argument(2); // type: long
2894 oldval = argument(4); // type: oop, int, or long
2895 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2896 } else if (kind == LS_xadd || kind == LS_xchg){
2897 receiver = argument(0); // type: oop
2898 base = argument(1); // type: oop
2899 offset = argument(2); // type: long
2900 oldval = NULL;
2901 newval = argument(4); // type: oop, int, or long
2902 }
2903
2904 // Null check receiver.
2905 receiver = null_check(receiver);
2906 if (stopped()) {
2907 return true;
2908 }
2909
2910 // Build field offset expression.
2911 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2912 // to be plain byte offsets, which are also the same as those accepted
2913 // by oopDesc::field_base.
2914 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2915 // 32-bit machines ignore the high half of long offsets
2916 offset = ConvL2X(offset);
2917 Node* adr = make_unsafe_address(base, offset);
2918 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2919
2920 // For CAS, unlike inline_unsafe_access, there seems no point in
2921 // trying to refine types. Just use the coarse types here.
2922 const Type *value_type = Type::get_const_basic_type(type);
2923 Compile::AliasType* alias_type = C->alias_type(adr_type);
2924 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2925
2926 if (kind == LS_xchg && type == T_OBJECT) {
2927 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2928 if (tjp != NULL) {
2929 value_type = tjp;
2930 }
2931 }
2932
2933 int alias_idx = C->get_alias_index(adr_type);
2934
2935 // Memory-model-wise, a LoadStore acts like a little synchronized
2936 // block, so needs barriers on each side. These don't translate
2937 // into actual barriers on most machines, but we still need rest of
2938 // compiler to respect ordering.
2939
2940 insert_mem_bar(Op_MemBarRelease);
2941 insert_mem_bar(Op_MemBarCPUOrder);
2942
2943 // 4984716: MemBars must be inserted before this
2944 // memory node in order to avoid a false
2945 // dependency which will confuse the scheduler.
2946 Node *mem = memory(alias_idx);
2947
2948 // For now, we handle only those cases that actually exist: ints,
2949 // longs, and Object. Adding others should be straightforward.
2950 Node* load_store;
2951 switch(type) {
2952 case T_INT:
2953 if (kind == LS_xadd) {
2954 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2955 } else if (kind == LS_xchg) {
2956 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2957 } else if (kind == LS_cmpxchg) {
2958 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2959 } else {
2960 ShouldNotReachHere();
2961 }
2962 break;
2963 case T_LONG:
2964 if (kind == LS_xadd) {
2965 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2966 } else if (kind == LS_xchg) {
2967 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2968 } else if (kind == LS_cmpxchg) {
2969 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2970 } else {
2971 ShouldNotReachHere();
2972 }
2973 break;
2974 case T_OBJECT:
2975 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2976 // could be delayed during Parse (for example, in adjust_map_after_if()).
2977 // Execute transformation here to avoid barrier generation in such case.
2978 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2979 newval = _gvn.makecon(TypePtr::NULL_PTR);
2980
2981 // Reference stores need a store barrier.
2982 if (kind == LS_xchg) {
2983 // If pre-barrier must execute before the oop store, old value will require do_load here.
2984 if (!can_move_pre_barrier()) {
2985 pre_barrier(true /* do_load*/,
2986 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2987 NULL /* pre_val*/,
2988 T_OBJECT);
2989 } // Else move pre_barrier to use load_store value, see below.
2990 } else if (kind == LS_cmpxchg) {
2991 // Same as for newval above:
2992 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2993 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2994 }
2995 // The only known value which might get overwritten is oldval.
2996 pre_barrier(false /* do_load */,
2997 control(), NULL, NULL, max_juint, NULL, NULL,
2998 oldval /* pre_val */,
2999 T_OBJECT);
3000 } else {
3001 ShouldNotReachHere();
3002 }
3003
3004 #ifdef _LP64
3005 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3006 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3007 if (kind == LS_xchg) {
3008 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3009 newval_enc, adr_type, value_type->make_narrowoop()));
3010 } else {
3011 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3012 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3013 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3014 newval_enc, oldval_enc));
3015 }
3016 } else
3017 #endif
3018 {
3019 if (kind == LS_xchg) {
3020 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3021 } else {
3022 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3023 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3024 }
3025 }
3026 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3027 break;
3028 default:
3029 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3030 break;
3031 }
3032
3033 // SCMemProjNodes represent the memory state of a LoadStore. Their
3034 // main role is to prevent LoadStore nodes from being optimized away
3035 // when their results aren't used.
3036 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3037 set_memory(proj, alias_idx);
3038
3039 if (type == T_OBJECT && kind == LS_xchg) {
3040 #ifdef _LP64
3041 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3042 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3043 }
3044 #endif
3045 if (can_move_pre_barrier()) {
3046 // Don't need to load pre_val. The old value is returned by load_store.
3047 // The pre_barrier can execute after the xchg as long as no safepoint
3048 // gets inserted between them.
3049 pre_barrier(false /* do_load */,
3050 control(), NULL, NULL, max_juint, NULL, NULL,
3051 load_store /* pre_val */,
3052 T_OBJECT);
3053 }
3054 }
3055
3056 // Add the trailing membar surrounding the access
3057 insert_mem_bar(Op_MemBarCPUOrder);
3058 insert_mem_bar(Op_MemBarAcquire);
3059
3060 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3061 set_result(load_store);
3062 return true;
3063 }
3064
3065 //----------------------------inline_unsafe_ordered_store----------------------
3066 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3067 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3068 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3069 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3070 // This is another variant of inline_unsafe_access, differing in
3071 // that it always issues store-store ("release") barrier and ensures
3072 // store-atomicity (which only matters for "long").
3073
3074 if (callee()->is_static()) return false; // caller must have the capability!
3075
3076 #ifndef PRODUCT
3077 {
3078 ResourceMark rm;
3079 // Check the signatures.
3080 ciSignature* sig = callee()->signature();
3081 #ifdef ASSERT
3082 BasicType rtype = sig->return_type()->basic_type();
3083 assert(rtype == T_VOID, "must return void");
3084 assert(sig->count() == 3, "has 3 arguments");
3085 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3086 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3087 #endif // ASSERT
3088 }
3089 #endif //PRODUCT
3090
3091 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3092
3093 // Get arguments:
3094 Node* receiver = argument(0); // type: oop
3095 Node* base = argument(1); // type: oop
3096 Node* offset = argument(2); // type: long
3097 Node* val = argument(4); // type: oop, int, or long
3098
3099 // Null check receiver.
3100 receiver = null_check(receiver);
3101 if (stopped()) {
3102 return true;
3103 }
3104
3105 // Build field offset expression.
3106 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3107 // 32-bit machines ignore the high half of long offsets
3108 offset = ConvL2X(offset);
3109 Node* adr = make_unsafe_address(base, offset);
3110 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3111 const Type *value_type = Type::get_const_basic_type(type);
3112 Compile::AliasType* alias_type = C->alias_type(adr_type);
3113
3114 insert_mem_bar(Op_MemBarRelease);
3115 insert_mem_bar(Op_MemBarCPUOrder);
3116 // Ensure that the store is atomic for longs:
3117 const bool require_atomic_access = true;
3118 Node* store;
3119 if (type == T_OBJECT) // reference stores need a store barrier.
3120 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3121 else {
3122 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3123 }
3124 insert_mem_bar(Op_MemBarCPUOrder);
3125 return true;
3126 }
3127
3128 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3129 // Regardless of form, don't allow previous ld/st to move down,
3130 // then issue acquire, release, or volatile mem_bar.
3131 insert_mem_bar(Op_MemBarCPUOrder);
3132 switch(id) {
3133 case vmIntrinsics::_loadFence:
3134 insert_mem_bar(Op_LoadFence);
3135 return true;
3136 case vmIntrinsics::_storeFence:
3137 insert_mem_bar(Op_StoreFence);
3138 return true;
3139 case vmIntrinsics::_fullFence:
3140 insert_mem_bar(Op_MemBarVolatile);
3141 return true;
3142 default:
3143 fatal_unexpected_iid(id);
3144 return false;
3145 }
3146 }
3147
3148 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3149 if (!kls->is_Con()) {
3150 return true;
3151 }
3152 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3153 if (klsptr == NULL) {
3154 return true;
3155 }
3156 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3157 // don't need a guard for a klass that is already initialized
3158 return !ik->is_initialized();
3159 }
3160
3161 //----------------------------inline_unsafe_allocate---------------------------
3162 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3163 bool LibraryCallKit::inline_unsafe_allocate() {
3164 if (callee()->is_static()) return false; // caller must have the capability!
3165
3166 null_check_receiver(); // null-check, then ignore
3167 Node* cls = null_check(argument(1));
3168 if (stopped()) return true;
3169
3170 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3171 kls = null_check(kls);
3172 if (stopped()) return true; // argument was like int.class
3173
3174 Node* test = NULL;
3175 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3176 // Note: The argument might still be an illegal value like
3177 // Serializable.class or Object[].class. The runtime will handle it.
3178 // But we must make an explicit check for initialization.
3179 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3180 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3181 // can generate code to load it as unsigned byte.
3182 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3183 Node* bits = intcon(InstanceKlass::fully_initialized);
3184 test = _gvn.transform(new (C) SubINode(inst, bits));
3185 // The 'test' is non-zero if we need to take a slow path.
3186 }
3187
3188 Node* obj = new_instance(kls, test);
3189 set_result(obj);
3190 return true;
3191 }
3192
3193 #ifdef TRACE_HAVE_INTRINSICS
3194 /*
3195 * oop -> myklass
3196 * myklass->trace_id |= USED
3197 * return myklass->trace_id & ~0x3
3198 */
3199 bool LibraryCallKit::inline_native_classID() {
3200 null_check_receiver(); // null-check, then ignore
3201 Node* cls = null_check(argument(1), T_OBJECT);
3202 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3203 kls = null_check(kls, T_OBJECT);
3204 ByteSize offset = TRACE_ID_OFFSET;
3205 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3206 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3207 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3208 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3209 Node* clsused = longcon(0x01l); // set the class bit
3210 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3211
3212 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3213 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3214 set_result(andl);
3215 return true;
3216 }
3217
3218 bool LibraryCallKit::inline_native_threadID() {
3219 Node* tls_ptr = NULL;
3220 Node* cur_thr = generate_current_thread(tls_ptr);
3221 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3222 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3223 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3224
3225 Node* threadid = NULL;
3226 size_t thread_id_size = OSThread::thread_id_size();
3227 if (thread_id_size == (size_t) BytesPerLong) {
3228 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3229 } else if (thread_id_size == (size_t) BytesPerInt) {
3230 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3231 } else {
3232 ShouldNotReachHere();
3233 }
3234 set_result(threadid);
3235 return true;
3236 }
3237 #endif
3238
3239 //------------------------inline_native_time_funcs--------------
3240 // inline code for System.currentTimeMillis() and System.nanoTime()
3241 // these have the same type and signature
3242 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3243 const TypeFunc* tf = OptoRuntime::void_long_Type();
3244 const TypePtr* no_memory_effects = NULL;
3245 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3246 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3247 #ifdef ASSERT
3248 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3249 assert(value_top == top(), "second value must be top");
3250 #endif
3251 set_result(value);
3252 return true;
3253 }
3254
3255 //------------------------inline_native_currentThread------------------
3256 bool LibraryCallKit::inline_native_currentThread() {
3257 Node* junk = NULL;
3258 set_result(generate_current_thread(junk));
3259 return true;
3260 }
3261
3262 //------------------------inline_native_isInterrupted------------------
3263 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3264 bool LibraryCallKit::inline_native_isInterrupted() {
3265 // Add a fast path to t.isInterrupted(clear_int):
3266 // (t == Thread.current() &&
3267 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3268 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3269 // So, in the common case that the interrupt bit is false,
3270 // we avoid making a call into the VM. Even if the interrupt bit
3271 // is true, if the clear_int argument is false, we avoid the VM call.
3272 // However, if the receiver is not currentThread, we must call the VM,
3273 // because there must be some locking done around the operation.
3274
3275 // We only go to the fast case code if we pass two guards.
3276 // Paths which do not pass are accumulated in the slow_region.
3277
3278 enum {
3279 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3280 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3281 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3282 PATH_LIMIT
3283 };
3284
3285 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3286 // out of the function.
3287 insert_mem_bar(Op_MemBarCPUOrder);
3288
3289 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3290 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3291
3292 RegionNode* slow_region = new (C) RegionNode(1);
3293 record_for_igvn(slow_region);
3294
3295 // (a) Receiving thread must be the current thread.
3296 Node* rec_thr = argument(0);
3297 Node* tls_ptr = NULL;
3298 Node* cur_thr = generate_current_thread(tls_ptr);
3299 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3300 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3301
3302 generate_slow_guard(bol_thr, slow_region);
3303
3304 // (b) Interrupt bit on TLS must be false.
3305 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3306 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3307 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3308
3309 // Set the control input on the field _interrupted read to prevent it floating up.
3310 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3311 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3312 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3313
3314 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3315
3316 // First fast path: if (!TLS._interrupted) return false;
3317 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3318 result_rgn->init_req(no_int_result_path, false_bit);
3319 result_val->init_req(no_int_result_path, intcon(0));
3320
3321 // drop through to next case
3322 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3323
3324 #ifndef TARGET_OS_FAMILY_windows
3325 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3326 Node* clr_arg = argument(1);
3327 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3328 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3329 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3330
3331 // Second fast path: ... else if (!clear_int) return true;
3332 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3333 result_rgn->init_req(no_clear_result_path, false_arg);
3334 result_val->init_req(no_clear_result_path, intcon(1));
3335
3336 // drop through to next case
3337 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3338 #else
3339 // To return true on Windows you must read the _interrupted field
3340 // and check the the event state i.e. take the slow path.
3341 #endif // TARGET_OS_FAMILY_windows
3342
3343 // (d) Otherwise, go to the slow path.
3344 slow_region->add_req(control());
3345 set_control( _gvn.transform(slow_region));
3346
3347 if (stopped()) {
3348 // There is no slow path.
3349 result_rgn->init_req(slow_result_path, top());
3350 result_val->init_req(slow_result_path, top());
3351 } else {
3352 // non-virtual because it is a private non-static
3353 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3354
3355 Node* slow_val = set_results_for_java_call(slow_call);
3356 // this->control() comes from set_results_for_java_call
3357
3358 Node* fast_io = slow_call->in(TypeFunc::I_O);
3359 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3360
3361 // These two phis are pre-filled with copies of of the fast IO and Memory
3362 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3363 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3364
3365 result_rgn->init_req(slow_result_path, control());
3366 result_io ->init_req(slow_result_path, i_o());
3367 result_mem->init_req(slow_result_path, reset_memory());
3368 result_val->init_req(slow_result_path, slow_val);
3369
3370 set_all_memory(_gvn.transform(result_mem));
3371 set_i_o( _gvn.transform(result_io));
3372 }
3373
3374 C->set_has_split_ifs(true); // Has chance for split-if optimization
3375 set_result(result_rgn, result_val);
3376 return true;
3377 }
3378
3379 //---------------------------load_mirror_from_klass----------------------------
3380 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3381 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3382 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3383 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3384 }
3385
3386 //-----------------------load_klass_from_mirror_common-------------------------
3387 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3388 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3389 // and branch to the given path on the region.
3390 // If never_see_null, take an uncommon trap on null, so we can optimistically
3391 // compile for the non-null case.
3392 // If the region is NULL, force never_see_null = true.
3393 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3394 bool never_see_null,
3395 RegionNode* region,
3396 int null_path,
3397 int offset) {
3398 if (region == NULL) never_see_null = true;
3399 Node* p = basic_plus_adr(mirror, offset);
3400 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3401 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3402 Node* null_ctl = top();
3403 kls = null_check_oop(kls, &null_ctl, never_see_null);
3404 if (region != NULL) {
3405 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3406 region->init_req(null_path, null_ctl);
3407 } else {
3408 assert(null_ctl == top(), "no loose ends");
3409 }
3410 return kls;
3411 }
3412
3413 //--------------------(inline_native_Class_query helpers)---------------------
3414 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3415 // Fall through if (mods & mask) == bits, take the guard otherwise.
3416 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3417 // Branch around if the given klass has the given modifier bit set.
3418 // Like generate_guard, adds a new path onto the region.
3419 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3420 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3421 Node* mask = intcon(modifier_mask);
3422 Node* bits = intcon(modifier_bits);
3423 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3424 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3425 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3426 return generate_fair_guard(bol, region);
3427 }
3428 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3429 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3430 }
3431
3432 //-------------------------inline_native_Class_query-------------------
3433 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3434 const Type* return_type = TypeInt::BOOL;
3435 Node* prim_return_value = top(); // what happens if it's a primitive class?
3436 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3437 bool expect_prim = false; // most of these guys expect to work on refs
3438
3439 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3440
3441 Node* mirror = argument(0);
3442 Node* obj = top();
3443
3444 switch (id) {
3445 case vmIntrinsics::_isInstance:
3446 // nothing is an instance of a primitive type
3447 prim_return_value = intcon(0);
3448 obj = argument(1);
3449 break;
3450 case vmIntrinsics::_getModifiers:
3451 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3452 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3453 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3454 break;
3455 case vmIntrinsics::_isInterface:
3456 prim_return_value = intcon(0);
3457 break;
3458 case vmIntrinsics::_isArray:
3459 prim_return_value = intcon(0);
3460 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3461 break;
3462 case vmIntrinsics::_isPrimitive:
3463 prim_return_value = intcon(1);
3464 expect_prim = true; // obviously
3465 break;
3466 case vmIntrinsics::_getSuperclass:
3467 prim_return_value = null();
3468 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3469 break;
3470 case vmIntrinsics::_getComponentType:
3471 prim_return_value = null();
3472 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3473 break;
3474 case vmIntrinsics::_getClassAccessFlags:
3475 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3476 return_type = TypeInt::INT; // not bool! 6297094
3477 break;
3478 default:
3479 fatal_unexpected_iid(id);
3480 break;
3481 }
3482
3483 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3484 if (mirror_con == NULL) return false; // cannot happen?
3485
3486 #ifndef PRODUCT
3487 if (C->print_intrinsics() || C->print_inlining()) {
3488 ciType* k = mirror_con->java_mirror_type();
3489 if (k) {
3490 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3491 k->print_name();
3492 tty->cr();
3493 }
3494 }
3495 #endif
3496
3497 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3498 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3499 record_for_igvn(region);
3500 PhiNode* phi = new (C) PhiNode(region, return_type);
3501
3502 // The mirror will never be null of Reflection.getClassAccessFlags, however
3503 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3504 // if it is. See bug 4774291.
3505
3506 // For Reflection.getClassAccessFlags(), the null check occurs in
3507 // the wrong place; see inline_unsafe_access(), above, for a similar
3508 // situation.
3509 mirror = null_check(mirror);
3510 // If mirror or obj is dead, only null-path is taken.
3511 if (stopped()) return true;
3512
3513 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3514
3515 // Now load the mirror's klass metaobject, and null-check it.
3516 // Side-effects region with the control path if the klass is null.
3517 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3518 // If kls is null, we have a primitive mirror.
3519 phi->init_req(_prim_path, prim_return_value);
3520 if (stopped()) { set_result(region, phi); return true; }
3521 bool safe_for_replace = (region->in(_prim_path) == top());
3522
3523 Node* p; // handy temp
3524 Node* null_ctl;
3525
3526 // Now that we have the non-null klass, we can perform the real query.
3527 // For constant classes, the query will constant-fold in LoadNode::Value.
3528 Node* query_value = top();
3529 switch (id) {
3530 case vmIntrinsics::_isInstance:
3531 // nothing is an instance of a primitive type
3532 query_value = gen_instanceof(obj, kls, safe_for_replace);
3533 break;
3534
3535 case vmIntrinsics::_getModifiers:
3536 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3537 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3538 break;
3539
3540 case vmIntrinsics::_isInterface:
3541 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3542 if (generate_interface_guard(kls, region) != NULL)
3543 // A guard was added. If the guard is taken, it was an interface.
3544 phi->add_req(intcon(1));
3545 // If we fall through, it's a plain class.
3546 query_value = intcon(0);
3547 break;
3548
3549 case vmIntrinsics::_isArray:
3550 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3551 if (generate_array_guard(kls, region) != NULL)
3552 // A guard was added. If the guard is taken, it was an array.
3553 phi->add_req(intcon(1));
3554 // If we fall through, it's a plain class.
3555 query_value = intcon(0);
3556 break;
3557
3558 case vmIntrinsics::_isPrimitive:
3559 query_value = intcon(0); // "normal" path produces false
3560 break;
3561
3562 case vmIntrinsics::_getSuperclass:
3563 // The rules here are somewhat unfortunate, but we can still do better
3564 // with random logic than with a JNI call.
3565 // Interfaces store null or Object as _super, but must report null.
3566 // Arrays store an intermediate super as _super, but must report Object.
3567 // Other types can report the actual _super.
3568 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3569 if (generate_interface_guard(kls, region) != NULL)
3570 // A guard was added. If the guard is taken, it was an interface.
3571 phi->add_req(null());
3572 if (generate_array_guard(kls, region) != NULL)
3573 // A guard was added. If the guard is taken, it was an array.
3574 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3575 // If we fall through, it's a plain class. Get its _super.
3576 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3577 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3578 null_ctl = top();
3579 kls = null_check_oop(kls, &null_ctl);
3580 if (null_ctl != top()) {
3581 // If the guard is taken, Object.superClass is null (both klass and mirror).
3582 region->add_req(null_ctl);
3583 phi ->add_req(null());
3584 }
3585 if (!stopped()) {
3586 query_value = load_mirror_from_klass(kls);
3587 }
3588 break;
3589
3590 case vmIntrinsics::_getComponentType:
3591 if (generate_array_guard(kls, region) != NULL) {
3592 // Be sure to pin the oop load to the guard edge just created:
3593 Node* is_array_ctrl = region->in(region->req()-1);
3594 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3595 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3596 phi->add_req(cmo);
3597 }
3598 query_value = null(); // non-array case is null
3599 break;
3600
3601 case vmIntrinsics::_getClassAccessFlags:
3602 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3603 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3604 break;
3605
3606 default:
3607 fatal_unexpected_iid(id);
3608 break;
3609 }
3610
3611 // Fall-through is the normal case of a query to a real class.
3612 phi->init_req(1, query_value);
3613 region->init_req(1, control());
3614
3615 C->set_has_split_ifs(true); // Has chance for split-if optimization
3616 set_result(region, phi);
3617 return true;
3618 }
3619
3620 //--------------------------inline_native_subtype_check------------------------
3621 // This intrinsic takes the JNI calls out of the heart of
3622 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3623 bool LibraryCallKit::inline_native_subtype_check() {
3624 // Pull both arguments off the stack.
3625 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3626 args[0] = argument(0);
3627 args[1] = argument(1);
3628 Node* klasses[2]; // corresponding Klasses: superk, subk
3629 klasses[0] = klasses[1] = top();
3630
3631 enum {
3632 // A full decision tree on {superc is prim, subc is prim}:
3633 _prim_0_path = 1, // {P,N} => false
3634 // {P,P} & superc!=subc => false
3635 _prim_same_path, // {P,P} & superc==subc => true
3636 _prim_1_path, // {N,P} => false
3637 _ref_subtype_path, // {N,N} & subtype check wins => true
3638 _both_ref_path, // {N,N} & subtype check loses => false
3639 PATH_LIMIT
3640 };
3641
3642 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3643 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3644 record_for_igvn(region);
3645
3646 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3647 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3648 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3649
3650 // First null-check both mirrors and load each mirror's klass metaobject.
3651 int which_arg;
3652 for (which_arg = 0; which_arg <= 1; which_arg++) {
3653 Node* arg = args[which_arg];
3654 arg = null_check(arg);
3655 if (stopped()) break;
3656 args[which_arg] = arg;
3657
3658 Node* p = basic_plus_adr(arg, class_klass_offset);
3659 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3660 klasses[which_arg] = _gvn.transform(kls);
3661 }
3662
3663 // Having loaded both klasses, test each for null.
3664 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3665 for (which_arg = 0; which_arg <= 1; which_arg++) {
3666 Node* kls = klasses[which_arg];
3667 Node* null_ctl = top();
3668 kls = null_check_oop(kls, &null_ctl, never_see_null);
3669 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3670 region->init_req(prim_path, null_ctl);
3671 if (stopped()) break;
3672 klasses[which_arg] = kls;
3673 }
3674
3675 if (!stopped()) {
3676 // now we have two reference types, in klasses[0..1]
3677 Node* subk = klasses[1]; // the argument to isAssignableFrom
3678 Node* superk = klasses[0]; // the receiver
3679 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3680 // now we have a successful reference subtype check
3681 region->set_req(_ref_subtype_path, control());
3682 }
3683
3684 // If both operands are primitive (both klasses null), then
3685 // we must return true when they are identical primitives.
3686 // It is convenient to test this after the first null klass check.
3687 set_control(region->in(_prim_0_path)); // go back to first null check
3688 if (!stopped()) {
3689 // Since superc is primitive, make a guard for the superc==subc case.
3690 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3691 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3692 generate_guard(bol_eq, region, PROB_FAIR);
3693 if (region->req() == PATH_LIMIT+1) {
3694 // A guard was added. If the added guard is taken, superc==subc.
3695 region->swap_edges(PATH_LIMIT, _prim_same_path);
3696 region->del_req(PATH_LIMIT);
3697 }
3698 region->set_req(_prim_0_path, control()); // Not equal after all.
3699 }
3700
3701 // these are the only paths that produce 'true':
3702 phi->set_req(_prim_same_path, intcon(1));
3703 phi->set_req(_ref_subtype_path, intcon(1));
3704
3705 // pull together the cases:
3706 assert(region->req() == PATH_LIMIT, "sane region");
3707 for (uint i = 1; i < region->req(); i++) {
3708 Node* ctl = region->in(i);
3709 if (ctl == NULL || ctl == top()) {
3710 region->set_req(i, top());
3711 phi ->set_req(i, top());
3712 } else if (phi->in(i) == NULL) {
3713 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3714 }
3715 }
3716
3717 set_control(_gvn.transform(region));
3718 set_result(_gvn.transform(phi));
3719 return true;
3720 }
3721
3722 //---------------------generate_array_guard_common------------------------
3723 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3724 bool obj_array, bool not_array) {
3725 // If obj_array/non_array==false/false:
3726 // Branch around if the given klass is in fact an array (either obj or prim).
3727 // If obj_array/non_array==false/true:
3728 // Branch around if the given klass is not an array klass of any kind.
3729 // If obj_array/non_array==true/true:
3730 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3731 // If obj_array/non_array==true/false:
3732 // Branch around if the kls is an oop array (Object[] or subtype)
3733 //
3734 // Like generate_guard, adds a new path onto the region.
3735 jint layout_con = 0;
3736 Node* layout_val = get_layout_helper(kls, layout_con);
3737 if (layout_val == NULL) {
3738 bool query = (obj_array
3739 ? Klass::layout_helper_is_objArray(layout_con)
3740 : Klass::layout_helper_is_array(layout_con));
3741 if (query == not_array) {
3742 return NULL; // never a branch
3743 } else { // always a branch
3744 Node* always_branch = control();
3745 if (region != NULL)
3746 region->add_req(always_branch);
3747 set_control(top());
3748 return always_branch;
3749 }
3750 }
3751 // Now test the correct condition.
3752 jint nval = (obj_array
3753 ? ((jint)Klass::_lh_array_tag_type_value
3754 << Klass::_lh_array_tag_shift)
3755 : Klass::_lh_neutral_value);
3756 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3757 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3758 // invert the test if we are looking for a non-array
3759 if (not_array) btest = BoolTest(btest).negate();
3760 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3761 return generate_fair_guard(bol, region);
3762 }
3763
3764
3765 //-----------------------inline_native_newArray--------------------------
3766 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3767 bool LibraryCallKit::inline_native_newArray() {
3768 Node* mirror = argument(0);
3769 Node* count_val = argument(1);
3770
3771 mirror = null_check(mirror);
3772 // If mirror or obj is dead, only null-path is taken.
3773 if (stopped()) return true;
3774
3775 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3776 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3777 PhiNode* result_val = new(C) PhiNode(result_reg,
3778 TypeInstPtr::NOTNULL);
3779 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3780 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3781 TypePtr::BOTTOM);
3782
3783 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3784 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3785 result_reg, _slow_path);
3786 Node* normal_ctl = control();
3787 Node* no_array_ctl = result_reg->in(_slow_path);
3788
3789 // Generate code for the slow case. We make a call to newArray().
3790 set_control(no_array_ctl);
3791 if (!stopped()) {
3792 // Either the input type is void.class, or else the
3793 // array klass has not yet been cached. Either the
3794 // ensuing call will throw an exception, or else it
3795 // will cache the array klass for next time.
3796 PreserveJVMState pjvms(this);
3797 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3798 Node* slow_result = set_results_for_java_call(slow_call);
3799 // this->control() comes from set_results_for_java_call
3800 result_reg->set_req(_slow_path, control());
3801 result_val->set_req(_slow_path, slow_result);
3802 result_io ->set_req(_slow_path, i_o());
3803 result_mem->set_req(_slow_path, reset_memory());
3804 }
3805
3806 set_control(normal_ctl);
3807 if (!stopped()) {
3808 // Normal case: The array type has been cached in the java.lang.Class.
3809 // The following call works fine even if the array type is polymorphic.
3810 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3811 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3812 result_reg->init_req(_normal_path, control());
3813 result_val->init_req(_normal_path, obj);
3814 result_io ->init_req(_normal_path, i_o());
3815 result_mem->init_req(_normal_path, reset_memory());
3816 }
3817
3818 // Return the combined state.
3819 set_i_o( _gvn.transform(result_io) );
3820 set_all_memory( _gvn.transform(result_mem));
3821
3822 C->set_has_split_ifs(true); // Has chance for split-if optimization
3823 set_result(result_reg, result_val);
3824 return true;
3825 }
3826
3827 //----------------------inline_native_getLength--------------------------
3828 // public static native int java.lang.reflect.Array.getLength(Object array);
3829 bool LibraryCallKit::inline_native_getLength() {
3830 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3831
3832 Node* array = null_check(argument(0));
3833 // If array is dead, only null-path is taken.
3834 if (stopped()) return true;
3835
3836 // Deoptimize if it is a non-array.
3837 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3838
3839 if (non_array != NULL) {
3840 PreserveJVMState pjvms(this);
3841 set_control(non_array);
3842 uncommon_trap(Deoptimization::Reason_intrinsic,
3843 Deoptimization::Action_maybe_recompile);
3844 }
3845
3846 // If control is dead, only non-array-path is taken.
3847 if (stopped()) return true;
3848
3849 // The works fine even if the array type is polymorphic.
3850 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3851 Node* result = load_array_length(array);
3852
3853 C->set_has_split_ifs(true); // Has chance for split-if optimization
3854 set_result(result);
3855 return true;
3856 }
3857
3858 //------------------------inline_array_copyOf----------------------------
3859 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3860 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3861 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3862 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3863
3864 // Get the arguments.
3865 Node* original = argument(0);
3866 Node* start = is_copyOfRange? argument(1): intcon(0);
3867 Node* end = is_copyOfRange? argument(2): argument(1);
3868 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3869
3870 Node* newcopy;
3871
3872 // Set the original stack and the reexecute bit for the interpreter to reexecute
3873 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3874 { PreserveReexecuteState preexecs(this);
3875 jvms()->set_should_reexecute(true);
3876
3877 array_type_mirror = null_check(array_type_mirror);
3878 original = null_check(original);
3879
3880 // Check if a null path was taken unconditionally.
3881 if (stopped()) return true;
3882
3883 Node* orig_length = load_array_length(original);
3884
3885 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3886 klass_node = null_check(klass_node);
3887
3888 RegionNode* bailout = new (C) RegionNode(1);
3889 record_for_igvn(bailout);
3890
3891 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3892 // Bail out if that is so.
3893 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3894 if (not_objArray != NULL) {
3895 // Improve the klass node's type from the new optimistic assumption:
3896 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3897 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3898 Node* cast = new (C) CastPPNode(klass_node, akls);
3899 cast->init_req(0, control());
3900 klass_node = _gvn.transform(cast);
3901 }
3902
3903 // Bail out if either start or end is negative.
3904 generate_negative_guard(start, bailout, &start);
3905 generate_negative_guard(end, bailout, &end);
3906
3907 Node* length = end;
3908 if (_gvn.type(start) != TypeInt::ZERO) {
3909 length = _gvn.transform(new (C) SubINode(end, start));
3910 }
3911
3912 // Bail out if length is negative.
3913 // Without this the new_array would throw
3914 // NegativeArraySizeException but IllegalArgumentException is what
3915 // should be thrown
3916 generate_negative_guard(length, bailout, &length);
3917
3918 if (bailout->req() > 1) {
3919 PreserveJVMState pjvms(this);
3920 set_control(_gvn.transform(bailout));
3921 uncommon_trap(Deoptimization::Reason_intrinsic,
3922 Deoptimization::Action_maybe_recompile);
3923 }
3924
3925 if (!stopped()) {
3926 // How many elements will we copy from the original?
3927 // The answer is MinI(orig_length - start, length).
3928 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3929 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3930
3931 newcopy = new_array(klass_node, length, 0); // no argments to push
3932
3933 // Generate a direct call to the right arraycopy function(s).
3934 // We know the copy is disjoint but we might not know if the
3935 // oop stores need checking.
3936 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3937 // This will fail a store-check if x contains any non-nulls.
3938 bool disjoint_bases = true;
3939 // if start > orig_length then the length of the copy may be
3940 // negative.
3941 bool length_never_negative = !is_copyOfRange;
3942 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3943 original, start, newcopy, intcon(0), moved,
3944 disjoint_bases, length_never_negative);
3945 }
3946 } // original reexecute is set back here
3947
3948 C->set_has_split_ifs(true); // Has chance for split-if optimization
3949 if (!stopped()) {
3950 set_result(newcopy);
3951 }
3952 return true;
3953 }
3954
3955
3956 //----------------------generate_virtual_guard---------------------------
3957 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3958 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3959 RegionNode* slow_region) {
3960 ciMethod* method = callee();
3961 int vtable_index = method->vtable_index();
3962 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3963 err_msg_res("bad index %d", vtable_index));
3964 // Get the Method* out of the appropriate vtable entry.
3965 int entry_offset = (InstanceKlass::vtable_start_offset() +
3966 vtable_index*vtableEntry::size()) * wordSize +
3967 vtableEntry::method_offset_in_bytes();
3968 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3969 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3970
3971 // Compare the target method with the expected method (e.g., Object.hashCode).
3972 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3973
3974 Node* native_call = makecon(native_call_addr);
3975 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3976 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3977
3978 return generate_slow_guard(test_native, slow_region);
3979 }
3980
3981 //-----------------------generate_method_call----------------------------
3982 // Use generate_method_call to make a slow-call to the real
3983 // method if the fast path fails. An alternative would be to
3984 // use a stub like OptoRuntime::slow_arraycopy_Java.
3985 // This only works for expanding the current library call,
3986 // not another intrinsic. (E.g., don't use this for making an
3987 // arraycopy call inside of the copyOf intrinsic.)
3988 CallJavaNode*
3989 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3990 // When compiling the intrinsic method itself, do not use this technique.
3991 guarantee(callee() != C->method(), "cannot make slow-call to self");
3992
3993 ciMethod* method = callee();
3994 // ensure the JVMS we have will be correct for this call
3995 guarantee(method_id == method->intrinsic_id(), "must match");
3996
3997 const TypeFunc* tf = TypeFunc::make(method);
3998 CallJavaNode* slow_call;
3999 if (is_static) {
4000 assert(!is_virtual, "");
4001 slow_call = new(C) CallStaticJavaNode(C, tf,
4002 SharedRuntime::get_resolve_static_call_stub(),
4003 method, bci());
4004 } else if (is_virtual) {
4005 null_check_receiver();
4006 int vtable_index = Method::invalid_vtable_index;
4007 if (UseInlineCaches) {
4008 // Suppress the vtable call
4009 } else {
4010 // hashCode and clone are not a miranda methods,
4011 // so the vtable index is fixed.
4012 // No need to use the linkResolver to get it.
4013 vtable_index = method->vtable_index();
4014 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4015 err_msg_res("bad index %d", vtable_index));
4016 }
4017 slow_call = new(C) CallDynamicJavaNode(tf,
4018 SharedRuntime::get_resolve_virtual_call_stub(),
4019 method, vtable_index, bci());
4020 } else { // neither virtual nor static: opt_virtual
4021 null_check_receiver();
4022 slow_call = new(C) CallStaticJavaNode(C, tf,
4023 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4024 method, bci());
4025 slow_call->set_optimized_virtual(true);
4026 }
4027 set_arguments_for_java_call(slow_call);
4028 set_edges_for_java_call(slow_call);
4029 return slow_call;
4030 }
4031
4032
4033 /**
4034 * Build special case code for calls to hashCode on an object. This call may
4035 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4036 * slightly different code.
4037 */
4038 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4039 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4040 assert(!(is_virtual && is_static), "either virtual, special, or static");
4041
4042 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4043
4044 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4045 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4046 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4047 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4048 Node* obj = NULL;
4049 if (!is_static) {
4050 // Check for hashing null object
4051 obj = null_check_receiver();
4052 if (stopped()) return true; // unconditionally null
4053 result_reg->init_req(_null_path, top());
4054 result_val->init_req(_null_path, top());
4055 } else {
4056 // Do a null check, and return zero if null.
4057 // System.identityHashCode(null) == 0
4058 obj = argument(0);
4059 Node* null_ctl = top();
4060 obj = null_check_oop(obj, &null_ctl);
4061 result_reg->init_req(_null_path, null_ctl);
4062 result_val->init_req(_null_path, _gvn.intcon(0));
4063 }
4064
4065 // Unconditionally null? Then return right away.
4066 if (stopped()) {
4067 set_control( result_reg->in(_null_path));
4068 if (!stopped())
4069 set_result(result_val->in(_null_path));
4070 return true;
4071 }
4072
4073 // We only go to the fast case code if we pass a number of guards. The
4074 // paths which do not pass are accumulated in the slow_region.
4075 RegionNode* slow_region = new (C) RegionNode(1);
4076 record_for_igvn(slow_region);
4077
4078 // If this is a virtual call, we generate a funny guard. We pull out
4079 // the vtable entry corresponding to hashCode() from the target object.
4080 // If the target method which we are calling happens to be the native
4081 // Object hashCode() method, we pass the guard. We do not need this
4082 // guard for non-virtual calls -- the caller is known to be the native
4083 // Object hashCode().
4084 if (is_virtual) {
4085 // After null check, get the object's klass.
4086 Node* obj_klass = load_object_klass(obj);
4087 generate_virtual_guard(obj_klass, slow_region);
4088 }
4089
4090 // Get the header out of the object, use LoadMarkNode when available
4091 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4092 // The control of the load must be NULL. Otherwise, the load can move before
4093 // the null check after castPP removal.
4094 Node* no_ctrl = NULL;
4095 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4096
4097 // Test the header to see if it is unlocked.
4098 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4099 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4100 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4101 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4102 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4103
4104 generate_slow_guard(test_unlocked, slow_region);
4105
4106 // Get the hash value and check to see that it has been properly assigned.
4107 // We depend on hash_mask being at most 32 bits and avoid the use of
4108 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4109 // vm: see markOop.hpp.
4110 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4111 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4112 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4113 // This hack lets the hash bits live anywhere in the mark object now, as long
4114 // as the shift drops the relevant bits into the low 32 bits. Note that
4115 // Java spec says that HashCode is an int so there's no point in capturing
4116 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4117 hshifted_header = ConvX2I(hshifted_header);
4118 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4119
4120 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4121 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4122 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4123
4124 generate_slow_guard(test_assigned, slow_region);
4125
4126 Node* init_mem = reset_memory();
4127 // fill in the rest of the null path:
4128 result_io ->init_req(_null_path, i_o());
4129 result_mem->init_req(_null_path, init_mem);
4130
4131 result_val->init_req(_fast_path, hash_val);
4132 result_reg->init_req(_fast_path, control());
4133 result_io ->init_req(_fast_path, i_o());
4134 result_mem->init_req(_fast_path, init_mem);
4135
4136 // Generate code for the slow case. We make a call to hashCode().
4137 set_control(_gvn.transform(slow_region));
4138 if (!stopped()) {
4139 // No need for PreserveJVMState, because we're using up the present state.
4140 set_all_memory(init_mem);
4141 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4142 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4143 Node* slow_result = set_results_for_java_call(slow_call);
4144 // this->control() comes from set_results_for_java_call
4145 result_reg->init_req(_slow_path, control());
4146 result_val->init_req(_slow_path, slow_result);
4147 result_io ->set_req(_slow_path, i_o());
4148 result_mem ->set_req(_slow_path, reset_memory());
4149 }
4150
4151 // Return the combined state.
4152 set_i_o( _gvn.transform(result_io) );
4153 set_all_memory( _gvn.transform(result_mem));
4154
4155 set_result(result_reg, result_val);
4156 return true;
4157 }
4158
4159 //---------------------------inline_native_getClass----------------------------
4160 // public final native Class<?> java.lang.Object.getClass();
4161 //
4162 // Build special case code for calls to getClass on an object.
4163 bool LibraryCallKit::inline_native_getClass() {
4164 Node* obj = null_check_receiver();
4165 if (stopped()) return true;
4166 set_result(load_mirror_from_klass(load_object_klass(obj)));
4167 return true;
4168 }
4169
4170 //-----------------inline_native_Reflection_getCallerClass---------------------
4171 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4172 //
4173 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4174 //
4175 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4176 // in that it must skip particular security frames and checks for
4177 // caller sensitive methods.
4178 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4179 #ifndef PRODUCT
4180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4181 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4182 }
4183 #endif
4184
4185 if (!jvms()->has_method()) {
4186 #ifndef PRODUCT
4187 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4188 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4189 }
4190 #endif
4191 return false;
4192 }
4193
4194 // Walk back up the JVM state to find the caller at the required
4195 // depth.
4196 JVMState* caller_jvms = jvms();
4197
4198 // Cf. JVM_GetCallerClass
4199 // NOTE: Start the loop at depth 1 because the current JVM state does
4200 // not include the Reflection.getCallerClass() frame.
4201 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4202 ciMethod* m = caller_jvms->method();
4203 switch (n) {
4204 case 0:
4205 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4206 break;
4207 case 1:
4208 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4209 if (!m->caller_sensitive()) {
4210 #ifndef PRODUCT
4211 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4212 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4213 }
4214 #endif
4215 return false; // bail-out; let JVM_GetCallerClass do the work
4216 }
4217 break;
4218 default:
4219 if (!m->is_ignored_by_security_stack_walk()) {
4220 // We have reached the desired frame; return the holder class.
4221 // Acquire method holder as java.lang.Class and push as constant.
4222 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4223 ciInstance* caller_mirror = caller_klass->java_mirror();
4224 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4225
4226 #ifndef PRODUCT
4227 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4228 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());
4229 tty->print_cr(" JVM state at this point:");
4230 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4231 ciMethod* m = jvms()->of_depth(i)->method();
4232 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4233 }
4234 }
4235 #endif
4236 return true;
4237 }
4238 break;
4239 }
4240 }
4241
4242 #ifndef PRODUCT
4243 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4244 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4245 tty->print_cr(" JVM state at this point:");
4246 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4247 ciMethod* m = jvms()->of_depth(i)->method();
4248 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4249 }
4250 }
4251 #endif
4252
4253 return false; // bail-out; let JVM_GetCallerClass do the work
4254 }
4255
4256 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4257 Node* arg = argument(0);
4258 Node* result;
4259
4260 switch (id) {
4261 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4262 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4263 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4264 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4265
4266 case vmIntrinsics::_doubleToLongBits: {
4267 // two paths (plus control) merge in a wood
4268 RegionNode *r = new (C) RegionNode(3);
4269 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4270
4271 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4272 // Build the boolean node
4273 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4274
4275 // Branch either way.
4276 // NaN case is less traveled, which makes all the difference.
4277 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4278 Node *opt_isnan = _gvn.transform(ifisnan);
4279 assert( opt_isnan->is_If(), "Expect an IfNode");
4280 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4281 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4282
4283 set_control(iftrue);
4284
4285 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4286 Node *slow_result = longcon(nan_bits); // return NaN
4287 phi->init_req(1, _gvn.transform( slow_result ));
4288 r->init_req(1, iftrue);
4289
4290 // Else fall through
4291 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4292 set_control(iffalse);
4293
4294 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4295 r->init_req(2, iffalse);
4296
4297 // Post merge
4298 set_control(_gvn.transform(r));
4299 record_for_igvn(r);
4300
4301 C->set_has_split_ifs(true); // Has chance for split-if optimization
4302 result = phi;
4303 assert(result->bottom_type()->isa_long(), "must be");
4304 break;
4305 }
4306
4307 case vmIntrinsics::_floatToIntBits: {
4308 // two paths (plus control) merge in a wood
4309 RegionNode *r = new (C) RegionNode(3);
4310 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4311
4312 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4313 // Build the boolean node
4314 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4315
4316 // Branch either way.
4317 // NaN case is less traveled, which makes all the difference.
4318 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4319 Node *opt_isnan = _gvn.transform(ifisnan);
4320 assert( opt_isnan->is_If(), "Expect an IfNode");
4321 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4322 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4323
4324 set_control(iftrue);
4325
4326 static const jint nan_bits = 0x7fc00000;
4327 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4328 phi->init_req(1, _gvn.transform( slow_result ));
4329 r->init_req(1, iftrue);
4330
4331 // Else fall through
4332 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4333 set_control(iffalse);
4334
4335 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4336 r->init_req(2, iffalse);
4337
4338 // Post merge
4339 set_control(_gvn.transform(r));
4340 record_for_igvn(r);
4341
4342 C->set_has_split_ifs(true); // Has chance for split-if optimization
4343 result = phi;
4344 assert(result->bottom_type()->isa_int(), "must be");
4345 break;
4346 }
4347
4348 default:
4349 fatal_unexpected_iid(id);
4350 break;
4351 }
4352 set_result(_gvn.transform(result));
4353 return true;
4354 }
4355
4356 #ifdef _LP64
4357 #define XTOP ,top() /*additional argument*/
4358 #else //_LP64
4359 #define XTOP /*no additional argument*/
4360 #endif //_LP64
4361
4362 //----------------------inline_unsafe_copyMemory-------------------------
4363 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4364 bool LibraryCallKit::inline_unsafe_copyMemory() {
4365 if (callee()->is_static()) return false; // caller must have the capability!
4366 null_check_receiver(); // null-check receiver
4367 if (stopped()) return true;
4368
4369 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4370
4371 Node* src_ptr = argument(1); // type: oop
4372 Node* src_off = ConvL2X(argument(2)); // type: long
4373 Node* dst_ptr = argument(4); // type: oop
4374 Node* dst_off = ConvL2X(argument(5)); // type: long
4375 Node* size = ConvL2X(argument(7)); // type: long
4376
4377 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4378 "fieldOffset must be byte-scaled");
4379
4380 Node* src = make_unsafe_address(src_ptr, src_off);
4381 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4382
4383 // Conservatively insert a memory barrier on all memory slices.
4384 // Do not let writes of the copy source or destination float below the copy.
4385 insert_mem_bar(Op_MemBarCPUOrder);
4386
4387 // Call it. Note that the length argument is not scaled.
4388 make_runtime_call(RC_LEAF|RC_NO_FP,
4389 OptoRuntime::fast_arraycopy_Type(),
4390 StubRoutines::unsafe_arraycopy(),
4391 "unsafe_arraycopy",
4392 TypeRawPtr::BOTTOM,
4393 src, dst, size XTOP);
4394
4395 // Do not let reads of the copy destination float above the copy.
4396 insert_mem_bar(Op_MemBarCPUOrder);
4397
4398 return true;
4399 }
4400
4401 //------------------------clone_coping-----------------------------------
4402 // Helper function for inline_native_clone.
4403 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4404 assert(obj_size != NULL, "");
4405 Node* raw_obj = alloc_obj->in(1);
4406 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4407
4408 AllocateNode* alloc = NULL;
4409 if (ReduceBulkZeroing) {
4410 // We will be completely responsible for initializing this object -
4411 // mark Initialize node as complete.
4412 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4413 // The object was just allocated - there should be no any stores!
4414 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4415 // Mark as complete_with_arraycopy so that on AllocateNode
4416 // expansion, we know this AllocateNode is initialized by an array
4417 // copy and a StoreStore barrier exists after the array copy.
4418 alloc->initialization()->set_complete_with_arraycopy();
4419 }
4420
4421 // Copy the fastest available way.
4422 // TODO: generate fields copies for small objects instead.
4423 Node* src = obj;
4424 Node* dest = alloc_obj;
4425 Node* size = _gvn.transform(obj_size);
4426
4427 // Exclude the header but include array length to copy by 8 bytes words.
4428 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4429 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4430 instanceOopDesc::base_offset_in_bytes();
4431 // base_off:
4432 // 8 - 32-bit VM
4433 // 12 - 64-bit VM, compressed klass
4434 // 16 - 64-bit VM, normal klass
4435 if (base_off % BytesPerLong != 0) {
4436 assert(UseCompressedClassPointers, "");
4437 if (is_array) {
4438 // Exclude length to copy by 8 bytes words.
4439 base_off += sizeof(int);
4440 } else {
4441 // Include klass to copy by 8 bytes words.
4442 base_off = instanceOopDesc::klass_offset_in_bytes();
4443 }
4444 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4445 }
4446 src = basic_plus_adr(src, base_off);
4447 dest = basic_plus_adr(dest, base_off);
4448
4449 // Compute the length also, if needed:
4450 Node* countx = size;
4451 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4452 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4453
4454 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4455 bool disjoint_bases = true;
4456 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4457 src, NULL, dest, NULL, countx,
4458 /*dest_uninitialized*/true);
4459
4460 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4461 if (card_mark) {
4462 assert(!is_array, "");
4463 // Put in store barrier for any and all oops we are sticking
4464 // into this object. (We could avoid this if we could prove
4465 // that the object type contains no oop fields at all.)
4466 Node* no_particular_value = NULL;
4467 Node* no_particular_field = NULL;
4468 int raw_adr_idx = Compile::AliasIdxRaw;
4469 post_barrier(control(),
4470 memory(raw_adr_type),
4471 alloc_obj,
4472 no_particular_field,
4473 raw_adr_idx,
4474 no_particular_value,
4475 T_OBJECT,
4476 false);
4477 }
4478
4479 // Do not let reads from the cloned object float above the arraycopy.
4480 if (alloc != NULL) {
4481 // Do not let stores that initialize this object be reordered with
4482 // a subsequent store that would make this object accessible by
4483 // other threads.
4484 // Record what AllocateNode this StoreStore protects so that
4485 // escape analysis can go from the MemBarStoreStoreNode to the
4486 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4487 // based on the escape status of the AllocateNode.
4488 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4489 } else {
4490 insert_mem_bar(Op_MemBarCPUOrder);
4491 }
4492 }
4493
4494 //------------------------inline_native_clone----------------------------
4495 // protected native Object java.lang.Object.clone();
4496 //
4497 // Here are the simple edge cases:
4498 // null receiver => normal trap
4499 // virtual and clone was overridden => slow path to out-of-line clone
4500 // not cloneable or finalizer => slow path to out-of-line Object.clone
4501 //
4502 // The general case has two steps, allocation and copying.
4503 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4504 //
4505 // Copying also has two cases, oop arrays and everything else.
4506 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4507 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4508 //
4509 // These steps fold up nicely if and when the cloned object's klass
4510 // can be sharply typed as an object array, a type array, or an instance.
4511 //
4512 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4513 PhiNode* result_val;
4514
4515 // Set the reexecute bit for the interpreter to reexecute
4516 // the bytecode that invokes Object.clone if deoptimization happens.
4517 { PreserveReexecuteState preexecs(this);
4518 jvms()->set_should_reexecute(true);
4519
4520 Node* obj = null_check_receiver();
4521 if (stopped()) return true;
4522
4523 Node* obj_klass = load_object_klass(obj);
4524 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4525 const TypeOopPtr* toop = ((tklass != NULL)
4526 ? tklass->as_instance_type()
4527 : TypeInstPtr::NOTNULL);
4528
4529 // Conservatively insert a memory barrier on all memory slices.
4530 // Do not let writes into the original float below the clone.
4531 insert_mem_bar(Op_MemBarCPUOrder);
4532
4533 // paths into result_reg:
4534 enum {
4535 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4536 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4537 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4538 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4539 PATH_LIMIT
4540 };
4541 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4542 result_val = new(C) PhiNode(result_reg,
4543 TypeInstPtr::NOTNULL);
4544 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4545 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4546 TypePtr::BOTTOM);
4547 record_for_igvn(result_reg);
4548
4549 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4550 int raw_adr_idx = Compile::AliasIdxRaw;
4551
4552 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4553 if (array_ctl != NULL) {
4554 // It's an array.
4555 PreserveJVMState pjvms(this);
4556 set_control(array_ctl);
4557 Node* obj_length = load_array_length(obj);
4558 Node* obj_size = NULL;
4559 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4560
4561 if (!use_ReduceInitialCardMarks()) {
4562 // If it is an oop array, it requires very special treatment,
4563 // because card marking is required on each card of the array.
4564 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4565 if (is_obja != NULL) {
4566 PreserveJVMState pjvms2(this);
4567 set_control(is_obja);
4568 // Generate a direct call to the right arraycopy function(s).
4569 bool disjoint_bases = true;
4570 bool length_never_negative = true;
4571 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4572 obj, intcon(0), alloc_obj, intcon(0),
4573 obj_length,
4574 disjoint_bases, length_never_negative);
4575 result_reg->init_req(_objArray_path, control());
4576 result_val->init_req(_objArray_path, alloc_obj);
4577 result_i_o ->set_req(_objArray_path, i_o());
4578 result_mem ->set_req(_objArray_path, reset_memory());
4579 }
4580 }
4581 // Otherwise, there are no card marks to worry about.
4582 // (We can dispense with card marks if we know the allocation
4583 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4584 // causes the non-eden paths to take compensating steps to
4585 // simulate a fresh allocation, so that no further
4586 // card marks are required in compiled code to initialize
4587 // the object.)
4588
4589 if (!stopped()) {
4590 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4591
4592 // Present the results of the copy.
4593 result_reg->init_req(_array_path, control());
4594 result_val->init_req(_array_path, alloc_obj);
4595 result_i_o ->set_req(_array_path, i_o());
4596 result_mem ->set_req(_array_path, reset_memory());
4597 }
4598 }
4599
4600 // We only go to the instance fast case code if we pass a number of guards.
4601 // The paths which do not pass are accumulated in the slow_region.
4602 RegionNode* slow_region = new (C) RegionNode(1);
4603 record_for_igvn(slow_region);
4604 if (!stopped()) {
4605 // It's an instance (we did array above). Make the slow-path tests.
4606 // If this is a virtual call, we generate a funny guard. We grab
4607 // the vtable entry corresponding to clone() from the target object.
4608 // If the target method which we are calling happens to be the
4609 // Object clone() method, we pass the guard. We do not need this
4610 // guard for non-virtual calls; the caller is known to be the native
4611 // Object clone().
4612 if (is_virtual) {
4613 generate_virtual_guard(obj_klass, slow_region);
4614 }
4615
4616 // The object must be cloneable and must not have a finalizer.
4617 // Both of these conditions may be checked in a single test.
4618 // We could optimize the cloneable test further, but we don't care.
4619 generate_access_flags_guard(obj_klass,
4620 // Test both conditions:
4621 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4622 // Must be cloneable but not finalizer:
4623 JVM_ACC_IS_CLONEABLE,
4624 slow_region);
4625 }
4626
4627 if (!stopped()) {
4628 // It's an instance, and it passed the slow-path tests.
4629 PreserveJVMState pjvms(this);
4630 Node* obj_size = NULL;
4631 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4632 // is reexecuted if deoptimization occurs and there could be problems when merging
4633 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4634 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4635
4636 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4637
4638 // Present the results of the slow call.
4639 result_reg->init_req(_instance_path, control());
4640 result_val->init_req(_instance_path, alloc_obj);
4641 result_i_o ->set_req(_instance_path, i_o());
4642 result_mem ->set_req(_instance_path, reset_memory());
4643 }
4644
4645 // Generate code for the slow case. We make a call to clone().
4646 set_control(_gvn.transform(slow_region));
4647 if (!stopped()) {
4648 PreserveJVMState pjvms(this);
4649 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4650 Node* slow_result = set_results_for_java_call(slow_call);
4651 // this->control() comes from set_results_for_java_call
4652 result_reg->init_req(_slow_path, control());
4653 result_val->init_req(_slow_path, slow_result);
4654 result_i_o ->set_req(_slow_path, i_o());
4655 result_mem ->set_req(_slow_path, reset_memory());
4656 }
4657
4658 // Return the combined state.
4659 set_control( _gvn.transform(result_reg));
4660 set_i_o( _gvn.transform(result_i_o));
4661 set_all_memory( _gvn.transform(result_mem));
4662 } // original reexecute is set back here
4663
4664 set_result(_gvn.transform(result_val));
4665 return true;
4666 }
4667
4668 //------------------------------basictype2arraycopy----------------------------
4669 address LibraryCallKit::basictype2arraycopy(BasicType t,
4670 Node* src_offset,
4671 Node* dest_offset,
4672 bool disjoint_bases,
4673 const char* &name,
4674 bool dest_uninitialized) {
4675 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4676 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4677
4678 bool aligned = false;
4679 bool disjoint = disjoint_bases;
4680
4681 // if the offsets are the same, we can treat the memory regions as
4682 // disjoint, because either the memory regions are in different arrays,
4683 // or they are identical (which we can treat as disjoint.) We can also
4684 // treat a copy with a destination index less that the source index
4685 // as disjoint since a low->high copy will work correctly in this case.
4686 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4687 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4688 // both indices are constants
4689 int s_offs = src_offset_inttype->get_con();
4690 int d_offs = dest_offset_inttype->get_con();
4691 int element_size = type2aelembytes(t);
4692 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4693 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4694 if (s_offs >= d_offs) disjoint = true;
4695 } else if (src_offset == dest_offset && src_offset != NULL) {
4696 // This can occur if the offsets are identical non-constants.
4697 disjoint = true;
4698 }
4699
4700 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4701 }
4702
4703
4704 //------------------------------inline_arraycopy-----------------------
4705 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4706 // Object dest, int destPos,
4707 // int length);
4708 bool LibraryCallKit::inline_arraycopy() {
4709 // Get the arguments.
4710 Node* src = argument(0); // type: oop
4711 Node* src_offset = argument(1); // type: int
4712 Node* dest = argument(2); // type: oop
4713 Node* dest_offset = argument(3); // type: int
4714 Node* length = argument(4); // type: int
4715
4716 // Compile time checks. If any of these checks cannot be verified at compile time,
4717 // we do not make a fast path for this call. Instead, we let the call remain as it
4718 // is. The checks we choose to mandate at compile time are:
4719 //
4720 // (1) src and dest are arrays.
4721 const Type* src_type = src->Value(&_gvn);
4722 const Type* dest_type = dest->Value(&_gvn);
4723 const TypeAryPtr* top_src = src_type->isa_aryptr();
4724 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4725
4726 // Do we have the type of src?
4727 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4728 // Do we have the type of dest?
4729 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4730 // Is the type for src from speculation?
4731 bool src_spec = false;
4732 // Is the type for dest from speculation?
4733 bool dest_spec = false;
4734
4735 if (!has_src || !has_dest) {
4736 // We don't have sufficient type information, let's see if
4737 // speculative types can help. We need to have types for both src
4738 // and dest so that it pays off.
4739
4740 // Do we already have or could we have type information for src
4741 bool could_have_src = has_src;
4742 // Do we already have or could we have type information for dest
4743 bool could_have_dest = has_dest;
4744
4745 ciKlass* src_k = NULL;
4746 if (!has_src) {
4747 src_k = src_type->speculative_type();
4748 if (src_k != NULL && src_k->is_array_klass()) {
4749 could_have_src = true;
4750 }
4751 }
4752
4753 ciKlass* dest_k = NULL;
4754 if (!has_dest) {
4755 dest_k = dest_type->speculative_type();
4756 if (dest_k != NULL && dest_k->is_array_klass()) {
4757 could_have_dest = true;
4758 }
4759 }
4760
4761 if (could_have_src && could_have_dest) {
4762 // This is going to pay off so emit the required guards
4763 if (!has_src) {
4764 src = maybe_cast_profiled_obj(src, src_k);
4765 src_type = _gvn.type(src);
4766 top_src = src_type->isa_aryptr();
4767 has_src = (top_src != NULL && top_src->klass() != NULL);
4768 src_spec = true;
4769 }
4770 if (!has_dest) {
4771 dest = maybe_cast_profiled_obj(dest, dest_k);
4772 dest_type = _gvn.type(dest);
4773 top_dest = dest_type->isa_aryptr();
4774 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4775 dest_spec = true;
4776 }
4777 }
4778 }
4779
4780 if (!has_src || !has_dest) {
4781 // Conservatively insert a memory barrier on all memory slices.
4782 // Do not let writes into the source float below the arraycopy.
4783 insert_mem_bar(Op_MemBarCPUOrder);
4784
4785 // Call StubRoutines::generic_arraycopy stub.
4786 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4787 src, src_offset, dest, dest_offset, length);
4788
4789 // Do not let reads from the destination float above the arraycopy.
4790 // Since we cannot type the arrays, we don't know which slices
4791 // might be affected. We could restrict this barrier only to those
4792 // memory slices which pertain to array elements--but don't bother.
4793 if (!InsertMemBarAfterArraycopy)
4794 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4795 insert_mem_bar(Op_MemBarCPUOrder);
4796 return true;
4797 }
4798
4799 // (2) src and dest arrays must have elements of the same BasicType
4800 // Figure out the size and type of the elements we will be copying.
4801 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4802 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4803 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4804 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4805
4806 if (src_elem != dest_elem || dest_elem == T_VOID) {
4807 // The component types are not the same or are not recognized. Punt.
4808 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4809 generate_slow_arraycopy(TypePtr::BOTTOM,
4810 src, src_offset, dest, dest_offset, length,
4811 /*dest_uninitialized*/false);
4812 return true;
4813 }
4814
4815 if (src_elem == T_OBJECT) {
4816 // If both arrays are object arrays then having the exact types
4817 // for both will remove the need for a subtype check at runtime
4818 // before the call and may make it possible to pick a faster copy
4819 // routine (without a subtype check on every element)
4820 // Do we have the exact type of src?
4821 bool could_have_src = src_spec;
4822 // Do we have the exact type of dest?
4823 bool could_have_dest = dest_spec;
4824 ciKlass* src_k = top_src->klass();
4825 ciKlass* dest_k = top_dest->klass();
4826 if (!src_spec) {
4827 src_k = src_type->speculative_type();
4828 if (src_k != NULL && src_k->is_array_klass()) {
4829 could_have_src = true;
4830 }
4831 }
4832 if (!dest_spec) {
4833 dest_k = dest_type->speculative_type();
4834 if (dest_k != NULL && dest_k->is_array_klass()) {
4835 could_have_dest = true;
4836 }
4837 }
4838 if (could_have_src && could_have_dest) {
4839 // If we can have both exact types, emit the missing guards
4840 if (could_have_src && !src_spec) {
4841 src = maybe_cast_profiled_obj(src, src_k);
4842 }
4843 if (could_have_dest && !dest_spec) {
4844 dest = maybe_cast_profiled_obj(dest, dest_k);
4845 }
4846 }
4847 }
4848
4849 //---------------------------------------------------------------------------
4850 // We will make a fast path for this call to arraycopy.
4851
4852 // We have the following tests left to perform:
4853 //
4854 // (3) src and dest must not be null.
4855 // (4) src_offset must not be negative.
4856 // (5) dest_offset must not be negative.
4857 // (6) length must not be negative.
4858 // (7) src_offset + length must not exceed length of src.
4859 // (8) dest_offset + length must not exceed length of dest.
4860 // (9) each element of an oop array must be assignable
4861
4862 RegionNode* slow_region = new (C) RegionNode(1);
4863 record_for_igvn(slow_region);
4864
4865 // (3) operands must not be null
4866 // We currently perform our null checks with the null_check routine.
4867 // This means that the null exceptions will be reported in the caller
4868 // rather than (correctly) reported inside of the native arraycopy call.
4869 // This should be corrected, given time. We do our null check with the
4870 // stack pointer restored.
4871 src = null_check(src, T_ARRAY);
4872 dest = null_check(dest, T_ARRAY);
4873
4874 // (4) src_offset must not be negative.
4875 generate_negative_guard(src_offset, slow_region);
4876
4877 // (5) dest_offset must not be negative.
4878 generate_negative_guard(dest_offset, slow_region);
4879
4880 // (6) length must not be negative (moved to generate_arraycopy()).
4881 // generate_negative_guard(length, slow_region);
4882
4883 // (7) src_offset + length must not exceed length of src.
4884 generate_limit_guard(src_offset, length,
4885 load_array_length(src),
4886 slow_region);
4887
4888 // (8) dest_offset + length must not exceed length of dest.
4889 generate_limit_guard(dest_offset, length,
4890 load_array_length(dest),
4891 slow_region);
4892
4893 // (9) each element of an oop array must be assignable
4894 // The generate_arraycopy subroutine checks this.
4895
4896 // This is where the memory effects are placed:
4897 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4898 generate_arraycopy(adr_type, dest_elem,
4899 src, src_offset, dest, dest_offset, length,
4900 false, false, slow_region);
4901
4902 return true;
4903 }
4904
4905 //-----------------------------generate_arraycopy----------------------
4906 // Generate an optimized call to arraycopy.
4907 // Caller must guard against non-arrays.
4908 // Caller must determine a common array basic-type for both arrays.
4909 // Caller must validate offsets against array bounds.
4910 // The slow_region has already collected guard failure paths
4911 // (such as out of bounds length or non-conformable array types).
4912 // The generated code has this shape, in general:
4913 //
4914 // if (length == 0) return // via zero_path
4915 // slowval = -1
4916 // if (types unknown) {
4917 // slowval = call generic copy loop
4918 // if (slowval == 0) return // via checked_path
4919 // } else if (indexes in bounds) {
4920 // if ((is object array) && !(array type check)) {
4921 // slowval = call checked copy loop
4922 // if (slowval == 0) return // via checked_path
4923 // } else {
4924 // call bulk copy loop
4925 // return // via fast_path
4926 // }
4927 // }
4928 // // adjust params for remaining work:
4929 // if (slowval != -1) {
4930 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4931 // }
4932 // slow_region:
4933 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4934 // return // via slow_call_path
4935 //
4936 // This routine is used from several intrinsics: System.arraycopy,
4937 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4938 //
4939 void
4940 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4941 BasicType basic_elem_type,
4942 Node* src, Node* src_offset,
4943 Node* dest, Node* dest_offset,
4944 Node* copy_length,
4945 bool disjoint_bases,
4946 bool length_never_negative,
4947 RegionNode* slow_region) {
4948
4949 if (slow_region == NULL) {
4950 slow_region = new(C) RegionNode(1);
4951 record_for_igvn(slow_region);
4952 }
4953
4954 Node* original_dest = dest;
4955 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4956 bool dest_uninitialized = false;
4957
4958 // See if this is the initialization of a newly-allocated array.
4959 // If so, we will take responsibility here for initializing it to zero.
4960 // (Note: Because tightly_coupled_allocation performs checks on the
4961 // out-edges of the dest, we need to avoid making derived pointers
4962 // from it until we have checked its uses.)
4963 if (ReduceBulkZeroing
4964 && !ZeroTLAB // pointless if already zeroed
4965 && basic_elem_type != T_CONFLICT // avoid corner case
4966 && !src->eqv_uncast(dest)
4967 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4968 != NULL)
4969 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4970 && alloc->maybe_set_complete(&_gvn)) {
4971 // "You break it, you buy it."
4972 InitializeNode* init = alloc->initialization();
4973 assert(init->is_complete(), "we just did this");
4974 init->set_complete_with_arraycopy();
4975 assert(dest->is_CheckCastPP(), "sanity");
4976 assert(dest->in(0)->in(0) == init, "dest pinned");
4977 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4978 // From this point on, every exit path is responsible for
4979 // initializing any non-copied parts of the object to zero.
4980 // Also, if this flag is set we make sure that arraycopy interacts properly
4981 // with G1, eliding pre-barriers. See CR 6627983.
4982 dest_uninitialized = true;
4983 } else {
4984 // No zeroing elimination here.
4985 alloc = NULL;
4986 //original_dest = dest;
4987 //dest_uninitialized = false;
4988 }
4989
4990 // Results are placed here:
4991 enum { fast_path = 1, // normal void-returning assembly stub
4992 checked_path = 2, // special assembly stub with cleanup
4993 slow_call_path = 3, // something went wrong; call the VM
4994 zero_path = 4, // bypass when length of copy is zero
4995 bcopy_path = 5, // copy primitive array by 64-bit blocks
4996 PATH_LIMIT = 6
4997 };
4998 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4999 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5000 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5001 record_for_igvn(result_region);
5002 _gvn.set_type_bottom(result_i_o);
5003 _gvn.set_type_bottom(result_memory);
5004 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5005
5006 // The slow_control path:
5007 Node* slow_control;
5008 Node* slow_i_o = i_o();
5009 Node* slow_mem = memory(adr_type);
5010 debug_only(slow_control = (Node*) badAddress);
5011
5012 // Checked control path:
5013 Node* checked_control = top();
5014 Node* checked_mem = NULL;
5015 Node* checked_i_o = NULL;
5016 Node* checked_value = NULL;
5017
5018 if (basic_elem_type == T_CONFLICT) {
5019 assert(!dest_uninitialized, "");
5020 Node* cv = generate_generic_arraycopy(adr_type,
5021 src, src_offset, dest, dest_offset,
5022 copy_length, dest_uninitialized);
5023 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5024 checked_control = control();
5025 checked_i_o = i_o();
5026 checked_mem = memory(adr_type);
5027 checked_value = cv;
5028 set_control(top()); // no fast path
5029 }
5030
5031 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5032 if (not_pos != NULL) {
5033 PreserveJVMState pjvms(this);
5034 set_control(not_pos);
5035
5036 // (6) length must not be negative.
5037 if (!length_never_negative) {
5038 generate_negative_guard(copy_length, slow_region);
5039 }
5040
5041 // copy_length is 0.
5042 if (!stopped() && dest_uninitialized) {
5043 Node* dest_length = alloc->in(AllocateNode::ALength);
5044 if (copy_length->eqv_uncast(dest_length)
5045 || _gvn.find_int_con(dest_length, 1) <= 0) {
5046 // There is no zeroing to do. No need for a secondary raw memory barrier.
5047 } else {
5048 // Clear the whole thing since there are no source elements to copy.
5049 generate_clear_array(adr_type, dest, basic_elem_type,
5050 intcon(0), NULL,
5051 alloc->in(AllocateNode::AllocSize));
5052 // Use a secondary InitializeNode as raw memory barrier.
5053 // Currently it is needed only on this path since other
5054 // paths have stub or runtime calls as raw memory barriers.
5055 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5056 Compile::AliasIdxRaw,
5057 top())->as_Initialize();
5058 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5059 }
5060 }
5061
5062 // Present the results of the fast call.
5063 result_region->init_req(zero_path, control());
5064 result_i_o ->init_req(zero_path, i_o());
5065 result_memory->init_req(zero_path, memory(adr_type));
5066 }
5067
5068 if (!stopped() && dest_uninitialized) {
5069 // We have to initialize the *uncopied* part of the array to zero.
5070 // The copy destination is the slice dest[off..off+len]. The other slices
5071 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5072 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5073 Node* dest_length = alloc->in(AllocateNode::ALength);
5074 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5075 copy_length));
5076
5077 // If there is a head section that needs zeroing, do it now.
5078 if (find_int_con(dest_offset, -1) != 0) {
5079 generate_clear_array(adr_type, dest, basic_elem_type,
5080 intcon(0), dest_offset,
5081 NULL);
5082 }
5083
5084 // Next, perform a dynamic check on the tail length.
5085 // It is often zero, and we can win big if we prove this.
5086 // There are two wins: Avoid generating the ClearArray
5087 // with its attendant messy index arithmetic, and upgrade
5088 // the copy to a more hardware-friendly word size of 64 bits.
5089 Node* tail_ctl = NULL;
5090 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5091 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5092 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5093 tail_ctl = generate_slow_guard(bol_lt, NULL);
5094 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5095 }
5096
5097 // At this point, let's assume there is no tail.
5098 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5099 // There is no tail. Try an upgrade to a 64-bit copy.
5100 bool didit = false;
5101 { PreserveJVMState pjvms(this);
5102 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5103 src, src_offset, dest, dest_offset,
5104 dest_size, dest_uninitialized);
5105 if (didit) {
5106 // Present the results of the block-copying fast call.
5107 result_region->init_req(bcopy_path, control());
5108 result_i_o ->init_req(bcopy_path, i_o());
5109 result_memory->init_req(bcopy_path, memory(adr_type));
5110 }
5111 }
5112 if (didit)
5113 set_control(top()); // no regular fast path
5114 }
5115
5116 // Clear the tail, if any.
5117 if (tail_ctl != NULL) {
5118 Node* notail_ctl = stopped() ? NULL : control();
5119 set_control(tail_ctl);
5120 if (notail_ctl == NULL) {
5121 generate_clear_array(adr_type, dest, basic_elem_type,
5122 dest_tail, NULL,
5123 dest_size);
5124 } else {
5125 // Make a local merge.
5126 Node* done_ctl = new(C) RegionNode(3);
5127 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5128 done_ctl->init_req(1, notail_ctl);
5129 done_mem->init_req(1, memory(adr_type));
5130 generate_clear_array(adr_type, dest, basic_elem_type,
5131 dest_tail, NULL,
5132 dest_size);
5133 done_ctl->init_req(2, control());
5134 done_mem->init_req(2, memory(adr_type));
5135 set_control( _gvn.transform(done_ctl));
5136 set_memory( _gvn.transform(done_mem), adr_type );
5137 }
5138 }
5139 }
5140
5141 BasicType copy_type = basic_elem_type;
5142 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5143 if (!stopped() && copy_type == T_OBJECT) {
5144 // If src and dest have compatible element types, we can copy bits.
5145 // Types S[] and D[] are compatible if D is a supertype of S.
5146 //
5147 // If they are not, we will use checked_oop_disjoint_arraycopy,
5148 // which performs a fast optimistic per-oop check, and backs off
5149 // further to JVM_ArrayCopy on the first per-oop check that fails.
5150 // (Actually, we don't move raw bits only; the GC requires card marks.)
5151
5152 // Get the Klass* for both src and dest
5153 Node* src_klass = load_object_klass(src);
5154 Node* dest_klass = load_object_klass(dest);
5155
5156 // Generate the subtype check.
5157 // This might fold up statically, or then again it might not.
5158 //
5159 // Non-static example: Copying List<String>.elements to a new String[].
5160 // The backing store for a List<String> is always an Object[],
5161 // but its elements are always type String, if the generic types
5162 // are correct at the source level.
5163 //
5164 // Test S[] against D[], not S against D, because (probably)
5165 // the secondary supertype cache is less busy for S[] than S.
5166 // This usually only matters when D is an interface.
5167 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5168 // Plug failing path into checked_oop_disjoint_arraycopy
5169 if (not_subtype_ctrl != top()) {
5170 PreserveJVMState pjvms(this);
5171 set_control(not_subtype_ctrl);
5172 // (At this point we can assume disjoint_bases, since types differ.)
5173 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5174 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5175 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5176 Node* dest_elem_klass = _gvn.transform(n1);
5177 Node* cv = generate_checkcast_arraycopy(adr_type,
5178 dest_elem_klass,
5179 src, src_offset, dest, dest_offset,
5180 ConvI2X(copy_length), dest_uninitialized);
5181 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5182 checked_control = control();
5183 checked_i_o = i_o();
5184 checked_mem = memory(adr_type);
5185 checked_value = cv;
5186 }
5187 // At this point we know we do not need type checks on oop stores.
5188
5189 // Let's see if we need card marks:
5190 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5191 // If we do not need card marks, copy using the jint or jlong stub.
5192 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5193 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5194 "sizes agree");
5195 }
5196 }
5197
5198 if (!stopped()) {
5199 // Generate the fast path, if possible.
5200 PreserveJVMState pjvms(this);
5201 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5202 src, src_offset, dest, dest_offset,
5203 ConvI2X(copy_length), dest_uninitialized);
5204
5205 // Present the results of the fast call.
5206 result_region->init_req(fast_path, control());
5207 result_i_o ->init_req(fast_path, i_o());
5208 result_memory->init_req(fast_path, memory(adr_type));
5209 }
5210
5211 // Here are all the slow paths up to this point, in one bundle:
5212 slow_control = top();
5213 if (slow_region != NULL)
5214 slow_control = _gvn.transform(slow_region);
5215 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5216
5217 set_control(checked_control);
5218 if (!stopped()) {
5219 // Clean up after the checked call.
5220 // The returned value is either 0 or -1^K,
5221 // where K = number of partially transferred array elements.
5222 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5223 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5224 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5225
5226 // If it is 0, we are done, so transfer to the end.
5227 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5228 result_region->init_req(checked_path, checks_done);
5229 result_i_o ->init_req(checked_path, checked_i_o);
5230 result_memory->init_req(checked_path, checked_mem);
5231
5232 // If it is not zero, merge into the slow call.
5233 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5234 RegionNode* slow_reg2 = new(C) RegionNode(3);
5235 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5236 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5237 record_for_igvn(slow_reg2);
5238 slow_reg2 ->init_req(1, slow_control);
5239 slow_i_o2 ->init_req(1, slow_i_o);
5240 slow_mem2 ->init_req(1, slow_mem);
5241 slow_reg2 ->init_req(2, control());
5242 slow_i_o2 ->init_req(2, checked_i_o);
5243 slow_mem2 ->init_req(2, checked_mem);
5244
5245 slow_control = _gvn.transform(slow_reg2);
5246 slow_i_o = _gvn.transform(slow_i_o2);
5247 slow_mem = _gvn.transform(slow_mem2);
5248
5249 if (alloc != NULL) {
5250 // We'll restart from the very beginning, after zeroing the whole thing.
5251 // This can cause double writes, but that's OK since dest is brand new.
5252 // So we ignore the low 31 bits of the value returned from the stub.
5253 } else {
5254 // We must continue the copy exactly where it failed, or else
5255 // another thread might see the wrong number of writes to dest.
5256 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5257 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5258 slow_offset->init_req(1, intcon(0));
5259 slow_offset->init_req(2, checked_offset);
5260 slow_offset = _gvn.transform(slow_offset);
5261
5262 // Adjust the arguments by the conditionally incoming offset.
5263 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5264 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5265 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5266
5267 // Tweak the node variables to adjust the code produced below:
5268 src_offset = src_off_plus;
5269 dest_offset = dest_off_plus;
5270 copy_length = length_minus;
5271 }
5272 }
5273
5274 set_control(slow_control);
5275 if (!stopped()) {
5276 // Generate the slow path, if needed.
5277 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5278
5279 set_memory(slow_mem, adr_type);
5280 set_i_o(slow_i_o);
5281
5282 if (dest_uninitialized) {
5283 generate_clear_array(adr_type, dest, basic_elem_type,
5284 intcon(0), NULL,
5285 alloc->in(AllocateNode::AllocSize));
5286 }
5287
5288 generate_slow_arraycopy(adr_type,
5289 src, src_offset, dest, dest_offset,
5290 copy_length, /*dest_uninitialized*/false);
5291
5292 result_region->init_req(slow_call_path, control());
5293 result_i_o ->init_req(slow_call_path, i_o());
5294 result_memory->init_req(slow_call_path, memory(adr_type));
5295 }
5296
5297 // Remove unused edges.
5298 for (uint i = 1; i < result_region->req(); i++) {
5299 if (result_region->in(i) == NULL)
5300 result_region->init_req(i, top());
5301 }
5302
5303 // Finished; return the combined state.
5304 set_control( _gvn.transform(result_region));
5305 set_i_o( _gvn.transform(result_i_o) );
5306 set_memory( _gvn.transform(result_memory), adr_type );
5307
5308 // The memory edges above are precise in order to model effects around
5309 // array copies accurately to allow value numbering of field loads around
5310 // arraycopy. Such field loads, both before and after, are common in Java
5311 // collections and similar classes involving header/array data structures.
5312 //
5313 // But with low number of register or when some registers are used or killed
5314 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5315 // The next memory barrier is added to avoid it. If the arraycopy can be
5316 // optimized away (which it can, sometimes) then we can manually remove
5317 // the membar also.
5318 //
5319 // Do not let reads from the cloned object float above the arraycopy.
5320 if (alloc != NULL) {
5321 // Do not let stores that initialize this object be reordered with
5322 // a subsequent store that would make this object accessible by
5323 // other threads.
5324 // Record what AllocateNode this StoreStore protects so that
5325 // escape analysis can go from the MemBarStoreStoreNode to the
5326 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5327 // based on the escape status of the AllocateNode.
5328 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5329 } else if (InsertMemBarAfterArraycopy)
5330 insert_mem_bar(Op_MemBarCPUOrder);
5331 }
5332
5333
5334 // Helper function which determines if an arraycopy immediately follows
5335 // an allocation, with no intervening tests or other escapes for the object.
5336 AllocateArrayNode*
5337 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5338 RegionNode* slow_region) {
5339 if (stopped()) return NULL; // no fast path
5340 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5341
5342 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5343 if (alloc == NULL) return NULL;
5344
5345 Node* rawmem = memory(Compile::AliasIdxRaw);
5346 // Is the allocation's memory state untouched?
5347 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5348 // Bail out if there have been raw-memory effects since the allocation.
5349 // (Example: There might have been a call or safepoint.)
5350 return NULL;
5351 }
5352 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5353 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5354 return NULL;
5355 }
5356
5357 // There must be no unexpected observers of this allocation.
5358 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5359 Node* obs = ptr->fast_out(i);
5360 if (obs != this->map()) {
5361 return NULL;
5362 }
5363 }
5364
5365 // This arraycopy must unconditionally follow the allocation of the ptr.
5366 Node* alloc_ctl = ptr->in(0);
5367 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5368
5369 Node* ctl = control();
5370 while (ctl != alloc_ctl) {
5371 // There may be guards which feed into the slow_region.
5372 // Any other control flow means that we might not get a chance
5373 // to finish initializing the allocated object.
5374 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5375 IfNode* iff = ctl->in(0)->as_If();
5376 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5377 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5378 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5379 ctl = iff->in(0); // This test feeds the known slow_region.
5380 continue;
5381 }
5382 // One more try: Various low-level checks bottom out in
5383 // uncommon traps. If the debug-info of the trap omits
5384 // any reference to the allocation, as we've already
5385 // observed, then there can be no objection to the trap.
5386 bool found_trap = false;
5387 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5388 Node* obs = not_ctl->fast_out(j);
5389 if (obs->in(0) == not_ctl && obs->is_Call() &&
5390 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5391 found_trap = true; break;
5392 }
5393 }
5394 if (found_trap) {
5395 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5396 continue;
5397 }
5398 }
5399 return NULL;
5400 }
5401
5402 // If we get this far, we have an allocation which immediately
5403 // precedes the arraycopy, and we can take over zeroing the new object.
5404 // The arraycopy will finish the initialization, and provide
5405 // a new control state to which we will anchor the destination pointer.
5406
5407 return alloc;
5408 }
5409
5410 // Helper for initialization of arrays, creating a ClearArray.
5411 // It writes zero bits in [start..end), within the body of an array object.
5412 // The memory effects are all chained onto the 'adr_type' alias category.
5413 //
5414 // Since the object is otherwise uninitialized, we are free
5415 // to put a little "slop" around the edges of the cleared area,
5416 // as long as it does not go back into the array's header,
5417 // or beyond the array end within the heap.
5418 //
5419 // The lower edge can be rounded down to the nearest jint and the
5420 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5421 //
5422 // Arguments:
5423 // adr_type memory slice where writes are generated
5424 // dest oop of the destination array
5425 // basic_elem_type element type of the destination
5426 // slice_idx array index of first element to store
5427 // slice_len number of elements to store (or NULL)
5428 // dest_size total size in bytes of the array object
5429 //
5430 // Exactly one of slice_len or dest_size must be non-NULL.
5431 // If dest_size is non-NULL, zeroing extends to the end of the object.
5432 // If slice_len is non-NULL, the slice_idx value must be a constant.
5433 void
5434 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5435 Node* dest,
5436 BasicType basic_elem_type,
5437 Node* slice_idx,
5438 Node* slice_len,
5439 Node* dest_size) {
5440 // one or the other but not both of slice_len and dest_size:
5441 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5442 if (slice_len == NULL) slice_len = top();
5443 if (dest_size == NULL) dest_size = top();
5444
5445 // operate on this memory slice:
5446 Node* mem = memory(adr_type); // memory slice to operate on
5447
5448 // scaling and rounding of indexes:
5449 int scale = exact_log2(type2aelembytes(basic_elem_type));
5450 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5451 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5452 int bump_bit = (-1 << scale) & BytesPerInt;
5453
5454 // determine constant starts and ends
5455 const intptr_t BIG_NEG = -128;
5456 assert(BIG_NEG + 2*abase < 0, "neg enough");
5457 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5458 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5459 if (slice_len_con == 0) {
5460 return; // nothing to do here
5461 }
5462 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5463 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5464 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5465 assert(end_con < 0, "not two cons");
5466 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5467 BytesPerLong);
5468 }
5469
5470 if (start_con >= 0 && end_con >= 0) {
5471 // Constant start and end. Simple.
5472 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5473 start_con, end_con, &_gvn);
5474 } else if (start_con >= 0 && dest_size != top()) {
5475 // Constant start, pre-rounded end after the tail of the array.
5476 Node* end = dest_size;
5477 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5478 start_con, end, &_gvn);
5479 } else if (start_con >= 0 && slice_len != top()) {
5480 // Constant start, non-constant end. End needs rounding up.
5481 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5482 intptr_t end_base = abase + (slice_idx_con << scale);
5483 int end_round = (-1 << scale) & (BytesPerLong - 1);
5484 Node* end = ConvI2X(slice_len);
5485 if (scale != 0)
5486 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5487 end_base += end_round;
5488 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5489 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5490 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5491 start_con, end, &_gvn);
5492 } else if (start_con < 0 && dest_size != top()) {
5493 // Non-constant start, pre-rounded end after the tail of the array.
5494 // This is almost certainly a "round-to-end" operation.
5495 Node* start = slice_idx;
5496 start = ConvI2X(start);
5497 if (scale != 0)
5498 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5499 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5500 if ((bump_bit | clear_low) != 0) {
5501 int to_clear = (bump_bit | clear_low);
5502 // Align up mod 8, then store a jint zero unconditionally
5503 // just before the mod-8 boundary.
5504 if (((abase + bump_bit) & ~to_clear) - bump_bit
5505 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5506 bump_bit = 0;
5507 assert((abase & to_clear) == 0, "array base must be long-aligned");
5508 } else {
5509 // Bump 'start' up to (or past) the next jint boundary:
5510 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5511 assert((abase & clear_low) == 0, "array base must be int-aligned");
5512 }
5513 // Round bumped 'start' down to jlong boundary in body of array.
5514 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5515 if (bump_bit != 0) {
5516 // Store a zero to the immediately preceding jint:
5517 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5518 Node* p1 = basic_plus_adr(dest, x1);
5519 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5520 mem = _gvn.transform(mem);
5521 }
5522 }
5523 Node* end = dest_size; // pre-rounded
5524 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5525 start, end, &_gvn);
5526 } else {
5527 // Non-constant start, unrounded non-constant end.
5528 // (Nobody zeroes a random midsection of an array using this routine.)
5529 ShouldNotReachHere(); // fix caller
5530 }
5531
5532 // Done.
5533 set_memory(mem, adr_type);
5534 }
5535
5536
5537 bool
5538 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5539 BasicType basic_elem_type,
5540 AllocateNode* alloc,
5541 Node* src, Node* src_offset,
5542 Node* dest, Node* dest_offset,
5543 Node* dest_size, bool dest_uninitialized) {
5544 // See if there is an advantage from block transfer.
5545 int scale = exact_log2(type2aelembytes(basic_elem_type));
5546 if (scale >= LogBytesPerLong)
5547 return false; // it is already a block transfer
5548
5549 // Look at the alignment of the starting offsets.
5550 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5551
5552 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5553 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5554 if (src_off_con < 0 || dest_off_con < 0)
5555 // At present, we can only understand constants.
5556 return false;
5557
5558 intptr_t src_off = abase + (src_off_con << scale);
5559 intptr_t dest_off = abase + (dest_off_con << scale);
5560
5561 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5562 // Non-aligned; too bad.
5563 // One more chance: Pick off an initial 32-bit word.
5564 // This is a common case, since abase can be odd mod 8.
5565 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5566 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5567 Node* sptr = basic_plus_adr(src, src_off);
5568 Node* dptr = basic_plus_adr(dest, dest_off);
5569 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5570 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5571 src_off += BytesPerInt;
5572 dest_off += BytesPerInt;
5573 } else {
5574 return false;
5575 }
5576 }
5577 assert(src_off % BytesPerLong == 0, "");
5578 assert(dest_off % BytesPerLong == 0, "");
5579
5580 // Do this copy by giant steps.
5581 Node* sptr = basic_plus_adr(src, src_off);
5582 Node* dptr = basic_plus_adr(dest, dest_off);
5583 Node* countx = dest_size;
5584 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5585 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5586
5587 bool disjoint_bases = true; // since alloc != NULL
5588 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5589 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5590
5591 return true;
5592 }
5593
5594
5595 // Helper function; generates code for the slow case.
5596 // We make a call to a runtime method which emulates the native method,
5597 // but without the native wrapper overhead.
5598 void
5599 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5600 Node* src, Node* src_offset,
5601 Node* dest, Node* dest_offset,
5602 Node* copy_length, bool dest_uninitialized) {
5603 assert(!dest_uninitialized, "Invariant");
5604 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5605 OptoRuntime::slow_arraycopy_Type(),
5606 OptoRuntime::slow_arraycopy_Java(),
5607 "slow_arraycopy", adr_type,
5608 src, src_offset, dest, dest_offset,
5609 copy_length);
5610
5611 // Handle exceptions thrown by this fellow:
5612 make_slow_call_ex(call, env()->Throwable_klass(), false);
5613 }
5614
5615 // Helper function; generates code for cases requiring runtime checks.
5616 Node*
5617 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5618 Node* dest_elem_klass,
5619 Node* src, Node* src_offset,
5620 Node* dest, Node* dest_offset,
5621 Node* copy_length, bool dest_uninitialized) {
5622 if (stopped()) return NULL;
5623
5624 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5625 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5626 return NULL;
5627 }
5628
5629 // Pick out the parameters required to perform a store-check
5630 // for the target array. This is an optimistic check. It will
5631 // look in each non-null element's class, at the desired klass's
5632 // super_check_offset, for the desired klass.
5633 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5634 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5635 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5636 Node* check_offset = ConvI2X(_gvn.transform(n3));
5637 Node* check_value = dest_elem_klass;
5638
5639 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5640 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5641
5642 // (We know the arrays are never conjoint, because their types differ.)
5643 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5644 OptoRuntime::checkcast_arraycopy_Type(),
5645 copyfunc_addr, "checkcast_arraycopy", adr_type,
5646 // five arguments, of which two are
5647 // intptr_t (jlong in LP64)
5648 src_start, dest_start,
5649 copy_length XTOP,
5650 check_offset XTOP,
5651 check_value);
5652
5653 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5654 }
5655
5656
5657 // Helper function; generates code for cases requiring runtime checks.
5658 Node*
5659 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5660 Node* src, Node* src_offset,
5661 Node* dest, Node* dest_offset,
5662 Node* copy_length, bool dest_uninitialized) {
5663 assert(!dest_uninitialized, "Invariant");
5664 if (stopped()) return NULL;
5665 address copyfunc_addr = StubRoutines::generic_arraycopy();
5666 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5667 return NULL;
5668 }
5669
5670 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5671 OptoRuntime::generic_arraycopy_Type(),
5672 copyfunc_addr, "generic_arraycopy", adr_type,
5673 src, src_offset, dest, dest_offset, copy_length);
5674
5675 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5676 }
5677
5678 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5679 void
5680 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5681 BasicType basic_elem_type,
5682 bool disjoint_bases,
5683 Node* src, Node* src_offset,
5684 Node* dest, Node* dest_offset,
5685 Node* copy_length, bool dest_uninitialized) {
5686 if (stopped()) return; // nothing to do
5687
5688 Node* src_start = src;
5689 Node* dest_start = dest;
5690 if (src_offset != NULL || dest_offset != NULL) {
5691 assert(src_offset != NULL && dest_offset != NULL, "");
5692 src_start = array_element_address(src, src_offset, basic_elem_type);
5693 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5694 }
5695
5696 // Figure out which arraycopy runtime method to call.
5697 const char* copyfunc_name = "arraycopy";
5698 address copyfunc_addr =
5699 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5700 disjoint_bases, copyfunc_name, dest_uninitialized);
5701
5702 // Call it. Note that the count_ix value is not scaled to a byte-size.
5703 make_runtime_call(RC_LEAF|RC_NO_FP,
5704 OptoRuntime::fast_arraycopy_Type(),
5705 copyfunc_addr, copyfunc_name, adr_type,
5706 src_start, dest_start, copy_length XTOP);
5707 }
5708
5709 //-------------inline_encodeISOArray-----------------------------------
5710 // encode char[] to byte[] in ISO_8859_1
5711 bool LibraryCallKit::inline_encodeISOArray() {
5712 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5713 // no receiver since it is static method
5714 Node *src = argument(0);
5715 Node *src_offset = argument(1);
5716 Node *dst = argument(2);
5717 Node *dst_offset = argument(3);
5718 Node *length = argument(4);
5719
5720 const Type* src_type = src->Value(&_gvn);
5721 const Type* dst_type = dst->Value(&_gvn);
5722 const TypeAryPtr* top_src = src_type->isa_aryptr();
5723 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5724 if (top_src == NULL || top_src->klass() == NULL ||
5725 top_dest == NULL || top_dest->klass() == NULL) {
5726 // failed array check
5727 return false;
5728 }
5729
5730 // Figure out the size and type of the elements we will be copying.
5731 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5732 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5733 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5734 return false;
5735 }
5736 Node* src_start = array_element_address(src, src_offset, src_elem);
5737 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5738 // 'src_start' points to src array + scaled offset
5739 // 'dst_start' points to dst array + scaled offset
5740
5741 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5742 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5743 enc = _gvn.transform(enc);
5744 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5745 set_memory(res_mem, mtype);
5746 set_result(enc);
5747 return true;
5748 }
5749
5750 //-------------inline_multiplyToLen-----------------------------------
5751 bool LibraryCallKit::inline_multiplyToLen() {
5752 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5753
5754 address stubAddr = StubRoutines::multiplyToLen();
5755 if (stubAddr == NULL) {
5756 return false; // Intrinsic's stub is not implemented on this platform
5757 }
5758 const char* stubName = "multiplyToLen";
5759
5760 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5761
5762 Node* x = argument(1);
5763 Node* xlen = argument(2);
5764 Node* y = argument(3);
5765 Node* ylen = argument(4);
5766 Node* z = argument(5);
5767
5768 const Type* x_type = x->Value(&_gvn);
5769 const Type* y_type = y->Value(&_gvn);
5770 const TypeAryPtr* top_x = x_type->isa_aryptr();
5771 const TypeAryPtr* top_y = y_type->isa_aryptr();
5772 if (top_x == NULL || top_x->klass() == NULL ||
5773 top_y == NULL || top_y->klass() == NULL) {
5774 // failed array check
5775 return false;
5776 }
5777
5778 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5779 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5780 if (x_elem != T_INT || y_elem != T_INT) {
5781 return false;
5782 }
5783
5784 // Set the original stack and the reexecute bit for the interpreter to reexecute
5785 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5786 // on the return from z array allocation in runtime.
5787 { PreserveReexecuteState preexecs(this);
5788 jvms()->set_should_reexecute(true);
5789
5790 Node* x_start = array_element_address(x, intcon(0), x_elem);
5791 Node* y_start = array_element_address(y, intcon(0), y_elem);
5792 // 'x_start' points to x array + scaled xlen
5793 // 'y_start' points to y array + scaled ylen
5794
5795 // Allocate the result array
5796 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5797 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5798 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5799
5800 IdealKit ideal(this);
5801
5802 #define __ ideal.
5803 Node* one = __ ConI(1);
5804 Node* zero = __ ConI(0);
5805 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5806 __ set(need_alloc, zero);
5807 __ set(z_alloc, z);
5808 __ if_then(z, BoolTest::eq, null()); {
5809 __ increment (need_alloc, one);
5810 } __ else_(); {
5811 // Update graphKit memory and control from IdealKit.
5812 sync_kit(ideal);
5813 Node* zlen_arg = load_array_length(z);
5814 // Update IdealKit memory and control from graphKit.
5815 __ sync_kit(this);
5816 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5817 __ increment (need_alloc, one);
5818 } __ end_if();
5819 } __ end_if();
5820
5821 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5822 // Update graphKit memory and control from IdealKit.
5823 sync_kit(ideal);
5824 Node * narr = new_array(klass_node, zlen, 1);
5825 // Update IdealKit memory and control from graphKit.
5826 __ sync_kit(this);
5827 __ set(z_alloc, narr);
5828 } __ end_if();
5829
5830 sync_kit(ideal);
5831 z = __ value(z_alloc);
5832 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5833 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5834 // Final sync IdealKit and GraphKit.
5835 final_sync(ideal);
5836 #undef __
5837
5838 Node* z_start = array_element_address(z, intcon(0), T_INT);
5839
5840 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5841 OptoRuntime::multiplyToLen_Type(),
5842 stubAddr, stubName, TypePtr::BOTTOM,
5843 x_start, xlen, y_start, ylen, z_start, zlen);
5844 } // original reexecute is set back here
5845
5846 C->set_has_split_ifs(true); // Has chance for split-if optimization
5847 set_result(z);
5848 return true;
5849 }
5850
5851
5852 /**
5853 * Calculate CRC32 for byte.
5854 * int java.util.zip.CRC32.update(int crc, int b)
5855 */
5856 bool LibraryCallKit::inline_updateCRC32() {
5857 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5858 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5859 // no receiver since it is static method
5860 Node* crc = argument(0); // type: int
5861 Node* b = argument(1); // type: int
5862
5863 /*
5864 * int c = ~ crc;
5865 * b = timesXtoThe32[(b ^ c) & 0xFF];
5866 * b = b ^ (c >>> 8);
5867 * crc = ~b;
5868 */
5869
5870 Node* M1 = intcon(-1);
5871 crc = _gvn.transform(new (C) XorINode(crc, M1));
5872 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5873 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5874
5875 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5876 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5877 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5878 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5879
5880 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5881 result = _gvn.transform(new (C) XorINode(crc, result));
5882 result = _gvn.transform(new (C) XorINode(result, M1));
5883 set_result(result);
5884 return true;
5885 }
5886
5887 /**
5888 * Calculate CRC32 for byte[] array.
5889 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5890 */
5891 bool LibraryCallKit::inline_updateBytesCRC32() {
5892 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5893 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5894 // no receiver since it is static method
5895 Node* crc = argument(0); // type: int
5896 Node* src = argument(1); // type: oop
5897 Node* offset = argument(2); // type: int
5898 Node* length = argument(3); // type: int
5899
5900 const Type* src_type = src->Value(&_gvn);
5901 const TypeAryPtr* top_src = src_type->isa_aryptr();
5902 if (top_src == NULL || top_src->klass() == NULL) {
5903 // failed array check
5904 return false;
5905 }
5906
5907 // Figure out the size and type of the elements we will be copying.
5908 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5909 if (src_elem != T_BYTE) {
5910 return false;
5911 }
5912
5913 // 'src_start' points to src array + scaled offset
5914 Node* src_start = array_element_address(src, offset, src_elem);
5915
5916 // We assume that range check is done by caller.
5917 // TODO: generate range check (offset+length < src.length) in debug VM.
5918
5919 // Call the stub.
5920 address stubAddr = StubRoutines::updateBytesCRC32();
5921 const char *stubName = "updateBytesCRC32";
5922
5923 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5924 stubAddr, stubName, TypePtr::BOTTOM,
5925 crc, src_start, length);
5926 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5927 set_result(result);
5928 return true;
5929 }
5930
5931 /**
5932 * Calculate CRC32 for ByteBuffer.
5933 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5934 */
5935 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5936 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5937 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5938 // no receiver since it is static method
5939 Node* crc = argument(0); // type: int
5940 Node* src = argument(1); // type: long
5941 Node* offset = argument(3); // type: int
5942 Node* length = argument(4); // type: int
5943
5944 src = ConvL2X(src); // adjust Java long to machine word
5945 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5946 offset = ConvI2X(offset);
5947
5948 // 'src_start' points to src array + scaled offset
5949 Node* src_start = basic_plus_adr(top(), base, offset);
5950
5951 // Call the stub.
5952 address stubAddr = StubRoutines::updateBytesCRC32();
5953 const char *stubName = "updateBytesCRC32";
5954
5955 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5956 stubAddr, stubName, TypePtr::BOTTOM,
5957 crc, src_start, length);
5958 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5959 set_result(result);
5960 return true;
5961 }
5962
5963 //----------------------------inline_reference_get----------------------------
5964 // public T java.lang.ref.Reference.get();
5965 bool LibraryCallKit::inline_reference_get() {
5966 const int referent_offset = java_lang_ref_Reference::referent_offset;
5967 guarantee(referent_offset > 0, "should have already been set");
5968
5969 // Get the argument:
5970 Node* reference_obj = null_check_receiver();
5971 if (stopped()) return true;
5972
5973 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5974
5975 ciInstanceKlass* klass = env()->Object_klass();
5976 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5977
5978 Node* no_ctrl = NULL;
5979 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5980
5981 // Use the pre-barrier to record the value in the referent field
5982 pre_barrier(false /* do_load */,
5983 control(),
5984 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5985 result /* pre_val */,
5986 T_OBJECT);
5987
5988 // Add memory barrier to prevent commoning reads from this field
5989 // across safepoint since GC can change its value.
5990 insert_mem_bar(Op_MemBarCPUOrder);
5991
5992 set_result(result);
5993 return true;
5994 }
5995
5996
5997 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5998 bool is_exact=true, bool is_static=false) {
5999
6000 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6001 assert(tinst != NULL, "obj is null");
6002 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6003 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6004
6005 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6006 ciSymbol::make(fieldTypeString),
6007 is_static);
6008 if (field == NULL) return (Node *) NULL;
6009 assert (field != NULL, "undefined field");
6010
6011 // Next code copied from Parse::do_get_xxx():
6012
6013 // Compute address and memory type.
6014 int offset = field->offset_in_bytes();
6015 bool is_vol = field->is_volatile();
6016 ciType* field_klass = field->type();
6017 assert(field_klass->is_loaded(), "should be loaded");
6018 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6019 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6020 BasicType bt = field->layout_type();
6021
6022 // Build the resultant type of the load
6023 const Type *type;
6024 if (bt == T_OBJECT) {
6025 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6026 } else {
6027 type = Type::get_const_basic_type(bt);
6028 }
6029
6030 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6031 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6032 }
6033 // Build the load.
6034 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6035 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
6036 // If reference is volatile, prevent following memory ops from
6037 // floating up past the volatile read. Also prevents commoning
6038 // another volatile read.
6039 if (is_vol) {
6040 // Memory barrier includes bogus read of value to force load BEFORE membar
6041 insert_mem_bar(Op_MemBarAcquire, loadedField);
6042 }
6043 return loadedField;
6044 }
6045
6046
6047 //------------------------------inline_aescrypt_Block-----------------------
6048 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6049 address stubAddr;
6050 const char *stubName;
6051 assert(UseAES, "need AES instruction support");
6052
6053 switch(id) {
6054 case vmIntrinsics::_aescrypt_encryptBlock:
6055 stubAddr = StubRoutines::aescrypt_encryptBlock();
6056 stubName = "aescrypt_encryptBlock";
6057 break;
6058 case vmIntrinsics::_aescrypt_decryptBlock:
6059 stubAddr = StubRoutines::aescrypt_decryptBlock();
6060 stubName = "aescrypt_decryptBlock";
6061 break;
6062 }
6063 if (stubAddr == NULL) return false;
6064
6065 Node* aescrypt_object = argument(0);
6066 Node* src = argument(1);
6067 Node* src_offset = argument(2);
6068 Node* dest = argument(3);
6069 Node* dest_offset = argument(4);
6070
6071 // (1) src and dest are arrays.
6072 const Type* src_type = src->Value(&_gvn);
6073 const Type* dest_type = dest->Value(&_gvn);
6074 const TypeAryPtr* top_src = src_type->isa_aryptr();
6075 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6076 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6077
6078 // for the quick and dirty code we will skip all the checks.
6079 // we are just trying to get the call to be generated.
6080 Node* src_start = src;
6081 Node* dest_start = dest;
6082 if (src_offset != NULL || dest_offset != NULL) {
6083 assert(src_offset != NULL && dest_offset != NULL, "");
6084 src_start = array_element_address(src, src_offset, T_BYTE);
6085 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6086 }
6087
6088 // now need to get the start of its expanded key array
6089 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6090 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6091 if (k_start == NULL) return false;
6092
6093 if (Matcher::pass_original_key_for_aes()) {
6094 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6095 // compatibility issues between Java key expansion and SPARC crypto instructions
6096 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6097 if (original_k_start == NULL) return false;
6098
6099 // Call the stub.
6100 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6101 stubAddr, stubName, TypePtr::BOTTOM,
6102 src_start, dest_start, k_start, original_k_start);
6103 } else {
6104 // Call the stub.
6105 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6106 stubAddr, stubName, TypePtr::BOTTOM,
6107 src_start, dest_start, k_start);
6108 }
6109
6110 return true;
6111 }
6112
6113 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6114 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6115 address stubAddr;
6116 const char *stubName;
6117
6118 assert(UseAES, "need AES instruction support");
6119
6120 switch(id) {
6121 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6122 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6123 stubName = "cipherBlockChaining_encryptAESCrypt";
6124 break;
6125 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6126 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6127 stubName = "cipherBlockChaining_decryptAESCrypt";
6128 break;
6129 }
6130 if (stubAddr == NULL) return false;
6131
6132 Node* cipherBlockChaining_object = argument(0);
6133 Node* src = argument(1);
6134 Node* src_offset = argument(2);
6135 Node* len = argument(3);
6136 Node* dest = argument(4);
6137 Node* dest_offset = argument(5);
6138
6139 // (1) src and dest are arrays.
6140 const Type* src_type = src->Value(&_gvn);
6141 const Type* dest_type = dest->Value(&_gvn);
6142 const TypeAryPtr* top_src = src_type->isa_aryptr();
6143 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6144 assert (top_src != NULL && top_src->klass() != NULL
6145 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6146
6147 // checks are the responsibility of the caller
6148 Node* src_start = src;
6149 Node* dest_start = dest;
6150 if (src_offset != NULL || dest_offset != NULL) {
6151 assert(src_offset != NULL && dest_offset != NULL, "");
6152 src_start = array_element_address(src, src_offset, T_BYTE);
6153 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6154 }
6155
6156 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6157 // (because of the predicated logic executed earlier).
6158 // so we cast it here safely.
6159 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6160
6161 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6162 if (embeddedCipherObj == NULL) return false;
6163
6164 // cast it to what we know it will be at runtime
6165 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6166 assert(tinst != NULL, "CBC obj is null");
6167 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6168 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6169 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6170
6171 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6172 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6173 const TypeOopPtr* xtype = aklass->as_instance_type();
6174 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6175 aescrypt_object = _gvn.transform(aescrypt_object);
6176
6177 // we need to get the start of the aescrypt_object's expanded key array
6178 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6179 if (k_start == NULL) return false;
6180
6181 // similarly, get the start address of the r vector
6182 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6183 if (objRvec == NULL) return false;
6184 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6185
6186 Node* cbcCrypt;
6187 if (Matcher::pass_original_key_for_aes()) {
6188 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6189 // compatibility issues between Java key expansion and SPARC crypto instructions
6190 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6191 if (original_k_start == NULL) return false;
6192
6193 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6194 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6195 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6196 stubAddr, stubName, TypePtr::BOTTOM,
6197 src_start, dest_start, k_start, r_start, len, original_k_start);
6198 } else {
6199 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6200 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6201 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6202 stubAddr, stubName, TypePtr::BOTTOM,
6203 src_start, dest_start, k_start, r_start, len);
6204 }
6205
6206 // return cipher length (int)
6207 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6208 set_result(retvalue);
6209 return true;
6210 }
6211
6212 //------------------------------get_key_start_from_aescrypt_object-----------------------
6213 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6214 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6215 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6216 if (objAESCryptKey == NULL) return (Node *) NULL;
6217
6218 // now have the array, need to get the start address of the K array
6219 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6220 return k_start;
6221 }
6222
6223 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6224 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6225 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6226 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6227 if (objAESCryptKey == NULL) return (Node *) NULL;
6228
6229 // now have the array, need to get the start address of the lastKey array
6230 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6231 return original_k_start;
6232 }
6233
6234 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6235 // Return node representing slow path of predicate check.
6236 // the pseudo code we want to emulate with this predicate is:
6237 // for encryption:
6238 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6239 // for decryption:
6240 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6241 // note cipher==plain is more conservative than the original java code but that's OK
6242 //
6243 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6244 // The receiver was checked for NULL already.
6245 Node* objCBC = argument(0);
6246
6247 // Load embeddedCipher field of CipherBlockChaining object.
6248 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6249
6250 // get AESCrypt klass for instanceOf check
6251 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6252 // will have same classloader as CipherBlockChaining object
6253 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6254 assert(tinst != NULL, "CBCobj is null");
6255 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6256
6257 // we want to do an instanceof comparison against the AESCrypt class
6258 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6259 if (!klass_AESCrypt->is_loaded()) {
6260 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6261 Node* ctrl = control();
6262 set_control(top()); // no regular fast path
6263 return ctrl;
6264 }
6265 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6266
6267 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6268 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6269 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6270
6271 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6272
6273 // for encryption, we are done
6274 if (!decrypting)
6275 return instof_false; // even if it is NULL
6276
6277 // for decryption, we need to add a further check to avoid
6278 // taking the intrinsic path when cipher and plain are the same
6279 // see the original java code for why.
6280 RegionNode* region = new(C) RegionNode(3);
6281 region->init_req(1, instof_false);
6282 Node* src = argument(1);
6283 Node* dest = argument(4);
6284 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6285 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6286 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6287 region->init_req(2, src_dest_conjoint);
6288
6289 record_for_igvn(region);
6290 return _gvn.transform(region);
6291 }
6292
6293 //------------------------------inline_sha_implCompress-----------------------
6294 //
6295 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6296 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6297 //
6298 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6299 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6300 //
6301 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6302 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6303 //
6304 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6305 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6306
6307 Node* sha_obj = argument(0);
6308 Node* src = argument(1); // type oop
6309 Node* ofs = argument(2); // type int
6310
6311 const Type* src_type = src->Value(&_gvn);
6312 const TypeAryPtr* top_src = src_type->isa_aryptr();
6313 if (top_src == NULL || top_src->klass() == NULL) {
6314 // failed array check
6315 return false;
6316 }
6317 // Figure out the size and type of the elements we will be copying.
6318 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6319 if (src_elem != T_BYTE) {
6320 return false;
6321 }
6322 // 'src_start' points to src array + offset
6323 Node* src_start = array_element_address(src, ofs, src_elem);
6324 Node* state = NULL;
6325 address stubAddr;
6326 const char *stubName;
6327
6328 switch(id) {
6329 case vmIntrinsics::_sha_implCompress:
6330 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6331 state = get_state_from_sha_object(sha_obj);
6332 stubAddr = StubRoutines::sha1_implCompress();
6333 stubName = "sha1_implCompress";
6334 break;
6335 case vmIntrinsics::_sha2_implCompress:
6336 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6337 state = get_state_from_sha_object(sha_obj);
6338 stubAddr = StubRoutines::sha256_implCompress();
6339 stubName = "sha256_implCompress";
6340 break;
6341 case vmIntrinsics::_sha5_implCompress:
6342 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6343 state = get_state_from_sha5_object(sha_obj);
6344 stubAddr = StubRoutines::sha512_implCompress();
6345 stubName = "sha512_implCompress";
6346 break;
6347 default:
6348 fatal_unexpected_iid(id);
6349 return false;
6350 }
6351 if (state == NULL) return false;
6352
6353 // Call the stub.
6354 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6355 stubAddr, stubName, TypePtr::BOTTOM,
6356 src_start, state);
6357
6358 return true;
6359 }
6360
6361 //------------------------------inline_digestBase_implCompressMB-----------------------
6362 //
6363 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6364 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6365 //
6366 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6367 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6368 "need SHA1/SHA256/SHA512 instruction support");
6369 assert((uint)predicate < 3, "sanity");
6370 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6371
6372 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6373 Node* src = argument(1); // byte[] array
6374 Node* ofs = argument(2); // type int
6375 Node* limit = argument(3); // type int
6376
6377 const Type* src_type = src->Value(&_gvn);
6378 const TypeAryPtr* top_src = src_type->isa_aryptr();
6379 if (top_src == NULL || top_src->klass() == NULL) {
6380 // failed array check
6381 return false;
6382 }
6383 // Figure out the size and type of the elements we will be copying.
6384 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6385 if (src_elem != T_BYTE) {
6386 return false;
6387 }
6388 // 'src_start' points to src array + offset
6389 Node* src_start = array_element_address(src, ofs, src_elem);
6390
6391 const char* klass_SHA_name = NULL;
6392 const char* stub_name = NULL;
6393 address stub_addr = NULL;
6394 bool long_state = false;
6395
6396 switch (predicate) {
6397 case 0:
6398 if (UseSHA1Intrinsics) {
6399 klass_SHA_name = "sun/security/provider/SHA";
6400 stub_name = "sha1_implCompressMB";
6401 stub_addr = StubRoutines::sha1_implCompressMB();
6402 }
6403 break;
6404 case 1:
6405 if (UseSHA256Intrinsics) {
6406 klass_SHA_name = "sun/security/provider/SHA2";
6407 stub_name = "sha256_implCompressMB";
6408 stub_addr = StubRoutines::sha256_implCompressMB();
6409 }
6410 break;
6411 case 2:
6412 if (UseSHA512Intrinsics) {
6413 klass_SHA_name = "sun/security/provider/SHA5";
6414 stub_name = "sha512_implCompressMB";
6415 stub_addr = StubRoutines::sha512_implCompressMB();
6416 long_state = true;
6417 }
6418 break;
6419 default:
6420 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6421 }
6422 if (klass_SHA_name != NULL) {
6423 // get DigestBase klass to lookup for SHA klass
6424 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6425 assert(tinst != NULL, "digestBase_obj is not instance???");
6426 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6427
6428 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6429 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6430 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6431 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6432 }
6433 return false;
6434 }
6435 //------------------------------inline_sha_implCompressMB-----------------------
6436 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6437 bool long_state, address stubAddr, const char *stubName,
6438 Node* src_start, Node* ofs, Node* limit) {
6439 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6440 const TypeOopPtr* xtype = aklass->as_instance_type();
6441 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6442 sha_obj = _gvn.transform(sha_obj);
6443
6444 Node* state;
6445 if (long_state) {
6446 state = get_state_from_sha5_object(sha_obj);
6447 } else {
6448 state = get_state_from_sha_object(sha_obj);
6449 }
6450 if (state == NULL) return false;
6451
6452 // Call the stub.
6453 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6454 OptoRuntime::digestBase_implCompressMB_Type(),
6455 stubAddr, stubName, TypePtr::BOTTOM,
6456 src_start, state, ofs, limit);
6457 // return ofs (int)
6458 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6459 set_result(result);
6460
6461 return true;
6462 }
6463
6464 //------------------------------get_state_from_sha_object-----------------------
6465 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6466 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6467 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6468 if (sha_state == NULL) return (Node *) NULL;
6469
6470 // now have the array, need to get the start address of the state array
6471 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6472 return state;
6473 }
6474
6475 //------------------------------get_state_from_sha5_object-----------------------
6476 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6477 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6478 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6479 if (sha_state == NULL) return (Node *) NULL;
6480
6481 // now have the array, need to get the start address of the state array
6482 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6483 return state;
6484 }
6485
6486 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6487 // Return node representing slow path of predicate check.
6488 // the pseudo code we want to emulate with this predicate is:
6489 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6490 //
6491 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6492 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6493 "need SHA1/SHA256/SHA512 instruction support");
6494 assert((uint)predicate < 3, "sanity");
6495
6496 // The receiver was checked for NULL already.
6497 Node* digestBaseObj = argument(0);
6498
6499 // get DigestBase klass for instanceOf check
6500 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6501 assert(tinst != NULL, "digestBaseObj is null");
6502 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6503
6504 const char* klass_SHA_name = NULL;
6505 switch (predicate) {
6506 case 0:
6507 if (UseSHA1Intrinsics) {
6508 // we want to do an instanceof comparison against the SHA class
6509 klass_SHA_name = "sun/security/provider/SHA";
6510 }
6511 break;
6512 case 1:
6513 if (UseSHA256Intrinsics) {
6514 // we want to do an instanceof comparison against the SHA2 class
6515 klass_SHA_name = "sun/security/provider/SHA2";
6516 }
6517 break;
6518 case 2:
6519 if (UseSHA512Intrinsics) {
6520 // we want to do an instanceof comparison against the SHA5 class
6521 klass_SHA_name = "sun/security/provider/SHA5";
6522 }
6523 break;
6524 default:
6525 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6526 }
6527
6528 ciKlass* klass_SHA = NULL;
6529 if (klass_SHA_name != NULL) {
6530 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6531 }
6532 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6533 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6534 Node* ctrl = control();
6535 set_control(top()); // no intrinsic path
6536 return ctrl;
6537 }
6538 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6539
6540 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6541 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6542 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6543 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6544
6545 return instof_false; // even if it is NULL
6546 }