1 /*
2 * Copyright (c) 1999, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/countbitsnode.hpp"
37 #include "opto/intrinsicnode.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/mathexactnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "prims/nativeLookup.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "trace/traceMacros.hpp"
49
50 class LibraryIntrinsic : public InlineCallGenerator {
51 // Extend the set of intrinsics known to the runtime:
52 public:
53 private:
54 bool _is_virtual;
55 bool _does_virtual_dispatch;
56 int8_t _predicates_count; // Intrinsic is predicated by several conditions
57 int8_t _last_predicate; // Last generated predicate
58 vmIntrinsics::ID _intrinsic_id;
59
60 public:
61 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
62 : InlineCallGenerator(m),
63 _is_virtual(is_virtual),
64 _does_virtual_dispatch(does_virtual_dispatch),
65 _predicates_count((int8_t)predicates_count),
66 _last_predicate((int8_t)-1),
67 _intrinsic_id(id)
68 {
69 }
70 virtual bool is_intrinsic() const { return true; }
71 virtual bool is_virtual() const { return _is_virtual; }
72 virtual bool is_predicated() const { return _predicates_count > 0; }
73 virtual int predicates_count() const { return _predicates_count; }
74 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
75 virtual JVMState* generate(JVMState* jvms);
76 virtual Node* generate_predicate(JVMState* jvms, int predicate);
77 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
78 };
79
80
81 // Local helper class for LibraryIntrinsic:
82 class LibraryCallKit : public GraphKit {
83 private:
84 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
85 Node* _result; // the result node, if any
86 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
87
88 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
89
90 public:
91 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
92 : GraphKit(jvms),
93 _intrinsic(intrinsic),
94 _result(NULL)
95 {
96 // Check if this is a root compile. In that case we don't have a caller.
97 if (!jvms->has_method()) {
98 _reexecute_sp = sp();
99 } else {
100 // Find out how many arguments the interpreter needs when deoptimizing
101 // and save the stack pointer value so it can used by uncommon_trap.
102 // We find the argument count by looking at the declared signature.
103 bool ignored_will_link;
104 ciSignature* declared_signature = NULL;
105 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
106 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
107 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
108 }
109 }
110
111 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
112
113 ciMethod* caller() const { return jvms()->method(); }
114 int bci() const { return jvms()->bci(); }
115 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
116 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
117 ciMethod* callee() const { return _intrinsic->method(); }
118
119 bool try_to_inline(int predicate);
120 Node* try_to_predicate(int predicate);
121
122 void push_result() {
123 // Push the result onto the stack.
124 if (!stopped() && result() != NULL) {
125 BasicType bt = result()->bottom_type()->basic_type();
126 push_node(bt, result());
127 }
128 }
129
130 private:
131 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
132 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
133 }
134
135 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
136 void set_result(RegionNode* region, PhiNode* value);
137 Node* result() { return _result; }
138
139 virtual int reexecute_sp() { return _reexecute_sp; }
140
141 // Helper functions to inline natives
142 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
143 Node* generate_slow_guard(Node* test, RegionNode* region);
144 Node* generate_fair_guard(Node* test, RegionNode* region);
145 Node* generate_negative_guard(Node* index, RegionNode* region,
146 // resulting CastII of index:
147 Node* *pos_index = NULL);
148 Node* generate_nonpositive_guard(Node* index, bool never_negative,
149 // resulting CastII of index:
150 Node* *pos_index = NULL);
151 Node* generate_limit_guard(Node* offset, Node* subseq_length,
152 Node* array_length,
153 RegionNode* region);
154 Node* generate_current_thread(Node* &tls_output);
155 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
156 bool disjoint_bases, const char* &name, bool dest_uninitialized);
157 Node* load_mirror_from_klass(Node* klass);
158 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
159 RegionNode* region, int null_path,
160 int offset);
161 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
162 RegionNode* region, int null_path) {
163 int offset = java_lang_Class::klass_offset_in_bytes();
164 return load_klass_from_mirror_common(mirror, never_see_null,
165 region, null_path,
166 offset);
167 }
168 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
169 RegionNode* region, int null_path) {
170 int offset = java_lang_Class::array_klass_offset_in_bytes();
171 return load_klass_from_mirror_common(mirror, never_see_null,
172 region, null_path,
173 offset);
174 }
175 Node* generate_access_flags_guard(Node* kls,
176 int modifier_mask, int modifier_bits,
177 RegionNode* region);
178 Node* generate_interface_guard(Node* kls, RegionNode* region);
179 Node* generate_array_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, false, false);
181 }
182 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
183 return generate_array_guard_common(kls, region, false, true);
184 }
185 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, true, false);
187 }
188 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
189 return generate_array_guard_common(kls, region, true, true);
190 }
191 Node* generate_array_guard_common(Node* kls, RegionNode* region,
192 bool obj_array, bool not_array);
193 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
194 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
195 bool is_virtual = false, bool is_static = false);
196 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
197 return generate_method_call(method_id, false, true);
198 }
199 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
200 return generate_method_call(method_id, true, false);
201 }
202 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
203
204 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
205 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
206 bool inline_string_compareTo();
207 bool inline_string_indexOf();
208 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
209 bool inline_string_equals();
210 Node* round_double_node(Node* n);
211 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
212 bool inline_math_native(vmIntrinsics::ID id);
213 bool inline_trig(vmIntrinsics::ID id);
214 bool inline_math(vmIntrinsics::ID id);
215 template <typename OverflowOp>
216 bool inline_math_overflow(Node* arg1, Node* arg2);
217 void inline_math_mathExact(Node* math, Node* test);
218 bool inline_math_addExactI(bool is_increment);
219 bool inline_math_addExactL(bool is_increment);
220 bool inline_math_multiplyExactI();
221 bool inline_math_multiplyExactL();
222 bool inline_math_negateExactI();
223 bool inline_math_negateExactL();
224 bool inline_math_subtractExactI(bool is_decrement);
225 bool inline_math_subtractExactL(bool is_decrement);
226 bool inline_exp();
227 bool inline_pow();
228 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
229 bool inline_min_max(vmIntrinsics::ID id);
230 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
231 // This returns Type::AnyPtr, RawPtr, or OopPtr.
232 int classify_unsafe_addr(Node* &base, Node* &offset);
233 Node* make_unsafe_address(Node* base, Node* offset);
234 // Helper for inline_unsafe_access.
235 // Generates the guards that check whether the result of
236 // Unsafe.getObject should be recorded in an SATB log buffer.
237 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
238 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
239 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
240 static bool klass_needs_init_guard(Node* kls);
241 bool inline_unsafe_allocate();
242 bool inline_unsafe_copyMemory();
243 bool inline_native_currentThread();
244 #ifdef TRACE_HAVE_INTRINSICS
245 bool inline_native_classID();
246 bool inline_native_threadID();
247 #endif
248 bool inline_native_time_funcs(address method, const char* funcName);
249 bool inline_native_isInterrupted();
250 bool inline_native_Class_query(vmIntrinsics::ID id);
251 bool inline_native_subtype_check();
252
253 bool inline_native_newArray();
254 bool inline_native_getLength();
255 bool inline_array_copyOf(bool is_copyOfRange);
256 bool inline_array_equals();
257 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
258 bool inline_native_clone(bool is_virtual);
259 bool inline_native_Reflection_getCallerClass();
260 // Helper function for inlining native object hash method
261 bool inline_native_hashcode(bool is_virtual, bool is_static);
262 bool inline_native_getClass();
263
264 // Helper functions for inlining arraycopy
265 bool inline_arraycopy();
266 void generate_arraycopy(const TypePtr* adr_type,
267 BasicType basic_elem_type,
268 Node* src, Node* src_offset,
269 Node* dest, Node* dest_offset,
270 Node* copy_length,
271 bool disjoint_bases = false,
272 bool length_never_negative = false,
273 RegionNode* slow_region = NULL);
274 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
275 RegionNode* slow_region);
276 void generate_clear_array(const TypePtr* adr_type,
277 Node* dest,
278 BasicType basic_elem_type,
279 Node* slice_off,
280 Node* slice_len,
281 Node* slice_end);
282 bool generate_block_arraycopy(const TypePtr* adr_type,
283 BasicType basic_elem_type,
284 AllocateNode* alloc,
285 Node* src, Node* src_offset,
286 Node* dest, Node* dest_offset,
287 Node* dest_size, bool dest_uninitialized);
288 void generate_slow_arraycopy(const TypePtr* adr_type,
289 Node* src, Node* src_offset,
290 Node* dest, Node* dest_offset,
291 Node* copy_length, bool dest_uninitialized);
292 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
293 Node* dest_elem_klass,
294 Node* src, Node* src_offset,
295 Node* dest, Node* dest_offset,
296 Node* copy_length, bool dest_uninitialized);
297 Node* generate_generic_arraycopy(const TypePtr* adr_type,
298 Node* src, Node* src_offset,
299 Node* dest, Node* dest_offset,
300 Node* copy_length, bool dest_uninitialized);
301 void generate_unchecked_arraycopy(const TypePtr* adr_type,
302 BasicType basic_elem_type,
303 bool disjoint_bases,
304 Node* src, Node* src_offset,
305 Node* dest, Node* dest_offset,
306 Node* copy_length, bool dest_uninitialized);
307 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
308 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
309 bool inline_unsafe_ordered_store(BasicType type);
310 bool inline_unsafe_fence(vmIntrinsics::ID id);
311 bool inline_fp_conversions(vmIntrinsics::ID id);
312 bool inline_number_methods(vmIntrinsics::ID id);
313 bool inline_reference_get();
314 bool inline_aescrypt_Block(vmIntrinsics::ID id);
315 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
316 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
317 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
318 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
319 bool inline_sha_implCompress(vmIntrinsics::ID id);
320 bool inline_digestBase_implCompressMB(int predicate);
321 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
322 bool long_state, address stubAddr, const char *stubName,
323 Node* src_start, Node* ofs, Node* limit);
324 Node* get_state_from_sha_object(Node *sha_object);
325 Node* get_state_from_sha5_object(Node *sha_object);
326 Node* inline_digestBase_implCompressMB_predicate(int predicate);
327 bool inline_encodeISOArray();
328 bool inline_updateCRC32();
329 bool inline_updateBytesCRC32();
330 bool inline_updateByteBufferCRC32();
331 };
332
333
334 //---------------------------make_vm_intrinsic----------------------------
335 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
336 vmIntrinsics::ID id = m->intrinsic_id();
337 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
338
339 if (DisableIntrinsic[0] != '\0'
340 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
341 // disabled by a user request on the command line:
342 // example: -XX:DisableIntrinsic=_hashCode,_getClass
343 return NULL;
344 }
345
346 if (!m->is_loaded()) {
347 // do not attempt to inline unloaded methods
348 return NULL;
349 }
350
351 // Only a few intrinsics implement a virtual dispatch.
352 // They are expensive calls which are also frequently overridden.
353 if (is_virtual) {
354 switch (id) {
355 case vmIntrinsics::_hashCode:
356 case vmIntrinsics::_clone:
357 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
358 break;
359 default:
360 return NULL;
361 }
362 }
363
364 // -XX:-InlineNatives disables nearly all intrinsics:
365 if (!InlineNatives) {
366 switch (id) {
367 case vmIntrinsics::_indexOf:
368 case vmIntrinsics::_compareTo:
369 case vmIntrinsics::_equals:
370 case vmIntrinsics::_equalsC:
371 case vmIntrinsics::_getAndAddInt:
372 case vmIntrinsics::_getAndAddLong:
373 case vmIntrinsics::_getAndSetInt:
374 case vmIntrinsics::_getAndSetLong:
375 case vmIntrinsics::_getAndSetObject:
376 case vmIntrinsics::_loadFence:
377 case vmIntrinsics::_storeFence:
378 case vmIntrinsics::_fullFence:
379 break; // InlineNatives does not control String.compareTo
380 case vmIntrinsics::_Reference_get:
381 break; // InlineNatives does not control Reference.get
382 default:
383 return NULL;
384 }
385 }
386
387 int predicates = 0;
388 bool does_virtual_dispatch = false;
389
390 switch (id) {
391 case vmIntrinsics::_compareTo:
392 if (!SpecialStringCompareTo) return NULL;
393 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
394 break;
395 case vmIntrinsics::_indexOf:
396 if (!SpecialStringIndexOf) return NULL;
397 break;
398 case vmIntrinsics::_equals:
399 if (!SpecialStringEquals) return NULL;
400 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
401 break;
402 case vmIntrinsics::_equalsC:
403 if (!SpecialArraysEquals) return NULL;
404 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
405 break;
406 case vmIntrinsics::_arraycopy:
407 if (!InlineArrayCopy) return NULL;
408 break;
409 case vmIntrinsics::_copyMemory:
410 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
411 if (!InlineArrayCopy) return NULL;
412 break;
413 case vmIntrinsics::_hashCode:
414 if (!InlineObjectHash) return NULL;
415 does_virtual_dispatch = true;
416 break;
417 case vmIntrinsics::_clone:
418 does_virtual_dispatch = true;
419 case vmIntrinsics::_copyOf:
420 case vmIntrinsics::_copyOfRange:
421 if (!InlineObjectCopy) return NULL;
422 // These also use the arraycopy intrinsic mechanism:
423 if (!InlineArrayCopy) return NULL;
424 break;
425 case vmIntrinsics::_encodeISOArray:
426 if (!SpecialEncodeISOArray) return NULL;
427 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
428 break;
429 case vmIntrinsics::_checkIndex:
430 // We do not intrinsify this. The optimizer does fine with it.
431 return NULL;
432
433 case vmIntrinsics::_getCallerClass:
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::_cipherBlockChaining_encryptAESCrypt:
524 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
525 if (!UseAESIntrinsics) return NULL;
526 // these two require the predicated logic
527 predicates = 1;
528 break;
529
530 case vmIntrinsics::_sha_implCompress:
531 if (!UseSHA1Intrinsics) return NULL;
532 break;
533
534 case vmIntrinsics::_sha2_implCompress:
535 if (!UseSHA256Intrinsics) return NULL;
536 break;
537
538 case vmIntrinsics::_sha5_implCompress:
539 if (!UseSHA512Intrinsics) return NULL;
540 break;
541
542 case vmIntrinsics::_digestBase_implCompressMB:
543 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
544 predicates = 3;
545 break;
546
547 case vmIntrinsics::_updateCRC32:
548 case vmIntrinsics::_updateBytesCRC32:
549 case vmIntrinsics::_updateByteBufferCRC32:
550 if (!UseCRC32Intrinsics) return NULL;
551 break;
552
553 case vmIntrinsics::_incrementExactI:
554 case vmIntrinsics::_addExactI:
555 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
556 break;
557 case vmIntrinsics::_incrementExactL:
558 case vmIntrinsics::_addExactL:
559 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
560 break;
561 case vmIntrinsics::_decrementExactI:
562 case vmIntrinsics::_subtractExactI:
563 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
564 break;
565 case vmIntrinsics::_decrementExactL:
566 case vmIntrinsics::_subtractExactL:
567 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
568 break;
569 case vmIntrinsics::_negateExactI:
570 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
571 break;
572 case vmIntrinsics::_negateExactL:
573 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
574 break;
575 case vmIntrinsics::_multiplyExactI:
576 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
577 break;
578 case vmIntrinsics::_multiplyExactL:
579 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
580 break;
581
582 default:
583 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
584 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
585 break;
586 }
587
588 // -XX:-InlineClassNatives disables natives from the Class class.
589 // The flag applies to all reflective calls, notably Array.newArray
590 // (visible to Java programmers as Array.newInstance).
591 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
592 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
593 if (!InlineClassNatives) return NULL;
594 }
595
596 // -XX:-InlineThreadNatives disables natives from the Thread class.
597 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
598 if (!InlineThreadNatives) return NULL;
599 }
600
601 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
602 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
603 m->holder()->name() == ciSymbol::java_lang_Float() ||
604 m->holder()->name() == ciSymbol::java_lang_Double()) {
605 if (!InlineMathNatives) return NULL;
606 }
607
608 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
609 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
610 if (!InlineUnsafeOps) return NULL;
611 }
612
613 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
614 }
615
616 //----------------------register_library_intrinsics-----------------------
617 // Initialize this file's data structures, for each Compile instance.
618 void Compile::register_library_intrinsics() {
619 // Nothing to do here.
620 }
621
622 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
623 LibraryCallKit kit(jvms, this);
624 Compile* C = kit.C;
625 int nodes = C->unique();
626 #ifndef PRODUCT
627 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
628 char buf[1000];
629 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
630 tty->print_cr("Intrinsic %s", str);
631 }
632 #endif
633 ciMethod* callee = kit.callee();
634 const int bci = kit.bci();
635
636 // Try to inline the intrinsic.
637 if (kit.try_to_inline(_last_predicate)) {
638 if (C->print_intrinsics() || C->print_inlining()) {
639 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
640 }
641 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
642 if (C->log()) {
643 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
644 vmIntrinsics::name_at(intrinsic_id()),
645 (is_virtual() ? " virtual='1'" : ""),
646 C->unique() - nodes);
647 }
648 // Push the result from the inlined method onto the stack.
649 kit.push_result();
650 C->print_inlining_update(this);
651 return kit.transfer_exceptions_into_jvms();
652 }
653
654 // The intrinsic bailed out
655 if (C->print_intrinsics() || C->print_inlining()) {
656 if (jvms->has_method()) {
657 // Not a root compile.
658 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
659 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
660 } else {
661 // Root compile
662 tty->print("Did not generate intrinsic %s%s at bci:%d in",
663 vmIntrinsics::name_at(intrinsic_id()),
664 (is_virtual() ? " (virtual)" : ""), bci);
665 }
666 }
667 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
668 C->print_inlining_update(this);
669 return NULL;
670 }
671
672 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
673 LibraryCallKit kit(jvms, this);
674 Compile* C = kit.C;
675 int nodes = C->unique();
676 _last_predicate = predicate;
677 #ifndef PRODUCT
678 assert(is_predicated() && predicate < predicates_count(), "sanity");
679 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
680 char buf[1000];
681 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
682 tty->print_cr("Predicate for intrinsic %s", str);
683 }
684 #endif
685 ciMethod* callee = kit.callee();
686 const int bci = kit.bci();
687
688 Node* slow_ctl = kit.try_to_predicate(predicate);
689 if (!kit.failing()) {
690 if (C->print_intrinsics() || C->print_inlining()) {
691 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
692 }
693 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
694 if (C->log()) {
695 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
696 vmIntrinsics::name_at(intrinsic_id()),
697 (is_virtual() ? " virtual='1'" : ""),
698 C->unique() - nodes);
699 }
700 return slow_ctl; // Could be NULL if the check folds.
701 }
702
703 // The intrinsic bailed out
704 if (C->print_intrinsics() || C->print_inlining()) {
705 if (jvms->has_method()) {
706 // Not a root compile.
707 const char* msg = "failed to generate predicate for intrinsic";
708 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
709 } else {
710 // Root compile
711 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
712 vmIntrinsics::name_at(intrinsic_id()),
713 (is_virtual() ? " (virtual)" : ""), bci);
714 }
715 }
716 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
717 return NULL;
718 }
719
720 bool LibraryCallKit::try_to_inline(int predicate) {
721 // Handle symbolic names for otherwise undistinguished boolean switches:
722 const bool is_store = true;
723 const bool is_native_ptr = true;
724 const bool is_static = true;
725 const bool is_volatile = true;
726
727 if (!jvms()->has_method()) {
728 // Root JVMState has a null method.
729 assert(map()->memory()->Opcode() == Op_Parm, "");
730 // Insert the memory aliasing node
731 set_all_memory(reset_memory());
732 }
733 assert(merged_memory(), "");
734
735
736 switch (intrinsic_id()) {
737 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
738 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
739 case vmIntrinsics::_getClass: return inline_native_getClass();
740
741 case vmIntrinsics::_dsin:
742 case vmIntrinsics::_dcos:
743 case vmIntrinsics::_dtan:
744 case vmIntrinsics::_dabs:
745 case vmIntrinsics::_datan2:
746 case vmIntrinsics::_dsqrt:
747 case vmIntrinsics::_dexp:
748 case vmIntrinsics::_dlog:
749 case vmIntrinsics::_dlog10:
750 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
751
752 case vmIntrinsics::_min:
753 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
754
755 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
756 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
757 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
758 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
759 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
760 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
761 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
762 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
763 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
764 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
765 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
766 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
767
768 case vmIntrinsics::_arraycopy: return inline_arraycopy();
769
770 case vmIntrinsics::_compareTo: return inline_string_compareTo();
771 case vmIntrinsics::_indexOf: return inline_string_indexOf();
772 case vmIntrinsics::_equals: return inline_string_equals();
773
774 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
775 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
776 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
777 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
778 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
779 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
780 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
781 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
782 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
783
784 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
785 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
786 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
787 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
788 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
789 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
790 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
791 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
792 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
793
794 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
795 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
796 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
797 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
798 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
799 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
800 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
801 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
802
803 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
804 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
805 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
806 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
807 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
808 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
809 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
810 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
811
812 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
813 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
814 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
815 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
816 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
817 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
818 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
819 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
820 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
821
822 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
823 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
824 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
825 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
826 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
827 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
828 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
829 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
830 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
831
832 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
833 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
834 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
835 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
836
837 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
838 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
839 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
840
841 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
842 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
843 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
844
845 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
846 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
847 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
848 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
849 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
850
851 case vmIntrinsics::_loadFence:
852 case vmIntrinsics::_storeFence:
853 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
854
855 case vmIntrinsics::_currentThread: return inline_native_currentThread();
856 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
857
858 #ifdef TRACE_HAVE_INTRINSICS
859 case vmIntrinsics::_classID: return inline_native_classID();
860 case vmIntrinsics::_threadID: return inline_native_threadID();
861 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
862 #endif
863 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
864 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
865 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
866 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
867 case vmIntrinsics::_newArray: return inline_native_newArray();
868 case vmIntrinsics::_getLength: return inline_native_getLength();
869 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
870 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
871 case vmIntrinsics::_equalsC: return inline_array_equals();
872 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
873
874 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
875
876 case vmIntrinsics::_isInstance:
877 case vmIntrinsics::_getModifiers:
878 case vmIntrinsics::_isInterface:
879 case vmIntrinsics::_isArray:
880 case vmIntrinsics::_isPrimitive:
881 case vmIntrinsics::_getSuperclass:
882 case vmIntrinsics::_getComponentType:
883 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
884
885 case vmIntrinsics::_floatToRawIntBits:
886 case vmIntrinsics::_floatToIntBits:
887 case vmIntrinsics::_intBitsToFloat:
888 case vmIntrinsics::_doubleToRawLongBits:
889 case vmIntrinsics::_doubleToLongBits:
890 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
891
892 case vmIntrinsics::_numberOfLeadingZeros_i:
893 case vmIntrinsics::_numberOfLeadingZeros_l:
894 case vmIntrinsics::_numberOfTrailingZeros_i:
895 case vmIntrinsics::_numberOfTrailingZeros_l:
896 case vmIntrinsics::_bitCount_i:
897 case vmIntrinsics::_bitCount_l:
898 case vmIntrinsics::_reverseBytes_i:
899 case vmIntrinsics::_reverseBytes_l:
900 case vmIntrinsics::_reverseBytes_s:
901 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
902
903 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
904
905 case vmIntrinsics::_Reference_get: return inline_reference_get();
906
907 case vmIntrinsics::_aescrypt_encryptBlock:
908 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
909
910 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
911 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
912 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
913
914 case vmIntrinsics::_sha_implCompress:
915 case vmIntrinsics::_sha2_implCompress:
916 case vmIntrinsics::_sha5_implCompress:
917 return inline_sha_implCompress(intrinsic_id());
918
919 case vmIntrinsics::_digestBase_implCompressMB:
920 return inline_digestBase_implCompressMB(predicate);
921
922 case vmIntrinsics::_encodeISOArray:
923 return inline_encodeISOArray();
924
925 case vmIntrinsics::_updateCRC32:
926 return inline_updateCRC32();
927 case vmIntrinsics::_updateBytesCRC32:
928 return inline_updateBytesCRC32();
929 case vmIntrinsics::_updateByteBufferCRC32:
930 return inline_updateByteBufferCRC32();
931
932 default:
933 // If you get here, it may be that someone has added a new intrinsic
934 // to the list in vmSymbols.hpp without implementing it here.
935 #ifndef PRODUCT
936 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
937 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
938 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
939 }
940 #endif
941 return false;
942 }
943 }
944
945 Node* LibraryCallKit::try_to_predicate(int predicate) {
946 if (!jvms()->has_method()) {
947 // Root JVMState has a null method.
948 assert(map()->memory()->Opcode() == Op_Parm, "");
949 // Insert the memory aliasing node
950 set_all_memory(reset_memory());
951 }
952 assert(merged_memory(), "");
953
954 switch (intrinsic_id()) {
955 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
956 return inline_cipherBlockChaining_AESCrypt_predicate(false);
957 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
958 return inline_cipherBlockChaining_AESCrypt_predicate(true);
959 case vmIntrinsics::_digestBase_implCompressMB:
960 return inline_digestBase_implCompressMB_predicate(predicate);
961
962 default:
963 // If you get here, it may be that someone has added a new intrinsic
964 // to the list in vmSymbols.hpp without implementing it here.
965 #ifndef PRODUCT
966 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
967 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
968 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
969 }
970 #endif
971 Node* slow_ctl = control();
972 set_control(top()); // No fast path instrinsic
973 return slow_ctl;
974 }
975 }
976
977 //------------------------------set_result-------------------------------
978 // Helper function for finishing intrinsics.
979 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
980 record_for_igvn(region);
981 set_control(_gvn.transform(region));
982 set_result( _gvn.transform(value));
983 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
984 }
985
986 //------------------------------generate_guard---------------------------
987 // Helper function for generating guarded fast-slow graph structures.
988 // The given 'test', if true, guards a slow path. If the test fails
989 // then a fast path can be taken. (We generally hope it fails.)
990 // In all cases, GraphKit::control() is updated to the fast path.
991 // The returned value represents the control for the slow path.
992 // The return value is never 'top'; it is either a valid control
993 // or NULL if it is obvious that the slow path can never be taken.
994 // Also, if region and the slow control are not NULL, the slow edge
995 // is appended to the region.
996 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
997 if (stopped()) {
998 // Already short circuited.
999 return NULL;
1000 }
1001
1002 // Build an if node and its projections.
1003 // If test is true we take the slow path, which we assume is uncommon.
1004 if (_gvn.type(test) == TypeInt::ZERO) {
1005 // The slow branch is never taken. No need to build this guard.
1006 return NULL;
1007 }
1008
1009 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1010
1011 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1012 if (if_slow == top()) {
1013 // The slow branch is never taken. No need to build this guard.
1014 return NULL;
1015 }
1016
1017 if (region != NULL)
1018 region->add_req(if_slow);
1019
1020 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1021 set_control(if_fast);
1022
1023 return if_slow;
1024 }
1025
1026 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1027 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1028 }
1029 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1030 return generate_guard(test, region, PROB_FAIR);
1031 }
1032
1033 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1034 Node* *pos_index) {
1035 if (stopped())
1036 return NULL; // already stopped
1037 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1038 return NULL; // index is already adequately typed
1039 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1040 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1041 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1042 if (is_neg != NULL && pos_index != NULL) {
1043 // Emulate effect of Parse::adjust_map_after_if.
1044 Node* ccast = new CastIINode(index, TypeInt::POS);
1045 ccast->set_req(0, control());
1046 (*pos_index) = _gvn.transform(ccast);
1047 }
1048 return is_neg;
1049 }
1050
1051 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1052 Node* *pos_index) {
1053 if (stopped())
1054 return NULL; // already stopped
1055 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1056 return NULL; // index is already adequately typed
1057 Node* cmp_le = _gvn.transform(new CmpINode(index, intcon(0)));
1058 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1059 Node* bol_le = _gvn.transform(new BoolNode(cmp_le, le_or_eq));
1060 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1061 if (is_notp != NULL && pos_index != NULL) {
1062 // Emulate effect of Parse::adjust_map_after_if.
1063 Node* ccast = new CastIINode(index, TypeInt::POS1);
1064 ccast->set_req(0, control());
1065 (*pos_index) = _gvn.transform(ccast);
1066 }
1067 return is_notp;
1068 }
1069
1070 // Make sure that 'position' is a valid limit index, in [0..length].
1071 // There are two equivalent plans for checking this:
1072 // A. (offset + copyLength) unsigned<= arrayLength
1073 // B. offset <= (arrayLength - copyLength)
1074 // We require that all of the values above, except for the sum and
1075 // difference, are already known to be non-negative.
1076 // Plan A is robust in the face of overflow, if offset and copyLength
1077 // are both hugely positive.
1078 //
1079 // Plan B is less direct and intuitive, but it does not overflow at
1080 // all, since the difference of two non-negatives is always
1081 // representable. Whenever Java methods must perform the equivalent
1082 // check they generally use Plan B instead of Plan A.
1083 // For the moment we use Plan A.
1084 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1085 Node* subseq_length,
1086 Node* array_length,
1087 RegionNode* region) {
1088 if (stopped())
1089 return NULL; // already stopped
1090 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1091 if (zero_offset && subseq_length->eqv_uncast(array_length))
1092 return NULL; // common case of whole-array copy
1093 Node* last = subseq_length;
1094 if (!zero_offset) // last += offset
1095 last = _gvn.transform(new AddINode(last, offset));
1096 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1097 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1098 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1099 return is_over;
1100 }
1101
1102
1103 //--------------------------generate_current_thread--------------------
1104 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1105 ciKlass* thread_klass = env()->Thread_klass();
1106 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1107 Node* thread = _gvn.transform(new ThreadLocalNode());
1108 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1109 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1110 tls_output = thread;
1111 return threadObj;
1112 }
1113
1114
1115 //------------------------------make_string_method_node------------------------
1116 // Helper method for String intrinsic functions. This version is called
1117 // with str1 and str2 pointing to String object nodes.
1118 //
1119 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1120 Node* no_ctrl = NULL;
1121
1122 // Get start addr of string
1123 Node* str1_value = load_String_value(no_ctrl, str1);
1124 Node* str1_offset = load_String_offset(no_ctrl, str1);
1125 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1126
1127 // Get length of string 1
1128 Node* str1_len = load_String_length(no_ctrl, str1);
1129
1130 Node* str2_value = load_String_value(no_ctrl, str2);
1131 Node* str2_offset = load_String_offset(no_ctrl, str2);
1132 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1133
1134 Node* str2_len = NULL;
1135 Node* result = NULL;
1136
1137 switch (opcode) {
1138 case Op_StrIndexOf:
1139 // Get length of string 2
1140 str2_len = load_String_length(no_ctrl, str2);
1141
1142 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1143 str1_start, str1_len, str2_start, str2_len);
1144 break;
1145 case Op_StrComp:
1146 // Get length of string 2
1147 str2_len = load_String_length(no_ctrl, str2);
1148
1149 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1150 str1_start, str1_len, str2_start, str2_len);
1151 break;
1152 case Op_StrEquals:
1153 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1154 str1_start, str2_start, str1_len);
1155 break;
1156 default:
1157 ShouldNotReachHere();
1158 return NULL;
1159 }
1160
1161 // All these intrinsics have checks.
1162 C->set_has_split_ifs(true); // Has chance for split-if optimization
1163
1164 return _gvn.transform(result);
1165 }
1166
1167 // Helper method for String intrinsic functions. This version is called
1168 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1169 // to Int nodes containing the lenghts of str1 and str2.
1170 //
1171 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1172 Node* result = NULL;
1173 switch (opcode) {
1174 case Op_StrIndexOf:
1175 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1176 str1_start, cnt1, str2_start, cnt2);
1177 break;
1178 case Op_StrComp:
1179 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1180 str1_start, cnt1, str2_start, cnt2);
1181 break;
1182 case Op_StrEquals:
1183 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1184 str1_start, str2_start, cnt1);
1185 break;
1186 default:
1187 ShouldNotReachHere();
1188 return NULL;
1189 }
1190
1191 // All these intrinsics have checks.
1192 C->set_has_split_ifs(true); // Has chance for split-if optimization
1193
1194 return _gvn.transform(result);
1195 }
1196
1197 //------------------------------inline_string_compareTo------------------------
1198 // public int java.lang.String.compareTo(String anotherString);
1199 bool LibraryCallKit::inline_string_compareTo() {
1200 Node* receiver = null_check(argument(0));
1201 Node* arg = null_check(argument(1));
1202 if (stopped()) {
1203 return true;
1204 }
1205 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1206 return true;
1207 }
1208
1209 //------------------------------inline_string_equals------------------------
1210 bool LibraryCallKit::inline_string_equals() {
1211 Node* receiver = null_check_receiver();
1212 // NOTE: Do not null check argument for String.equals() because spec
1213 // allows to specify NULL as argument.
1214 Node* argument = this->argument(1);
1215 if (stopped()) {
1216 return true;
1217 }
1218
1219 // paths (plus control) merge
1220 RegionNode* region = new RegionNode(5);
1221 Node* phi = new PhiNode(region, TypeInt::BOOL);
1222
1223 // does source == target string?
1224 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1225 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1226
1227 Node* if_eq = generate_slow_guard(bol, NULL);
1228 if (if_eq != NULL) {
1229 // receiver == argument
1230 phi->init_req(2, intcon(1));
1231 region->init_req(2, if_eq);
1232 }
1233
1234 // get String klass for instanceOf
1235 ciInstanceKlass* klass = env()->String_klass();
1236
1237 if (!stopped()) {
1238 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1239 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1240 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1241
1242 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1243 //instanceOf == true, fallthrough
1244
1245 if (inst_false != NULL) {
1246 phi->init_req(3, intcon(0));
1247 region->init_req(3, inst_false);
1248 }
1249 }
1250
1251 if (!stopped()) {
1252 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1253
1254 // Properly cast the argument to String
1255 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1256 // This path is taken only when argument's type is String:NotNull.
1257 argument = cast_not_null(argument, false);
1258
1259 Node* no_ctrl = NULL;
1260
1261 // Get start addr of receiver
1262 Node* receiver_val = load_String_value(no_ctrl, receiver);
1263 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1264 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1265
1266 // Get length of receiver
1267 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1268
1269 // Get start addr of argument
1270 Node* argument_val = load_String_value(no_ctrl, argument);
1271 Node* argument_offset = load_String_offset(no_ctrl, argument);
1272 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1273
1274 // Get length of argument
1275 Node* argument_cnt = load_String_length(no_ctrl, argument);
1276
1277 // Check for receiver count != argument count
1278 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1279 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1280 Node* if_ne = generate_slow_guard(bol, NULL);
1281 if (if_ne != NULL) {
1282 phi->init_req(4, intcon(0));
1283 region->init_req(4, if_ne);
1284 }
1285
1286 // Check for count == 0 is done by assembler code for StrEquals.
1287
1288 if (!stopped()) {
1289 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1290 phi->init_req(1, equals);
1291 region->init_req(1, control());
1292 }
1293 }
1294
1295 // post merge
1296 set_control(_gvn.transform(region));
1297 record_for_igvn(region);
1298
1299 set_result(_gvn.transform(phi));
1300 return true;
1301 }
1302
1303 //------------------------------inline_array_equals----------------------------
1304 bool LibraryCallKit::inline_array_equals() {
1305 Node* arg1 = argument(0);
1306 Node* arg2 = argument(1);
1307 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1308 return true;
1309 }
1310
1311 // Java version of String.indexOf(constant string)
1312 // class StringDecl {
1313 // StringDecl(char[] ca) {
1314 // offset = 0;
1315 // count = ca.length;
1316 // value = ca;
1317 // }
1318 // int offset;
1319 // int count;
1320 // char[] value;
1321 // }
1322 //
1323 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1324 // int targetOffset, int cache_i, int md2) {
1325 // int cache = cache_i;
1326 // int sourceOffset = string_object.offset;
1327 // int sourceCount = string_object.count;
1328 // int targetCount = target_object.length;
1329 //
1330 // int targetCountLess1 = targetCount - 1;
1331 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1332 //
1333 // char[] source = string_object.value;
1334 // char[] target = target_object;
1335 // int lastChar = target[targetCountLess1];
1336 //
1337 // outer_loop:
1338 // for (int i = sourceOffset; i < sourceEnd; ) {
1339 // int src = source[i + targetCountLess1];
1340 // if (src == lastChar) {
1341 // // With random strings and a 4-character alphabet,
1342 // // reverse matching at this point sets up 0.8% fewer
1343 // // frames, but (paradoxically) makes 0.3% more probes.
1344 // // Since those probes are nearer the lastChar probe,
1345 // // there is may be a net D$ win with reverse matching.
1346 // // But, reversing loop inhibits unroll of inner loop
1347 // // for unknown reason. So, does running outer loop from
1348 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1349 // for (int j = 0; j < targetCountLess1; j++) {
1350 // if (target[targetOffset + j] != source[i+j]) {
1351 // if ((cache & (1 << source[i+j])) == 0) {
1352 // if (md2 < j+1) {
1353 // i += j+1;
1354 // continue outer_loop;
1355 // }
1356 // }
1357 // i += md2;
1358 // continue outer_loop;
1359 // }
1360 // }
1361 // return i - sourceOffset;
1362 // }
1363 // if ((cache & (1 << src)) == 0) {
1364 // i += targetCountLess1;
1365 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1366 // i++;
1367 // }
1368 // return -1;
1369 // }
1370
1371 //------------------------------string_indexOf------------------------
1372 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1373 jint cache_i, jint md2_i) {
1374
1375 Node* no_ctrl = NULL;
1376 float likely = PROB_LIKELY(0.9);
1377 float unlikely = PROB_UNLIKELY(0.9);
1378
1379 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1380
1381 Node* source = load_String_value(no_ctrl, string_object);
1382 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1383 Node* sourceCount = load_String_length(no_ctrl, string_object);
1384
1385 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1386 jint target_length = target_array->length();
1387 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1388 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1389
1390 // String.value field is known to be @Stable.
1391 if (UseImplicitStableValues) {
1392 target = cast_array_to_stable(target, target_type);
1393 }
1394
1395 IdealKit kit(this, false, true);
1396 #define __ kit.
1397 Node* zero = __ ConI(0);
1398 Node* one = __ ConI(1);
1399 Node* cache = __ ConI(cache_i);
1400 Node* md2 = __ ConI(md2_i);
1401 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1402 Node* targetCount = __ ConI(target_length);
1403 Node* targetCountLess1 = __ ConI(target_length - 1);
1404 Node* targetOffset = __ ConI(targetOffset_i);
1405 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1406
1407 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1408 Node* outer_loop = __ make_label(2 /* goto */);
1409 Node* return_ = __ make_label(1);
1410
1411 __ set(rtn,__ ConI(-1));
1412 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1413 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1414 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1415 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1416 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1417 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1418 Node* tpj = __ AddI(targetOffset, __ value(j));
1419 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1420 Node* ipj = __ AddI(__ value(i), __ value(j));
1421 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1422 __ if_then(targ, BoolTest::ne, src2); {
1423 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1424 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1425 __ increment(i, __ AddI(__ value(j), one));
1426 __ goto_(outer_loop);
1427 } __ end_if(); __ dead(j);
1428 }__ end_if(); __ dead(j);
1429 __ increment(i, md2);
1430 __ goto_(outer_loop);
1431 }__ end_if();
1432 __ increment(j, one);
1433 }__ end_loop(); __ dead(j);
1434 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1435 __ goto_(return_);
1436 }__ end_if();
1437 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1438 __ increment(i, targetCountLess1);
1439 }__ end_if();
1440 __ increment(i, one);
1441 __ bind(outer_loop);
1442 }__ end_loop(); __ dead(i);
1443 __ bind(return_);
1444
1445 // Final sync IdealKit and GraphKit.
1446 final_sync(kit);
1447 Node* result = __ value(rtn);
1448 #undef __
1449 C->set_has_loops(true);
1450 return result;
1451 }
1452
1453 //------------------------------inline_string_indexOf------------------------
1454 bool LibraryCallKit::inline_string_indexOf() {
1455 Node* receiver = argument(0);
1456 Node* arg = argument(1);
1457
1458 Node* result;
1459 // Disable the use of pcmpestri until it can be guaranteed that
1460 // the load doesn't cross into the uncommited space.
1461 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1462 UseSSE42Intrinsics) {
1463 // Generate SSE4.2 version of indexOf
1464 // We currently only have match rules that use SSE4.2
1465
1466 receiver = null_check(receiver);
1467 arg = null_check(arg);
1468 if (stopped()) {
1469 return true;
1470 }
1471
1472 ciInstanceKlass* str_klass = env()->String_klass();
1473 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1474
1475 // Make the merge point
1476 RegionNode* result_rgn = new RegionNode(4);
1477 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1478 Node* no_ctrl = NULL;
1479
1480 // Get start addr of source string
1481 Node* source = load_String_value(no_ctrl, receiver);
1482 Node* source_offset = load_String_offset(no_ctrl, receiver);
1483 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1484
1485 // Get length of source string
1486 Node* source_cnt = load_String_length(no_ctrl, receiver);
1487
1488 // Get start addr of substring
1489 Node* substr = load_String_value(no_ctrl, arg);
1490 Node* substr_offset = load_String_offset(no_ctrl, arg);
1491 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1492
1493 // Get length of source string
1494 Node* substr_cnt = load_String_length(no_ctrl, arg);
1495
1496 // Check for substr count > string count
1497 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1498 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1499 Node* if_gt = generate_slow_guard(bol, NULL);
1500 if (if_gt != NULL) {
1501 result_phi->init_req(2, intcon(-1));
1502 result_rgn->init_req(2, if_gt);
1503 }
1504
1505 if (!stopped()) {
1506 // Check for substr count == 0
1507 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1508 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1509 Node* if_zero = generate_slow_guard(bol, NULL);
1510 if (if_zero != NULL) {
1511 result_phi->init_req(3, intcon(0));
1512 result_rgn->init_req(3, if_zero);
1513 }
1514 }
1515
1516 if (!stopped()) {
1517 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1518 result_phi->init_req(1, result);
1519 result_rgn->init_req(1, control());
1520 }
1521 set_control(_gvn.transform(result_rgn));
1522 record_for_igvn(result_rgn);
1523 result = _gvn.transform(result_phi);
1524
1525 } else { // Use LibraryCallKit::string_indexOf
1526 // don't intrinsify if argument isn't a constant string.
1527 if (!arg->is_Con()) {
1528 return false;
1529 }
1530 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1531 if (str_type == NULL) {
1532 return false;
1533 }
1534 ciInstanceKlass* klass = env()->String_klass();
1535 ciObject* str_const = str_type->const_oop();
1536 if (str_const == NULL || str_const->klass() != klass) {
1537 return false;
1538 }
1539 ciInstance* str = str_const->as_instance();
1540 assert(str != NULL, "must be instance");
1541
1542 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1543 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1544
1545 int o;
1546 int c;
1547 if (java_lang_String::has_offset_field()) {
1548 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1549 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1550 } else {
1551 o = 0;
1552 c = pat->length();
1553 }
1554
1555 // constant strings have no offset and count == length which
1556 // simplifies the resulting code somewhat so lets optimize for that.
1557 if (o != 0 || c != pat->length()) {
1558 return false;
1559 }
1560
1561 receiver = null_check(receiver, T_OBJECT);
1562 // NOTE: No null check on the argument is needed since it's a constant String oop.
1563 if (stopped()) {
1564 return true;
1565 }
1566
1567 // The null string as a pattern always returns 0 (match at beginning of string)
1568 if (c == 0) {
1569 set_result(intcon(0));
1570 return true;
1571 }
1572
1573 // Generate default indexOf
1574 jchar lastChar = pat->char_at(o + (c - 1));
1575 int cache = 0;
1576 int i;
1577 for (i = 0; i < c - 1; i++) {
1578 assert(i < pat->length(), "out of range");
1579 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1580 }
1581
1582 int md2 = c;
1583 for (i = 0; i < c - 1; i++) {
1584 assert(i < pat->length(), "out of range");
1585 if (pat->char_at(o + i) == lastChar) {
1586 md2 = (c - 1) - i;
1587 }
1588 }
1589
1590 result = string_indexOf(receiver, pat, o, cache, md2);
1591 }
1592 set_result(result);
1593 return true;
1594 }
1595
1596 //--------------------------round_double_node--------------------------------
1597 // Round a double node if necessary.
1598 Node* LibraryCallKit::round_double_node(Node* n) {
1599 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1600 n = _gvn.transform(new RoundDoubleNode(0, n));
1601 return n;
1602 }
1603
1604 //------------------------------inline_math-----------------------------------
1605 // public static double Math.abs(double)
1606 // public static double Math.sqrt(double)
1607 // public static double Math.log(double)
1608 // public static double Math.log10(double)
1609 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1610 Node* arg = round_double_node(argument(0));
1611 Node* n;
1612 switch (id) {
1613 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1614 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1615 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1616 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1617 default: fatal_unexpected_iid(id); break;
1618 }
1619 set_result(_gvn.transform(n));
1620 return true;
1621 }
1622
1623 //------------------------------inline_trig----------------------------------
1624 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1625 // argument reduction which will turn into a fast/slow diamond.
1626 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1627 Node* arg = round_double_node(argument(0));
1628 Node* n = NULL;
1629
1630 switch (id) {
1631 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1632 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1633 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1634 default: fatal_unexpected_iid(id); break;
1635 }
1636 n = _gvn.transform(n);
1637
1638 // Rounding required? Check for argument reduction!
1639 if (Matcher::strict_fp_requires_explicit_rounding) {
1640 static const double pi_4 = 0.7853981633974483;
1641 static const double neg_pi_4 = -0.7853981633974483;
1642 // pi/2 in 80-bit extended precision
1643 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1644 // -pi/2 in 80-bit extended precision
1645 // 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};
1646 // Cutoff value for using this argument reduction technique
1647 //static const double pi_2_minus_epsilon = 1.564660403643354;
1648 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1649
1650 // Pseudocode for sin:
1651 // if (x <= Math.PI / 4.0) {
1652 // if (x >= -Math.PI / 4.0) return fsin(x);
1653 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1654 // } else {
1655 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1656 // }
1657 // return StrictMath.sin(x);
1658
1659 // Pseudocode for cos:
1660 // if (x <= Math.PI / 4.0) {
1661 // if (x >= -Math.PI / 4.0) return fcos(x);
1662 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1663 // } else {
1664 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1665 // }
1666 // return StrictMath.cos(x);
1667
1668 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1669 // requires a special machine instruction to load it. Instead we'll try
1670 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1671 // probably do the math inside the SIN encoding.
1672
1673 // Make the merge point
1674 RegionNode* r = new RegionNode(3);
1675 Node* phi = new PhiNode(r, Type::DOUBLE);
1676
1677 // Flatten arg so we need only 1 test
1678 Node *abs = _gvn.transform(new AbsDNode(arg));
1679 // Node for PI/4 constant
1680 Node *pi4 = makecon(TypeD::make(pi_4));
1681 // Check PI/4 : abs(arg)
1682 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1683 // Check: If PI/4 < abs(arg) then go slow
1684 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1685 // Branch either way
1686 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1687 set_control(opt_iff(r,iff));
1688
1689 // Set fast path result
1690 phi->init_req(2, n);
1691
1692 // Slow path - non-blocking leaf call
1693 Node* call = NULL;
1694 switch (id) {
1695 case vmIntrinsics::_dsin:
1696 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1697 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1698 "Sin", NULL, arg, top());
1699 break;
1700 case vmIntrinsics::_dcos:
1701 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1702 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1703 "Cos", NULL, arg, top());
1704 break;
1705 case vmIntrinsics::_dtan:
1706 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1707 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1708 "Tan", NULL, arg, top());
1709 break;
1710 }
1711 assert(control()->in(0) == call, "");
1712 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1713 r->init_req(1, control());
1714 phi->init_req(1, slow_result);
1715
1716 // Post-merge
1717 set_control(_gvn.transform(r));
1718 record_for_igvn(r);
1719 n = _gvn.transform(phi);
1720
1721 C->set_has_split_ifs(true); // Has chance for split-if optimization
1722 }
1723 set_result(n);
1724 return true;
1725 }
1726
1727 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1728 //-------------------
1729 //result=(result.isNaN())? funcAddr():result;
1730 // Check: If isNaN() by checking result!=result? then either trap
1731 // or go to runtime
1732 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1733 // Build the boolean node
1734 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1735
1736 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1737 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1738 // The pow or exp intrinsic returned a NaN, which requires a call
1739 // to the runtime. Recompile with the runtime call.
1740 uncommon_trap(Deoptimization::Reason_intrinsic,
1741 Deoptimization::Action_make_not_entrant);
1742 }
1743 return result;
1744 } else {
1745 // If this inlining ever returned NaN in the past, we compile a call
1746 // to the runtime to properly handle corner cases
1747
1748 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1749 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1750 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1751
1752 if (!if_slow->is_top()) {
1753 RegionNode* result_region = new RegionNode(3);
1754 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1755
1756 result_region->init_req(1, if_fast);
1757 result_val->init_req(1, result);
1758
1759 set_control(if_slow);
1760
1761 const TypePtr* no_memory_effects = NULL;
1762 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1763 no_memory_effects,
1764 x, top(), y, y ? top() : NULL);
1765 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1766 #ifdef ASSERT
1767 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1768 assert(value_top == top(), "second value must be top");
1769 #endif
1770
1771 result_region->init_req(2, control());
1772 result_val->init_req(2, value);
1773 set_control(_gvn.transform(result_region));
1774 return _gvn.transform(result_val);
1775 } else {
1776 return result;
1777 }
1778 }
1779 }
1780
1781 //------------------------------inline_exp-------------------------------------
1782 // Inline exp instructions, if possible. The Intel hardware only misses
1783 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1784 bool LibraryCallKit::inline_exp() {
1785 Node* arg = round_double_node(argument(0));
1786 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1787
1788 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1789 set_result(n);
1790
1791 C->set_has_split_ifs(true); // Has chance for split-if optimization
1792 return true;
1793 }
1794
1795 //------------------------------inline_pow-------------------------------------
1796 // Inline power instructions, if possible.
1797 bool LibraryCallKit::inline_pow() {
1798 // Pseudocode for pow
1799 // if (y == 2) {
1800 // return x * x;
1801 // } else {
1802 // if (x <= 0.0) {
1803 // long longy = (long)y;
1804 // if ((double)longy == y) { // if y is long
1805 // if (y + 1 == y) longy = 0; // huge number: even
1806 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1807 // } else {
1808 // result = NaN;
1809 // }
1810 // } else {
1811 // result = DPow(x,y);
1812 // }
1813 // if (result != result)? {
1814 // result = uncommon_trap() or runtime_call();
1815 // }
1816 // return result;
1817 // }
1818
1819 Node* x = round_double_node(argument(0));
1820 Node* y = round_double_node(argument(2));
1821
1822 Node* result = NULL;
1823
1824 Node* const_two_node = makecon(TypeD::make(2.0));
1825 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1826 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1827 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1828 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1829 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1830
1831 RegionNode* region_node = new RegionNode(3);
1832 region_node->init_req(1, if_true);
1833
1834 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1835 // special case for x^y where y == 2, we can convert it to x * x
1836 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1837
1838 // set control to if_false since we will now process the false branch
1839 set_control(if_false);
1840
1841 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1842 // Short form: skip the fancy tests and just check for NaN result.
1843 result = _gvn.transform(new PowDNode(C, control(), x, y));
1844 } else {
1845 // If this inlining ever returned NaN in the past, include all
1846 // checks + call to the runtime.
1847
1848 // Set the merge point for If node with condition of (x <= 0.0)
1849 // There are four possible paths to region node and phi node
1850 RegionNode *r = new RegionNode(4);
1851 Node *phi = new PhiNode(r, Type::DOUBLE);
1852
1853 // Build the first if node: if (x <= 0.0)
1854 // Node for 0 constant
1855 Node *zeronode = makecon(TypeD::ZERO);
1856 // Check x:0
1857 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1858 // Check: If (x<=0) then go complex path
1859 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1860 // Branch either way
1861 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1862 // Fast path taken; set region slot 3
1863 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1864 r->init_req(3,fast_taken); // Capture fast-control
1865
1866 // Fast path not-taken, i.e. slow path
1867 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1868
1869 // Set fast path result
1870 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1871 phi->init_req(3, fast_result);
1872
1873 // Complex path
1874 // Build the second if node (if y is long)
1875 // Node for (long)y
1876 Node *longy = _gvn.transform(new ConvD2LNode(y));
1877 // Node for (double)((long) y)
1878 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1879 // Check (double)((long) y) : y
1880 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1881 // Check if (y isn't long) then go to slow path
1882
1883 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1884 // Branch either way
1885 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1886 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1887
1888 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1889
1890 // Calculate DPow(abs(x), y)*(1 & (long)y)
1891 // Node for constant 1
1892 Node *conone = longcon(1);
1893 // 1& (long)y
1894 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1895
1896 // A huge number is always even. Detect a huge number by checking
1897 // if y + 1 == y and set integer to be tested for parity to 0.
1898 // Required for corner case:
1899 // (long)9.223372036854776E18 = max_jlong
1900 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1901 // max_jlong is odd but 9.223372036854776E18 is even
1902 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1903 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1904 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1905 Node* correctedsign = NULL;
1906 if (ConditionalMoveLimit != 0) {
1907 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1908 } else {
1909 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1910 RegionNode *r = new RegionNode(3);
1911 Node *phi = new PhiNode(r, TypeLong::LONG);
1912 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1913 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1914 phi->init_req(1, signnode);
1915 phi->init_req(2, longcon(0));
1916 correctedsign = _gvn.transform(phi);
1917 ylong_path = _gvn.transform(r);
1918 record_for_igvn(r);
1919 }
1920
1921 // zero node
1922 Node *conzero = longcon(0);
1923 // Check (1&(long)y)==0?
1924 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1925 // Check if (1&(long)y)!=0?, if so the result is negative
1926 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1927 // abs(x)
1928 Node *absx=_gvn.transform(new AbsDNode(x));
1929 // abs(x)^y
1930 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1931 // -abs(x)^y
1932 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1933 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1934 Node *signresult = NULL;
1935 if (ConditionalMoveLimit != 0) {
1936 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1937 } else {
1938 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1939 RegionNode *r = new RegionNode(3);
1940 Node *phi = new PhiNode(r, Type::DOUBLE);
1941 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1942 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1943 phi->init_req(1, absxpowy);
1944 phi->init_req(2, negabsxpowy);
1945 signresult = _gvn.transform(phi);
1946 ylong_path = _gvn.transform(r);
1947 record_for_igvn(r);
1948 }
1949 // Set complex path fast result
1950 r->init_req(2, ylong_path);
1951 phi->init_req(2, signresult);
1952
1953 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1954 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1955 r->init_req(1,slow_path);
1956 phi->init_req(1,slow_result);
1957
1958 // Post merge
1959 set_control(_gvn.transform(r));
1960 record_for_igvn(r);
1961 result = _gvn.transform(phi);
1962 }
1963
1964 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1965
1966 // control from finish_pow_exp is now input to the region node
1967 region_node->set_req(2, control());
1968 // the result from finish_pow_exp is now input to the phi node
1969 phi_node->init_req(2, result);
1970 set_control(_gvn.transform(region_node));
1971 record_for_igvn(region_node);
1972 set_result(_gvn.transform(phi_node));
1973
1974 C->set_has_split_ifs(true); // Has chance for split-if optimization
1975 return true;
1976 }
1977
1978 //------------------------------runtime_math-----------------------------
1979 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1980 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1981 "must be (DD)D or (D)D type");
1982
1983 // Inputs
1984 Node* a = round_double_node(argument(0));
1985 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1986
1987 const TypePtr* no_memory_effects = NULL;
1988 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1989 no_memory_effects,
1990 a, top(), b, b ? top() : NULL);
1991 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1992 #ifdef ASSERT
1993 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1994 assert(value_top == top(), "second value must be top");
1995 #endif
1996
1997 set_result(value);
1998 return true;
1999 }
2000
2001 //------------------------------inline_math_native-----------------------------
2002 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2003 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2004 switch (id) {
2005 // These intrinsics are not properly supported on all hardware
2006 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2007 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2008 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2009 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2010 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2011 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2012
2013 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2014 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2015 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2016 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2017
2018 // These intrinsics are supported on all hardware
2019 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2020 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2021
2022 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2023 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2024 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2025 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2026 #undef FN_PTR
2027
2028 // These intrinsics are not yet correctly implemented
2029 case vmIntrinsics::_datan2:
2030 return false;
2031
2032 default:
2033 fatal_unexpected_iid(id);
2034 return false;
2035 }
2036 }
2037
2038 static bool is_simple_name(Node* n) {
2039 return (n->req() == 1 // constant
2040 || (n->is_Type() && n->as_Type()->type()->singleton())
2041 || n->is_Proj() // parameter or return value
2042 || n->is_Phi() // local of some sort
2043 );
2044 }
2045
2046 //----------------------------inline_min_max-----------------------------------
2047 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2048 set_result(generate_min_max(id, argument(0), argument(1)));
2049 return true;
2050 }
2051
2052 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2053 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2054 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2055 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2056 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2057
2058 {
2059 PreserveJVMState pjvms(this);
2060 PreserveReexecuteState preexecs(this);
2061 jvms()->set_should_reexecute(true);
2062
2063 set_control(slow_path);
2064 set_i_o(i_o());
2065
2066 uncommon_trap(Deoptimization::Reason_intrinsic,
2067 Deoptimization::Action_none);
2068 }
2069
2070 set_control(fast_path);
2071 set_result(math);
2072 }
2073
2074 template <typename OverflowOp>
2075 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2076 typedef typename OverflowOp::MathOp MathOp;
2077
2078 MathOp* mathOp = new MathOp(arg1, arg2);
2079 Node* operation = _gvn.transform( mathOp );
2080 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2081 inline_math_mathExact(operation, ofcheck);
2082 return true;
2083 }
2084
2085 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2086 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2087 }
2088
2089 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2090 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2091 }
2092
2093 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2094 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2095 }
2096
2097 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2098 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2099 }
2100
2101 bool LibraryCallKit::inline_math_negateExactI() {
2102 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2103 }
2104
2105 bool LibraryCallKit::inline_math_negateExactL() {
2106 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2107 }
2108
2109 bool LibraryCallKit::inline_math_multiplyExactI() {
2110 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2111 }
2112
2113 bool LibraryCallKit::inline_math_multiplyExactL() {
2114 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2115 }
2116
2117 Node*
2118 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2119 // These are the candidate return value:
2120 Node* xvalue = x0;
2121 Node* yvalue = y0;
2122
2123 if (xvalue == yvalue) {
2124 return xvalue;
2125 }
2126
2127 bool want_max = (id == vmIntrinsics::_max);
2128
2129 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2130 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2131 if (txvalue == NULL || tyvalue == NULL) return top();
2132 // This is not really necessary, but it is consistent with a
2133 // hypothetical MaxINode::Value method:
2134 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2135
2136 // %%% This folding logic should (ideally) be in a different place.
2137 // Some should be inside IfNode, and there to be a more reliable
2138 // transformation of ?: style patterns into cmoves. We also want
2139 // more powerful optimizations around cmove and min/max.
2140
2141 // Try to find a dominating comparison of these guys.
2142 // It can simplify the index computation for Arrays.copyOf
2143 // and similar uses of System.arraycopy.
2144 // First, compute the normalized version of CmpI(x, y).
2145 int cmp_op = Op_CmpI;
2146 Node* xkey = xvalue;
2147 Node* ykey = yvalue;
2148 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2149 if (ideal_cmpxy->is_Cmp()) {
2150 // E.g., if we have CmpI(length - offset, count),
2151 // it might idealize to CmpI(length, count + offset)
2152 cmp_op = ideal_cmpxy->Opcode();
2153 xkey = ideal_cmpxy->in(1);
2154 ykey = ideal_cmpxy->in(2);
2155 }
2156
2157 // Start by locating any relevant comparisons.
2158 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2159 Node* cmpxy = NULL;
2160 Node* cmpyx = NULL;
2161 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2162 Node* cmp = start_from->fast_out(k);
2163 if (cmp->outcnt() > 0 && // must have prior uses
2164 cmp->in(0) == NULL && // must be context-independent
2165 cmp->Opcode() == cmp_op) { // right kind of compare
2166 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2167 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2168 }
2169 }
2170
2171 const int NCMPS = 2;
2172 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2173 int cmpn;
2174 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2175 if (cmps[cmpn] != NULL) break; // find a result
2176 }
2177 if (cmpn < NCMPS) {
2178 // Look for a dominating test that tells us the min and max.
2179 int depth = 0; // Limit search depth for speed
2180 Node* dom = control();
2181 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2182 if (++depth >= 100) break;
2183 Node* ifproj = dom;
2184 if (!ifproj->is_Proj()) continue;
2185 Node* iff = ifproj->in(0);
2186 if (!iff->is_If()) continue;
2187 Node* bol = iff->in(1);
2188 if (!bol->is_Bool()) continue;
2189 Node* cmp = bol->in(1);
2190 if (cmp == NULL) continue;
2191 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2192 if (cmps[cmpn] == cmp) break;
2193 if (cmpn == NCMPS) continue;
2194 BoolTest::mask btest = bol->as_Bool()->_test._test;
2195 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2196 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2197 // At this point, we know that 'x btest y' is true.
2198 switch (btest) {
2199 case BoolTest::eq:
2200 // They are proven equal, so we can collapse the min/max.
2201 // Either value is the answer. Choose the simpler.
2202 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2203 return yvalue;
2204 return xvalue;
2205 case BoolTest::lt: // x < y
2206 case BoolTest::le: // x <= y
2207 return (want_max ? yvalue : xvalue);
2208 case BoolTest::gt: // x > y
2209 case BoolTest::ge: // x >= y
2210 return (want_max ? xvalue : yvalue);
2211 }
2212 }
2213 }
2214
2215 // We failed to find a dominating test.
2216 // Let's pick a test that might GVN with prior tests.
2217 Node* best_bol = NULL;
2218 BoolTest::mask best_btest = BoolTest::illegal;
2219 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2220 Node* cmp = cmps[cmpn];
2221 if (cmp == NULL) continue;
2222 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2223 Node* bol = cmp->fast_out(j);
2224 if (!bol->is_Bool()) continue;
2225 BoolTest::mask btest = bol->as_Bool()->_test._test;
2226 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2227 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2228 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2229 best_bol = bol->as_Bool();
2230 best_btest = btest;
2231 }
2232 }
2233 }
2234
2235 Node* answer_if_true = NULL;
2236 Node* answer_if_false = NULL;
2237 switch (best_btest) {
2238 default:
2239 if (cmpxy == NULL)
2240 cmpxy = ideal_cmpxy;
2241 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2242 // and fall through:
2243 case BoolTest::lt: // x < y
2244 case BoolTest::le: // x <= y
2245 answer_if_true = (want_max ? yvalue : xvalue);
2246 answer_if_false = (want_max ? xvalue : yvalue);
2247 break;
2248 case BoolTest::gt: // x > y
2249 case BoolTest::ge: // x >= y
2250 answer_if_true = (want_max ? xvalue : yvalue);
2251 answer_if_false = (want_max ? yvalue : xvalue);
2252 break;
2253 }
2254
2255 jint hi, lo;
2256 if (want_max) {
2257 // We can sharpen the minimum.
2258 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2259 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2260 } else {
2261 // We can sharpen the maximum.
2262 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2263 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2264 }
2265
2266 // Use a flow-free graph structure, to avoid creating excess control edges
2267 // which could hinder other optimizations.
2268 // Since Math.min/max is often used with arraycopy, we want
2269 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2270 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2271 answer_if_false, answer_if_true,
2272 TypeInt::make(lo, hi, widen));
2273
2274 return _gvn.transform(cmov);
2275
2276 /*
2277 // This is not as desirable as it may seem, since Min and Max
2278 // nodes do not have a full set of optimizations.
2279 // And they would interfere, anyway, with 'if' optimizations
2280 // and with CMoveI canonical forms.
2281 switch (id) {
2282 case vmIntrinsics::_min:
2283 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2284 case vmIntrinsics::_max:
2285 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2286 default:
2287 ShouldNotReachHere();
2288 }
2289 */
2290 }
2291
2292 inline int
2293 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2294 const TypePtr* base_type = TypePtr::NULL_PTR;
2295 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2296 if (base_type == NULL) {
2297 // Unknown type.
2298 return Type::AnyPtr;
2299 } else if (base_type == TypePtr::NULL_PTR) {
2300 // Since this is a NULL+long form, we have to switch to a rawptr.
2301 base = _gvn.transform(new CastX2PNode(offset));
2302 offset = MakeConX(0);
2303 return Type::RawPtr;
2304 } else if (base_type->base() == Type::RawPtr) {
2305 return Type::RawPtr;
2306 } else if (base_type->isa_oopptr()) {
2307 // Base is never null => always a heap address.
2308 if (base_type->ptr() == TypePtr::NotNull) {
2309 return Type::OopPtr;
2310 }
2311 // Offset is small => always a heap address.
2312 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2313 if (offset_type != NULL &&
2314 base_type->offset() == 0 && // (should always be?)
2315 offset_type->_lo >= 0 &&
2316 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2317 return Type::OopPtr;
2318 }
2319 // Otherwise, it might either be oop+off or NULL+addr.
2320 return Type::AnyPtr;
2321 } else {
2322 // No information:
2323 return Type::AnyPtr;
2324 }
2325 }
2326
2327 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2328 int kind = classify_unsafe_addr(base, offset);
2329 if (kind == Type::RawPtr) {
2330 return basic_plus_adr(top(), base, offset);
2331 } else {
2332 return basic_plus_adr(base, offset);
2333 }
2334 }
2335
2336 //--------------------------inline_number_methods-----------------------------
2337 // inline int Integer.numberOfLeadingZeros(int)
2338 // inline int Long.numberOfLeadingZeros(long)
2339 //
2340 // inline int Integer.numberOfTrailingZeros(int)
2341 // inline int Long.numberOfTrailingZeros(long)
2342 //
2343 // inline int Integer.bitCount(int)
2344 // inline int Long.bitCount(long)
2345 //
2346 // inline char Character.reverseBytes(char)
2347 // inline short Short.reverseBytes(short)
2348 // inline int Integer.reverseBytes(int)
2349 // inline long Long.reverseBytes(long)
2350 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2351 Node* arg = argument(0);
2352 Node* n;
2353 switch (id) {
2354 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2355 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2356 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2357 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2358 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2359 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2360 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2361 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2362 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2363 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2364 default: fatal_unexpected_iid(id); break;
2365 }
2366 set_result(_gvn.transform(n));
2367 return true;
2368 }
2369
2370 //----------------------------inline_unsafe_access----------------------------
2371
2372 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2373
2374 // Helper that guards and inserts a pre-barrier.
2375 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2376 Node* pre_val, bool need_mem_bar) {
2377 // We could be accessing the referent field of a reference object. If so, when G1
2378 // is enabled, we need to log the value in the referent field in an SATB buffer.
2379 // This routine performs some compile time filters and generates suitable
2380 // runtime filters that guard the pre-barrier code.
2381 // Also add memory barrier for non volatile load from the referent field
2382 // to prevent commoning of loads across safepoint.
2383 if (!UseG1GC && !need_mem_bar)
2384 return;
2385
2386 // Some compile time checks.
2387
2388 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2389 const TypeX* otype = offset->find_intptr_t_type();
2390 if (otype != NULL && otype->is_con() &&
2391 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2392 // Constant offset but not the reference_offset so just return
2393 return;
2394 }
2395
2396 // We only need to generate the runtime guards for instances.
2397 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2398 if (btype != NULL) {
2399 if (btype->isa_aryptr()) {
2400 // Array type so nothing to do
2401 return;
2402 }
2403
2404 const TypeInstPtr* itype = btype->isa_instptr();
2405 if (itype != NULL) {
2406 // Can the klass of base_oop be statically determined to be
2407 // _not_ a sub-class of Reference and _not_ Object?
2408 ciKlass* klass = itype->klass();
2409 if ( klass->is_loaded() &&
2410 !klass->is_subtype_of(env()->Reference_klass()) &&
2411 !env()->Object_klass()->is_subtype_of(klass)) {
2412 return;
2413 }
2414 }
2415 }
2416
2417 // The compile time filters did not reject base_oop/offset so
2418 // we need to generate the following runtime filters
2419 //
2420 // if (offset == java_lang_ref_Reference::_reference_offset) {
2421 // if (instance_of(base, java.lang.ref.Reference)) {
2422 // pre_barrier(_, pre_val, ...);
2423 // }
2424 // }
2425
2426 float likely = PROB_LIKELY( 0.999);
2427 float unlikely = PROB_UNLIKELY(0.999);
2428
2429 IdealKit ideal(this);
2430 #define __ ideal.
2431
2432 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2433
2434 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2435 // Update graphKit memory and control from IdealKit.
2436 sync_kit(ideal);
2437
2438 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2439 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2440
2441 // Update IdealKit memory and control from graphKit.
2442 __ sync_kit(this);
2443
2444 Node* one = __ ConI(1);
2445 // is_instof == 0 if base_oop == NULL
2446 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2447
2448 // Update graphKit from IdeakKit.
2449 sync_kit(ideal);
2450
2451 // Use the pre-barrier to record the value in the referent field
2452 pre_barrier(false /* do_load */,
2453 __ ctrl(),
2454 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2455 pre_val /* pre_val */,
2456 T_OBJECT);
2457 if (need_mem_bar) {
2458 // Add memory barrier to prevent commoning reads from this field
2459 // across safepoint since GC can change its value.
2460 insert_mem_bar(Op_MemBarCPUOrder);
2461 }
2462 // Update IdealKit from graphKit.
2463 __ sync_kit(this);
2464
2465 } __ end_if(); // _ref_type != ref_none
2466 } __ end_if(); // offset == referent_offset
2467
2468 // Final sync IdealKit and GraphKit.
2469 final_sync(ideal);
2470 #undef __
2471 }
2472
2473
2474 // Interpret Unsafe.fieldOffset cookies correctly:
2475 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2476
2477 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2478 // Attempt to infer a sharper value type from the offset and base type.
2479 ciKlass* sharpened_klass = NULL;
2480
2481 // See if it is an instance field, with an object type.
2482 if (alias_type->field() != NULL) {
2483 assert(!is_native_ptr, "native pointer op cannot use a java address");
2484 if (alias_type->field()->type()->is_klass()) {
2485 sharpened_klass = alias_type->field()->type()->as_klass();
2486 }
2487 }
2488
2489 // See if it is a narrow oop array.
2490 if (adr_type->isa_aryptr()) {
2491 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2492 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2493 if (elem_type != NULL) {
2494 sharpened_klass = elem_type->klass();
2495 }
2496 }
2497 }
2498
2499 // The sharpened class might be unloaded if there is no class loader
2500 // contraint in place.
2501 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2502 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2503
2504 #ifndef PRODUCT
2505 if (C->print_intrinsics() || C->print_inlining()) {
2506 tty->print(" from base type: "); adr_type->dump();
2507 tty->print(" sharpened value: "); tjp->dump();
2508 }
2509 #endif
2510 // Sharpen the value type.
2511 return tjp;
2512 }
2513 return NULL;
2514 }
2515
2516 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2517 if (callee()->is_static()) return false; // caller must have the capability!
2518
2519 #ifndef PRODUCT
2520 {
2521 ResourceMark rm;
2522 // Check the signatures.
2523 ciSignature* sig = callee()->signature();
2524 #ifdef ASSERT
2525 if (!is_store) {
2526 // Object getObject(Object base, int/long offset), etc.
2527 BasicType rtype = sig->return_type()->basic_type();
2528 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2529 rtype = T_ADDRESS; // it is really a C void*
2530 assert(rtype == type, "getter must return the expected value");
2531 if (!is_native_ptr) {
2532 assert(sig->count() == 2, "oop getter has 2 arguments");
2533 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2534 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2535 } else {
2536 assert(sig->count() == 1, "native getter has 1 argument");
2537 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2538 }
2539 } else {
2540 // void putObject(Object base, int/long offset, Object x), etc.
2541 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2542 if (!is_native_ptr) {
2543 assert(sig->count() == 3, "oop putter has 3 arguments");
2544 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2545 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2546 } else {
2547 assert(sig->count() == 2, "native putter has 2 arguments");
2548 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2549 }
2550 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2551 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2552 vtype = T_ADDRESS; // it is really a C void*
2553 assert(vtype == type, "putter must accept the expected value");
2554 }
2555 #endif // ASSERT
2556 }
2557 #endif //PRODUCT
2558
2559 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2560
2561 Node* receiver = argument(0); // type: oop
2562
2563 // Build address expression. See the code in inline_unsafe_prefetch.
2564 Node* adr;
2565 Node* heap_base_oop = top();
2566 Node* offset = top();
2567 Node* val;
2568
2569 if (!is_native_ptr) {
2570 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2571 Node* base = argument(1); // type: oop
2572 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2573 offset = argument(2); // type: long
2574 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2575 // to be plain byte offsets, which are also the same as those accepted
2576 // by oopDesc::field_base.
2577 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2578 "fieldOffset must be byte-scaled");
2579 // 32-bit machines ignore the high half!
2580 offset = ConvL2X(offset);
2581 adr = make_unsafe_address(base, offset);
2582 heap_base_oop = base;
2583 val = is_store ? argument(4) : NULL;
2584 } else {
2585 Node* ptr = argument(1); // type: long
2586 ptr = ConvL2X(ptr); // adjust Java long to machine word
2587 adr = make_unsafe_address(NULL, ptr);
2588 val = is_store ? argument(3) : NULL;
2589 }
2590
2591 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2592
2593 // First guess at the value type.
2594 const Type *value_type = Type::get_const_basic_type(type);
2595
2596 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2597 // there was not enough information to nail it down.
2598 Compile::AliasType* alias_type = C->alias_type(adr_type);
2599 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2600
2601 // We will need memory barriers unless we can determine a unique
2602 // alias category for this reference. (Note: If for some reason
2603 // the barriers get omitted and the unsafe reference begins to "pollute"
2604 // the alias analysis of the rest of the graph, either Compile::can_alias
2605 // or Compile::must_alias will throw a diagnostic assert.)
2606 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2607
2608 // If we are reading the value of the referent field of a Reference
2609 // object (either by using Unsafe directly or through reflection)
2610 // then, if G1 is enabled, we need to record the referent in an
2611 // SATB log buffer using the pre-barrier mechanism.
2612 // Also we need to add memory barrier to prevent commoning reads
2613 // from this field across safepoint since GC can change its value.
2614 bool need_read_barrier = !is_native_ptr && !is_store &&
2615 offset != top() && heap_base_oop != top();
2616
2617 if (!is_store && type == T_OBJECT) {
2618 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2619 if (tjp != NULL) {
2620 value_type = tjp;
2621 }
2622 }
2623
2624 receiver = null_check(receiver);
2625 if (stopped()) {
2626 return true;
2627 }
2628 // Heap pointers get a null-check from the interpreter,
2629 // as a courtesy. However, this is not guaranteed by Unsafe,
2630 // and it is not possible to fully distinguish unintended nulls
2631 // from intended ones in this API.
2632
2633 if (is_volatile) {
2634 // We need to emit leading and trailing CPU membars (see below) in
2635 // addition to memory membars when is_volatile. This is a little
2636 // too strong, but avoids the need to insert per-alias-type
2637 // volatile membars (for stores; compare Parse::do_put_xxx), which
2638 // we cannot do effectively here because we probably only have a
2639 // rough approximation of type.
2640 need_mem_bar = true;
2641 // For Stores, place a memory ordering barrier now.
2642 if (is_store) {
2643 insert_mem_bar(Op_MemBarRelease);
2644 } else {
2645 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2646 insert_mem_bar(Op_MemBarVolatile);
2647 }
2648 }
2649 }
2650
2651 // Memory barrier to prevent normal and 'unsafe' accesses from
2652 // bypassing each other. Happens after null checks, so the
2653 // exception paths do not take memory state from the memory barrier,
2654 // so there's no problems making a strong assert about mixing users
2655 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2656 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2657 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2658
2659 if (!is_store) {
2660 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2661 // load value
2662 switch (type) {
2663 case T_BOOLEAN:
2664 case T_CHAR:
2665 case T_BYTE:
2666 case T_SHORT:
2667 case T_INT:
2668 case T_LONG:
2669 case T_FLOAT:
2670 case T_DOUBLE:
2671 break;
2672 case T_OBJECT:
2673 if (need_read_barrier) {
2674 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2675 }
2676 break;
2677 case T_ADDRESS:
2678 // Cast to an int type.
2679 p = _gvn.transform(new CastP2XNode(NULL, p));
2680 p = ConvX2UL(p);
2681 break;
2682 default:
2683 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2684 break;
2685 }
2686 // The load node has the control of the preceding MemBarCPUOrder. All
2687 // following nodes will have the control of the MemBarCPUOrder inserted at
2688 // the end of this method. So, pushing the load onto the stack at a later
2689 // point is fine.
2690 set_result(p);
2691 } else {
2692 // place effect of store into memory
2693 switch (type) {
2694 case T_DOUBLE:
2695 val = dstore_rounding(val);
2696 break;
2697 case T_ADDRESS:
2698 // Repackage the long as a pointer.
2699 val = ConvL2X(val);
2700 val = _gvn.transform(new CastX2PNode(val));
2701 break;
2702 }
2703
2704 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2705 if (type != T_OBJECT ) {
2706 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2707 } else {
2708 // Possibly an oop being stored to Java heap or native memory
2709 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2710 // oop to Java heap.
2711 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2712 } else {
2713 // We can't tell at compile time if we are storing in the Java heap or outside
2714 // of it. So we need to emit code to conditionally do the proper type of
2715 // store.
2716
2717 IdealKit ideal(this);
2718 #define __ ideal.
2719 // QQQ who knows what probability is here??
2720 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2721 // Sync IdealKit and graphKit.
2722 sync_kit(ideal);
2723 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2724 // Update IdealKit memory.
2725 __ sync_kit(this);
2726 } __ else_(); {
2727 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2728 } __ end_if();
2729 // Final sync IdealKit and GraphKit.
2730 final_sync(ideal);
2731 #undef __
2732 }
2733 }
2734 }
2735
2736 if (is_volatile) {
2737 if (!is_store) {
2738 insert_mem_bar(Op_MemBarAcquire);
2739 } else {
2740 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2741 insert_mem_bar(Op_MemBarVolatile);
2742 }
2743 }
2744 }
2745
2746 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2747
2748 return true;
2749 }
2750
2751 //----------------------------inline_unsafe_prefetch----------------------------
2752
2753 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2754 #ifndef PRODUCT
2755 {
2756 ResourceMark rm;
2757 // Check the signatures.
2758 ciSignature* sig = callee()->signature();
2759 #ifdef ASSERT
2760 // Object getObject(Object base, int/long offset), etc.
2761 BasicType rtype = sig->return_type()->basic_type();
2762 if (!is_native_ptr) {
2763 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2764 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2765 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2766 } else {
2767 assert(sig->count() == 1, "native prefetch has 1 argument");
2768 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2769 }
2770 #endif // ASSERT
2771 }
2772 #endif // !PRODUCT
2773
2774 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2775
2776 const int idx = is_static ? 0 : 1;
2777 if (!is_static) {
2778 null_check_receiver();
2779 if (stopped()) {
2780 return true;
2781 }
2782 }
2783
2784 // Build address expression. See the code in inline_unsafe_access.
2785 Node *adr;
2786 if (!is_native_ptr) {
2787 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2788 Node* base = argument(idx + 0); // type: oop
2789 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2790 Node* offset = argument(idx + 1); // type: long
2791 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2792 // to be plain byte offsets, which are also the same as those accepted
2793 // by oopDesc::field_base.
2794 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2795 "fieldOffset must be byte-scaled");
2796 // 32-bit machines ignore the high half!
2797 offset = ConvL2X(offset);
2798 adr = make_unsafe_address(base, offset);
2799 } else {
2800 Node* ptr = argument(idx + 0); // type: long
2801 ptr = ConvL2X(ptr); // adjust Java long to machine word
2802 adr = make_unsafe_address(NULL, ptr);
2803 }
2804
2805 // Generate the read or write prefetch
2806 Node *prefetch;
2807 if (is_store) {
2808 prefetch = new PrefetchWriteNode(i_o(), adr);
2809 } else {
2810 prefetch = new PrefetchReadNode(i_o(), adr);
2811 }
2812 prefetch->init_req(0, control());
2813 set_i_o(_gvn.transform(prefetch));
2814
2815 return true;
2816 }
2817
2818 //----------------------------inline_unsafe_load_store----------------------------
2819 // This method serves a couple of different customers (depending on LoadStoreKind):
2820 //
2821 // LS_cmpxchg:
2822 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2823 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2824 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2825 //
2826 // LS_xadd:
2827 // public int getAndAddInt( Object o, long offset, int delta)
2828 // public long getAndAddLong(Object o, long offset, long delta)
2829 //
2830 // LS_xchg:
2831 // int getAndSet(Object o, long offset, int newValue)
2832 // long getAndSet(Object o, long offset, long newValue)
2833 // Object getAndSet(Object o, long offset, Object newValue)
2834 //
2835 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2836 // This basic scheme here is the same as inline_unsafe_access, but
2837 // differs in enough details that combining them would make the code
2838 // overly confusing. (This is a true fact! I originally combined
2839 // them, but even I was confused by it!) As much code/comments as
2840 // possible are retained from inline_unsafe_access though to make
2841 // the correspondences clearer. - dl
2842
2843 if (callee()->is_static()) return false; // caller must have the capability!
2844
2845 #ifndef PRODUCT
2846 BasicType rtype;
2847 {
2848 ResourceMark rm;
2849 // Check the signatures.
2850 ciSignature* sig = callee()->signature();
2851 rtype = sig->return_type()->basic_type();
2852 if (kind == LS_xadd || kind == LS_xchg) {
2853 // Check the signatures.
2854 #ifdef ASSERT
2855 assert(rtype == type, "get and set must return the expected type");
2856 assert(sig->count() == 3, "get and set has 3 arguments");
2857 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2858 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2859 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2860 #endif // ASSERT
2861 } else if (kind == LS_cmpxchg) {
2862 // Check the signatures.
2863 #ifdef ASSERT
2864 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2865 assert(sig->count() == 4, "CAS has 4 arguments");
2866 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2867 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2868 #endif // ASSERT
2869 } else {
2870 ShouldNotReachHere();
2871 }
2872 }
2873 #endif //PRODUCT
2874
2875 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2876
2877 // Get arguments:
2878 Node* receiver = NULL;
2879 Node* base = NULL;
2880 Node* offset = NULL;
2881 Node* oldval = NULL;
2882 Node* newval = NULL;
2883 if (kind == LS_cmpxchg) {
2884 const bool two_slot_type = type2size[type] == 2;
2885 receiver = argument(0); // type: oop
2886 base = argument(1); // type: oop
2887 offset = argument(2); // type: long
2888 oldval = argument(4); // type: oop, int, or long
2889 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2890 } else if (kind == LS_xadd || kind == LS_xchg){
2891 receiver = argument(0); // type: oop
2892 base = argument(1); // type: oop
2893 offset = argument(2); // type: long
2894 oldval = NULL;
2895 newval = argument(4); // type: oop, int, or long
2896 }
2897
2898 // Null check receiver.
2899 receiver = null_check(receiver);
2900 if (stopped()) {
2901 return true;
2902 }
2903
2904 // Build field offset expression.
2905 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2906 // to be plain byte offsets, which are also the same as those accepted
2907 // by oopDesc::field_base.
2908 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2909 // 32-bit machines ignore the high half of long offsets
2910 offset = ConvL2X(offset);
2911 Node* adr = make_unsafe_address(base, offset);
2912 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2913
2914 // For CAS, unlike inline_unsafe_access, there seems no point in
2915 // trying to refine types. Just use the coarse types here.
2916 const Type *value_type = Type::get_const_basic_type(type);
2917 Compile::AliasType* alias_type = C->alias_type(adr_type);
2918 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2919
2920 if (kind == LS_xchg && type == T_OBJECT) {
2921 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2922 if (tjp != NULL) {
2923 value_type = tjp;
2924 }
2925 }
2926
2927 int alias_idx = C->get_alias_index(adr_type);
2928
2929 // Memory-model-wise, a LoadStore acts like a little synchronized
2930 // block, so needs barriers on each side. These don't translate
2931 // into actual barriers on most machines, but we still need rest of
2932 // compiler to respect ordering.
2933
2934 insert_mem_bar(Op_MemBarRelease);
2935 insert_mem_bar(Op_MemBarCPUOrder);
2936
2937 // 4984716: MemBars must be inserted before this
2938 // memory node in order to avoid a false
2939 // dependency which will confuse the scheduler.
2940 Node *mem = memory(alias_idx);
2941
2942 // For now, we handle only those cases that actually exist: ints,
2943 // longs, and Object. Adding others should be straightforward.
2944 Node* load_store;
2945 switch(type) {
2946 case T_INT:
2947 if (kind == LS_xadd) {
2948 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2949 } else if (kind == LS_xchg) {
2950 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2951 } else if (kind == LS_cmpxchg) {
2952 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2953 } else {
2954 ShouldNotReachHere();
2955 }
2956 break;
2957 case T_LONG:
2958 if (kind == LS_xadd) {
2959 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2960 } else if (kind == LS_xchg) {
2961 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2962 } else if (kind == LS_cmpxchg) {
2963 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2964 } else {
2965 ShouldNotReachHere();
2966 }
2967 break;
2968 case T_OBJECT:
2969 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2970 // could be delayed during Parse (for example, in adjust_map_after_if()).
2971 // Execute transformation here to avoid barrier generation in such case.
2972 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2973 newval = _gvn.makecon(TypePtr::NULL_PTR);
2974
2975 // Reference stores need a store barrier.
2976 if (kind == LS_xchg) {
2977 // If pre-barrier must execute before the oop store, old value will require do_load here.
2978 if (!can_move_pre_barrier()) {
2979 pre_barrier(true /* do_load*/,
2980 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2981 NULL /* pre_val*/,
2982 T_OBJECT);
2983 } // Else move pre_barrier to use load_store value, see below.
2984 } else if (kind == LS_cmpxchg) {
2985 // Same as for newval above:
2986 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2987 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2988 }
2989 // The only known value which might get overwritten is oldval.
2990 pre_barrier(false /* do_load */,
2991 control(), NULL, NULL, max_juint, NULL, NULL,
2992 oldval /* pre_val */,
2993 T_OBJECT);
2994 } else {
2995 ShouldNotReachHere();
2996 }
2997
2998 #ifdef _LP64
2999 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3000 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3001 if (kind == LS_xchg) {
3002 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
3003 newval_enc, adr_type, value_type->make_narrowoop()));
3004 } else {
3005 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3006 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3007 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
3008 newval_enc, oldval_enc));
3009 }
3010 } else
3011 #endif
3012 {
3013 if (kind == LS_xchg) {
3014 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3015 } else {
3016 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3017 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3018 }
3019 }
3020 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3021 break;
3022 default:
3023 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3024 break;
3025 }
3026
3027 // SCMemProjNodes represent the memory state of a LoadStore. Their
3028 // main role is to prevent LoadStore nodes from being optimized away
3029 // when their results aren't used.
3030 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3031 set_memory(proj, alias_idx);
3032
3033 if (type == T_OBJECT && kind == LS_xchg) {
3034 #ifdef _LP64
3035 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3036 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3037 }
3038 #endif
3039 if (can_move_pre_barrier()) {
3040 // Don't need to load pre_val. The old value is returned by load_store.
3041 // The pre_barrier can execute after the xchg as long as no safepoint
3042 // gets inserted between them.
3043 pre_barrier(false /* do_load */,
3044 control(), NULL, NULL, max_juint, NULL, NULL,
3045 load_store /* pre_val */,
3046 T_OBJECT);
3047 }
3048 }
3049
3050 // Add the trailing membar surrounding the access
3051 insert_mem_bar(Op_MemBarCPUOrder);
3052 insert_mem_bar(Op_MemBarAcquire);
3053
3054 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3055 set_result(load_store);
3056 return true;
3057 }
3058
3059 //----------------------------inline_unsafe_ordered_store----------------------
3060 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3061 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3062 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3063 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3064 // This is another variant of inline_unsafe_access, differing in
3065 // that it always issues store-store ("release") barrier and ensures
3066 // store-atomicity (which only matters for "long").
3067
3068 if (callee()->is_static()) return false; // caller must have the capability!
3069
3070 #ifndef PRODUCT
3071 {
3072 ResourceMark rm;
3073 // Check the signatures.
3074 ciSignature* sig = callee()->signature();
3075 #ifdef ASSERT
3076 BasicType rtype = sig->return_type()->basic_type();
3077 assert(rtype == T_VOID, "must return void");
3078 assert(sig->count() == 3, "has 3 arguments");
3079 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3080 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3081 #endif // ASSERT
3082 }
3083 #endif //PRODUCT
3084
3085 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3086
3087 // Get arguments:
3088 Node* receiver = argument(0); // type: oop
3089 Node* base = argument(1); // type: oop
3090 Node* offset = argument(2); // type: long
3091 Node* val = argument(4); // type: oop, int, or long
3092
3093 // Null check receiver.
3094 receiver = null_check(receiver);
3095 if (stopped()) {
3096 return true;
3097 }
3098
3099 // Build field offset expression.
3100 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3101 // 32-bit machines ignore the high half of long offsets
3102 offset = ConvL2X(offset);
3103 Node* adr = make_unsafe_address(base, offset);
3104 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3105 const Type *value_type = Type::get_const_basic_type(type);
3106 Compile::AliasType* alias_type = C->alias_type(adr_type);
3107
3108 insert_mem_bar(Op_MemBarRelease);
3109 insert_mem_bar(Op_MemBarCPUOrder);
3110 // Ensure that the store is atomic for longs:
3111 const bool require_atomic_access = true;
3112 Node* store;
3113 if (type == T_OBJECT) // reference stores need a store barrier.
3114 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3115 else {
3116 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3117 }
3118 insert_mem_bar(Op_MemBarCPUOrder);
3119 return true;
3120 }
3121
3122 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3123 // Regardless of form, don't allow previous ld/st to move down,
3124 // then issue acquire, release, or volatile mem_bar.
3125 insert_mem_bar(Op_MemBarCPUOrder);
3126 switch(id) {
3127 case vmIntrinsics::_loadFence:
3128 insert_mem_bar(Op_LoadFence);
3129 return true;
3130 case vmIntrinsics::_storeFence:
3131 insert_mem_bar(Op_StoreFence);
3132 return true;
3133 case vmIntrinsics::_fullFence:
3134 insert_mem_bar(Op_MemBarVolatile);
3135 return true;
3136 default:
3137 fatal_unexpected_iid(id);
3138 return false;
3139 }
3140 }
3141
3142 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3143 if (!kls->is_Con()) {
3144 return true;
3145 }
3146 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3147 if (klsptr == NULL) {
3148 return true;
3149 }
3150 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3151 // don't need a guard for a klass that is already initialized
3152 return !ik->is_initialized();
3153 }
3154
3155 //----------------------------inline_unsafe_allocate---------------------------
3156 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3157 bool LibraryCallKit::inline_unsafe_allocate() {
3158 if (callee()->is_static()) return false; // caller must have the capability!
3159
3160 null_check_receiver(); // null-check, then ignore
3161 Node* cls = null_check(argument(1));
3162 if (stopped()) return true;
3163
3164 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3165 kls = null_check(kls);
3166 if (stopped()) return true; // argument was like int.class
3167
3168 Node* test = NULL;
3169 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3170 // Note: The argument might still be an illegal value like
3171 // Serializable.class or Object[].class. The runtime will handle it.
3172 // But we must make an explicit check for initialization.
3173 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3174 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3175 // can generate code to load it as unsigned byte.
3176 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3177 Node* bits = intcon(InstanceKlass::fully_initialized);
3178 test = _gvn.transform(new SubINode(inst, bits));
3179 // The 'test' is non-zero if we need to take a slow path.
3180 }
3181
3182 Node* obj = new_instance(kls, test);
3183 set_result(obj);
3184 return true;
3185 }
3186
3187 #ifdef TRACE_HAVE_INTRINSICS
3188 /*
3189 * oop -> myklass
3190 * myklass->trace_id |= USED
3191 * return myklass->trace_id & ~0x3
3192 */
3193 bool LibraryCallKit::inline_native_classID() {
3194 null_check_receiver(); // null-check, then ignore
3195 Node* cls = null_check(argument(1), T_OBJECT);
3196 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3197 kls = null_check(kls, T_OBJECT);
3198 ByteSize offset = TRACE_ID_OFFSET;
3199 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3200 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3201 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3202 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3203 Node* clsused = longcon(0x01l); // set the class bit
3204 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3205
3206 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3207 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3208 set_result(andl);
3209 return true;
3210 }
3211
3212 bool LibraryCallKit::inline_native_threadID() {
3213 Node* tls_ptr = NULL;
3214 Node* cur_thr = generate_current_thread(tls_ptr);
3215 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3216 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3217 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3218
3219 Node* threadid = NULL;
3220 size_t thread_id_size = OSThread::thread_id_size();
3221 if (thread_id_size == (size_t) BytesPerLong) {
3222 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3223 } else if (thread_id_size == (size_t) BytesPerInt) {
3224 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3225 } else {
3226 ShouldNotReachHere();
3227 }
3228 set_result(threadid);
3229 return true;
3230 }
3231 #endif
3232
3233 //------------------------inline_native_time_funcs--------------
3234 // inline code for System.currentTimeMillis() and System.nanoTime()
3235 // these have the same type and signature
3236 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3237 const TypeFunc* tf = OptoRuntime::void_long_Type();
3238 const TypePtr* no_memory_effects = NULL;
3239 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3240 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3241 #ifdef ASSERT
3242 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3243 assert(value_top == top(), "second value must be top");
3244 #endif
3245 set_result(value);
3246 return true;
3247 }
3248
3249 //------------------------inline_native_currentThread------------------
3250 bool LibraryCallKit::inline_native_currentThread() {
3251 Node* junk = NULL;
3252 set_result(generate_current_thread(junk));
3253 return true;
3254 }
3255
3256 //------------------------inline_native_isInterrupted------------------
3257 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3258 bool LibraryCallKit::inline_native_isInterrupted() {
3259 // Add a fast path to t.isInterrupted(clear_int):
3260 // (t == Thread.current() &&
3261 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3262 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3263 // So, in the common case that the interrupt bit is false,
3264 // we avoid making a call into the VM. Even if the interrupt bit
3265 // is true, if the clear_int argument is false, we avoid the VM call.
3266 // However, if the receiver is not currentThread, we must call the VM,
3267 // because there must be some locking done around the operation.
3268
3269 // We only go to the fast case code if we pass two guards.
3270 // Paths which do not pass are accumulated in the slow_region.
3271
3272 enum {
3273 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3274 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3275 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3276 PATH_LIMIT
3277 };
3278
3279 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3280 // out of the function.
3281 insert_mem_bar(Op_MemBarCPUOrder);
3282
3283 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3284 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3285
3286 RegionNode* slow_region = new RegionNode(1);
3287 record_for_igvn(slow_region);
3288
3289 // (a) Receiving thread must be the current thread.
3290 Node* rec_thr = argument(0);
3291 Node* tls_ptr = NULL;
3292 Node* cur_thr = generate_current_thread(tls_ptr);
3293 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3294 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3295
3296 generate_slow_guard(bol_thr, slow_region);
3297
3298 // (b) Interrupt bit on TLS must be false.
3299 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3300 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3301 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3302
3303 // Set the control input on the field _interrupted read to prevent it floating up.
3304 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3305 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3306 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3307
3308 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3309
3310 // First fast path: if (!TLS._interrupted) return false;
3311 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3312 result_rgn->init_req(no_int_result_path, false_bit);
3313 result_val->init_req(no_int_result_path, intcon(0));
3314
3315 // drop through to next case
3316 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3317
3318 #ifndef TARGET_OS_FAMILY_windows
3319 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3320 Node* clr_arg = argument(1);
3321 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3322 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3323 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3324
3325 // Second fast path: ... else if (!clear_int) return true;
3326 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3327 result_rgn->init_req(no_clear_result_path, false_arg);
3328 result_val->init_req(no_clear_result_path, intcon(1));
3329
3330 // drop through to next case
3331 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3332 #else
3333 // To return true on Windows you must read the _interrupted field
3334 // and check the the event state i.e. take the slow path.
3335 #endif // TARGET_OS_FAMILY_windows
3336
3337 // (d) Otherwise, go to the slow path.
3338 slow_region->add_req(control());
3339 set_control( _gvn.transform(slow_region));
3340
3341 if (stopped()) {
3342 // There is no slow path.
3343 result_rgn->init_req(slow_result_path, top());
3344 result_val->init_req(slow_result_path, top());
3345 } else {
3346 // non-virtual because it is a private non-static
3347 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3348
3349 Node* slow_val = set_results_for_java_call(slow_call);
3350 // this->control() comes from set_results_for_java_call
3351
3352 Node* fast_io = slow_call->in(TypeFunc::I_O);
3353 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3354
3355 // These two phis are pre-filled with copies of of the fast IO and Memory
3356 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3357 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3358
3359 result_rgn->init_req(slow_result_path, control());
3360 result_io ->init_req(slow_result_path, i_o());
3361 result_mem->init_req(slow_result_path, reset_memory());
3362 result_val->init_req(slow_result_path, slow_val);
3363
3364 set_all_memory(_gvn.transform(result_mem));
3365 set_i_o( _gvn.transform(result_io));
3366 }
3367
3368 C->set_has_split_ifs(true); // Has chance for split-if optimization
3369 set_result(result_rgn, result_val);
3370 return true;
3371 }
3372
3373 //---------------------------load_mirror_from_klass----------------------------
3374 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3375 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3376 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3377 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3378 }
3379
3380 //-----------------------load_klass_from_mirror_common-------------------------
3381 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3382 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3383 // and branch to the given path on the region.
3384 // If never_see_null, take an uncommon trap on null, so we can optimistically
3385 // compile for the non-null case.
3386 // If the region is NULL, force never_see_null = true.
3387 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3388 bool never_see_null,
3389 RegionNode* region,
3390 int null_path,
3391 int offset) {
3392 if (region == NULL) never_see_null = true;
3393 Node* p = basic_plus_adr(mirror, offset);
3394 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3395 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3396 Node* null_ctl = top();
3397 kls = null_check_oop(kls, &null_ctl, never_see_null);
3398 if (region != NULL) {
3399 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3400 region->init_req(null_path, null_ctl);
3401 } else {
3402 assert(null_ctl == top(), "no loose ends");
3403 }
3404 return kls;
3405 }
3406
3407 //--------------------(inline_native_Class_query helpers)---------------------
3408 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3409 // Fall through if (mods & mask) == bits, take the guard otherwise.
3410 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3411 // Branch around if the given klass has the given modifier bit set.
3412 // Like generate_guard, adds a new path onto the region.
3413 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3414 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3415 Node* mask = intcon(modifier_mask);
3416 Node* bits = intcon(modifier_bits);
3417 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3418 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3419 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3420 return generate_fair_guard(bol, region);
3421 }
3422 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3423 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3424 }
3425
3426 //-------------------------inline_native_Class_query-------------------
3427 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3428 const Type* return_type = TypeInt::BOOL;
3429 Node* prim_return_value = top(); // what happens if it's a primitive class?
3430 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3431 bool expect_prim = false; // most of these guys expect to work on refs
3432
3433 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3434
3435 Node* mirror = argument(0);
3436 Node* obj = top();
3437
3438 switch (id) {
3439 case vmIntrinsics::_isInstance:
3440 // nothing is an instance of a primitive type
3441 prim_return_value = intcon(0);
3442 obj = argument(1);
3443 break;
3444 case vmIntrinsics::_getModifiers:
3445 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3446 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3447 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3448 break;
3449 case vmIntrinsics::_isInterface:
3450 prim_return_value = intcon(0);
3451 break;
3452 case vmIntrinsics::_isArray:
3453 prim_return_value = intcon(0);
3454 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3455 break;
3456 case vmIntrinsics::_isPrimitive:
3457 prim_return_value = intcon(1);
3458 expect_prim = true; // obviously
3459 break;
3460 case vmIntrinsics::_getSuperclass:
3461 prim_return_value = null();
3462 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3463 break;
3464 case vmIntrinsics::_getComponentType:
3465 prim_return_value = null();
3466 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3467 break;
3468 case vmIntrinsics::_getClassAccessFlags:
3469 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3470 return_type = TypeInt::INT; // not bool! 6297094
3471 break;
3472 default:
3473 fatal_unexpected_iid(id);
3474 break;
3475 }
3476
3477 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3478 if (mirror_con == NULL) return false; // cannot happen?
3479
3480 #ifndef PRODUCT
3481 if (C->print_intrinsics() || C->print_inlining()) {
3482 ciType* k = mirror_con->java_mirror_type();
3483 if (k) {
3484 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3485 k->print_name();
3486 tty->cr();
3487 }
3488 }
3489 #endif
3490
3491 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3492 RegionNode* region = new RegionNode(PATH_LIMIT);
3493 record_for_igvn(region);
3494 PhiNode* phi = new PhiNode(region, return_type);
3495
3496 // The mirror will never be null of Reflection.getClassAccessFlags, however
3497 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3498 // if it is. See bug 4774291.
3499
3500 // For Reflection.getClassAccessFlags(), the null check occurs in
3501 // the wrong place; see inline_unsafe_access(), above, for a similar
3502 // situation.
3503 mirror = null_check(mirror);
3504 // If mirror or obj is dead, only null-path is taken.
3505 if (stopped()) return true;
3506
3507 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3508
3509 // Now load the mirror's klass metaobject, and null-check it.
3510 // Side-effects region with the control path if the klass is null.
3511 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3512 // If kls is null, we have a primitive mirror.
3513 phi->init_req(_prim_path, prim_return_value);
3514 if (stopped()) { set_result(region, phi); return true; }
3515 bool safe_for_replace = (region->in(_prim_path) == top());
3516
3517 Node* p; // handy temp
3518 Node* null_ctl;
3519
3520 // Now that we have the non-null klass, we can perform the real query.
3521 // For constant classes, the query will constant-fold in LoadNode::Value.
3522 Node* query_value = top();
3523 switch (id) {
3524 case vmIntrinsics::_isInstance:
3525 // nothing is an instance of a primitive type
3526 query_value = gen_instanceof(obj, kls, safe_for_replace);
3527 break;
3528
3529 case vmIntrinsics::_getModifiers:
3530 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3531 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3532 break;
3533
3534 case vmIntrinsics::_isInterface:
3535 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3536 if (generate_interface_guard(kls, region) != NULL)
3537 // A guard was added. If the guard is taken, it was an interface.
3538 phi->add_req(intcon(1));
3539 // If we fall through, it's a plain class.
3540 query_value = intcon(0);
3541 break;
3542
3543 case vmIntrinsics::_isArray:
3544 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3545 if (generate_array_guard(kls, region) != NULL)
3546 // A guard was added. If the guard is taken, it was an array.
3547 phi->add_req(intcon(1));
3548 // If we fall through, it's a plain class.
3549 query_value = intcon(0);
3550 break;
3551
3552 case vmIntrinsics::_isPrimitive:
3553 query_value = intcon(0); // "normal" path produces false
3554 break;
3555
3556 case vmIntrinsics::_getSuperclass:
3557 // The rules here are somewhat unfortunate, but we can still do better
3558 // with random logic than with a JNI call.
3559 // Interfaces store null or Object as _super, but must report null.
3560 // Arrays store an intermediate super as _super, but must report Object.
3561 // Other types can report the actual _super.
3562 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3563 if (generate_interface_guard(kls, region) != NULL)
3564 // A guard was added. If the guard is taken, it was an interface.
3565 phi->add_req(null());
3566 if (generate_array_guard(kls, region) != NULL)
3567 // A guard was added. If the guard is taken, it was an array.
3568 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3569 // If we fall through, it's a plain class. Get its _super.
3570 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3571 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3572 null_ctl = top();
3573 kls = null_check_oop(kls, &null_ctl);
3574 if (null_ctl != top()) {
3575 // If the guard is taken, Object.superClass is null (both klass and mirror).
3576 region->add_req(null_ctl);
3577 phi ->add_req(null());
3578 }
3579 if (!stopped()) {
3580 query_value = load_mirror_from_klass(kls);
3581 }
3582 break;
3583
3584 case vmIntrinsics::_getComponentType:
3585 if (generate_array_guard(kls, region) != NULL) {
3586 // Be sure to pin the oop load to the guard edge just created:
3587 Node* is_array_ctrl = region->in(region->req()-1);
3588 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3589 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3590 phi->add_req(cmo);
3591 }
3592 query_value = null(); // non-array case is null
3593 break;
3594
3595 case vmIntrinsics::_getClassAccessFlags:
3596 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3597 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3598 break;
3599
3600 default:
3601 fatal_unexpected_iid(id);
3602 break;
3603 }
3604
3605 // Fall-through is the normal case of a query to a real class.
3606 phi->init_req(1, query_value);
3607 region->init_req(1, control());
3608
3609 C->set_has_split_ifs(true); // Has chance for split-if optimization
3610 set_result(region, phi);
3611 return true;
3612 }
3613
3614 //--------------------------inline_native_subtype_check------------------------
3615 // This intrinsic takes the JNI calls out of the heart of
3616 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3617 bool LibraryCallKit::inline_native_subtype_check() {
3618 // Pull both arguments off the stack.
3619 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3620 args[0] = argument(0);
3621 args[1] = argument(1);
3622 Node* klasses[2]; // corresponding Klasses: superk, subk
3623 klasses[0] = klasses[1] = top();
3624
3625 enum {
3626 // A full decision tree on {superc is prim, subc is prim}:
3627 _prim_0_path = 1, // {P,N} => false
3628 // {P,P} & superc!=subc => false
3629 _prim_same_path, // {P,P} & superc==subc => true
3630 _prim_1_path, // {N,P} => false
3631 _ref_subtype_path, // {N,N} & subtype check wins => true
3632 _both_ref_path, // {N,N} & subtype check loses => false
3633 PATH_LIMIT
3634 };
3635
3636 RegionNode* region = new RegionNode(PATH_LIMIT);
3637 Node* phi = new PhiNode(region, TypeInt::BOOL);
3638 record_for_igvn(region);
3639
3640 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3641 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3642 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3643
3644 // First null-check both mirrors and load each mirror's klass metaobject.
3645 int which_arg;
3646 for (which_arg = 0; which_arg <= 1; which_arg++) {
3647 Node* arg = args[which_arg];
3648 arg = null_check(arg);
3649 if (stopped()) break;
3650 args[which_arg] = arg;
3651
3652 Node* p = basic_plus_adr(arg, class_klass_offset);
3653 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3654 klasses[which_arg] = _gvn.transform(kls);
3655 }
3656
3657 // Having loaded both klasses, test each for null.
3658 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3659 for (which_arg = 0; which_arg <= 1; which_arg++) {
3660 Node* kls = klasses[which_arg];
3661 Node* null_ctl = top();
3662 kls = null_check_oop(kls, &null_ctl, never_see_null);
3663 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3664 region->init_req(prim_path, null_ctl);
3665 if (stopped()) break;
3666 klasses[which_arg] = kls;
3667 }
3668
3669 if (!stopped()) {
3670 // now we have two reference types, in klasses[0..1]
3671 Node* subk = klasses[1]; // the argument to isAssignableFrom
3672 Node* superk = klasses[0]; // the receiver
3673 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3674 // now we have a successful reference subtype check
3675 region->set_req(_ref_subtype_path, control());
3676 }
3677
3678 // If both operands are primitive (both klasses null), then
3679 // we must return true when they are identical primitives.
3680 // It is convenient to test this after the first null klass check.
3681 set_control(region->in(_prim_0_path)); // go back to first null check
3682 if (!stopped()) {
3683 // Since superc is primitive, make a guard for the superc==subc case.
3684 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3685 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3686 generate_guard(bol_eq, region, PROB_FAIR);
3687 if (region->req() == PATH_LIMIT+1) {
3688 // A guard was added. If the added guard is taken, superc==subc.
3689 region->swap_edges(PATH_LIMIT, _prim_same_path);
3690 region->del_req(PATH_LIMIT);
3691 }
3692 region->set_req(_prim_0_path, control()); // Not equal after all.
3693 }
3694
3695 // these are the only paths that produce 'true':
3696 phi->set_req(_prim_same_path, intcon(1));
3697 phi->set_req(_ref_subtype_path, intcon(1));
3698
3699 // pull together the cases:
3700 assert(region->req() == PATH_LIMIT, "sane region");
3701 for (uint i = 1; i < region->req(); i++) {
3702 Node* ctl = region->in(i);
3703 if (ctl == NULL || ctl == top()) {
3704 region->set_req(i, top());
3705 phi ->set_req(i, top());
3706 } else if (phi->in(i) == NULL) {
3707 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3708 }
3709 }
3710
3711 set_control(_gvn.transform(region));
3712 set_result(_gvn.transform(phi));
3713 return true;
3714 }
3715
3716 //---------------------generate_array_guard_common------------------------
3717 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3718 bool obj_array, bool not_array) {
3719 // If obj_array/non_array==false/false:
3720 // Branch around if the given klass is in fact an array (either obj or prim).
3721 // If obj_array/non_array==false/true:
3722 // Branch around if the given klass is not an array klass of any kind.
3723 // If obj_array/non_array==true/true:
3724 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3725 // If obj_array/non_array==true/false:
3726 // Branch around if the kls is an oop array (Object[] or subtype)
3727 //
3728 // Like generate_guard, adds a new path onto the region.
3729 jint layout_con = 0;
3730 Node* layout_val = get_layout_helper(kls, layout_con);
3731 if (layout_val == NULL) {
3732 bool query = (obj_array
3733 ? Klass::layout_helper_is_objArray(layout_con)
3734 : Klass::layout_helper_is_array(layout_con));
3735 if (query == not_array) {
3736 return NULL; // never a branch
3737 } else { // always a branch
3738 Node* always_branch = control();
3739 if (region != NULL)
3740 region->add_req(always_branch);
3741 set_control(top());
3742 return always_branch;
3743 }
3744 }
3745 // Now test the correct condition.
3746 jint nval = (obj_array
3747 ? ((jint)Klass::_lh_array_tag_type_value
3748 << Klass::_lh_array_tag_shift)
3749 : Klass::_lh_neutral_value);
3750 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3751 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3752 // invert the test if we are looking for a non-array
3753 if (not_array) btest = BoolTest(btest).negate();
3754 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3755 return generate_fair_guard(bol, region);
3756 }
3757
3758
3759 //-----------------------inline_native_newArray--------------------------
3760 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3761 bool LibraryCallKit::inline_native_newArray() {
3762 Node* mirror = argument(0);
3763 Node* count_val = argument(1);
3764
3765 mirror = null_check(mirror);
3766 // If mirror or obj is dead, only null-path is taken.
3767 if (stopped()) return true;
3768
3769 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3770 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3771 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3772 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3773 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3774
3775 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3776 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3777 result_reg, _slow_path);
3778 Node* normal_ctl = control();
3779 Node* no_array_ctl = result_reg->in(_slow_path);
3780
3781 // Generate code for the slow case. We make a call to newArray().
3782 set_control(no_array_ctl);
3783 if (!stopped()) {
3784 // Either the input type is void.class, or else the
3785 // array klass has not yet been cached. Either the
3786 // ensuing call will throw an exception, or else it
3787 // will cache the array klass for next time.
3788 PreserveJVMState pjvms(this);
3789 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3790 Node* slow_result = set_results_for_java_call(slow_call);
3791 // this->control() comes from set_results_for_java_call
3792 result_reg->set_req(_slow_path, control());
3793 result_val->set_req(_slow_path, slow_result);
3794 result_io ->set_req(_slow_path, i_o());
3795 result_mem->set_req(_slow_path, reset_memory());
3796 }
3797
3798 set_control(normal_ctl);
3799 if (!stopped()) {
3800 // Normal case: The array type has been cached in the java.lang.Class.
3801 // The following call works fine even if the array type is polymorphic.
3802 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3803 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3804 result_reg->init_req(_normal_path, control());
3805 result_val->init_req(_normal_path, obj);
3806 result_io ->init_req(_normal_path, i_o());
3807 result_mem->init_req(_normal_path, reset_memory());
3808 }
3809
3810 // Return the combined state.
3811 set_i_o( _gvn.transform(result_io) );
3812 set_all_memory( _gvn.transform(result_mem));
3813
3814 C->set_has_split_ifs(true); // Has chance for split-if optimization
3815 set_result(result_reg, result_val);
3816 return true;
3817 }
3818
3819 //----------------------inline_native_getLength--------------------------
3820 // public static native int java.lang.reflect.Array.getLength(Object array);
3821 bool LibraryCallKit::inline_native_getLength() {
3822 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3823
3824 Node* array = null_check(argument(0));
3825 // If array is dead, only null-path is taken.
3826 if (stopped()) return true;
3827
3828 // Deoptimize if it is a non-array.
3829 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3830
3831 if (non_array != NULL) {
3832 PreserveJVMState pjvms(this);
3833 set_control(non_array);
3834 uncommon_trap(Deoptimization::Reason_intrinsic,
3835 Deoptimization::Action_maybe_recompile);
3836 }
3837
3838 // If control is dead, only non-array-path is taken.
3839 if (stopped()) return true;
3840
3841 // The works fine even if the array type is polymorphic.
3842 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3843 Node* result = load_array_length(array);
3844
3845 C->set_has_split_ifs(true); // Has chance for split-if optimization
3846 set_result(result);
3847 return true;
3848 }
3849
3850 //------------------------inline_array_copyOf----------------------------
3851 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3852 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3853 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3854 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3855
3856 // Get the arguments.
3857 Node* original = argument(0);
3858 Node* start = is_copyOfRange? argument(1): intcon(0);
3859 Node* end = is_copyOfRange? argument(2): argument(1);
3860 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3861
3862 Node* newcopy;
3863
3864 // Set the original stack and the reexecute bit for the interpreter to reexecute
3865 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3866 { PreserveReexecuteState preexecs(this);
3867 jvms()->set_should_reexecute(true);
3868
3869 array_type_mirror = null_check(array_type_mirror);
3870 original = null_check(original);
3871
3872 // Check if a null path was taken unconditionally.
3873 if (stopped()) return true;
3874
3875 Node* orig_length = load_array_length(original);
3876
3877 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3878 klass_node = null_check(klass_node);
3879
3880 RegionNode* bailout = new RegionNode(1);
3881 record_for_igvn(bailout);
3882
3883 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3884 // Bail out if that is so.
3885 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3886 if (not_objArray != NULL) {
3887 // Improve the klass node's type from the new optimistic assumption:
3888 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3889 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3890 Node* cast = new CastPPNode(klass_node, akls);
3891 cast->init_req(0, control());
3892 klass_node = _gvn.transform(cast);
3893 }
3894
3895 // Bail out if either start or end is negative.
3896 generate_negative_guard(start, bailout, &start);
3897 generate_negative_guard(end, bailout, &end);
3898
3899 Node* length = end;
3900 if (_gvn.type(start) != TypeInt::ZERO) {
3901 length = _gvn.transform(new SubINode(end, start));
3902 }
3903
3904 // Bail out if length is negative.
3905 // Without this the new_array would throw
3906 // NegativeArraySizeException but IllegalArgumentException is what
3907 // should be thrown
3908 generate_negative_guard(length, bailout, &length);
3909
3910 if (bailout->req() > 1) {
3911 PreserveJVMState pjvms(this);
3912 set_control(_gvn.transform(bailout));
3913 uncommon_trap(Deoptimization::Reason_intrinsic,
3914 Deoptimization::Action_maybe_recompile);
3915 }
3916
3917 if (!stopped()) {
3918 // How many elements will we copy from the original?
3919 // The answer is MinI(orig_length - start, length).
3920 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3921 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3922
3923 newcopy = new_array(klass_node, length, 0); // no argments to push
3924
3925 // Generate a direct call to the right arraycopy function(s).
3926 // We know the copy is disjoint but we might not know if the
3927 // oop stores need checking.
3928 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3929 // This will fail a store-check if x contains any non-nulls.
3930 bool disjoint_bases = true;
3931 // if start > orig_length then the length of the copy may be
3932 // negative.
3933 bool length_never_negative = !is_copyOfRange;
3934 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3935 original, start, newcopy, intcon(0), moved,
3936 disjoint_bases, length_never_negative);
3937 }
3938 } // original reexecute is set back here
3939
3940 C->set_has_split_ifs(true); // Has chance for split-if optimization
3941 if (!stopped()) {
3942 set_result(newcopy);
3943 }
3944 return true;
3945 }
3946
3947
3948 //----------------------generate_virtual_guard---------------------------
3949 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3950 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3951 RegionNode* slow_region) {
3952 ciMethod* method = callee();
3953 int vtable_index = method->vtable_index();
3954 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3955 err_msg_res("bad index %d", vtable_index));
3956 // Get the Method* out of the appropriate vtable entry.
3957 int entry_offset = (InstanceKlass::vtable_start_offset() +
3958 vtable_index*vtableEntry::size()) * wordSize +
3959 vtableEntry::method_offset_in_bytes();
3960 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3961 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3962
3963 // Compare the target method with the expected method (e.g., Object.hashCode).
3964 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3965
3966 Node* native_call = makecon(native_call_addr);
3967 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3968 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3969
3970 return generate_slow_guard(test_native, slow_region);
3971 }
3972
3973 //-----------------------generate_method_call----------------------------
3974 // Use generate_method_call to make a slow-call to the real
3975 // method if the fast path fails. An alternative would be to
3976 // use a stub like OptoRuntime::slow_arraycopy_Java.
3977 // This only works for expanding the current library call,
3978 // not another intrinsic. (E.g., don't use this for making an
3979 // arraycopy call inside of the copyOf intrinsic.)
3980 CallJavaNode*
3981 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3982 // When compiling the intrinsic method itself, do not use this technique.
3983 guarantee(callee() != C->method(), "cannot make slow-call to self");
3984
3985 ciMethod* method = callee();
3986 // ensure the JVMS we have will be correct for this call
3987 guarantee(method_id == method->intrinsic_id(), "must match");
3988
3989 const TypeFunc* tf = TypeFunc::make(method);
3990 CallJavaNode* slow_call;
3991 if (is_static) {
3992 assert(!is_virtual, "");
3993 slow_call = new CallStaticJavaNode(C, tf,
3994 SharedRuntime::get_resolve_static_call_stub(),
3995 method, bci());
3996 } else if (is_virtual) {
3997 null_check_receiver();
3998 int vtable_index = Method::invalid_vtable_index;
3999 if (UseInlineCaches) {
4000 // Suppress the vtable call
4001 } else {
4002 // hashCode and clone are not a miranda methods,
4003 // so the vtable index is fixed.
4004 // No need to use the linkResolver to get it.
4005 vtable_index = method->vtable_index();
4006 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4007 err_msg_res("bad index %d", vtable_index));
4008 }
4009 slow_call = new CallDynamicJavaNode(tf,
4010 SharedRuntime::get_resolve_virtual_call_stub(),
4011 method, vtable_index, bci());
4012 } else { // neither virtual nor static: opt_virtual
4013 null_check_receiver();
4014 slow_call = new CallStaticJavaNode(C, tf,
4015 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4016 method, bci());
4017 slow_call->set_optimized_virtual(true);
4018 }
4019 set_arguments_for_java_call(slow_call);
4020 set_edges_for_java_call(slow_call);
4021 return slow_call;
4022 }
4023
4024
4025 /**
4026 * Build special case code for calls to hashCode on an object. This call may
4027 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4028 * slightly different code.
4029 */
4030 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4031 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4032 assert(!(is_virtual && is_static), "either virtual, special, or static");
4033
4034 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4035
4036 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4037 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4038 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4039 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4040 Node* obj = NULL;
4041 if (!is_static) {
4042 // Check for hashing null object
4043 obj = null_check_receiver();
4044 if (stopped()) return true; // unconditionally null
4045 result_reg->init_req(_null_path, top());
4046 result_val->init_req(_null_path, top());
4047 } else {
4048 // Do a null check, and return zero if null.
4049 // System.identityHashCode(null) == 0
4050 obj = argument(0);
4051 Node* null_ctl = top();
4052 obj = null_check_oop(obj, &null_ctl);
4053 result_reg->init_req(_null_path, null_ctl);
4054 result_val->init_req(_null_path, _gvn.intcon(0));
4055 }
4056
4057 // Unconditionally null? Then return right away.
4058 if (stopped()) {
4059 set_control( result_reg->in(_null_path));
4060 if (!stopped())
4061 set_result(result_val->in(_null_path));
4062 return true;
4063 }
4064
4065 // We only go to the fast case code if we pass a number of guards. The
4066 // paths which do not pass are accumulated in the slow_region.
4067 RegionNode* slow_region = new RegionNode(1);
4068 record_for_igvn(slow_region);
4069
4070 // If this is a virtual call, we generate a funny guard. We pull out
4071 // the vtable entry corresponding to hashCode() from the target object.
4072 // If the target method which we are calling happens to be the native
4073 // Object hashCode() method, we pass the guard. We do not need this
4074 // guard for non-virtual calls -- the caller is known to be the native
4075 // Object hashCode().
4076 if (is_virtual) {
4077 // After null check, get the object's klass.
4078 Node* obj_klass = load_object_klass(obj);
4079 generate_virtual_guard(obj_klass, slow_region);
4080 }
4081
4082 // Get the header out of the object, use LoadMarkNode when available
4083 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4084 // The control of the load must be NULL. Otherwise, the load can move before
4085 // the null check after castPP removal.
4086 Node* no_ctrl = NULL;
4087 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4088
4089 // Test the header to see if it is unlocked.
4090 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4091 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4092 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4093 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4094 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4095
4096 generate_slow_guard(test_unlocked, slow_region);
4097
4098 // Get the hash value and check to see that it has been properly assigned.
4099 // We depend on hash_mask being at most 32 bits and avoid the use of
4100 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4101 // vm: see markOop.hpp.
4102 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4103 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4104 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4105 // This hack lets the hash bits live anywhere in the mark object now, as long
4106 // as the shift drops the relevant bits into the low 32 bits. Note that
4107 // Java spec says that HashCode is an int so there's no point in capturing
4108 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4109 hshifted_header = ConvX2I(hshifted_header);
4110 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4111
4112 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4113 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4114 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4115
4116 generate_slow_guard(test_assigned, slow_region);
4117
4118 Node* init_mem = reset_memory();
4119 // fill in the rest of the null path:
4120 result_io ->init_req(_null_path, i_o());
4121 result_mem->init_req(_null_path, init_mem);
4122
4123 result_val->init_req(_fast_path, hash_val);
4124 result_reg->init_req(_fast_path, control());
4125 result_io ->init_req(_fast_path, i_o());
4126 result_mem->init_req(_fast_path, init_mem);
4127
4128 // Generate code for the slow case. We make a call to hashCode().
4129 set_control(_gvn.transform(slow_region));
4130 if (!stopped()) {
4131 // No need for PreserveJVMState, because we're using up the present state.
4132 set_all_memory(init_mem);
4133 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4134 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4135 Node* slow_result = set_results_for_java_call(slow_call);
4136 // this->control() comes from set_results_for_java_call
4137 result_reg->init_req(_slow_path, control());
4138 result_val->init_req(_slow_path, slow_result);
4139 result_io ->set_req(_slow_path, i_o());
4140 result_mem ->set_req(_slow_path, reset_memory());
4141 }
4142
4143 // Return the combined state.
4144 set_i_o( _gvn.transform(result_io) );
4145 set_all_memory( _gvn.transform(result_mem));
4146
4147 set_result(result_reg, result_val);
4148 return true;
4149 }
4150
4151 //---------------------------inline_native_getClass----------------------------
4152 // public final native Class<?> java.lang.Object.getClass();
4153 //
4154 // Build special case code for calls to getClass on an object.
4155 bool LibraryCallKit::inline_native_getClass() {
4156 Node* obj = null_check_receiver();
4157 if (stopped()) return true;
4158 set_result(load_mirror_from_klass(load_object_klass(obj)));
4159 return true;
4160 }
4161
4162 //-----------------inline_native_Reflection_getCallerClass---------------------
4163 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4164 //
4165 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4166 //
4167 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4168 // in that it must skip particular security frames and checks for
4169 // caller sensitive methods.
4170 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4171 #ifndef PRODUCT
4172 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4173 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4174 }
4175 #endif
4176
4177 if (!jvms()->has_method()) {
4178 #ifndef PRODUCT
4179 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4180 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4181 }
4182 #endif
4183 return false;
4184 }
4185
4186 // Walk back up the JVM state to find the caller at the required
4187 // depth.
4188 JVMState* caller_jvms = jvms();
4189
4190 // Cf. JVM_GetCallerClass
4191 // NOTE: Start the loop at depth 1 because the current JVM state does
4192 // not include the Reflection.getCallerClass() frame.
4193 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4194 ciMethod* m = caller_jvms->method();
4195 switch (n) {
4196 case 0:
4197 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4198 break;
4199 case 1:
4200 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4201 if (!m->caller_sensitive()) {
4202 #ifndef PRODUCT
4203 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4204 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4205 }
4206 #endif
4207 return false; // bail-out; let JVM_GetCallerClass do the work
4208 }
4209 break;
4210 default:
4211 if (!m->is_ignored_by_security_stack_walk()) {
4212 // We have reached the desired frame; return the holder class.
4213 // Acquire method holder as java.lang.Class and push as constant.
4214 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4215 ciInstance* caller_mirror = caller_klass->java_mirror();
4216 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4217
4218 #ifndef PRODUCT
4219 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4220 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());
4221 tty->print_cr(" JVM state at this point:");
4222 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4223 ciMethod* m = jvms()->of_depth(i)->method();
4224 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4225 }
4226 }
4227 #endif
4228 return true;
4229 }
4230 break;
4231 }
4232 }
4233
4234 #ifndef PRODUCT
4235 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4236 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4237 tty->print_cr(" JVM state at this point:");
4238 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4239 ciMethod* m = jvms()->of_depth(i)->method();
4240 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4241 }
4242 }
4243 #endif
4244
4245 return false; // bail-out; let JVM_GetCallerClass do the work
4246 }
4247
4248 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4249 Node* arg = argument(0);
4250 Node* result;
4251
4252 switch (id) {
4253 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4254 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4255 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4256 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4257
4258 case vmIntrinsics::_doubleToLongBits: {
4259 // two paths (plus control) merge in a wood
4260 RegionNode *r = new RegionNode(3);
4261 Node *phi = new PhiNode(r, TypeLong::LONG);
4262
4263 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4264 // Build the boolean node
4265 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4266
4267 // Branch either way.
4268 // NaN case is less traveled, which makes all the difference.
4269 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4270 Node *opt_isnan = _gvn.transform(ifisnan);
4271 assert( opt_isnan->is_If(), "Expect an IfNode");
4272 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4273 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4274
4275 set_control(iftrue);
4276
4277 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4278 Node *slow_result = longcon(nan_bits); // return NaN
4279 phi->init_req(1, _gvn.transform( slow_result ));
4280 r->init_req(1, iftrue);
4281
4282 // Else fall through
4283 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4284 set_control(iffalse);
4285
4286 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4287 r->init_req(2, iffalse);
4288
4289 // Post merge
4290 set_control(_gvn.transform(r));
4291 record_for_igvn(r);
4292
4293 C->set_has_split_ifs(true); // Has chance for split-if optimization
4294 result = phi;
4295 assert(result->bottom_type()->isa_long(), "must be");
4296 break;
4297 }
4298
4299 case vmIntrinsics::_floatToIntBits: {
4300 // two paths (plus control) merge in a wood
4301 RegionNode *r = new RegionNode(3);
4302 Node *phi = new PhiNode(r, TypeInt::INT);
4303
4304 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4305 // Build the boolean node
4306 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4307
4308 // Branch either way.
4309 // NaN case is less traveled, which makes all the difference.
4310 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4311 Node *opt_isnan = _gvn.transform(ifisnan);
4312 assert( opt_isnan->is_If(), "Expect an IfNode");
4313 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4314 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4315
4316 set_control(iftrue);
4317
4318 static const jint nan_bits = 0x7fc00000;
4319 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4320 phi->init_req(1, _gvn.transform( slow_result ));
4321 r->init_req(1, iftrue);
4322
4323 // Else fall through
4324 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4325 set_control(iffalse);
4326
4327 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4328 r->init_req(2, iffalse);
4329
4330 // Post merge
4331 set_control(_gvn.transform(r));
4332 record_for_igvn(r);
4333
4334 C->set_has_split_ifs(true); // Has chance for split-if optimization
4335 result = phi;
4336 assert(result->bottom_type()->isa_int(), "must be");
4337 break;
4338 }
4339
4340 default:
4341 fatal_unexpected_iid(id);
4342 break;
4343 }
4344 set_result(_gvn.transform(result));
4345 return true;
4346 }
4347
4348 #ifdef _LP64
4349 #define XTOP ,top() /*additional argument*/
4350 #else //_LP64
4351 #define XTOP /*no additional argument*/
4352 #endif //_LP64
4353
4354 //----------------------inline_unsafe_copyMemory-------------------------
4355 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4356 bool LibraryCallKit::inline_unsafe_copyMemory() {
4357 if (callee()->is_static()) return false; // caller must have the capability!
4358 null_check_receiver(); // null-check receiver
4359 if (stopped()) return true;
4360
4361 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4362
4363 Node* src_ptr = argument(1); // type: oop
4364 Node* src_off = ConvL2X(argument(2)); // type: long
4365 Node* dst_ptr = argument(4); // type: oop
4366 Node* dst_off = ConvL2X(argument(5)); // type: long
4367 Node* size = ConvL2X(argument(7)); // type: long
4368
4369 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4370 "fieldOffset must be byte-scaled");
4371
4372 Node* src = make_unsafe_address(src_ptr, src_off);
4373 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4374
4375 // Conservatively insert a memory barrier on all memory slices.
4376 // Do not let writes of the copy source or destination float below the copy.
4377 insert_mem_bar(Op_MemBarCPUOrder);
4378
4379 // Call it. Note that the length argument is not scaled.
4380 make_runtime_call(RC_LEAF|RC_NO_FP,
4381 OptoRuntime::fast_arraycopy_Type(),
4382 StubRoutines::unsafe_arraycopy(),
4383 "unsafe_arraycopy",
4384 TypeRawPtr::BOTTOM,
4385 src, dst, size XTOP);
4386
4387 // Do not let reads of the copy destination float above the copy.
4388 insert_mem_bar(Op_MemBarCPUOrder);
4389
4390 return true;
4391 }
4392
4393 //------------------------clone_coping-----------------------------------
4394 // Helper function for inline_native_clone.
4395 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4396 assert(obj_size != NULL, "");
4397 Node* raw_obj = alloc_obj->in(1);
4398 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4399
4400 AllocateNode* alloc = NULL;
4401 if (ReduceBulkZeroing) {
4402 // We will be completely responsible for initializing this object -
4403 // mark Initialize node as complete.
4404 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4405 // The object was just allocated - there should be no any stores!
4406 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4407 // Mark as complete_with_arraycopy so that on AllocateNode
4408 // expansion, we know this AllocateNode is initialized by an array
4409 // copy and a StoreStore barrier exists after the array copy.
4410 alloc->initialization()->set_complete_with_arraycopy();
4411 }
4412
4413 // Copy the fastest available way.
4414 // TODO: generate fields copies for small objects instead.
4415 Node* src = obj;
4416 Node* dest = alloc_obj;
4417 Node* size = _gvn.transform(obj_size);
4418
4419 // Exclude the header but include array length to copy by 8 bytes words.
4420 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4421 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4422 instanceOopDesc::base_offset_in_bytes();
4423 // base_off:
4424 // 8 - 32-bit VM
4425 // 12 - 64-bit VM, compressed klass
4426 // 16 - 64-bit VM, normal klass
4427 if (base_off % BytesPerLong != 0) {
4428 assert(UseCompressedClassPointers, "");
4429 if (is_array) {
4430 // Exclude length to copy by 8 bytes words.
4431 base_off += sizeof(int);
4432 } else {
4433 // Include klass to copy by 8 bytes words.
4434 base_off = instanceOopDesc::klass_offset_in_bytes();
4435 }
4436 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4437 }
4438 src = basic_plus_adr(src, base_off);
4439 dest = basic_plus_adr(dest, base_off);
4440
4441 // Compute the length also, if needed:
4442 Node* countx = size;
4443 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4444 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4445
4446 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4447 bool disjoint_bases = true;
4448 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4449 src, NULL, dest, NULL, countx,
4450 /*dest_uninitialized*/true);
4451
4452 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4453 if (card_mark) {
4454 assert(!is_array, "");
4455 // Put in store barrier for any and all oops we are sticking
4456 // into this object. (We could avoid this if we could prove
4457 // that the object type contains no oop fields at all.)
4458 Node* no_particular_value = NULL;
4459 Node* no_particular_field = NULL;
4460 int raw_adr_idx = Compile::AliasIdxRaw;
4461 post_barrier(control(),
4462 memory(raw_adr_type),
4463 alloc_obj,
4464 no_particular_field,
4465 raw_adr_idx,
4466 no_particular_value,
4467 T_OBJECT,
4468 false);
4469 }
4470
4471 // Do not let reads from the cloned object float above the arraycopy.
4472 if (alloc != NULL) {
4473 // Do not let stores that initialize this object be reordered with
4474 // a subsequent store that would make this object accessible by
4475 // other threads.
4476 // Record what AllocateNode this StoreStore protects so that
4477 // escape analysis can go from the MemBarStoreStoreNode to the
4478 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4479 // based on the escape status of the AllocateNode.
4480 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4481 } else {
4482 insert_mem_bar(Op_MemBarCPUOrder);
4483 }
4484 }
4485
4486 //------------------------inline_native_clone----------------------------
4487 // protected native Object java.lang.Object.clone();
4488 //
4489 // Here are the simple edge cases:
4490 // null receiver => normal trap
4491 // virtual and clone was overridden => slow path to out-of-line clone
4492 // not cloneable or finalizer => slow path to out-of-line Object.clone
4493 //
4494 // The general case has two steps, allocation and copying.
4495 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4496 //
4497 // Copying also has two cases, oop arrays and everything else.
4498 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4499 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4500 //
4501 // These steps fold up nicely if and when the cloned object's klass
4502 // can be sharply typed as an object array, a type array, or an instance.
4503 //
4504 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4505 PhiNode* result_val;
4506
4507 // Set the reexecute bit for the interpreter to reexecute
4508 // the bytecode that invokes Object.clone if deoptimization happens.
4509 { PreserveReexecuteState preexecs(this);
4510 jvms()->set_should_reexecute(true);
4511
4512 Node* obj = null_check_receiver();
4513 if (stopped()) return true;
4514
4515 Node* obj_klass = load_object_klass(obj);
4516 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4517 const TypeOopPtr* toop = ((tklass != NULL)
4518 ? tklass->as_instance_type()
4519 : TypeInstPtr::NOTNULL);
4520
4521 // Conservatively insert a memory barrier on all memory slices.
4522 // Do not let writes into the original float below the clone.
4523 insert_mem_bar(Op_MemBarCPUOrder);
4524
4525 // paths into result_reg:
4526 enum {
4527 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4528 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4529 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4530 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4531 PATH_LIMIT
4532 };
4533 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4534 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4535 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4536 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4537 record_for_igvn(result_reg);
4538
4539 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4540 int raw_adr_idx = Compile::AliasIdxRaw;
4541
4542 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4543 if (array_ctl != NULL) {
4544 // It's an array.
4545 PreserveJVMState pjvms(this);
4546 set_control(array_ctl);
4547 Node* obj_length = load_array_length(obj);
4548 Node* obj_size = NULL;
4549 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4550
4551 if (!use_ReduceInitialCardMarks()) {
4552 // If it is an oop array, it requires very special treatment,
4553 // because card marking is required on each card of the array.
4554 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4555 if (is_obja != NULL) {
4556 PreserveJVMState pjvms2(this);
4557 set_control(is_obja);
4558 // Generate a direct call to the right arraycopy function(s).
4559 bool disjoint_bases = true;
4560 bool length_never_negative = true;
4561 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4562 obj, intcon(0), alloc_obj, intcon(0),
4563 obj_length,
4564 disjoint_bases, length_never_negative);
4565 result_reg->init_req(_objArray_path, control());
4566 result_val->init_req(_objArray_path, alloc_obj);
4567 result_i_o ->set_req(_objArray_path, i_o());
4568 result_mem ->set_req(_objArray_path, reset_memory());
4569 }
4570 }
4571 // Otherwise, there are no card marks to worry about.
4572 // (We can dispense with card marks if we know the allocation
4573 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4574 // causes the non-eden paths to take compensating steps to
4575 // simulate a fresh allocation, so that no further
4576 // card marks are required in compiled code to initialize
4577 // the object.)
4578
4579 if (!stopped()) {
4580 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4581
4582 // Present the results of the copy.
4583 result_reg->init_req(_array_path, control());
4584 result_val->init_req(_array_path, alloc_obj);
4585 result_i_o ->set_req(_array_path, i_o());
4586 result_mem ->set_req(_array_path, reset_memory());
4587 }
4588 }
4589
4590 // We only go to the instance fast case code if we pass a number of guards.
4591 // The paths which do not pass are accumulated in the slow_region.
4592 RegionNode* slow_region = new RegionNode(1);
4593 record_for_igvn(slow_region);
4594 if (!stopped()) {
4595 // It's an instance (we did array above). Make the slow-path tests.
4596 // If this is a virtual call, we generate a funny guard. We grab
4597 // the vtable entry corresponding to clone() from the target object.
4598 // If the target method which we are calling happens to be the
4599 // Object clone() method, we pass the guard. We do not need this
4600 // guard for non-virtual calls; the caller is known to be the native
4601 // Object clone().
4602 if (is_virtual) {
4603 generate_virtual_guard(obj_klass, slow_region);
4604 }
4605
4606 // The object must be cloneable and must not have a finalizer.
4607 // Both of these conditions may be checked in a single test.
4608 // We could optimize the cloneable test further, but we don't care.
4609 generate_access_flags_guard(obj_klass,
4610 // Test both conditions:
4611 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4612 // Must be cloneable but not finalizer:
4613 JVM_ACC_IS_CLONEABLE,
4614 slow_region);
4615 }
4616
4617 if (!stopped()) {
4618 // It's an instance, and it passed the slow-path tests.
4619 PreserveJVMState pjvms(this);
4620 Node* obj_size = NULL;
4621 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4622 // is reexecuted if deoptimization occurs and there could be problems when merging
4623 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4624 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4625
4626 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4627
4628 // Present the results of the slow call.
4629 result_reg->init_req(_instance_path, control());
4630 result_val->init_req(_instance_path, alloc_obj);
4631 result_i_o ->set_req(_instance_path, i_o());
4632 result_mem ->set_req(_instance_path, reset_memory());
4633 }
4634
4635 // Generate code for the slow case. We make a call to clone().
4636 set_control(_gvn.transform(slow_region));
4637 if (!stopped()) {
4638 PreserveJVMState pjvms(this);
4639 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4640 Node* slow_result = set_results_for_java_call(slow_call);
4641 // this->control() comes from set_results_for_java_call
4642 result_reg->init_req(_slow_path, control());
4643 result_val->init_req(_slow_path, slow_result);
4644 result_i_o ->set_req(_slow_path, i_o());
4645 result_mem ->set_req(_slow_path, reset_memory());
4646 }
4647
4648 // Return the combined state.
4649 set_control( _gvn.transform(result_reg));
4650 set_i_o( _gvn.transform(result_i_o));
4651 set_all_memory( _gvn.transform(result_mem));
4652 } // original reexecute is set back here
4653
4654 set_result(_gvn.transform(result_val));
4655 return true;
4656 }
4657
4658 //------------------------------basictype2arraycopy----------------------------
4659 address LibraryCallKit::basictype2arraycopy(BasicType t,
4660 Node* src_offset,
4661 Node* dest_offset,
4662 bool disjoint_bases,
4663 const char* &name,
4664 bool dest_uninitialized) {
4665 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4666 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4667
4668 bool aligned = false;
4669 bool disjoint = disjoint_bases;
4670
4671 // if the offsets are the same, we can treat the memory regions as
4672 // disjoint, because either the memory regions are in different arrays,
4673 // or they are identical (which we can treat as disjoint.) We can also
4674 // treat a copy with a destination index less that the source index
4675 // as disjoint since a low->high copy will work correctly in this case.
4676 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4677 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4678 // both indices are constants
4679 int s_offs = src_offset_inttype->get_con();
4680 int d_offs = dest_offset_inttype->get_con();
4681 int element_size = type2aelembytes(t);
4682 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4683 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4684 if (s_offs >= d_offs) disjoint = true;
4685 } else if (src_offset == dest_offset && src_offset != NULL) {
4686 // This can occur if the offsets are identical non-constants.
4687 disjoint = true;
4688 }
4689
4690 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4691 }
4692
4693
4694 //------------------------------inline_arraycopy-----------------------
4695 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4696 // Object dest, int destPos,
4697 // int length);
4698 bool LibraryCallKit::inline_arraycopy() {
4699 // Get the arguments.
4700 Node* src = argument(0); // type: oop
4701 Node* src_offset = argument(1); // type: int
4702 Node* dest = argument(2); // type: oop
4703 Node* dest_offset = argument(3); // type: int
4704 Node* length = argument(4); // type: int
4705
4706 // Compile time checks. If any of these checks cannot be verified at compile time,
4707 // we do not make a fast path for this call. Instead, we let the call remain as it
4708 // is. The checks we choose to mandate at compile time are:
4709 //
4710 // (1) src and dest are arrays.
4711 const Type* src_type = src->Value(&_gvn);
4712 const Type* dest_type = dest->Value(&_gvn);
4713 const TypeAryPtr* top_src = src_type->isa_aryptr();
4714 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4715
4716 // Do we have the type of src?
4717 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4718 // Do we have the type of dest?
4719 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4720 // Is the type for src from speculation?
4721 bool src_spec = false;
4722 // Is the type for dest from speculation?
4723 bool dest_spec = false;
4724
4725 if (!has_src || !has_dest) {
4726 // We don't have sufficient type information, let's see if
4727 // speculative types can help. We need to have types for both src
4728 // and dest so that it pays off.
4729
4730 // Do we already have or could we have type information for src
4731 bool could_have_src = has_src;
4732 // Do we already have or could we have type information for dest
4733 bool could_have_dest = has_dest;
4734
4735 ciKlass* src_k = NULL;
4736 if (!has_src) {
4737 src_k = src_type->speculative_type_not_null();
4738 if (src_k != NULL && src_k->is_array_klass()) {
4739 could_have_src = true;
4740 }
4741 }
4742
4743 ciKlass* dest_k = NULL;
4744 if (!has_dest) {
4745 dest_k = dest_type->speculative_type_not_null();
4746 if (dest_k != NULL && dest_k->is_array_klass()) {
4747 could_have_dest = true;
4748 }
4749 }
4750
4751 if (could_have_src && could_have_dest) {
4752 // This is going to pay off so emit the required guards
4753 if (!has_src) {
4754 src = maybe_cast_profiled_obj(src, src_k);
4755 src_type = _gvn.type(src);
4756 top_src = src_type->isa_aryptr();
4757 has_src = (top_src != NULL && top_src->klass() != NULL);
4758 src_spec = true;
4759 }
4760 if (!has_dest) {
4761 dest = maybe_cast_profiled_obj(dest, dest_k);
4762 dest_type = _gvn.type(dest);
4763 top_dest = dest_type->isa_aryptr();
4764 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4765 dest_spec = true;
4766 }
4767 }
4768 }
4769
4770 if (!has_src || !has_dest) {
4771 // Conservatively insert a memory barrier on all memory slices.
4772 // Do not let writes into the source float below the arraycopy.
4773 insert_mem_bar(Op_MemBarCPUOrder);
4774
4775 // Call StubRoutines::generic_arraycopy stub.
4776 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4777 src, src_offset, dest, dest_offset, length);
4778
4779 // Do not let reads from the destination float above the arraycopy.
4780 // Since we cannot type the arrays, we don't know which slices
4781 // might be affected. We could restrict this barrier only to those
4782 // memory slices which pertain to array elements--but don't bother.
4783 if (!InsertMemBarAfterArraycopy)
4784 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4785 insert_mem_bar(Op_MemBarCPUOrder);
4786 return true;
4787 }
4788
4789 // (2) src and dest arrays must have elements of the same BasicType
4790 // Figure out the size and type of the elements we will be copying.
4791 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4792 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4793 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4794 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4795
4796 if (src_elem != dest_elem || dest_elem == T_VOID) {
4797 // The component types are not the same or are not recognized. Punt.
4798 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4799 generate_slow_arraycopy(TypePtr::BOTTOM,
4800 src, src_offset, dest, dest_offset, length,
4801 /*dest_uninitialized*/false);
4802 return true;
4803 }
4804
4805 if (src_elem == T_OBJECT) {
4806 // If both arrays are object arrays then having the exact types
4807 // for both will remove the need for a subtype check at runtime
4808 // before the call and may make it possible to pick a faster copy
4809 // routine (without a subtype check on every element)
4810 // Do we have the exact type of src?
4811 bool could_have_src = src_spec;
4812 // Do we have the exact type of dest?
4813 bool could_have_dest = dest_spec;
4814 ciKlass* src_k = top_src->klass();
4815 ciKlass* dest_k = top_dest->klass();
4816 if (!src_spec) {
4817 src_k = src_type->speculative_type_not_null();
4818 if (src_k != NULL && src_k->is_array_klass()) {
4819 could_have_src = true;
4820 }
4821 }
4822 if (!dest_spec) {
4823 dest_k = dest_type->speculative_type_not_null();
4824 if (dest_k != NULL && dest_k->is_array_klass()) {
4825 could_have_dest = true;
4826 }
4827 }
4828 if (could_have_src && could_have_dest) {
4829 // If we can have both exact types, emit the missing guards
4830 if (could_have_src && !src_spec) {
4831 src = maybe_cast_profiled_obj(src, src_k);
4832 }
4833 if (could_have_dest && !dest_spec) {
4834 dest = maybe_cast_profiled_obj(dest, dest_k);
4835 }
4836 }
4837 }
4838
4839 //---------------------------------------------------------------------------
4840 // We will make a fast path for this call to arraycopy.
4841
4842 // We have the following tests left to perform:
4843 //
4844 // (3) src and dest must not be null.
4845 // (4) src_offset must not be negative.
4846 // (5) dest_offset must not be negative.
4847 // (6) length must not be negative.
4848 // (7) src_offset + length must not exceed length of src.
4849 // (8) dest_offset + length must not exceed length of dest.
4850 // (9) each element of an oop array must be assignable
4851
4852 RegionNode* slow_region = new RegionNode(1);
4853 record_for_igvn(slow_region);
4854
4855 // (3) operands must not be null
4856 // We currently perform our null checks with the null_check routine.
4857 // This means that the null exceptions will be reported in the caller
4858 // rather than (correctly) reported inside of the native arraycopy call.
4859 // This should be corrected, given time. We do our null check with the
4860 // stack pointer restored.
4861 src = null_check(src, T_ARRAY);
4862 dest = null_check(dest, T_ARRAY);
4863
4864 // (4) src_offset must not be negative.
4865 generate_negative_guard(src_offset, slow_region);
4866
4867 // (5) dest_offset must not be negative.
4868 generate_negative_guard(dest_offset, slow_region);
4869
4870 // (6) length must not be negative (moved to generate_arraycopy()).
4871 // generate_negative_guard(length, slow_region);
4872
4873 // (7) src_offset + length must not exceed length of src.
4874 generate_limit_guard(src_offset, length,
4875 load_array_length(src),
4876 slow_region);
4877
4878 // (8) dest_offset + length must not exceed length of dest.
4879 generate_limit_guard(dest_offset, length,
4880 load_array_length(dest),
4881 slow_region);
4882
4883 // (9) each element of an oop array must be assignable
4884 // The generate_arraycopy subroutine checks this.
4885
4886 // This is where the memory effects are placed:
4887 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4888 generate_arraycopy(adr_type, dest_elem,
4889 src, src_offset, dest, dest_offset, length,
4890 false, false, slow_region);
4891
4892 return true;
4893 }
4894
4895 //-----------------------------generate_arraycopy----------------------
4896 // Generate an optimized call to arraycopy.
4897 // Caller must guard against non-arrays.
4898 // Caller must determine a common array basic-type for both arrays.
4899 // Caller must validate offsets against array bounds.
4900 // The slow_region has already collected guard failure paths
4901 // (such as out of bounds length or non-conformable array types).
4902 // The generated code has this shape, in general:
4903 //
4904 // if (length == 0) return // via zero_path
4905 // slowval = -1
4906 // if (types unknown) {
4907 // slowval = call generic copy loop
4908 // if (slowval == 0) return // via checked_path
4909 // } else if (indexes in bounds) {
4910 // if ((is object array) && !(array type check)) {
4911 // slowval = call checked copy loop
4912 // if (slowval == 0) return // via checked_path
4913 // } else {
4914 // call bulk copy loop
4915 // return // via fast_path
4916 // }
4917 // }
4918 // // adjust params for remaining work:
4919 // if (slowval != -1) {
4920 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4921 // }
4922 // slow_region:
4923 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4924 // return // via slow_call_path
4925 //
4926 // This routine is used from several intrinsics: System.arraycopy,
4927 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4928 //
4929 void
4930 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4931 BasicType basic_elem_type,
4932 Node* src, Node* src_offset,
4933 Node* dest, Node* dest_offset,
4934 Node* copy_length,
4935 bool disjoint_bases,
4936 bool length_never_negative,
4937 RegionNode* slow_region) {
4938
4939 if (slow_region == NULL) {
4940 slow_region = new RegionNode(1);
4941 record_for_igvn(slow_region);
4942 }
4943
4944 Node* original_dest = dest;
4945 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4946 bool dest_uninitialized = false;
4947
4948 // See if this is the initialization of a newly-allocated array.
4949 // If so, we will take responsibility here for initializing it to zero.
4950 // (Note: Because tightly_coupled_allocation performs checks on the
4951 // out-edges of the dest, we need to avoid making derived pointers
4952 // from it until we have checked its uses.)
4953 if (ReduceBulkZeroing
4954 && !ZeroTLAB // pointless if already zeroed
4955 && basic_elem_type != T_CONFLICT // avoid corner case
4956 && !src->eqv_uncast(dest)
4957 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4958 != NULL)
4959 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4960 && alloc->maybe_set_complete(&_gvn)) {
4961 // "You break it, you buy it."
4962 InitializeNode* init = alloc->initialization();
4963 assert(init->is_complete(), "we just did this");
4964 init->set_complete_with_arraycopy();
4965 assert(dest->is_CheckCastPP(), "sanity");
4966 assert(dest->in(0)->in(0) == init, "dest pinned");
4967 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4968 // From this point on, every exit path is responsible for
4969 // initializing any non-copied parts of the object to zero.
4970 // Also, if this flag is set we make sure that arraycopy interacts properly
4971 // with G1, eliding pre-barriers. See CR 6627983.
4972 dest_uninitialized = true;
4973 } else {
4974 // No zeroing elimination here.
4975 alloc = NULL;
4976 //original_dest = dest;
4977 //dest_uninitialized = false;
4978 }
4979
4980 // Results are placed here:
4981 enum { fast_path = 1, // normal void-returning assembly stub
4982 checked_path = 2, // special assembly stub with cleanup
4983 slow_call_path = 3, // something went wrong; call the VM
4984 zero_path = 4, // bypass when length of copy is zero
4985 bcopy_path = 5, // copy primitive array by 64-bit blocks
4986 PATH_LIMIT = 6
4987 };
4988 RegionNode* result_region = new RegionNode(PATH_LIMIT);
4989 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
4990 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
4991 record_for_igvn(result_region);
4992 _gvn.set_type_bottom(result_i_o);
4993 _gvn.set_type_bottom(result_memory);
4994 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4995
4996 // The slow_control path:
4997 Node* slow_control;
4998 Node* slow_i_o = i_o();
4999 Node* slow_mem = memory(adr_type);
5000 debug_only(slow_control = (Node*) badAddress);
5001
5002 // Checked control path:
5003 Node* checked_control = top();
5004 Node* checked_mem = NULL;
5005 Node* checked_i_o = NULL;
5006 Node* checked_value = NULL;
5007
5008 if (basic_elem_type == T_CONFLICT) {
5009 assert(!dest_uninitialized, "");
5010 Node* cv = generate_generic_arraycopy(adr_type,
5011 src, src_offset, dest, dest_offset,
5012 copy_length, dest_uninitialized);
5013 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5014 checked_control = control();
5015 checked_i_o = i_o();
5016 checked_mem = memory(adr_type);
5017 checked_value = cv;
5018 set_control(top()); // no fast path
5019 }
5020
5021 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5022 if (not_pos != NULL) {
5023 PreserveJVMState pjvms(this);
5024 set_control(not_pos);
5025
5026 // (6) length must not be negative.
5027 if (!length_never_negative) {
5028 generate_negative_guard(copy_length, slow_region);
5029 }
5030
5031 // copy_length is 0.
5032 if (!stopped() && dest_uninitialized) {
5033 Node* dest_length = alloc->in(AllocateNode::ALength);
5034 if (copy_length->eqv_uncast(dest_length)
5035 || _gvn.find_int_con(dest_length, 1) <= 0) {
5036 // There is no zeroing to do. No need for a secondary raw memory barrier.
5037 } else {
5038 // Clear the whole thing since there are no source elements to copy.
5039 generate_clear_array(adr_type, dest, basic_elem_type,
5040 intcon(0), NULL,
5041 alloc->in(AllocateNode::AllocSize));
5042 // Use a secondary InitializeNode as raw memory barrier.
5043 // Currently it is needed only on this path since other
5044 // paths have stub or runtime calls as raw memory barriers.
5045 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5046 Compile::AliasIdxRaw,
5047 top())->as_Initialize();
5048 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5049 }
5050 }
5051
5052 // Present the results of the fast call.
5053 result_region->init_req(zero_path, control());
5054 result_i_o ->init_req(zero_path, i_o());
5055 result_memory->init_req(zero_path, memory(adr_type));
5056 }
5057
5058 if (!stopped() && dest_uninitialized) {
5059 // We have to initialize the *uncopied* part of the array to zero.
5060 // The copy destination is the slice dest[off..off+len]. The other slices
5061 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5062 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5063 Node* dest_length = alloc->in(AllocateNode::ALength);
5064 Node* dest_tail = _gvn.transform(new AddINode(dest_offset, copy_length));
5065
5066 // If there is a head section that needs zeroing, do it now.
5067 if (find_int_con(dest_offset, -1) != 0) {
5068 generate_clear_array(adr_type, dest, basic_elem_type,
5069 intcon(0), dest_offset,
5070 NULL);
5071 }
5072
5073 // Next, perform a dynamic check on the tail length.
5074 // It is often zero, and we can win big if we prove this.
5075 // There are two wins: Avoid generating the ClearArray
5076 // with its attendant messy index arithmetic, and upgrade
5077 // the copy to a more hardware-friendly word size of 64 bits.
5078 Node* tail_ctl = NULL;
5079 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5080 Node* cmp_lt = _gvn.transform(new CmpINode(dest_tail, dest_length));
5081 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
5082 tail_ctl = generate_slow_guard(bol_lt, NULL);
5083 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5084 }
5085
5086 // At this point, let's assume there is no tail.
5087 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5088 // There is no tail. Try an upgrade to a 64-bit copy.
5089 bool didit = false;
5090 { PreserveJVMState pjvms(this);
5091 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5092 src, src_offset, dest, dest_offset,
5093 dest_size, dest_uninitialized);
5094 if (didit) {
5095 // Present the results of the block-copying fast call.
5096 result_region->init_req(bcopy_path, control());
5097 result_i_o ->init_req(bcopy_path, i_o());
5098 result_memory->init_req(bcopy_path, memory(adr_type));
5099 }
5100 }
5101 if (didit)
5102 set_control(top()); // no regular fast path
5103 }
5104
5105 // Clear the tail, if any.
5106 if (tail_ctl != NULL) {
5107 Node* notail_ctl = stopped() ? NULL : control();
5108 set_control(tail_ctl);
5109 if (notail_ctl == NULL) {
5110 generate_clear_array(adr_type, dest, basic_elem_type,
5111 dest_tail, NULL,
5112 dest_size);
5113 } else {
5114 // Make a local merge.
5115 Node* done_ctl = new RegionNode(3);
5116 Node* done_mem = new PhiNode(done_ctl, Type::MEMORY, adr_type);
5117 done_ctl->init_req(1, notail_ctl);
5118 done_mem->init_req(1, memory(adr_type));
5119 generate_clear_array(adr_type, dest, basic_elem_type,
5120 dest_tail, NULL,
5121 dest_size);
5122 done_ctl->init_req(2, control());
5123 done_mem->init_req(2, memory(adr_type));
5124 set_control( _gvn.transform(done_ctl));
5125 set_memory( _gvn.transform(done_mem), adr_type );
5126 }
5127 }
5128 }
5129
5130 BasicType copy_type = basic_elem_type;
5131 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5132 if (!stopped() && copy_type == T_OBJECT) {
5133 // If src and dest have compatible element types, we can copy bits.
5134 // Types S[] and D[] are compatible if D is a supertype of S.
5135 //
5136 // If they are not, we will use checked_oop_disjoint_arraycopy,
5137 // which performs a fast optimistic per-oop check, and backs off
5138 // further to JVM_ArrayCopy on the first per-oop check that fails.
5139 // (Actually, we don't move raw bits only; the GC requires card marks.)
5140
5141 // Get the Klass* for both src and dest
5142 Node* src_klass = load_object_klass(src);
5143 Node* dest_klass = load_object_klass(dest);
5144
5145 // Generate the subtype check.
5146 // This might fold up statically, or then again it might not.
5147 //
5148 // Non-static example: Copying List<String>.elements to a new String[].
5149 // The backing store for a List<String> is always an Object[],
5150 // but its elements are always type String, if the generic types
5151 // are correct at the source level.
5152 //
5153 // Test S[] against D[], not S against D, because (probably)
5154 // the secondary supertype cache is less busy for S[] than S.
5155 // This usually only matters when D is an interface.
5156 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5157 // Plug failing path into checked_oop_disjoint_arraycopy
5158 if (not_subtype_ctrl != top()) {
5159 PreserveJVMState pjvms(this);
5160 set_control(not_subtype_ctrl);
5161 // (At this point we can assume disjoint_bases, since types differ.)
5162 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5163 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5164 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5165 Node* dest_elem_klass = _gvn.transform(n1);
5166 Node* cv = generate_checkcast_arraycopy(adr_type,
5167 dest_elem_klass,
5168 src, src_offset, dest, dest_offset,
5169 ConvI2X(copy_length), dest_uninitialized);
5170 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5171 checked_control = control();
5172 checked_i_o = i_o();
5173 checked_mem = memory(adr_type);
5174 checked_value = cv;
5175 }
5176 // At this point we know we do not need type checks on oop stores.
5177
5178 // Let's see if we need card marks:
5179 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5180 // If we do not need card marks, copy using the jint or jlong stub.
5181 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5182 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5183 "sizes agree");
5184 }
5185 }
5186
5187 if (!stopped()) {
5188 // Generate the fast path, if possible.
5189 PreserveJVMState pjvms(this);
5190 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5191 src, src_offset, dest, dest_offset,
5192 ConvI2X(copy_length), dest_uninitialized);
5193
5194 // Present the results of the fast call.
5195 result_region->init_req(fast_path, control());
5196 result_i_o ->init_req(fast_path, i_o());
5197 result_memory->init_req(fast_path, memory(adr_type));
5198 }
5199
5200 // Here are all the slow paths up to this point, in one bundle:
5201 slow_control = top();
5202 if (slow_region != NULL)
5203 slow_control = _gvn.transform(slow_region);
5204 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5205
5206 set_control(checked_control);
5207 if (!stopped()) {
5208 // Clean up after the checked call.
5209 // The returned value is either 0 or -1^K,
5210 // where K = number of partially transferred array elements.
5211 Node* cmp = _gvn.transform(new CmpINode(checked_value, intcon(0)));
5212 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
5213 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5214
5215 // If it is 0, we are done, so transfer to the end.
5216 Node* checks_done = _gvn.transform(new IfTrueNode(iff));
5217 result_region->init_req(checked_path, checks_done);
5218 result_i_o ->init_req(checked_path, checked_i_o);
5219 result_memory->init_req(checked_path, checked_mem);
5220
5221 // If it is not zero, merge into the slow call.
5222 set_control( _gvn.transform(new IfFalseNode(iff) ));
5223 RegionNode* slow_reg2 = new RegionNode(3);
5224 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
5225 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
5226 record_for_igvn(slow_reg2);
5227 slow_reg2 ->init_req(1, slow_control);
5228 slow_i_o2 ->init_req(1, slow_i_o);
5229 slow_mem2 ->init_req(1, slow_mem);
5230 slow_reg2 ->init_req(2, control());
5231 slow_i_o2 ->init_req(2, checked_i_o);
5232 slow_mem2 ->init_req(2, checked_mem);
5233
5234 slow_control = _gvn.transform(slow_reg2);
5235 slow_i_o = _gvn.transform(slow_i_o2);
5236 slow_mem = _gvn.transform(slow_mem2);
5237
5238 if (alloc != NULL) {
5239 // We'll restart from the very beginning, after zeroing the whole thing.
5240 // This can cause double writes, but that's OK since dest is brand new.
5241 // So we ignore the low 31 bits of the value returned from the stub.
5242 } else {
5243 // We must continue the copy exactly where it failed, or else
5244 // another thread might see the wrong number of writes to dest.
5245 Node* checked_offset = _gvn.transform(new XorINode(checked_value, intcon(-1)));
5246 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
5247 slow_offset->init_req(1, intcon(0));
5248 slow_offset->init_req(2, checked_offset);
5249 slow_offset = _gvn.transform(slow_offset);
5250
5251 // Adjust the arguments by the conditionally incoming offset.
5252 Node* src_off_plus = _gvn.transform(new AddINode(src_offset, slow_offset));
5253 Node* dest_off_plus = _gvn.transform(new AddINode(dest_offset, slow_offset));
5254 Node* length_minus = _gvn.transform(new SubINode(copy_length, slow_offset));
5255
5256 // Tweak the node variables to adjust the code produced below:
5257 src_offset = src_off_plus;
5258 dest_offset = dest_off_plus;
5259 copy_length = length_minus;
5260 }
5261 }
5262
5263 set_control(slow_control);
5264 if (!stopped()) {
5265 // Generate the slow path, if needed.
5266 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5267
5268 set_memory(slow_mem, adr_type);
5269 set_i_o(slow_i_o);
5270
5271 if (dest_uninitialized) {
5272 generate_clear_array(adr_type, dest, basic_elem_type,
5273 intcon(0), NULL,
5274 alloc->in(AllocateNode::AllocSize));
5275 }
5276
5277 generate_slow_arraycopy(adr_type,
5278 src, src_offset, dest, dest_offset,
5279 copy_length, /*dest_uninitialized*/false);
5280
5281 result_region->init_req(slow_call_path, control());
5282 result_i_o ->init_req(slow_call_path, i_o());
5283 result_memory->init_req(slow_call_path, memory(adr_type));
5284 }
5285
5286 // Remove unused edges.
5287 for (uint i = 1; i < result_region->req(); i++) {
5288 if (result_region->in(i) == NULL)
5289 result_region->init_req(i, top());
5290 }
5291
5292 // Finished; return the combined state.
5293 set_control( _gvn.transform(result_region));
5294 set_i_o( _gvn.transform(result_i_o) );
5295 set_memory( _gvn.transform(result_memory), adr_type );
5296
5297 // The memory edges above are precise in order to model effects around
5298 // array copies accurately to allow value numbering of field loads around
5299 // arraycopy. Such field loads, both before and after, are common in Java
5300 // collections and similar classes involving header/array data structures.
5301 //
5302 // But with low number of register or when some registers are used or killed
5303 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5304 // The next memory barrier is added to avoid it. If the arraycopy can be
5305 // optimized away (which it can, sometimes) then we can manually remove
5306 // the membar also.
5307 //
5308 // Do not let reads from the cloned object float above the arraycopy.
5309 if (alloc != NULL) {
5310 // Do not let stores that initialize this object be reordered with
5311 // a subsequent store that would make this object accessible by
5312 // other threads.
5313 // Record what AllocateNode this StoreStore protects so that
5314 // escape analysis can go from the MemBarStoreStoreNode to the
5315 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5316 // based on the escape status of the AllocateNode.
5317 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5318 } else if (InsertMemBarAfterArraycopy)
5319 insert_mem_bar(Op_MemBarCPUOrder);
5320 }
5321
5322
5323 // Helper function which determines if an arraycopy immediately follows
5324 // an allocation, with no intervening tests or other escapes for the object.
5325 AllocateArrayNode*
5326 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5327 RegionNode* slow_region) {
5328 if (stopped()) return NULL; // no fast path
5329 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5330
5331 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5332 if (alloc == NULL) return NULL;
5333
5334 Node* rawmem = memory(Compile::AliasIdxRaw);
5335 // Is the allocation's memory state untouched?
5336 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5337 // Bail out if there have been raw-memory effects since the allocation.
5338 // (Example: There might have been a call or safepoint.)
5339 return NULL;
5340 }
5341 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5342 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5343 return NULL;
5344 }
5345
5346 // There must be no unexpected observers of this allocation.
5347 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5348 Node* obs = ptr->fast_out(i);
5349 if (obs != this->map()) {
5350 return NULL;
5351 }
5352 }
5353
5354 // This arraycopy must unconditionally follow the allocation of the ptr.
5355 Node* alloc_ctl = ptr->in(0);
5356 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5357
5358 Node* ctl = control();
5359 while (ctl != alloc_ctl) {
5360 // There may be guards which feed into the slow_region.
5361 // Any other control flow means that we might not get a chance
5362 // to finish initializing the allocated object.
5363 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5364 IfNode* iff = ctl->in(0)->as_If();
5365 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5366 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5367 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5368 ctl = iff->in(0); // This test feeds the known slow_region.
5369 continue;
5370 }
5371 // One more try: Various low-level checks bottom out in
5372 // uncommon traps. If the debug-info of the trap omits
5373 // any reference to the allocation, as we've already
5374 // observed, then there can be no objection to the trap.
5375 bool found_trap = false;
5376 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5377 Node* obs = not_ctl->fast_out(j);
5378 if (obs->in(0) == not_ctl && obs->is_Call() &&
5379 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5380 found_trap = true; break;
5381 }
5382 }
5383 if (found_trap) {
5384 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5385 continue;
5386 }
5387 }
5388 return NULL;
5389 }
5390
5391 // If we get this far, we have an allocation which immediately
5392 // precedes the arraycopy, and we can take over zeroing the new object.
5393 // The arraycopy will finish the initialization, and provide
5394 // a new control state to which we will anchor the destination pointer.
5395
5396 return alloc;
5397 }
5398
5399 // Helper for initialization of arrays, creating a ClearArray.
5400 // It writes zero bits in [start..end), within the body of an array object.
5401 // The memory effects are all chained onto the 'adr_type' alias category.
5402 //
5403 // Since the object is otherwise uninitialized, we are free
5404 // to put a little "slop" around the edges of the cleared area,
5405 // as long as it does not go back into the array's header,
5406 // or beyond the array end within the heap.
5407 //
5408 // The lower edge can be rounded down to the nearest jint and the
5409 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5410 //
5411 // Arguments:
5412 // adr_type memory slice where writes are generated
5413 // dest oop of the destination array
5414 // basic_elem_type element type of the destination
5415 // slice_idx array index of first element to store
5416 // slice_len number of elements to store (or NULL)
5417 // dest_size total size in bytes of the array object
5418 //
5419 // Exactly one of slice_len or dest_size must be non-NULL.
5420 // If dest_size is non-NULL, zeroing extends to the end of the object.
5421 // If slice_len is non-NULL, the slice_idx value must be a constant.
5422 void
5423 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5424 Node* dest,
5425 BasicType basic_elem_type,
5426 Node* slice_idx,
5427 Node* slice_len,
5428 Node* dest_size) {
5429 // one or the other but not both of slice_len and dest_size:
5430 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5431 if (slice_len == NULL) slice_len = top();
5432 if (dest_size == NULL) dest_size = top();
5433
5434 // operate on this memory slice:
5435 Node* mem = memory(adr_type); // memory slice to operate on
5436
5437 // scaling and rounding of indexes:
5438 int scale = exact_log2(type2aelembytes(basic_elem_type));
5439 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5440 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5441 int bump_bit = (-1 << scale) & BytesPerInt;
5442
5443 // determine constant starts and ends
5444 const intptr_t BIG_NEG = -128;
5445 assert(BIG_NEG + 2*abase < 0, "neg enough");
5446 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5447 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5448 if (slice_len_con == 0) {
5449 return; // nothing to do here
5450 }
5451 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5452 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5453 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5454 assert(end_con < 0, "not two cons");
5455 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5456 BytesPerLong);
5457 }
5458
5459 if (start_con >= 0 && end_con >= 0) {
5460 // Constant start and end. Simple.
5461 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5462 start_con, end_con, &_gvn);
5463 } else if (start_con >= 0 && dest_size != top()) {
5464 // Constant start, pre-rounded end after the tail of the array.
5465 Node* end = dest_size;
5466 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5467 start_con, end, &_gvn);
5468 } else if (start_con >= 0 && slice_len != top()) {
5469 // Constant start, non-constant end. End needs rounding up.
5470 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5471 intptr_t end_base = abase + (slice_idx_con << scale);
5472 int end_round = (-1 << scale) & (BytesPerLong - 1);
5473 Node* end = ConvI2X(slice_len);
5474 if (scale != 0)
5475 end = _gvn.transform(new LShiftXNode(end, intcon(scale) ));
5476 end_base += end_round;
5477 end = _gvn.transform(new AddXNode(end, MakeConX(end_base)));
5478 end = _gvn.transform(new AndXNode(end, MakeConX(~end_round)));
5479 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5480 start_con, end, &_gvn);
5481 } else if (start_con < 0 && dest_size != top()) {
5482 // Non-constant start, pre-rounded end after the tail of the array.
5483 // This is almost certainly a "round-to-end" operation.
5484 Node* start = slice_idx;
5485 start = ConvI2X(start);
5486 if (scale != 0)
5487 start = _gvn.transform(new LShiftXNode( start, intcon(scale) ));
5488 start = _gvn.transform(new AddXNode(start, MakeConX(abase)));
5489 if ((bump_bit | clear_low) != 0) {
5490 int to_clear = (bump_bit | clear_low);
5491 // Align up mod 8, then store a jint zero unconditionally
5492 // just before the mod-8 boundary.
5493 if (((abase + bump_bit) & ~to_clear) - bump_bit
5494 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5495 bump_bit = 0;
5496 assert((abase & to_clear) == 0, "array base must be long-aligned");
5497 } else {
5498 // Bump 'start' up to (or past) the next jint boundary:
5499 start = _gvn.transform(new AddXNode(start, MakeConX(bump_bit)));
5500 assert((abase & clear_low) == 0, "array base must be int-aligned");
5501 }
5502 // Round bumped 'start' down to jlong boundary in body of array.
5503 start = _gvn.transform(new AndXNode(start, MakeConX(~to_clear)));
5504 if (bump_bit != 0) {
5505 // Store a zero to the immediately preceding jint:
5506 Node* x1 = _gvn.transform(new AddXNode(start, MakeConX(-bump_bit)));
5507 Node* p1 = basic_plus_adr(dest, x1);
5508 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5509 mem = _gvn.transform(mem);
5510 }
5511 }
5512 Node* end = dest_size; // pre-rounded
5513 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5514 start, end, &_gvn);
5515 } else {
5516 // Non-constant start, unrounded non-constant end.
5517 // (Nobody zeroes a random midsection of an array using this routine.)
5518 ShouldNotReachHere(); // fix caller
5519 }
5520
5521 // Done.
5522 set_memory(mem, adr_type);
5523 }
5524
5525
5526 bool
5527 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5528 BasicType basic_elem_type,
5529 AllocateNode* alloc,
5530 Node* src, Node* src_offset,
5531 Node* dest, Node* dest_offset,
5532 Node* dest_size, bool dest_uninitialized) {
5533 // See if there is an advantage from block transfer.
5534 int scale = exact_log2(type2aelembytes(basic_elem_type));
5535 if (scale >= LogBytesPerLong)
5536 return false; // it is already a block transfer
5537
5538 // Look at the alignment of the starting offsets.
5539 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5540
5541 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5542 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5543 if (src_off_con < 0 || dest_off_con < 0)
5544 // At present, we can only understand constants.
5545 return false;
5546
5547 intptr_t src_off = abase + (src_off_con << scale);
5548 intptr_t dest_off = abase + (dest_off_con << scale);
5549
5550 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5551 // Non-aligned; too bad.
5552 // One more chance: Pick off an initial 32-bit word.
5553 // This is a common case, since abase can be odd mod 8.
5554 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5555 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5556 Node* sptr = basic_plus_adr(src, src_off);
5557 Node* dptr = basic_plus_adr(dest, dest_off);
5558 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5559 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5560 src_off += BytesPerInt;
5561 dest_off += BytesPerInt;
5562 } else {
5563 return false;
5564 }
5565 }
5566 assert(src_off % BytesPerLong == 0, "");
5567 assert(dest_off % BytesPerLong == 0, "");
5568
5569 // Do this copy by giant steps.
5570 Node* sptr = basic_plus_adr(src, src_off);
5571 Node* dptr = basic_plus_adr(dest, dest_off);
5572 Node* countx = dest_size;
5573 countx = _gvn.transform(new SubXNode(countx, MakeConX(dest_off)));
5574 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong)));
5575
5576 bool disjoint_bases = true; // since alloc != NULL
5577 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5578 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5579
5580 return true;
5581 }
5582
5583
5584 // Helper function; generates code for the slow case.
5585 // We make a call to a runtime method which emulates the native method,
5586 // but without the native wrapper overhead.
5587 void
5588 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5589 Node* src, Node* src_offset,
5590 Node* dest, Node* dest_offset,
5591 Node* copy_length, bool dest_uninitialized) {
5592 assert(!dest_uninitialized, "Invariant");
5593 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5594 OptoRuntime::slow_arraycopy_Type(),
5595 OptoRuntime::slow_arraycopy_Java(),
5596 "slow_arraycopy", adr_type,
5597 src, src_offset, dest, dest_offset,
5598 copy_length);
5599
5600 // Handle exceptions thrown by this fellow:
5601 make_slow_call_ex(call, env()->Throwable_klass(), false);
5602 }
5603
5604 // Helper function; generates code for cases requiring runtime checks.
5605 Node*
5606 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5607 Node* dest_elem_klass,
5608 Node* src, Node* src_offset,
5609 Node* dest, Node* dest_offset,
5610 Node* copy_length, bool dest_uninitialized) {
5611 if (stopped()) return NULL;
5612
5613 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5614 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5615 return NULL;
5616 }
5617
5618 // Pick out the parameters required to perform a store-check
5619 // for the target array. This is an optimistic check. It will
5620 // look in each non-null element's class, at the desired klass's
5621 // super_check_offset, for the desired klass.
5622 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5623 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5624 Node* n3 = new LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5625 Node* check_offset = ConvI2X(_gvn.transform(n3));
5626 Node* check_value = dest_elem_klass;
5627
5628 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5629 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5630
5631 // (We know the arrays are never conjoint, because their types differ.)
5632 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5633 OptoRuntime::checkcast_arraycopy_Type(),
5634 copyfunc_addr, "checkcast_arraycopy", adr_type,
5635 // five arguments, of which two are
5636 // intptr_t (jlong in LP64)
5637 src_start, dest_start,
5638 copy_length XTOP,
5639 check_offset XTOP,
5640 check_value);
5641
5642 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5643 }
5644
5645
5646 // Helper function; generates code for cases requiring runtime checks.
5647 Node*
5648 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5649 Node* src, Node* src_offset,
5650 Node* dest, Node* dest_offset,
5651 Node* copy_length, bool dest_uninitialized) {
5652 assert(!dest_uninitialized, "Invariant");
5653 if (stopped()) return NULL;
5654 address copyfunc_addr = StubRoutines::generic_arraycopy();
5655 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5656 return NULL;
5657 }
5658
5659 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5660 OptoRuntime::generic_arraycopy_Type(),
5661 copyfunc_addr, "generic_arraycopy", adr_type,
5662 src, src_offset, dest, dest_offset, copy_length);
5663
5664 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5665 }
5666
5667 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5668 void
5669 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5670 BasicType basic_elem_type,
5671 bool disjoint_bases,
5672 Node* src, Node* src_offset,
5673 Node* dest, Node* dest_offset,
5674 Node* copy_length, bool dest_uninitialized) {
5675 if (stopped()) return; // nothing to do
5676
5677 Node* src_start = src;
5678 Node* dest_start = dest;
5679 if (src_offset != NULL || dest_offset != NULL) {
5680 assert(src_offset != NULL && dest_offset != NULL, "");
5681 src_start = array_element_address(src, src_offset, basic_elem_type);
5682 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5683 }
5684
5685 // Figure out which arraycopy runtime method to call.
5686 const char* copyfunc_name = "arraycopy";
5687 address copyfunc_addr =
5688 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5689 disjoint_bases, copyfunc_name, dest_uninitialized);
5690
5691 // Call it. Note that the count_ix value is not scaled to a byte-size.
5692 make_runtime_call(RC_LEAF|RC_NO_FP,
5693 OptoRuntime::fast_arraycopy_Type(),
5694 copyfunc_addr, copyfunc_name, adr_type,
5695 src_start, dest_start, copy_length XTOP);
5696 }
5697
5698 //-------------inline_encodeISOArray-----------------------------------
5699 // encode char[] to byte[] in ISO_8859_1
5700 bool LibraryCallKit::inline_encodeISOArray() {
5701 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5702 // no receiver since it is static method
5703 Node *src = argument(0);
5704 Node *src_offset = argument(1);
5705 Node *dst = argument(2);
5706 Node *dst_offset = argument(3);
5707 Node *length = argument(4);
5708
5709 const Type* src_type = src->Value(&_gvn);
5710 const Type* dst_type = dst->Value(&_gvn);
5711 const TypeAryPtr* top_src = src_type->isa_aryptr();
5712 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5713 if (top_src == NULL || top_src->klass() == NULL ||
5714 top_dest == NULL || top_dest->klass() == NULL) {
5715 // failed array check
5716 return false;
5717 }
5718
5719 // Figure out the size and type of the elements we will be copying.
5720 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5721 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5722 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5723 return false;
5724 }
5725 Node* src_start = array_element_address(src, src_offset, src_elem);
5726 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5727 // 'src_start' points to src array + scaled offset
5728 // 'dst_start' points to dst array + scaled offset
5729
5730 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5731 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5732 enc = _gvn.transform(enc);
5733 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5734 set_memory(res_mem, mtype);
5735 set_result(enc);
5736 return true;
5737 }
5738
5739 /**
5740 * Calculate CRC32 for byte.
5741 * int java.util.zip.CRC32.update(int crc, int b)
5742 */
5743 bool LibraryCallKit::inline_updateCRC32() {
5744 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5745 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5746 // no receiver since it is static method
5747 Node* crc = argument(0); // type: int
5748 Node* b = argument(1); // type: int
5749
5750 /*
5751 * int c = ~ crc;
5752 * b = timesXtoThe32[(b ^ c) & 0xFF];
5753 * b = b ^ (c >>> 8);
5754 * crc = ~b;
5755 */
5756
5757 Node* M1 = intcon(-1);
5758 crc = _gvn.transform(new XorINode(crc, M1));
5759 Node* result = _gvn.transform(new XorINode(crc, b));
5760 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5761
5762 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5763 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5764 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5765 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5766
5767 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5768 result = _gvn.transform(new XorINode(crc, result));
5769 result = _gvn.transform(new XorINode(result, M1));
5770 set_result(result);
5771 return true;
5772 }
5773
5774 /**
5775 * Calculate CRC32 for byte[] array.
5776 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5777 */
5778 bool LibraryCallKit::inline_updateBytesCRC32() {
5779 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5780 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5781 // no receiver since it is static method
5782 Node* crc = argument(0); // type: int
5783 Node* src = argument(1); // type: oop
5784 Node* offset = argument(2); // type: int
5785 Node* length = argument(3); // type: int
5786
5787 const Type* src_type = src->Value(&_gvn);
5788 const TypeAryPtr* top_src = src_type->isa_aryptr();
5789 if (top_src == NULL || top_src->klass() == NULL) {
5790 // failed array check
5791 return false;
5792 }
5793
5794 // Figure out the size and type of the elements we will be copying.
5795 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5796 if (src_elem != T_BYTE) {
5797 return false;
5798 }
5799
5800 // 'src_start' points to src array + scaled offset
5801 Node* src_start = array_element_address(src, offset, src_elem);
5802
5803 // We assume that range check is done by caller.
5804 // TODO: generate range check (offset+length < src.length) in debug VM.
5805
5806 // Call the stub.
5807 address stubAddr = StubRoutines::updateBytesCRC32();
5808 const char *stubName = "updateBytesCRC32";
5809
5810 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5811 stubAddr, stubName, TypePtr::BOTTOM,
5812 crc, src_start, length);
5813 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5814 set_result(result);
5815 return true;
5816 }
5817
5818 /**
5819 * Calculate CRC32 for ByteBuffer.
5820 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5821 */
5822 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5823 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5824 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5825 // no receiver since it is static method
5826 Node* crc = argument(0); // type: int
5827 Node* src = argument(1); // type: long
5828 Node* offset = argument(3); // type: int
5829 Node* length = argument(4); // type: int
5830
5831 src = ConvL2X(src); // adjust Java long to machine word
5832 Node* base = _gvn.transform(new CastX2PNode(src));
5833 offset = ConvI2X(offset);
5834
5835 // 'src_start' points to src array + scaled offset
5836 Node* src_start = basic_plus_adr(top(), base, offset);
5837
5838 // Call the stub.
5839 address stubAddr = StubRoutines::updateBytesCRC32();
5840 const char *stubName = "updateBytesCRC32";
5841
5842 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5843 stubAddr, stubName, TypePtr::BOTTOM,
5844 crc, src_start, length);
5845 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5846 set_result(result);
5847 return true;
5848 }
5849
5850 //----------------------------inline_reference_get----------------------------
5851 // public T java.lang.ref.Reference.get();
5852 bool LibraryCallKit::inline_reference_get() {
5853 const int referent_offset = java_lang_ref_Reference::referent_offset;
5854 guarantee(referent_offset > 0, "should have already been set");
5855
5856 // Get the argument:
5857 Node* reference_obj = null_check_receiver();
5858 if (stopped()) return true;
5859
5860 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5861
5862 ciInstanceKlass* klass = env()->Object_klass();
5863 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5864
5865 Node* no_ctrl = NULL;
5866 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5867
5868 // Use the pre-barrier to record the value in the referent field
5869 pre_barrier(false /* do_load */,
5870 control(),
5871 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5872 result /* pre_val */,
5873 T_OBJECT);
5874
5875 // Add memory barrier to prevent commoning reads from this field
5876 // across safepoint since GC can change its value.
5877 insert_mem_bar(Op_MemBarCPUOrder);
5878
5879 set_result(result);
5880 return true;
5881 }
5882
5883
5884 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5885 bool is_exact=true, bool is_static=false) {
5886
5887 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5888 assert(tinst != NULL, "obj is null");
5889 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5890 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5891
5892 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5893 ciSymbol::make(fieldTypeString),
5894 is_static);
5895 if (field == NULL) return (Node *) NULL;
5896 assert (field != NULL, "undefined field");
5897
5898 // Next code copied from Parse::do_get_xxx():
5899
5900 // Compute address and memory type.
5901 int offset = field->offset_in_bytes();
5902 bool is_vol = field->is_volatile();
5903 ciType* field_klass = field->type();
5904 assert(field_klass->is_loaded(), "should be loaded");
5905 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5906 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5907 BasicType bt = field->layout_type();
5908
5909 // Build the resultant type of the load
5910 const Type *type;
5911 if (bt == T_OBJECT) {
5912 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5913 } else {
5914 type = Type::get_const_basic_type(bt);
5915 }
5916
5917 // Build the load.
5918 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5919 return loadedField;
5920 }
5921
5922
5923 //------------------------------inline_aescrypt_Block-----------------------
5924 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5925 address stubAddr;
5926 const char *stubName;
5927 assert(UseAES, "need AES instruction support");
5928
5929 switch(id) {
5930 case vmIntrinsics::_aescrypt_encryptBlock:
5931 stubAddr = StubRoutines::aescrypt_encryptBlock();
5932 stubName = "aescrypt_encryptBlock";
5933 break;
5934 case vmIntrinsics::_aescrypt_decryptBlock:
5935 stubAddr = StubRoutines::aescrypt_decryptBlock();
5936 stubName = "aescrypt_decryptBlock";
5937 break;
5938 }
5939 if (stubAddr == NULL) return false;
5940
5941 Node* aescrypt_object = argument(0);
5942 Node* src = argument(1);
5943 Node* src_offset = argument(2);
5944 Node* dest = argument(3);
5945 Node* dest_offset = argument(4);
5946
5947 // (1) src and dest are arrays.
5948 const Type* src_type = src->Value(&_gvn);
5949 const Type* dest_type = dest->Value(&_gvn);
5950 const TypeAryPtr* top_src = src_type->isa_aryptr();
5951 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5952 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5953
5954 // for the quick and dirty code we will skip all the checks.
5955 // we are just trying to get the call to be generated.
5956 Node* src_start = src;
5957 Node* dest_start = dest;
5958 if (src_offset != NULL || dest_offset != NULL) {
5959 assert(src_offset != NULL && dest_offset != NULL, "");
5960 src_start = array_element_address(src, src_offset, T_BYTE);
5961 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5962 }
5963
5964 // now need to get the start of its expanded key array
5965 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5966 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5967 if (k_start == NULL) return false;
5968
5969 if (Matcher::pass_original_key_for_aes()) {
5970 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5971 // compatibility issues between Java key expansion and SPARC crypto instructions
5972 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5973 if (original_k_start == NULL) return false;
5974
5975 // Call the stub.
5976 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5977 stubAddr, stubName, TypePtr::BOTTOM,
5978 src_start, dest_start, k_start, original_k_start);
5979 } else {
5980 // Call the stub.
5981 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5982 stubAddr, stubName, TypePtr::BOTTOM,
5983 src_start, dest_start, k_start);
5984 }
5985
5986 return true;
5987 }
5988
5989 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5990 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5991 address stubAddr;
5992 const char *stubName;
5993
5994 assert(UseAES, "need AES instruction support");
5995
5996 switch(id) {
5997 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5998 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5999 stubName = "cipherBlockChaining_encryptAESCrypt";
6000 break;
6001 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6002 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6003 stubName = "cipherBlockChaining_decryptAESCrypt";
6004 break;
6005 }
6006 if (stubAddr == NULL) return false;
6007
6008 Node* cipherBlockChaining_object = argument(0);
6009 Node* src = argument(1);
6010 Node* src_offset = argument(2);
6011 Node* len = argument(3);
6012 Node* dest = argument(4);
6013 Node* dest_offset = argument(5);
6014
6015 // (1) src and dest are arrays.
6016 const Type* src_type = src->Value(&_gvn);
6017 const Type* dest_type = dest->Value(&_gvn);
6018 const TypeAryPtr* top_src = src_type->isa_aryptr();
6019 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6020 assert (top_src != NULL && top_src->klass() != NULL
6021 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6022
6023 // checks are the responsibility of the caller
6024 Node* src_start = src;
6025 Node* dest_start = dest;
6026 if (src_offset != NULL || dest_offset != NULL) {
6027 assert(src_offset != NULL && dest_offset != NULL, "");
6028 src_start = array_element_address(src, src_offset, T_BYTE);
6029 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6030 }
6031
6032 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6033 // (because of the predicated logic executed earlier).
6034 // so we cast it here safely.
6035 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6036
6037 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6038 if (embeddedCipherObj == NULL) return false;
6039
6040 // cast it to what we know it will be at runtime
6041 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6042 assert(tinst != NULL, "CBC obj is null");
6043 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6044 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6045 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6046
6047 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6048 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6049 const TypeOopPtr* xtype = aklass->as_instance_type();
6050 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6051 aescrypt_object = _gvn.transform(aescrypt_object);
6052
6053 // we need to get the start of the aescrypt_object's expanded key array
6054 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6055 if (k_start == NULL) return false;
6056
6057 // similarly, get the start address of the r vector
6058 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6059 if (objRvec == NULL) return false;
6060 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6061
6062 Node* cbcCrypt;
6063 if (Matcher::pass_original_key_for_aes()) {
6064 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6065 // compatibility issues between Java key expansion and SPARC crypto instructions
6066 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6067 if (original_k_start == NULL) return false;
6068
6069 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6070 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6071 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6072 stubAddr, stubName, TypePtr::BOTTOM,
6073 src_start, dest_start, k_start, r_start, len, original_k_start);
6074 } else {
6075 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6076 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6077 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6078 stubAddr, stubName, TypePtr::BOTTOM,
6079 src_start, dest_start, k_start, r_start, len);
6080 }
6081
6082 // return cipher length (int)
6083 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6084 set_result(retvalue);
6085 return true;
6086 }
6087
6088 //------------------------------get_key_start_from_aescrypt_object-----------------------
6089 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6090 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6091 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6092 if (objAESCryptKey == NULL) return (Node *) NULL;
6093
6094 // now have the array, need to get the start address of the K array
6095 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6096 return k_start;
6097 }
6098
6099 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6100 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6101 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6102 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6103 if (objAESCryptKey == NULL) return (Node *) NULL;
6104
6105 // now have the array, need to get the start address of the lastKey array
6106 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6107 return original_k_start;
6108 }
6109
6110 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6111 // Return node representing slow path of predicate check.
6112 // the pseudo code we want to emulate with this predicate is:
6113 // for encryption:
6114 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6115 // for decryption:
6116 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6117 // note cipher==plain is more conservative than the original java code but that's OK
6118 //
6119 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6120 // The receiver was checked for NULL already.
6121 Node* objCBC = argument(0);
6122
6123 // Load embeddedCipher field of CipherBlockChaining object.
6124 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6125
6126 // get AESCrypt klass for instanceOf check
6127 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6128 // will have same classloader as CipherBlockChaining object
6129 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6130 assert(tinst != NULL, "CBCobj is null");
6131 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6132
6133 // we want to do an instanceof comparison against the AESCrypt class
6134 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6135 if (!klass_AESCrypt->is_loaded()) {
6136 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6137 Node* ctrl = control();
6138 set_control(top()); // no regular fast path
6139 return ctrl;
6140 }
6141 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6142
6143 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6144 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6145 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6146
6147 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6148
6149 // for encryption, we are done
6150 if (!decrypting)
6151 return instof_false; // even if it is NULL
6152
6153 // for decryption, we need to add a further check to avoid
6154 // taking the intrinsic path when cipher and plain are the same
6155 // see the original java code for why.
6156 RegionNode* region = new RegionNode(3);
6157 region->init_req(1, instof_false);
6158 Node* src = argument(1);
6159 Node* dest = argument(4);
6160 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6161 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6162 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6163 region->init_req(2, src_dest_conjoint);
6164
6165 record_for_igvn(region);
6166 return _gvn.transform(region);
6167 }
6168
6169 //------------------------------inline_sha_implCompress-----------------------
6170 //
6171 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6172 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6173 //
6174 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6175 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6176 //
6177 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6178 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6179 //
6180 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6181 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6182
6183 Node* sha_obj = argument(0);
6184 Node* src = argument(1); // type oop
6185 Node* ofs = argument(2); // type int
6186
6187 const Type* src_type = src->Value(&_gvn);
6188 const TypeAryPtr* top_src = src_type->isa_aryptr();
6189 if (top_src == NULL || top_src->klass() == NULL) {
6190 // failed array check
6191 return false;
6192 }
6193 // Figure out the size and type of the elements we will be copying.
6194 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6195 if (src_elem != T_BYTE) {
6196 return false;
6197 }
6198 // 'src_start' points to src array + offset
6199 Node* src_start = array_element_address(src, ofs, src_elem);
6200 Node* state = NULL;
6201 address stubAddr;
6202 const char *stubName;
6203
6204 switch(id) {
6205 case vmIntrinsics::_sha_implCompress:
6206 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6207 state = get_state_from_sha_object(sha_obj);
6208 stubAddr = StubRoutines::sha1_implCompress();
6209 stubName = "sha1_implCompress";
6210 break;
6211 case vmIntrinsics::_sha2_implCompress:
6212 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6213 state = get_state_from_sha_object(sha_obj);
6214 stubAddr = StubRoutines::sha256_implCompress();
6215 stubName = "sha256_implCompress";
6216 break;
6217 case vmIntrinsics::_sha5_implCompress:
6218 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6219 state = get_state_from_sha5_object(sha_obj);
6220 stubAddr = StubRoutines::sha512_implCompress();
6221 stubName = "sha512_implCompress";
6222 break;
6223 default:
6224 fatal_unexpected_iid(id);
6225 return false;
6226 }
6227 if (state == NULL) return false;
6228
6229 // Call the stub.
6230 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6231 stubAddr, stubName, TypePtr::BOTTOM,
6232 src_start, state);
6233
6234 return true;
6235 }
6236
6237 //------------------------------inline_digestBase_implCompressMB-----------------------
6238 //
6239 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6240 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6241 //
6242 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6243 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6244 "need SHA1/SHA256/SHA512 instruction support");
6245 assert((uint)predicate < 3, "sanity");
6246 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6247
6248 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6249 Node* src = argument(1); // byte[] array
6250 Node* ofs = argument(2); // type int
6251 Node* limit = argument(3); // type int
6252
6253 const Type* src_type = src->Value(&_gvn);
6254 const TypeAryPtr* top_src = src_type->isa_aryptr();
6255 if (top_src == NULL || top_src->klass() == NULL) {
6256 // failed array check
6257 return false;
6258 }
6259 // Figure out the size and type of the elements we will be copying.
6260 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6261 if (src_elem != T_BYTE) {
6262 return false;
6263 }
6264 // 'src_start' points to src array + offset
6265 Node* src_start = array_element_address(src, ofs, src_elem);
6266
6267 const char* klass_SHA_name = NULL;
6268 const char* stub_name = NULL;
6269 address stub_addr = NULL;
6270 bool long_state = false;
6271
6272 switch (predicate) {
6273 case 0:
6274 if (UseSHA1Intrinsics) {
6275 klass_SHA_name = "sun/security/provider/SHA";
6276 stub_name = "sha1_implCompressMB";
6277 stub_addr = StubRoutines::sha1_implCompressMB();
6278 }
6279 break;
6280 case 1:
6281 if (UseSHA256Intrinsics) {
6282 klass_SHA_name = "sun/security/provider/SHA2";
6283 stub_name = "sha256_implCompressMB";
6284 stub_addr = StubRoutines::sha256_implCompressMB();
6285 }
6286 break;
6287 case 2:
6288 if (UseSHA512Intrinsics) {
6289 klass_SHA_name = "sun/security/provider/SHA5";
6290 stub_name = "sha512_implCompressMB";
6291 stub_addr = StubRoutines::sha512_implCompressMB();
6292 long_state = true;
6293 }
6294 break;
6295 default:
6296 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6297 }
6298 if (klass_SHA_name != NULL) {
6299 // get DigestBase klass to lookup for SHA klass
6300 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6301 assert(tinst != NULL, "digestBase_obj is not instance???");
6302 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6303
6304 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6305 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6306 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6307 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6308 }
6309 return false;
6310 }
6311 //------------------------------inline_sha_implCompressMB-----------------------
6312 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6313 bool long_state, address stubAddr, const char *stubName,
6314 Node* src_start, Node* ofs, Node* limit) {
6315 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6316 const TypeOopPtr* xtype = aklass->as_instance_type();
6317 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6318 sha_obj = _gvn.transform(sha_obj);
6319
6320 Node* state;
6321 if (long_state) {
6322 state = get_state_from_sha5_object(sha_obj);
6323 } else {
6324 state = get_state_from_sha_object(sha_obj);
6325 }
6326 if (state == NULL) return false;
6327
6328 // Call the stub.
6329 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6330 OptoRuntime::digestBase_implCompressMB_Type(),
6331 stubAddr, stubName, TypePtr::BOTTOM,
6332 src_start, state, ofs, limit);
6333 // return ofs (int)
6334 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6335 set_result(result);
6336
6337 return true;
6338 }
6339
6340 //------------------------------get_state_from_sha_object-----------------------
6341 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6342 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6343 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6344 if (sha_state == NULL) return (Node *) NULL;
6345
6346 // now have the array, need to get the start address of the state array
6347 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6348 return state;
6349 }
6350
6351 //------------------------------get_state_from_sha5_object-----------------------
6352 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6353 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6354 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6355 if (sha_state == NULL) return (Node *) NULL;
6356
6357 // now have the array, need to get the start address of the state array
6358 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6359 return state;
6360 }
6361
6362 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6363 // Return node representing slow path of predicate check.
6364 // the pseudo code we want to emulate with this predicate is:
6365 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6366 //
6367 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6368 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6369 "need SHA1/SHA256/SHA512 instruction support");
6370 assert((uint)predicate < 3, "sanity");
6371
6372 // The receiver was checked for NULL already.
6373 Node* digestBaseObj = argument(0);
6374
6375 // get DigestBase klass for instanceOf check
6376 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6377 assert(tinst != NULL, "digestBaseObj is null");
6378 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6379
6380 const char* klass_SHA_name = NULL;
6381 switch (predicate) {
6382 case 0:
6383 if (UseSHA1Intrinsics) {
6384 // we want to do an instanceof comparison against the SHA class
6385 klass_SHA_name = "sun/security/provider/SHA";
6386 }
6387 break;
6388 case 1:
6389 if (UseSHA256Intrinsics) {
6390 // we want to do an instanceof comparison against the SHA2 class
6391 klass_SHA_name = "sun/security/provider/SHA2";
6392 }
6393 break;
6394 case 2:
6395 if (UseSHA512Intrinsics) {
6396 // we want to do an instanceof comparison against the SHA5 class
6397 klass_SHA_name = "sun/security/provider/SHA5";
6398 }
6399 break;
6400 default:
6401 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6402 }
6403
6404 ciKlass* klass_SHA = NULL;
6405 if (klass_SHA_name != NULL) {
6406 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6407 }
6408 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6409 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6410 Node* ctrl = control();
6411 set_control(top()); // no intrinsic path
6412 return ctrl;
6413 }
6414 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6415
6416 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6417 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6418 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6419 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6420
6421 return instof_false; // even if it is NULL
6422 }