1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/connode.hpp"
35 #include "opto/idealKit.hpp"
36 #include "opto/mathexactnode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/parse.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "prims/nativeLookup.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "trace/traceMacros.hpp"
44
45 class LibraryIntrinsic : public InlineCallGenerator {
46 // Extend the set of intrinsics known to the runtime:
47 public:
48 private:
49 bool _is_virtual;
50 bool _does_virtual_dispatch;
51 int8_t _predicates_count; // Intrinsic is predicated by several conditions
52 int8_t _last_predicate; // Last generated predicate
53 vmIntrinsics::ID _intrinsic_id;
54
55 public:
56 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
57 : InlineCallGenerator(m),
58 _is_virtual(is_virtual),
59 _does_virtual_dispatch(does_virtual_dispatch),
60 _predicates_count((int8_t)predicates_count),
61 _last_predicate((int8_t)-1),
62 _intrinsic_id(id)
63 {
64 }
65 virtual bool is_intrinsic() const { return true; }
66 virtual bool is_virtual() const { return _is_virtual; }
67 virtual bool is_predicated() const { return _predicates_count > 0; }
68 virtual int predicates_count() const { return _predicates_count; }
69 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
70 virtual JVMState* generate(JVMState* jvms);
71 virtual Node* generate_predicate(JVMState* jvms, int predicate);
72 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
73 };
74
75
76 // Local helper class for LibraryIntrinsic:
77 class LibraryCallKit : public GraphKit {
78 private:
79 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
80 Node* _result; // the result node, if any
81 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
82
83 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
84
85 public:
86 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
87 : GraphKit(jvms),
88 _intrinsic(intrinsic),
89 _result(NULL)
90 {
91 // Check if this is a root compile. In that case we don't have a caller.
92 if (!jvms->has_method()) {
93 _reexecute_sp = sp();
94 } else {
95 // Find out how many arguments the interpreter needs when deoptimizing
96 // and save the stack pointer value so it can used by uncommon_trap.
97 // We find the argument count by looking at the declared signature.
98 bool ignored_will_link;
99 ciSignature* declared_signature = NULL;
100 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
101 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
102 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
103 }
104 }
105
106 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
107
108 ciMethod* caller() const { return jvms()->method(); }
109 int bci() const { return jvms()->bci(); }
110 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
111 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
112 ciMethod* callee() const { return _intrinsic->method(); }
113
114 bool try_to_inline(int predicate);
115 Node* try_to_predicate(int predicate);
116
117 void push_result() {
118 // Push the result onto the stack.
119 if (!stopped() && result() != NULL) {
120 BasicType bt = result()->bottom_type()->basic_type();
121 push_node(bt, result());
122 }
123 }
124
125 private:
126 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
127 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
128 }
129
130 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
131 void set_result(RegionNode* region, PhiNode* value);
132 Node* result() { return _result; }
133
134 virtual int reexecute_sp() { return _reexecute_sp; }
135
136 // Helper functions to inline natives
137 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
138 Node* generate_slow_guard(Node* test, RegionNode* region);
139 Node* generate_fair_guard(Node* test, RegionNode* region);
140 Node* generate_negative_guard(Node* index, RegionNode* region,
141 // resulting CastII of index:
142 Node* *pos_index = NULL);
143 Node* generate_nonpositive_guard(Node* index, bool never_negative,
144 // resulting CastII of index:
145 Node* *pos_index = NULL);
146 Node* generate_limit_guard(Node* offset, Node* subseq_length,
147 Node* array_length,
148 RegionNode* region);
149 Node* generate_current_thread(Node* &tls_output);
150 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
151 bool disjoint_bases, const char* &name, bool dest_uninitialized);
152 Node* load_mirror_from_klass(Node* klass);
153 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
154 RegionNode* region, int null_path,
155 int offset);
156 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
157 RegionNode* region, int null_path) {
158 int offset = java_lang_Class::klass_offset_in_bytes();
159 return load_klass_from_mirror_common(mirror, never_see_null,
160 region, null_path,
161 offset);
162 }
163 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path) {
165 int offset = java_lang_Class::array_klass_offset_in_bytes();
166 return load_klass_from_mirror_common(mirror, never_see_null,
167 region, null_path,
168 offset);
169 }
170 Node* generate_access_flags_guard(Node* kls,
171 int modifier_mask, int modifier_bits,
172 RegionNode* region);
173 Node* generate_interface_guard(Node* kls, RegionNode* region);
174 Node* generate_array_guard(Node* kls, RegionNode* region) {
175 return generate_array_guard_common(kls, region, false, false);
176 }
177 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
178 return generate_array_guard_common(kls, region, false, true);
179 }
180 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, true, false);
182 }
183 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, true, true);
185 }
186 Node* generate_array_guard_common(Node* kls, RegionNode* region,
187 bool obj_array, bool not_array);
188 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
189 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
190 bool is_virtual = false, bool is_static = false);
191 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
192 return generate_method_call(method_id, false, true);
193 }
194 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
195 return generate_method_call(method_id, true, false);
196 }
197 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
198
199 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
200 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
201 bool inline_string_compareTo();
202 bool inline_string_indexOf();
203 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
204 bool inline_string_equals();
205 Node* round_double_node(Node* n);
206 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
207 bool inline_math_native(vmIntrinsics::ID id);
208 bool inline_trig(vmIntrinsics::ID id);
209 bool inline_math(vmIntrinsics::ID id);
210 template <typename OverflowOp>
211 bool inline_math_overflow(Node* arg1, Node* arg2);
212 void inline_math_mathExact(Node* math, Node* test);
213 bool inline_math_addExactI(bool is_increment);
214 bool inline_math_addExactL(bool is_increment);
215 bool inline_math_multiplyExactI();
216 bool inline_math_multiplyExactL();
217 bool inline_math_negateExactI();
218 bool inline_math_negateExactL();
219 bool inline_math_subtractExactI(bool is_decrement);
220 bool inline_math_subtractExactL(bool is_decrement);
221 bool inline_exp();
222 bool inline_pow();
223 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_min_max(vmIntrinsics::ID id);
225 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
226 // This returns Type::AnyPtr, RawPtr, or OopPtr.
227 int classify_unsafe_addr(Node* &base, Node* &offset);
228 Node* make_unsafe_address(Node* base, Node* offset);
229 // Helper for inline_unsafe_access.
230 // Generates the guards that check whether the result of
231 // Unsafe.getObject should be recorded in an SATB log buffer.
232 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
233 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
234 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
235 static bool klass_needs_init_guard(Node* kls);
236 bool inline_unsafe_allocate();
237 bool inline_unsafe_copyMemory();
238 bool inline_native_currentThread();
239 #ifdef TRACE_HAVE_INTRINSICS
240 bool inline_native_classID();
241 bool inline_native_threadID();
242 #endif
243 bool inline_native_time_funcs(address method, const char* funcName);
244 bool inline_native_isInterrupted();
245 bool inline_native_Class_query(vmIntrinsics::ID id);
246 bool inline_native_subtype_check();
247
248 bool inline_native_newArray();
249 bool inline_native_getLength();
250 bool inline_array_copyOf(bool is_copyOfRange);
251 bool inline_array_equals();
252 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
253 bool inline_native_clone(bool is_virtual);
254 bool inline_native_Reflection_getCallerClass();
255 // Helper function for inlining native object hash method
256 bool inline_native_hashcode(bool is_virtual, bool is_static);
257 bool inline_native_getClass();
258
259 // Helper functions for inlining arraycopy
260 bool inline_arraycopy();
261 void generate_arraycopy(const TypePtr* adr_type,
262 BasicType basic_elem_type,
263 Node* src, Node* src_offset,
264 Node* dest, Node* dest_offset,
265 Node* copy_length,
266 bool disjoint_bases = false,
267 bool length_never_negative = false,
268 RegionNode* slow_region = NULL);
269 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
270 RegionNode* slow_region);
271 void generate_clear_array(const TypePtr* adr_type,
272 Node* dest,
273 BasicType basic_elem_type,
274 Node* slice_off,
275 Node* slice_len,
276 Node* slice_end);
277 bool generate_block_arraycopy(const TypePtr* adr_type,
278 BasicType basic_elem_type,
279 AllocateNode* alloc,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* dest_size, bool dest_uninitialized);
283 void generate_slow_arraycopy(const TypePtr* adr_type,
284 Node* src, Node* src_offset,
285 Node* dest, Node* dest_offset,
286 Node* copy_length, bool dest_uninitialized);
287 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
288 Node* dest_elem_klass,
289 Node* src, Node* src_offset,
290 Node* dest, Node* dest_offset,
291 Node* copy_length, bool dest_uninitialized);
292 Node* generate_generic_arraycopy(const TypePtr* adr_type,
293 Node* src, Node* src_offset,
294 Node* dest, Node* dest_offset,
295 Node* copy_length, bool dest_uninitialized);
296 void generate_unchecked_arraycopy(const TypePtr* adr_type,
297 BasicType basic_elem_type,
298 bool disjoint_bases,
299 Node* src, Node* src_offset,
300 Node* dest, Node* dest_offset,
301 Node* copy_length, bool dest_uninitialized);
302 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
303 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
304 bool inline_unsafe_ordered_store(BasicType type);
305 bool inline_unsafe_fence(vmIntrinsics::ID id);
306 bool inline_fp_conversions(vmIntrinsics::ID id);
307 bool inline_number_methods(vmIntrinsics::ID id);
308 bool inline_reference_get();
309 bool inline_aescrypt_Block(vmIntrinsics::ID id);
310 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
311 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
312 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
313 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
314 bool inline_sha_implCompress(vmIntrinsics::ID id);
315 bool inline_digestBase_implCompressMB(int predicate);
316 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
317 bool long_state, address stubAddr, const char *stubName,
318 Node* src_start, Node* ofs, Node* limit);
319 Node* get_state_from_sha_object(Node *sha_object);
320 Node* get_state_from_sha5_object(Node *sha_object);
321 Node* inline_digestBase_implCompressMB_predicate(int predicate);
322 bool inline_encodeISOArray();
323 bool inline_updateCRC32();
324 bool inline_updateBytesCRC32();
325 bool inline_updateByteBufferCRC32();
326 bool inline_multiplyToLen();
327
328 bool inline_profileBoolean();
329 };
330
331
332 //---------------------------make_vm_intrinsic----------------------------
333 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
334 vmIntrinsics::ID id = m->intrinsic_id();
335 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
336
337 ccstr disable_intr = NULL;
338
339 if ((DisableIntrinsic[0] != '\0'
340 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
341 (method_has_option_value("DisableIntrinsic", disable_intr)
342 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
343 // disabled by a user request on the command line:
344 // example: -XX:DisableIntrinsic=_hashCode,_getClass
345 return NULL;
346 }
347
348 if (!m->is_loaded()) {
349 // do not attempt to inline unloaded methods
350 return NULL;
351 }
352
353 // Only a few intrinsics implement a virtual dispatch.
354 // They are expensive calls which are also frequently overridden.
355 if (is_virtual) {
356 switch (id) {
357 case vmIntrinsics::_hashCode:
358 case vmIntrinsics::_clone:
359 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
360 break;
361 default:
362 return NULL;
363 }
364 }
365
366 // -XX:-InlineNatives disables nearly all intrinsics:
367 if (!InlineNatives) {
368 switch (id) {
369 case vmIntrinsics::_indexOf:
370 case vmIntrinsics::_compareTo:
371 case vmIntrinsics::_equals:
372 case vmIntrinsics::_equalsC:
373 case vmIntrinsics::_getAndAddInt:
374 case vmIntrinsics::_getAndAddLong:
375 case vmIntrinsics::_getAndSetInt:
376 case vmIntrinsics::_getAndSetLong:
377 case vmIntrinsics::_getAndSetObject:
378 case vmIntrinsics::_loadFence:
379 case vmIntrinsics::_storeFence:
380 case vmIntrinsics::_fullFence:
381 break; // InlineNatives does not control String.compareTo
382 case vmIntrinsics::_Reference_get:
383 break; // InlineNatives does not control Reference.get
384 default:
385 return NULL;
386 }
387 }
388
389 int predicates = 0;
390 bool does_virtual_dispatch = false;
391
392 switch (id) {
393 case vmIntrinsics::_compareTo:
394 if (!SpecialStringCompareTo) return NULL;
395 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
396 break;
397 case vmIntrinsics::_indexOf:
398 if (!SpecialStringIndexOf) return NULL;
399 break;
400 case vmIntrinsics::_equals:
401 if (!SpecialStringEquals) return NULL;
402 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
403 break;
404 case vmIntrinsics::_equalsC:
405 if (!SpecialArraysEquals) return NULL;
406 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
407 break;
408 case vmIntrinsics::_arraycopy:
409 if (!InlineArrayCopy) return NULL;
410 break;
411 case vmIntrinsics::_copyMemory:
412 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
413 if (!InlineArrayCopy) return NULL;
414 break;
415 case vmIntrinsics::_hashCode:
416 if (!InlineObjectHash) return NULL;
417 does_virtual_dispatch = true;
418 break;
419 case vmIntrinsics::_clone:
420 does_virtual_dispatch = true;
421 case vmIntrinsics::_copyOf:
422 case vmIntrinsics::_copyOfRange:
423 if (!InlineObjectCopy) return NULL;
424 // These also use the arraycopy intrinsic mechanism:
425 if (!InlineArrayCopy) return NULL;
426 break;
427 case vmIntrinsics::_encodeISOArray:
428 if (!SpecialEncodeISOArray) return NULL;
429 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
430 break;
431 case vmIntrinsics::_checkIndex:
432 // We do not intrinsify this. The optimizer does fine with it.
433 return NULL;
434
435 case vmIntrinsics::_getCallerClass:
436 if (!UseNewReflection) return NULL;
437 if (!InlineReflectionGetCallerClass) return NULL;
438 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
439 break;
440
441 case vmIntrinsics::_bitCount_i:
442 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
443 break;
444
445 case vmIntrinsics::_bitCount_l:
446 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
447 break;
448
449 case vmIntrinsics::_numberOfLeadingZeros_i:
450 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
451 break;
452
453 case vmIntrinsics::_numberOfLeadingZeros_l:
454 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
455 break;
456
457 case vmIntrinsics::_numberOfTrailingZeros_i:
458 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
459 break;
460
461 case vmIntrinsics::_numberOfTrailingZeros_l:
462 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
463 break;
464
465 case vmIntrinsics::_reverseBytes_c:
466 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
467 break;
468 case vmIntrinsics::_reverseBytes_s:
469 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
470 break;
471 case vmIntrinsics::_reverseBytes_i:
472 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
473 break;
474 case vmIntrinsics::_reverseBytes_l:
475 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
476 break;
477
478 case vmIntrinsics::_Reference_get:
479 // Use the intrinsic version of Reference.get() so that the value in
480 // the referent field can be registered by the G1 pre-barrier code.
481 // Also add memory barrier to prevent commoning reads from this field
482 // across safepoint since GC can change it value.
483 break;
484
485 case vmIntrinsics::_compareAndSwapObject:
486 #ifdef _LP64
487 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
488 #endif
489 break;
490
491 case vmIntrinsics::_compareAndSwapLong:
492 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
493 break;
494
495 case vmIntrinsics::_getAndAddInt:
496 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
497 break;
498
499 case vmIntrinsics::_getAndAddLong:
500 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
501 break;
502
503 case vmIntrinsics::_getAndSetInt:
504 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
505 break;
506
507 case vmIntrinsics::_getAndSetLong:
508 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
509 break;
510
511 case vmIntrinsics::_getAndSetObject:
512 #ifdef _LP64
513 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
514 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
515 break;
516 #else
517 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
518 break;
519 #endif
520
521 case vmIntrinsics::_aescrypt_encryptBlock:
522 case vmIntrinsics::_aescrypt_decryptBlock:
523 if (!UseAESIntrinsics) return NULL;
524 break;
525
526 case vmIntrinsics::_multiplyToLen:
527 if (!UseMultiplyToLenIntrinsic) return NULL;
528 break;
529
530 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
531 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
532 if (!UseAESIntrinsics) return NULL;
533 // these two require the predicated logic
534 predicates = 1;
535 break;
536
537 case vmIntrinsics::_sha_implCompress:
538 if (!UseSHA1Intrinsics) return NULL;
539 break;
540
541 case vmIntrinsics::_sha2_implCompress:
542 if (!UseSHA256Intrinsics) return NULL;
543 break;
544
545 case vmIntrinsics::_sha5_implCompress:
546 if (!UseSHA512Intrinsics) return NULL;
547 break;
548
549 case vmIntrinsics::_digestBase_implCompressMB:
550 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
551 predicates = 3;
552 break;
553
554 case vmIntrinsics::_updateCRC32:
555 case vmIntrinsics::_updateBytesCRC32:
556 case vmIntrinsics::_updateByteBufferCRC32:
557 if (!UseCRC32Intrinsics) return NULL;
558 break;
559
560 case vmIntrinsics::_incrementExactI:
561 case vmIntrinsics::_addExactI:
562 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
563 break;
564 case vmIntrinsics::_incrementExactL:
565 case vmIntrinsics::_addExactL:
566 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
567 break;
568 case vmIntrinsics::_decrementExactI:
569 case vmIntrinsics::_subtractExactI:
570 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
571 break;
572 case vmIntrinsics::_decrementExactL:
573 case vmIntrinsics::_subtractExactL:
574 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
575 break;
576 case vmIntrinsics::_negateExactI:
577 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
578 break;
579 case vmIntrinsics::_negateExactL:
580 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
581 break;
582 case vmIntrinsics::_multiplyExactI:
583 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
584 break;
585 case vmIntrinsics::_multiplyExactL:
586 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
587 break;
588
589 default:
590 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
591 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
592 break;
593 }
594
595 // -XX:-InlineClassNatives disables natives from the Class class.
596 // The flag applies to all reflective calls, notably Array.newArray
597 // (visible to Java programmers as Array.newInstance).
598 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
599 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
600 if (!InlineClassNatives) return NULL;
601 }
602
603 // -XX:-InlineThreadNatives disables natives from the Thread class.
604 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
605 if (!InlineThreadNatives) return NULL;
606 }
607
608 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
609 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
610 m->holder()->name() == ciSymbol::java_lang_Float() ||
611 m->holder()->name() == ciSymbol::java_lang_Double()) {
612 if (!InlineMathNatives) return NULL;
613 }
614
615 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
616 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
617 if (!InlineUnsafeOps) return NULL;
618 }
619
620 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
621 }
622
623 //----------------------register_library_intrinsics-----------------------
624 // Initialize this file's data structures, for each Compile instance.
625 void Compile::register_library_intrinsics() {
626 // Nothing to do here.
627 }
628
629 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
630 LibraryCallKit kit(jvms, this);
631 Compile* C = kit.C;
632 int nodes = C->unique();
633 #ifndef PRODUCT
634 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
635 char buf[1000];
636 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
637 tty->print_cr("Intrinsic %s", str);
638 }
639 #endif
640 ciMethod* callee = kit.callee();
641 const int bci = kit.bci();
642
643 // Try to inline the intrinsic.
644 if (kit.try_to_inline(_last_predicate)) {
645 if (C->print_intrinsics() || C->print_inlining()) {
646 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
647 }
648 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
649 if (C->log()) {
650 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
651 vmIntrinsics::name_at(intrinsic_id()),
652 (is_virtual() ? " virtual='1'" : ""),
653 C->unique() - nodes);
654 }
655 // Push the result from the inlined method onto the stack.
656 kit.push_result();
657 return kit.transfer_exceptions_into_jvms();
658 }
659
660 // The intrinsic bailed out
661 if (C->print_intrinsics() || C->print_inlining()) {
662 if (jvms->has_method()) {
663 // Not a root compile.
664 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
665 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
666 } else {
667 // Root compile
668 tty->print("Did not generate intrinsic %s%s at bci:%d in",
669 vmIntrinsics::name_at(intrinsic_id()),
670 (is_virtual() ? " (virtual)" : ""), bci);
671 }
672 }
673 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
674 return NULL;
675 }
676
677 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
678 LibraryCallKit kit(jvms, this);
679 Compile* C = kit.C;
680 int nodes = C->unique();
681 _last_predicate = predicate;
682 #ifndef PRODUCT
683 assert(is_predicated() && predicate < predicates_count(), "sanity");
684 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
685 char buf[1000];
686 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
687 tty->print_cr("Predicate for intrinsic %s", str);
688 }
689 #endif
690 ciMethod* callee = kit.callee();
691 const int bci = kit.bci();
692
693 Node* slow_ctl = kit.try_to_predicate(predicate);
694 if (!kit.failing()) {
695 if (C->print_intrinsics() || C->print_inlining()) {
696 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
697 }
698 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
699 if (C->log()) {
700 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
701 vmIntrinsics::name_at(intrinsic_id()),
702 (is_virtual() ? " virtual='1'" : ""),
703 C->unique() - nodes);
704 }
705 return slow_ctl; // Could be NULL if the check folds.
706 }
707
708 // The intrinsic bailed out
709 if (C->print_intrinsics() || C->print_inlining()) {
710 if (jvms->has_method()) {
711 // Not a root compile.
712 const char* msg = "failed to generate predicate for intrinsic";
713 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
714 } else {
715 // Root compile
716 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
717 vmIntrinsics::name_at(intrinsic_id()),
718 (is_virtual() ? " (virtual)" : ""), bci);
719 }
720 }
721 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
722 return NULL;
723 }
724
725 bool LibraryCallKit::try_to_inline(int predicate) {
726 // Handle symbolic names for otherwise undistinguished boolean switches:
727 const bool is_store = true;
728 const bool is_native_ptr = true;
729 const bool is_static = true;
730 const bool is_volatile = true;
731
732 if (!jvms()->has_method()) {
733 // Root JVMState has a null method.
734 assert(map()->memory()->Opcode() == Op_Parm, "");
735 // Insert the memory aliasing node
736 set_all_memory(reset_memory());
737 }
738 assert(merged_memory(), "");
739
740
741 switch (intrinsic_id()) {
742 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
743 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
744 case vmIntrinsics::_getClass: return inline_native_getClass();
745
746 case vmIntrinsics::_dsin:
747 case vmIntrinsics::_dcos:
748 case vmIntrinsics::_dtan:
749 case vmIntrinsics::_dabs:
750 case vmIntrinsics::_datan2:
751 case vmIntrinsics::_dsqrt:
752 case vmIntrinsics::_dexp:
753 case vmIntrinsics::_dlog:
754 case vmIntrinsics::_dlog10:
755 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
756
757 case vmIntrinsics::_min:
758 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
759
760 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
761 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
762 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
763 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
764 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
765 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
766 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
767 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
768 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
769 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
770 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
771 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
772
773 case vmIntrinsics::_arraycopy: return inline_arraycopy();
774
775 case vmIntrinsics::_compareTo: return inline_string_compareTo();
776 case vmIntrinsics::_indexOf: return inline_string_indexOf();
777 case vmIntrinsics::_equals: return inline_string_equals();
778
779 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
780 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
781 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
782 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
783 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
784 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
785 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
786 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
787 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
788
789 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
790 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
791 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
792 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
793 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
794 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
795 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
796 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
797 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
798
799 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
800 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
801 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
802 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
803 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
804 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
805 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
806 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
807
808 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
809 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
810 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
811 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
812 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
813 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
814 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
815 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
816
817 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
818 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
819 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
820 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
821 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
822 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
823 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
824 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
825 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
826
827 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
828 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
829 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
830 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
831 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
832 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
833 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
834 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
835 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
836
837 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
838 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
839 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
840 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
841
842 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
843 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
844 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
845
846 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
847 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
848 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
849
850 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
851 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
852 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
853 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
854 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
855
856 case vmIntrinsics::_loadFence:
857 case vmIntrinsics::_storeFence:
858 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
859
860 case vmIntrinsics::_currentThread: return inline_native_currentThread();
861 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
862
863 #ifdef TRACE_HAVE_INTRINSICS
864 case vmIntrinsics::_classID: return inline_native_classID();
865 case vmIntrinsics::_threadID: return inline_native_threadID();
866 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
867 #endif
868 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
869 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
870 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
871 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
872 case vmIntrinsics::_newArray: return inline_native_newArray();
873 case vmIntrinsics::_getLength: return inline_native_getLength();
874 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
875 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
876 case vmIntrinsics::_equalsC: return inline_array_equals();
877 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
878
879 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
880
881 case vmIntrinsics::_isInstance:
882 case vmIntrinsics::_getModifiers:
883 case vmIntrinsics::_isInterface:
884 case vmIntrinsics::_isArray:
885 case vmIntrinsics::_isPrimitive:
886 case vmIntrinsics::_getSuperclass:
887 case vmIntrinsics::_getComponentType:
888 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
889
890 case vmIntrinsics::_floatToRawIntBits:
891 case vmIntrinsics::_floatToIntBits:
892 case vmIntrinsics::_intBitsToFloat:
893 case vmIntrinsics::_doubleToRawLongBits:
894 case vmIntrinsics::_doubleToLongBits:
895 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
896
897 case vmIntrinsics::_numberOfLeadingZeros_i:
898 case vmIntrinsics::_numberOfLeadingZeros_l:
899 case vmIntrinsics::_numberOfTrailingZeros_i:
900 case vmIntrinsics::_numberOfTrailingZeros_l:
901 case vmIntrinsics::_bitCount_i:
902 case vmIntrinsics::_bitCount_l:
903 case vmIntrinsics::_reverseBytes_i:
904 case vmIntrinsics::_reverseBytes_l:
905 case vmIntrinsics::_reverseBytes_s:
906 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
907
908 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
909
910 case vmIntrinsics::_Reference_get: return inline_reference_get();
911
912 case vmIntrinsics::_aescrypt_encryptBlock:
913 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
914
915 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
916 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
917 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
918
919 case vmIntrinsics::_sha_implCompress:
920 case vmIntrinsics::_sha2_implCompress:
921 case vmIntrinsics::_sha5_implCompress:
922 return inline_sha_implCompress(intrinsic_id());
923
924 case vmIntrinsics::_digestBase_implCompressMB:
925 return inline_digestBase_implCompressMB(predicate);
926
927 case vmIntrinsics::_multiplyToLen:
928 return inline_multiplyToLen();
929
930 case vmIntrinsics::_encodeISOArray:
931 return inline_encodeISOArray();
932
933 case vmIntrinsics::_updateCRC32:
934 return inline_updateCRC32();
935 case vmIntrinsics::_updateBytesCRC32:
936 return inline_updateBytesCRC32();
937 case vmIntrinsics::_updateByteBufferCRC32:
938 return inline_updateByteBufferCRC32();
939
940 case vmIntrinsics::_profileBoolean:
941 return inline_profileBoolean();
942
943 default:
944 // If you get here, it may be that someone has added a new intrinsic
945 // to the list in vmSymbols.hpp without implementing it here.
946 #ifndef PRODUCT
947 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
948 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
949 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
950 }
951 #endif
952 return false;
953 }
954 }
955
956 Node* LibraryCallKit::try_to_predicate(int predicate) {
957 if (!jvms()->has_method()) {
958 // Root JVMState has a null method.
959 assert(map()->memory()->Opcode() == Op_Parm, "");
960 // Insert the memory aliasing node
961 set_all_memory(reset_memory());
962 }
963 assert(merged_memory(), "");
964
965 switch (intrinsic_id()) {
966 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
967 return inline_cipherBlockChaining_AESCrypt_predicate(false);
968 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
969 return inline_cipherBlockChaining_AESCrypt_predicate(true);
970 case vmIntrinsics::_digestBase_implCompressMB:
971 return inline_digestBase_implCompressMB_predicate(predicate);
972
973 default:
974 // If you get here, it may be that someone has added a new intrinsic
975 // to the list in vmSymbols.hpp without implementing it here.
976 #ifndef PRODUCT
977 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
978 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
979 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
980 }
981 #endif
982 Node* slow_ctl = control();
983 set_control(top()); // No fast path instrinsic
984 return slow_ctl;
985 }
986 }
987
988 //------------------------------set_result-------------------------------
989 // Helper function for finishing intrinsics.
990 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
991 record_for_igvn(region);
992 set_control(_gvn.transform(region));
993 set_result( _gvn.transform(value));
994 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
995 }
996
997 //------------------------------generate_guard---------------------------
998 // Helper function for generating guarded fast-slow graph structures.
999 // The given 'test', if true, guards a slow path. If the test fails
1000 // then a fast path can be taken. (We generally hope it fails.)
1001 // In all cases, GraphKit::control() is updated to the fast path.
1002 // The returned value represents the control for the slow path.
1003 // The return value is never 'top'; it is either a valid control
1004 // or NULL if it is obvious that the slow path can never be taken.
1005 // Also, if region and the slow control are not NULL, the slow edge
1006 // is appended to the region.
1007 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1008 if (stopped()) {
1009 // Already short circuited.
1010 return NULL;
1011 }
1012
1013 // Build an if node and its projections.
1014 // If test is true we take the slow path, which we assume is uncommon.
1015 if (_gvn.type(test) == TypeInt::ZERO) {
1016 // The slow branch is never taken. No need to build this guard.
1017 return NULL;
1018 }
1019
1020 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1021
1022 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1023 if (if_slow == top()) {
1024 // The slow branch is never taken. No need to build this guard.
1025 return NULL;
1026 }
1027
1028 if (region != NULL)
1029 region->add_req(if_slow);
1030
1031 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1032 set_control(if_fast);
1033
1034 return if_slow;
1035 }
1036
1037 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1038 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1039 }
1040 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1041 return generate_guard(test, region, PROB_FAIR);
1042 }
1043
1044 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1045 Node* *pos_index) {
1046 if (stopped())
1047 return NULL; // already stopped
1048 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1049 return NULL; // index is already adequately typed
1050 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1051 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1052 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1053 if (is_neg != NULL && pos_index != NULL) {
1054 // Emulate effect of Parse::adjust_map_after_if.
1055 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1056 ccast->set_req(0, control());
1057 (*pos_index) = _gvn.transform(ccast);
1058 }
1059 return is_neg;
1060 }
1061
1062 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1063 Node* *pos_index) {
1064 if (stopped())
1065 return NULL; // already stopped
1066 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1067 return NULL; // index is already adequately typed
1068 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1069 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1070 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1071 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1072 if (is_notp != NULL && pos_index != NULL) {
1073 // Emulate effect of Parse::adjust_map_after_if.
1074 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1075 ccast->set_req(0, control());
1076 (*pos_index) = _gvn.transform(ccast);
1077 }
1078 return is_notp;
1079 }
1080
1081 // Make sure that 'position' is a valid limit index, in [0..length].
1082 // There are two equivalent plans for checking this:
1083 // A. (offset + copyLength) unsigned<= arrayLength
1084 // B. offset <= (arrayLength - copyLength)
1085 // We require that all of the values above, except for the sum and
1086 // difference, are already known to be non-negative.
1087 // Plan A is robust in the face of overflow, if offset and copyLength
1088 // are both hugely positive.
1089 //
1090 // Plan B is less direct and intuitive, but it does not overflow at
1091 // all, since the difference of two non-negatives is always
1092 // representable. Whenever Java methods must perform the equivalent
1093 // check they generally use Plan B instead of Plan A.
1094 // For the moment we use Plan A.
1095 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1096 Node* subseq_length,
1097 Node* array_length,
1098 RegionNode* region) {
1099 if (stopped())
1100 return NULL; // already stopped
1101 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1102 if (zero_offset && subseq_length->eqv_uncast(array_length))
1103 return NULL; // common case of whole-array copy
1104 Node* last = subseq_length;
1105 if (!zero_offset) // last += offset
1106 last = _gvn.transform(new (C) AddINode(last, offset));
1107 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1108 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1109 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1110 return is_over;
1111 }
1112
1113
1114 //--------------------------generate_current_thread--------------------
1115 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1116 ciKlass* thread_klass = env()->Thread_klass();
1117 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1118 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1119 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1120 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1121 tls_output = thread;
1122 return threadObj;
1123 }
1124
1125
1126 //------------------------------make_string_method_node------------------------
1127 // Helper method for String intrinsic functions. This version is called
1128 // with str1 and str2 pointing to String object nodes.
1129 //
1130 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1131 Node* no_ctrl = NULL;
1132
1133 // Get start addr of string
1134 Node* str1_value = load_String_value(no_ctrl, str1);
1135 Node* str1_offset = load_String_offset(no_ctrl, str1);
1136 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1137
1138 // Get length of string 1
1139 Node* str1_len = load_String_length(no_ctrl, str1);
1140
1141 Node* str2_value = load_String_value(no_ctrl, str2);
1142 Node* str2_offset = load_String_offset(no_ctrl, str2);
1143 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1144
1145 Node* str2_len = NULL;
1146 Node* result = NULL;
1147
1148 switch (opcode) {
1149 case Op_StrIndexOf:
1150 // Get length of string 2
1151 str2_len = load_String_length(no_ctrl, str2);
1152
1153 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1154 str1_start, str1_len, str2_start, str2_len);
1155 break;
1156 case Op_StrComp:
1157 // Get length of string 2
1158 str2_len = load_String_length(no_ctrl, str2);
1159
1160 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1161 str1_start, str1_len, str2_start, str2_len);
1162 break;
1163 case Op_StrEquals:
1164 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1165 str1_start, str2_start, str1_len);
1166 break;
1167 default:
1168 ShouldNotReachHere();
1169 return NULL;
1170 }
1171
1172 // All these intrinsics have checks.
1173 C->set_has_split_ifs(true); // Has chance for split-if optimization
1174
1175 return _gvn.transform(result);
1176 }
1177
1178 // Helper method for String intrinsic functions. This version is called
1179 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1180 // to Int nodes containing the lenghts of str1 and str2.
1181 //
1182 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1183 Node* result = NULL;
1184 switch (opcode) {
1185 case Op_StrIndexOf:
1186 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1187 str1_start, cnt1, str2_start, cnt2);
1188 break;
1189 case Op_StrComp:
1190 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1191 str1_start, cnt1, str2_start, cnt2);
1192 break;
1193 case Op_StrEquals:
1194 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1195 str1_start, str2_start, cnt1);
1196 break;
1197 default:
1198 ShouldNotReachHere();
1199 return NULL;
1200 }
1201
1202 // All these intrinsics have checks.
1203 C->set_has_split_ifs(true); // Has chance for split-if optimization
1204
1205 return _gvn.transform(result);
1206 }
1207
1208 //------------------------------inline_string_compareTo------------------------
1209 // public int java.lang.String.compareTo(String anotherString);
1210 bool LibraryCallKit::inline_string_compareTo() {
1211 Node* receiver = null_check(argument(0));
1212 Node* arg = null_check(argument(1));
1213 if (stopped()) {
1214 return true;
1215 }
1216 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1217 return true;
1218 }
1219
1220 //------------------------------inline_string_equals------------------------
1221 bool LibraryCallKit::inline_string_equals() {
1222 Node* receiver = null_check_receiver();
1223 // NOTE: Do not null check argument for String.equals() because spec
1224 // allows to specify NULL as argument.
1225 Node* argument = this->argument(1);
1226 if (stopped()) {
1227 return true;
1228 }
1229
1230 // paths (plus control) merge
1231 RegionNode* region = new (C) RegionNode(5);
1232 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1233
1234 // does source == target string?
1235 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1236 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1237
1238 Node* if_eq = generate_slow_guard(bol, NULL);
1239 if (if_eq != NULL) {
1240 // receiver == argument
1241 phi->init_req(2, intcon(1));
1242 region->init_req(2, if_eq);
1243 }
1244
1245 // get String klass for instanceOf
1246 ciInstanceKlass* klass = env()->String_klass();
1247
1248 if (!stopped()) {
1249 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1250 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1251 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1252
1253 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1254 //instanceOf == true, fallthrough
1255
1256 if (inst_false != NULL) {
1257 phi->init_req(3, intcon(0));
1258 region->init_req(3, inst_false);
1259 }
1260 }
1261
1262 if (!stopped()) {
1263 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1264
1265 // Properly cast the argument to String
1266 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1267 // This path is taken only when argument's type is String:NotNull.
1268 argument = cast_not_null(argument, false);
1269
1270 Node* no_ctrl = NULL;
1271
1272 // Get start addr of receiver
1273 Node* receiver_val = load_String_value(no_ctrl, receiver);
1274 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1275 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1276
1277 // Get length of receiver
1278 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1279
1280 // Get start addr of argument
1281 Node* argument_val = load_String_value(no_ctrl, argument);
1282 Node* argument_offset = load_String_offset(no_ctrl, argument);
1283 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1284
1285 // Get length of argument
1286 Node* argument_cnt = load_String_length(no_ctrl, argument);
1287
1288 // Check for receiver count != argument count
1289 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1290 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1291 Node* if_ne = generate_slow_guard(bol, NULL);
1292 if (if_ne != NULL) {
1293 phi->init_req(4, intcon(0));
1294 region->init_req(4, if_ne);
1295 }
1296
1297 // Check for count == 0 is done by assembler code for StrEquals.
1298
1299 if (!stopped()) {
1300 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1301 phi->init_req(1, equals);
1302 region->init_req(1, control());
1303 }
1304 }
1305
1306 // post merge
1307 set_control(_gvn.transform(region));
1308 record_for_igvn(region);
1309
1310 set_result(_gvn.transform(phi));
1311 return true;
1312 }
1313
1314 //------------------------------inline_array_equals----------------------------
1315 bool LibraryCallKit::inline_array_equals() {
1316 Node* arg1 = argument(0);
1317 Node* arg2 = argument(1);
1318 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1319 return true;
1320 }
1321
1322 // Java version of String.indexOf(constant string)
1323 // class StringDecl {
1324 // StringDecl(char[] ca) {
1325 // offset = 0;
1326 // count = ca.length;
1327 // value = ca;
1328 // }
1329 // int offset;
1330 // int count;
1331 // char[] value;
1332 // }
1333 //
1334 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1335 // int targetOffset, int cache_i, int md2) {
1336 // int cache = cache_i;
1337 // int sourceOffset = string_object.offset;
1338 // int sourceCount = string_object.count;
1339 // int targetCount = target_object.length;
1340 //
1341 // int targetCountLess1 = targetCount - 1;
1342 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1343 //
1344 // char[] source = string_object.value;
1345 // char[] target = target_object;
1346 // int lastChar = target[targetCountLess1];
1347 //
1348 // outer_loop:
1349 // for (int i = sourceOffset; i < sourceEnd; ) {
1350 // int src = source[i + targetCountLess1];
1351 // if (src == lastChar) {
1352 // // With random strings and a 4-character alphabet,
1353 // // reverse matching at this point sets up 0.8% fewer
1354 // // frames, but (paradoxically) makes 0.3% more probes.
1355 // // Since those probes are nearer the lastChar probe,
1356 // // there is may be a net D$ win with reverse matching.
1357 // // But, reversing loop inhibits unroll of inner loop
1358 // // for unknown reason. So, does running outer loop from
1359 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1360 // for (int j = 0; j < targetCountLess1; j++) {
1361 // if (target[targetOffset + j] != source[i+j]) {
1362 // if ((cache & (1 << source[i+j])) == 0) {
1363 // if (md2 < j+1) {
1364 // i += j+1;
1365 // continue outer_loop;
1366 // }
1367 // }
1368 // i += md2;
1369 // continue outer_loop;
1370 // }
1371 // }
1372 // return i - sourceOffset;
1373 // }
1374 // if ((cache & (1 << src)) == 0) {
1375 // i += targetCountLess1;
1376 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1377 // i++;
1378 // }
1379 // return -1;
1380 // }
1381
1382 //------------------------------string_indexOf------------------------
1383 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1384 jint cache_i, jint md2_i) {
1385
1386 Node* no_ctrl = NULL;
1387 float likely = PROB_LIKELY(0.9);
1388 float unlikely = PROB_UNLIKELY(0.9);
1389
1390 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1391
1392 Node* source = load_String_value(no_ctrl, string_object);
1393 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1394 Node* sourceCount = load_String_length(no_ctrl, string_object);
1395
1396 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1397 jint target_length = target_array->length();
1398 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1399 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1400
1401 // String.value field is known to be @Stable.
1402 if (UseImplicitStableValues) {
1403 target = cast_array_to_stable(target, target_type);
1404 }
1405
1406 IdealKit kit(this, false, true);
1407 #define __ kit.
1408 Node* zero = __ ConI(0);
1409 Node* one = __ ConI(1);
1410 Node* cache = __ ConI(cache_i);
1411 Node* md2 = __ ConI(md2_i);
1412 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1413 Node* targetCount = __ ConI(target_length);
1414 Node* targetCountLess1 = __ ConI(target_length - 1);
1415 Node* targetOffset = __ ConI(targetOffset_i);
1416 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1417
1418 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1419 Node* outer_loop = __ make_label(2 /* goto */);
1420 Node* return_ = __ make_label(1);
1421
1422 __ set(rtn,__ ConI(-1));
1423 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1424 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1425 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1426 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1427 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1428 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1429 Node* tpj = __ AddI(targetOffset, __ value(j));
1430 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1431 Node* ipj = __ AddI(__ value(i), __ value(j));
1432 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1433 __ if_then(targ, BoolTest::ne, src2); {
1434 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1435 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1436 __ increment(i, __ AddI(__ value(j), one));
1437 __ goto_(outer_loop);
1438 } __ end_if(); __ dead(j);
1439 }__ end_if(); __ dead(j);
1440 __ increment(i, md2);
1441 __ goto_(outer_loop);
1442 }__ end_if();
1443 __ increment(j, one);
1444 }__ end_loop(); __ dead(j);
1445 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1446 __ goto_(return_);
1447 }__ end_if();
1448 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1449 __ increment(i, targetCountLess1);
1450 }__ end_if();
1451 __ increment(i, one);
1452 __ bind(outer_loop);
1453 }__ end_loop(); __ dead(i);
1454 __ bind(return_);
1455
1456 // Final sync IdealKit and GraphKit.
1457 final_sync(kit);
1458 Node* result = __ value(rtn);
1459 #undef __
1460 C->set_has_loops(true);
1461 return result;
1462 }
1463
1464 //------------------------------inline_string_indexOf------------------------
1465 bool LibraryCallKit::inline_string_indexOf() {
1466 Node* receiver = argument(0);
1467 Node* arg = argument(1);
1468
1469 Node* result;
1470 // Disable the use of pcmpestri until it can be guaranteed that
1471 // the load doesn't cross into the uncommited space.
1472 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1473 UseSSE42Intrinsics) {
1474 // Generate SSE4.2 version of indexOf
1475 // We currently only have match rules that use SSE4.2
1476
1477 receiver = null_check(receiver);
1478 arg = null_check(arg);
1479 if (stopped()) {
1480 return true;
1481 }
1482
1483 ciInstanceKlass* str_klass = env()->String_klass();
1484 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1485
1486 // Make the merge point
1487 RegionNode* result_rgn = new (C) RegionNode(4);
1488 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1489 Node* no_ctrl = NULL;
1490
1491 // Get start addr of source string
1492 Node* source = load_String_value(no_ctrl, receiver);
1493 Node* source_offset = load_String_offset(no_ctrl, receiver);
1494 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1495
1496 // Get length of source string
1497 Node* source_cnt = load_String_length(no_ctrl, receiver);
1498
1499 // Get start addr of substring
1500 Node* substr = load_String_value(no_ctrl, arg);
1501 Node* substr_offset = load_String_offset(no_ctrl, arg);
1502 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1503
1504 // Get length of source string
1505 Node* substr_cnt = load_String_length(no_ctrl, arg);
1506
1507 // Check for substr count > string count
1508 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1509 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1510 Node* if_gt = generate_slow_guard(bol, NULL);
1511 if (if_gt != NULL) {
1512 result_phi->init_req(2, intcon(-1));
1513 result_rgn->init_req(2, if_gt);
1514 }
1515
1516 if (!stopped()) {
1517 // Check for substr count == 0
1518 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1519 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1520 Node* if_zero = generate_slow_guard(bol, NULL);
1521 if (if_zero != NULL) {
1522 result_phi->init_req(3, intcon(0));
1523 result_rgn->init_req(3, if_zero);
1524 }
1525 }
1526
1527 if (!stopped()) {
1528 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1529 result_phi->init_req(1, result);
1530 result_rgn->init_req(1, control());
1531 }
1532 set_control(_gvn.transform(result_rgn));
1533 record_for_igvn(result_rgn);
1534 result = _gvn.transform(result_phi);
1535
1536 } else { // Use LibraryCallKit::string_indexOf
1537 // don't intrinsify if argument isn't a constant string.
1538 if (!arg->is_Con()) {
1539 return false;
1540 }
1541 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1542 if (str_type == NULL) {
1543 return false;
1544 }
1545 ciInstanceKlass* klass = env()->String_klass();
1546 ciObject* str_const = str_type->const_oop();
1547 if (str_const == NULL || str_const->klass() != klass) {
1548 return false;
1549 }
1550 ciInstance* str = str_const->as_instance();
1551 assert(str != NULL, "must be instance");
1552
1553 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1554 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1555
1556 int o;
1557 int c;
1558 if (java_lang_String::has_offset_field()) {
1559 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1560 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1561 } else {
1562 o = 0;
1563 c = pat->length();
1564 }
1565
1566 // constant strings have no offset and count == length which
1567 // simplifies the resulting code somewhat so lets optimize for that.
1568 if (o != 0 || c != pat->length()) {
1569 return false;
1570 }
1571
1572 receiver = null_check(receiver, T_OBJECT);
1573 // NOTE: No null check on the argument is needed since it's a constant String oop.
1574 if (stopped()) {
1575 return true;
1576 }
1577
1578 // The null string as a pattern always returns 0 (match at beginning of string)
1579 if (c == 0) {
1580 set_result(intcon(0));
1581 return true;
1582 }
1583
1584 // Generate default indexOf
1585 jchar lastChar = pat->char_at(o + (c - 1));
1586 int cache = 0;
1587 int i;
1588 for (i = 0; i < c - 1; i++) {
1589 assert(i < pat->length(), "out of range");
1590 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1591 }
1592
1593 int md2 = c;
1594 for (i = 0; i < c - 1; i++) {
1595 assert(i < pat->length(), "out of range");
1596 if (pat->char_at(o + i) == lastChar) {
1597 md2 = (c - 1) - i;
1598 }
1599 }
1600
1601 result = string_indexOf(receiver, pat, o, cache, md2);
1602 }
1603 set_result(result);
1604 return true;
1605 }
1606
1607 //--------------------------round_double_node--------------------------------
1608 // Round a double node if necessary.
1609 Node* LibraryCallKit::round_double_node(Node* n) {
1610 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1611 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1612 return n;
1613 }
1614
1615 //------------------------------inline_math-----------------------------------
1616 // public static double Math.abs(double)
1617 // public static double Math.sqrt(double)
1618 // public static double Math.log(double)
1619 // public static double Math.log10(double)
1620 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1621 Node* arg = round_double_node(argument(0));
1622 Node* n;
1623 switch (id) {
1624 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1625 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1626 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1627 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1628 default: fatal_unexpected_iid(id); break;
1629 }
1630 set_result(_gvn.transform(n));
1631 return true;
1632 }
1633
1634 //------------------------------inline_trig----------------------------------
1635 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1636 // argument reduction which will turn into a fast/slow diamond.
1637 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1638 Node* arg = round_double_node(argument(0));
1639 Node* n = NULL;
1640
1641 switch (id) {
1642 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1643 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1644 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1645 default: fatal_unexpected_iid(id); break;
1646 }
1647 n = _gvn.transform(n);
1648
1649 // Rounding required? Check for argument reduction!
1650 if (Matcher::strict_fp_requires_explicit_rounding) {
1651 static const double pi_4 = 0.7853981633974483;
1652 static const double neg_pi_4 = -0.7853981633974483;
1653 // pi/2 in 80-bit extended precision
1654 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1655 // -pi/2 in 80-bit extended precision
1656 // 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};
1657 // Cutoff value for using this argument reduction technique
1658 //static const double pi_2_minus_epsilon = 1.564660403643354;
1659 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1660
1661 // Pseudocode for sin:
1662 // if (x <= Math.PI / 4.0) {
1663 // if (x >= -Math.PI / 4.0) return fsin(x);
1664 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1665 // } else {
1666 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1667 // }
1668 // return StrictMath.sin(x);
1669
1670 // Pseudocode for cos:
1671 // if (x <= Math.PI / 4.0) {
1672 // if (x >= -Math.PI / 4.0) return fcos(x);
1673 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1674 // } else {
1675 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1676 // }
1677 // return StrictMath.cos(x);
1678
1679 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1680 // requires a special machine instruction to load it. Instead we'll try
1681 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1682 // probably do the math inside the SIN encoding.
1683
1684 // Make the merge point
1685 RegionNode* r = new (C) RegionNode(3);
1686 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1687
1688 // Flatten arg so we need only 1 test
1689 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1690 // Node for PI/4 constant
1691 Node *pi4 = makecon(TypeD::make(pi_4));
1692 // Check PI/4 : abs(arg)
1693 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1694 // Check: If PI/4 < abs(arg) then go slow
1695 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1696 // Branch either way
1697 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1698 set_control(opt_iff(r,iff));
1699
1700 // Set fast path result
1701 phi->init_req(2, n);
1702
1703 // Slow path - non-blocking leaf call
1704 Node* call = NULL;
1705 switch (id) {
1706 case vmIntrinsics::_dsin:
1707 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1708 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1709 "Sin", NULL, arg, top());
1710 break;
1711 case vmIntrinsics::_dcos:
1712 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1713 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1714 "Cos", NULL, arg, top());
1715 break;
1716 case vmIntrinsics::_dtan:
1717 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1718 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1719 "Tan", NULL, arg, top());
1720 break;
1721 }
1722 assert(control()->in(0) == call, "");
1723 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1724 r->init_req(1, control());
1725 phi->init_req(1, slow_result);
1726
1727 // Post-merge
1728 set_control(_gvn.transform(r));
1729 record_for_igvn(r);
1730 n = _gvn.transform(phi);
1731
1732 C->set_has_split_ifs(true); // Has chance for split-if optimization
1733 }
1734 set_result(n);
1735 return true;
1736 }
1737
1738 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1739 //-------------------
1740 //result=(result.isNaN())? funcAddr():result;
1741 // Check: If isNaN() by checking result!=result? then either trap
1742 // or go to runtime
1743 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1744 // Build the boolean node
1745 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1746
1747 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1748 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1749 // The pow or exp intrinsic returned a NaN, which requires a call
1750 // to the runtime. Recompile with the runtime call.
1751 uncommon_trap(Deoptimization::Reason_intrinsic,
1752 Deoptimization::Action_make_not_entrant);
1753 }
1754 return result;
1755 } else {
1756 // If this inlining ever returned NaN in the past, we compile a call
1757 // to the runtime to properly handle corner cases
1758
1759 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1760 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1761 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1762
1763 if (!if_slow->is_top()) {
1764 RegionNode* result_region = new (C) RegionNode(3);
1765 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1766
1767 result_region->init_req(1, if_fast);
1768 result_val->init_req(1, result);
1769
1770 set_control(if_slow);
1771
1772 const TypePtr* no_memory_effects = NULL;
1773 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1774 no_memory_effects,
1775 x, top(), y, y ? top() : NULL);
1776 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1777 #ifdef ASSERT
1778 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1779 assert(value_top == top(), "second value must be top");
1780 #endif
1781
1782 result_region->init_req(2, control());
1783 result_val->init_req(2, value);
1784 set_control(_gvn.transform(result_region));
1785 return _gvn.transform(result_val);
1786 } else {
1787 return result;
1788 }
1789 }
1790 }
1791
1792 //------------------------------inline_exp-------------------------------------
1793 // Inline exp instructions, if possible. The Intel hardware only misses
1794 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1795 bool LibraryCallKit::inline_exp() {
1796 Node* arg = round_double_node(argument(0));
1797 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1798
1799 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1800 set_result(n);
1801
1802 C->set_has_split_ifs(true); // Has chance for split-if optimization
1803 return true;
1804 }
1805
1806 //------------------------------inline_pow-------------------------------------
1807 // Inline power instructions, if possible.
1808 bool LibraryCallKit::inline_pow() {
1809 // Pseudocode for pow
1810 // if (y == 2) {
1811 // return x * x;
1812 // } else {
1813 // if (x <= 0.0) {
1814 // long longy = (long)y;
1815 // if ((double)longy == y) { // if y is long
1816 // if (y + 1 == y) longy = 0; // huge number: even
1817 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1818 // } else {
1819 // result = NaN;
1820 // }
1821 // } else {
1822 // result = DPow(x,y);
1823 // }
1824 // if (result != result)? {
1825 // result = uncommon_trap() or runtime_call();
1826 // }
1827 // return result;
1828 // }
1829
1830 Node* x = round_double_node(argument(0));
1831 Node* y = round_double_node(argument(2));
1832
1833 Node* result = NULL;
1834
1835 Node* const_two_node = makecon(TypeD::make(2.0));
1836 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1837 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1838 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1839 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1840 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1841
1842 RegionNode* region_node = new (C) RegionNode(3);
1843 region_node->init_req(1, if_true);
1844
1845 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1846 // special case for x^y where y == 2, we can convert it to x * x
1847 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1848
1849 // set control to if_false since we will now process the false branch
1850 set_control(if_false);
1851
1852 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1853 // Short form: skip the fancy tests and just check for NaN result.
1854 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1855 } else {
1856 // If this inlining ever returned NaN in the past, include all
1857 // checks + call to the runtime.
1858
1859 // Set the merge point for If node with condition of (x <= 0.0)
1860 // There are four possible paths to region node and phi node
1861 RegionNode *r = new (C) RegionNode(4);
1862 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1863
1864 // Build the first if node: if (x <= 0.0)
1865 // Node for 0 constant
1866 Node *zeronode = makecon(TypeD::ZERO);
1867 // Check x:0
1868 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1869 // Check: If (x<=0) then go complex path
1870 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1871 // Branch either way
1872 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1873 // Fast path taken; set region slot 3
1874 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1875 r->init_req(3,fast_taken); // Capture fast-control
1876
1877 // Fast path not-taken, i.e. slow path
1878 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1879
1880 // Set fast path result
1881 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1882 phi->init_req(3, fast_result);
1883
1884 // Complex path
1885 // Build the second if node (if y is long)
1886 // Node for (long)y
1887 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1888 // Node for (double)((long) y)
1889 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1890 // Check (double)((long) y) : y
1891 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1892 // Check if (y isn't long) then go to slow path
1893
1894 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1895 // Branch either way
1896 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1897 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1898
1899 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1900
1901 // Calculate DPow(abs(x), y)*(1 & (long)y)
1902 // Node for constant 1
1903 Node *conone = longcon(1);
1904 // 1& (long)y
1905 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1906
1907 // A huge number is always even. Detect a huge number by checking
1908 // if y + 1 == y and set integer to be tested for parity to 0.
1909 // Required for corner case:
1910 // (long)9.223372036854776E18 = max_jlong
1911 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1912 // max_jlong is odd but 9.223372036854776E18 is even
1913 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1914 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1915 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1916 Node* correctedsign = NULL;
1917 if (ConditionalMoveLimit != 0) {
1918 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1919 } else {
1920 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1921 RegionNode *r = new (C) RegionNode(3);
1922 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1923 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1924 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1925 phi->init_req(1, signnode);
1926 phi->init_req(2, longcon(0));
1927 correctedsign = _gvn.transform(phi);
1928 ylong_path = _gvn.transform(r);
1929 record_for_igvn(r);
1930 }
1931
1932 // zero node
1933 Node *conzero = longcon(0);
1934 // Check (1&(long)y)==0?
1935 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1936 // Check if (1&(long)y)!=0?, if so the result is negative
1937 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1938 // abs(x)
1939 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1940 // abs(x)^y
1941 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1942 // -abs(x)^y
1943 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1944 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1945 Node *signresult = NULL;
1946 if (ConditionalMoveLimit != 0) {
1947 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1948 } else {
1949 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1950 RegionNode *r = new (C) RegionNode(3);
1951 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1952 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1953 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1954 phi->init_req(1, absxpowy);
1955 phi->init_req(2, negabsxpowy);
1956 signresult = _gvn.transform(phi);
1957 ylong_path = _gvn.transform(r);
1958 record_for_igvn(r);
1959 }
1960 // Set complex path fast result
1961 r->init_req(2, ylong_path);
1962 phi->init_req(2, signresult);
1963
1964 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1965 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1966 r->init_req(1,slow_path);
1967 phi->init_req(1,slow_result);
1968
1969 // Post merge
1970 set_control(_gvn.transform(r));
1971 record_for_igvn(r);
1972 result = _gvn.transform(phi);
1973 }
1974
1975 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1976
1977 // control from finish_pow_exp is now input to the region node
1978 region_node->set_req(2, control());
1979 // the result from finish_pow_exp is now input to the phi node
1980 phi_node->init_req(2, result);
1981 set_control(_gvn.transform(region_node));
1982 record_for_igvn(region_node);
1983 set_result(_gvn.transform(phi_node));
1984
1985 C->set_has_split_ifs(true); // Has chance for split-if optimization
1986 return true;
1987 }
1988
1989 //------------------------------runtime_math-----------------------------
1990 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1991 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1992 "must be (DD)D or (D)D type");
1993
1994 // Inputs
1995 Node* a = round_double_node(argument(0));
1996 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1997
1998 const TypePtr* no_memory_effects = NULL;
1999 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2000 no_memory_effects,
2001 a, top(), b, b ? top() : NULL);
2002 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2003 #ifdef ASSERT
2004 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2005 assert(value_top == top(), "second value must be top");
2006 #endif
2007
2008 set_result(value);
2009 return true;
2010 }
2011
2012 //------------------------------inline_math_native-----------------------------
2013 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2014 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2015 switch (id) {
2016 // These intrinsics are not properly supported on all hardware
2017 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2018 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2019 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2020 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2021 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2022 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2023
2024 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2025 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2026 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2027 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2028
2029 // These intrinsics are supported on all hardware
2030 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2031 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2032
2033 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2034 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2035 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2036 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2037 #undef FN_PTR
2038
2039 // These intrinsics are not yet correctly implemented
2040 case vmIntrinsics::_datan2:
2041 return false;
2042
2043 default:
2044 fatal_unexpected_iid(id);
2045 return false;
2046 }
2047 }
2048
2049 static bool is_simple_name(Node* n) {
2050 return (n->req() == 1 // constant
2051 || (n->is_Type() && n->as_Type()->type()->singleton())
2052 || n->is_Proj() // parameter or return value
2053 || n->is_Phi() // local of some sort
2054 );
2055 }
2056
2057 //----------------------------inline_min_max-----------------------------------
2058 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2059 set_result(generate_min_max(id, argument(0), argument(1)));
2060 return true;
2061 }
2062
2063 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2064 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2065 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2066 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2067 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2068
2069 {
2070 PreserveJVMState pjvms(this);
2071 PreserveReexecuteState preexecs(this);
2072 jvms()->set_should_reexecute(true);
2073
2074 set_control(slow_path);
2075 set_i_o(i_o());
2076
2077 uncommon_trap(Deoptimization::Reason_intrinsic,
2078 Deoptimization::Action_none);
2079 }
2080
2081 set_control(fast_path);
2082 set_result(math);
2083 }
2084
2085 template <typename OverflowOp>
2086 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2087 typedef typename OverflowOp::MathOp MathOp;
2088
2089 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2090 Node* operation = _gvn.transform( mathOp );
2091 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2092 inline_math_mathExact(operation, ofcheck);
2093 return true;
2094 }
2095
2096 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2097 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2098 }
2099
2100 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2101 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2102 }
2103
2104 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2105 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2106 }
2107
2108 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2109 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2110 }
2111
2112 bool LibraryCallKit::inline_math_negateExactI() {
2113 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2114 }
2115
2116 bool LibraryCallKit::inline_math_negateExactL() {
2117 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2118 }
2119
2120 bool LibraryCallKit::inline_math_multiplyExactI() {
2121 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2122 }
2123
2124 bool LibraryCallKit::inline_math_multiplyExactL() {
2125 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2126 }
2127
2128 Node*
2129 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2130 // These are the candidate return value:
2131 Node* xvalue = x0;
2132 Node* yvalue = y0;
2133
2134 if (xvalue == yvalue) {
2135 return xvalue;
2136 }
2137
2138 bool want_max = (id == vmIntrinsics::_max);
2139
2140 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2141 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2142 if (txvalue == NULL || tyvalue == NULL) return top();
2143 // This is not really necessary, but it is consistent with a
2144 // hypothetical MaxINode::Value method:
2145 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2146
2147 // %%% This folding logic should (ideally) be in a different place.
2148 // Some should be inside IfNode, and there to be a more reliable
2149 // transformation of ?: style patterns into cmoves. We also want
2150 // more powerful optimizations around cmove and min/max.
2151
2152 // Try to find a dominating comparison of these guys.
2153 // It can simplify the index computation for Arrays.copyOf
2154 // and similar uses of System.arraycopy.
2155 // First, compute the normalized version of CmpI(x, y).
2156 int cmp_op = Op_CmpI;
2157 Node* xkey = xvalue;
2158 Node* ykey = yvalue;
2159 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2160 if (ideal_cmpxy->is_Cmp()) {
2161 // E.g., if we have CmpI(length - offset, count),
2162 // it might idealize to CmpI(length, count + offset)
2163 cmp_op = ideal_cmpxy->Opcode();
2164 xkey = ideal_cmpxy->in(1);
2165 ykey = ideal_cmpxy->in(2);
2166 }
2167
2168 // Start by locating any relevant comparisons.
2169 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2170 Node* cmpxy = NULL;
2171 Node* cmpyx = NULL;
2172 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2173 Node* cmp = start_from->fast_out(k);
2174 if (cmp->outcnt() > 0 && // must have prior uses
2175 cmp->in(0) == NULL && // must be context-independent
2176 cmp->Opcode() == cmp_op) { // right kind of compare
2177 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2178 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2179 }
2180 }
2181
2182 const int NCMPS = 2;
2183 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2184 int cmpn;
2185 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2186 if (cmps[cmpn] != NULL) break; // find a result
2187 }
2188 if (cmpn < NCMPS) {
2189 // Look for a dominating test that tells us the min and max.
2190 int depth = 0; // Limit search depth for speed
2191 Node* dom = control();
2192 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2193 if (++depth >= 100) break;
2194 Node* ifproj = dom;
2195 if (!ifproj->is_Proj()) continue;
2196 Node* iff = ifproj->in(0);
2197 if (!iff->is_If()) continue;
2198 Node* bol = iff->in(1);
2199 if (!bol->is_Bool()) continue;
2200 Node* cmp = bol->in(1);
2201 if (cmp == NULL) continue;
2202 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2203 if (cmps[cmpn] == cmp) break;
2204 if (cmpn == NCMPS) continue;
2205 BoolTest::mask btest = bol->as_Bool()->_test._test;
2206 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2207 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2208 // At this point, we know that 'x btest y' is true.
2209 switch (btest) {
2210 case BoolTest::eq:
2211 // They are proven equal, so we can collapse the min/max.
2212 // Either value is the answer. Choose the simpler.
2213 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2214 return yvalue;
2215 return xvalue;
2216 case BoolTest::lt: // x < y
2217 case BoolTest::le: // x <= y
2218 return (want_max ? yvalue : xvalue);
2219 case BoolTest::gt: // x > y
2220 case BoolTest::ge: // x >= y
2221 return (want_max ? xvalue : yvalue);
2222 }
2223 }
2224 }
2225
2226 // We failed to find a dominating test.
2227 // Let's pick a test that might GVN with prior tests.
2228 Node* best_bol = NULL;
2229 BoolTest::mask best_btest = BoolTest::illegal;
2230 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2231 Node* cmp = cmps[cmpn];
2232 if (cmp == NULL) continue;
2233 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2234 Node* bol = cmp->fast_out(j);
2235 if (!bol->is_Bool()) continue;
2236 BoolTest::mask btest = bol->as_Bool()->_test._test;
2237 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2238 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2239 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2240 best_bol = bol->as_Bool();
2241 best_btest = btest;
2242 }
2243 }
2244 }
2245
2246 Node* answer_if_true = NULL;
2247 Node* answer_if_false = NULL;
2248 switch (best_btest) {
2249 default:
2250 if (cmpxy == NULL)
2251 cmpxy = ideal_cmpxy;
2252 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2253 // and fall through:
2254 case BoolTest::lt: // x < y
2255 case BoolTest::le: // x <= y
2256 answer_if_true = (want_max ? yvalue : xvalue);
2257 answer_if_false = (want_max ? xvalue : yvalue);
2258 break;
2259 case BoolTest::gt: // x > y
2260 case BoolTest::ge: // x >= y
2261 answer_if_true = (want_max ? xvalue : yvalue);
2262 answer_if_false = (want_max ? yvalue : xvalue);
2263 break;
2264 }
2265
2266 jint hi, lo;
2267 if (want_max) {
2268 // We can sharpen the minimum.
2269 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2270 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2271 } else {
2272 // We can sharpen the maximum.
2273 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2274 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2275 }
2276
2277 // Use a flow-free graph structure, to avoid creating excess control edges
2278 // which could hinder other optimizations.
2279 // Since Math.min/max is often used with arraycopy, we want
2280 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2281 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2282 answer_if_false, answer_if_true,
2283 TypeInt::make(lo, hi, widen));
2284
2285 return _gvn.transform(cmov);
2286
2287 /*
2288 // This is not as desirable as it may seem, since Min and Max
2289 // nodes do not have a full set of optimizations.
2290 // And they would interfere, anyway, with 'if' optimizations
2291 // and with CMoveI canonical forms.
2292 switch (id) {
2293 case vmIntrinsics::_min:
2294 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2295 case vmIntrinsics::_max:
2296 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2297 default:
2298 ShouldNotReachHere();
2299 }
2300 */
2301 }
2302
2303 inline int
2304 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2305 const TypePtr* base_type = TypePtr::NULL_PTR;
2306 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2307 if (base_type == NULL) {
2308 // Unknown type.
2309 return Type::AnyPtr;
2310 } else if (base_type == TypePtr::NULL_PTR) {
2311 // Since this is a NULL+long form, we have to switch to a rawptr.
2312 base = _gvn.transform(new (C) CastX2PNode(offset));
2313 offset = MakeConX(0);
2314 return Type::RawPtr;
2315 } else if (base_type->base() == Type::RawPtr) {
2316 return Type::RawPtr;
2317 } else if (base_type->isa_oopptr()) {
2318 // Base is never null => always a heap address.
2319 if (base_type->ptr() == TypePtr::NotNull) {
2320 return Type::OopPtr;
2321 }
2322 // Offset is small => always a heap address.
2323 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2324 if (offset_type != NULL &&
2325 base_type->offset() == 0 && // (should always be?)
2326 offset_type->_lo >= 0 &&
2327 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2328 return Type::OopPtr;
2329 }
2330 // Otherwise, it might either be oop+off or NULL+addr.
2331 return Type::AnyPtr;
2332 } else {
2333 // No information:
2334 return Type::AnyPtr;
2335 }
2336 }
2337
2338 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2339 int kind = classify_unsafe_addr(base, offset);
2340 if (kind == Type::RawPtr) {
2341 return basic_plus_adr(top(), base, offset);
2342 } else {
2343 return basic_plus_adr(base, offset);
2344 }
2345 }
2346
2347 //--------------------------inline_number_methods-----------------------------
2348 // inline int Integer.numberOfLeadingZeros(int)
2349 // inline int Long.numberOfLeadingZeros(long)
2350 //
2351 // inline int Integer.numberOfTrailingZeros(int)
2352 // inline int Long.numberOfTrailingZeros(long)
2353 //
2354 // inline int Integer.bitCount(int)
2355 // inline int Long.bitCount(long)
2356 //
2357 // inline char Character.reverseBytes(char)
2358 // inline short Short.reverseBytes(short)
2359 // inline int Integer.reverseBytes(int)
2360 // inline long Long.reverseBytes(long)
2361 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2362 Node* arg = argument(0);
2363 Node* n;
2364 switch (id) {
2365 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2366 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2367 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2368 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2369 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2370 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2371 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2372 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2373 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2374 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2375 default: fatal_unexpected_iid(id); break;
2376 }
2377 set_result(_gvn.transform(n));
2378 return true;
2379 }
2380
2381 //----------------------------inline_unsafe_access----------------------------
2382
2383 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2384
2385 // Helper that guards and inserts a pre-barrier.
2386 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2387 Node* pre_val, bool need_mem_bar) {
2388 // We could be accessing the referent field of a reference object. If so, when G1
2389 // is enabled, we need to log the value in the referent field in an SATB buffer.
2390 // This routine performs some compile time filters and generates suitable
2391 // runtime filters that guard the pre-barrier code.
2392 // Also add memory barrier for non volatile load from the referent field
2393 // to prevent commoning of loads across safepoint.
2394 if (!UseG1GC && !need_mem_bar)
2395 return;
2396
2397 // Some compile time checks.
2398
2399 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2400 const TypeX* otype = offset->find_intptr_t_type();
2401 if (otype != NULL && otype->is_con() &&
2402 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2403 // Constant offset but not the reference_offset so just return
2404 return;
2405 }
2406
2407 // We only need to generate the runtime guards for instances.
2408 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2409 if (btype != NULL) {
2410 if (btype->isa_aryptr()) {
2411 // Array type so nothing to do
2412 return;
2413 }
2414
2415 const TypeInstPtr* itype = btype->isa_instptr();
2416 if (itype != NULL) {
2417 // Can the klass of base_oop be statically determined to be
2418 // _not_ a sub-class of Reference and _not_ Object?
2419 ciKlass* klass = itype->klass();
2420 if ( klass->is_loaded() &&
2421 !klass->is_subtype_of(env()->Reference_klass()) &&
2422 !env()->Object_klass()->is_subtype_of(klass)) {
2423 return;
2424 }
2425 }
2426 }
2427
2428 // The compile time filters did not reject base_oop/offset so
2429 // we need to generate the following runtime filters
2430 //
2431 // if (offset == java_lang_ref_Reference::_reference_offset) {
2432 // if (instance_of(base, java.lang.ref.Reference)) {
2433 // pre_barrier(_, pre_val, ...);
2434 // }
2435 // }
2436
2437 float likely = PROB_LIKELY( 0.999);
2438 float unlikely = PROB_UNLIKELY(0.999);
2439
2440 IdealKit ideal(this);
2441 #define __ ideal.
2442
2443 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2444
2445 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2446 // Update graphKit memory and control from IdealKit.
2447 sync_kit(ideal);
2448
2449 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2450 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2451
2452 // Update IdealKit memory and control from graphKit.
2453 __ sync_kit(this);
2454
2455 Node* one = __ ConI(1);
2456 // is_instof == 0 if base_oop == NULL
2457 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2458
2459 // Update graphKit from IdeakKit.
2460 sync_kit(ideal);
2461
2462 // Use the pre-barrier to record the value in the referent field
2463 pre_barrier(false /* do_load */,
2464 __ ctrl(),
2465 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2466 pre_val /* pre_val */,
2467 T_OBJECT);
2468 if (need_mem_bar) {
2469 // Add memory barrier to prevent commoning reads from this field
2470 // across safepoint since GC can change its value.
2471 insert_mem_bar(Op_MemBarCPUOrder);
2472 }
2473 // Update IdealKit from graphKit.
2474 __ sync_kit(this);
2475
2476 } __ end_if(); // _ref_type != ref_none
2477 } __ end_if(); // offset == referent_offset
2478
2479 // Final sync IdealKit and GraphKit.
2480 final_sync(ideal);
2481 #undef __
2482 }
2483
2484
2485 // Interpret Unsafe.fieldOffset cookies correctly:
2486 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2487
2488 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2489 // Attempt to infer a sharper value type from the offset and base type.
2490 ciKlass* sharpened_klass = NULL;
2491
2492 // See if it is an instance field, with an object type.
2493 if (alias_type->field() != NULL) {
2494 assert(!is_native_ptr, "native pointer op cannot use a java address");
2495 if (alias_type->field()->type()->is_klass()) {
2496 sharpened_klass = alias_type->field()->type()->as_klass();
2497 }
2498 }
2499
2500 // See if it is a narrow oop array.
2501 if (adr_type->isa_aryptr()) {
2502 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2503 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2504 if (elem_type != NULL) {
2505 sharpened_klass = elem_type->klass();
2506 }
2507 }
2508 }
2509
2510 // The sharpened class might be unloaded if there is no class loader
2511 // contraint in place.
2512 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2513 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2514
2515 #ifndef PRODUCT
2516 if (C->print_intrinsics() || C->print_inlining()) {
2517 tty->print(" from base type: "); adr_type->dump();
2518 tty->print(" sharpened value: "); tjp->dump();
2519 }
2520 #endif
2521 // Sharpen the value type.
2522 return tjp;
2523 }
2524 return NULL;
2525 }
2526
2527 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2528 if (callee()->is_static()) return false; // caller must have the capability!
2529
2530 #ifndef PRODUCT
2531 {
2532 ResourceMark rm;
2533 // Check the signatures.
2534 ciSignature* sig = callee()->signature();
2535 #ifdef ASSERT
2536 if (!is_store) {
2537 // Object getObject(Object base, int/long offset), etc.
2538 BasicType rtype = sig->return_type()->basic_type();
2539 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2540 rtype = T_ADDRESS; // it is really a C void*
2541 assert(rtype == type, "getter must return the expected value");
2542 if (!is_native_ptr) {
2543 assert(sig->count() == 2, "oop getter has 2 arguments");
2544 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2545 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2546 } else {
2547 assert(sig->count() == 1, "native getter has 1 argument");
2548 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2549 }
2550 } else {
2551 // void putObject(Object base, int/long offset, Object x), etc.
2552 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2553 if (!is_native_ptr) {
2554 assert(sig->count() == 3, "oop putter has 3 arguments");
2555 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2556 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2557 } else {
2558 assert(sig->count() == 2, "native putter has 2 arguments");
2559 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2560 }
2561 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2562 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2563 vtype = T_ADDRESS; // it is really a C void*
2564 assert(vtype == type, "putter must accept the expected value");
2565 }
2566 #endif // ASSERT
2567 }
2568 #endif //PRODUCT
2569
2570 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2571
2572 Node* receiver = argument(0); // type: oop
2573
2574 // Build address expression. See the code in inline_unsafe_prefetch.
2575 Node* adr;
2576 Node* heap_base_oop = top();
2577 Node* offset = top();
2578 Node* val;
2579
2580 if (!is_native_ptr) {
2581 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2582 Node* base = argument(1); // type: oop
2583 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2584 offset = argument(2); // type: long
2585 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2586 // to be plain byte offsets, which are also the same as those accepted
2587 // by oopDesc::field_base.
2588 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2589 "fieldOffset must be byte-scaled");
2590 // 32-bit machines ignore the high half!
2591 offset = ConvL2X(offset);
2592 adr = make_unsafe_address(base, offset);
2593 heap_base_oop = base;
2594 val = is_store ? argument(4) : NULL;
2595 } else {
2596 Node* ptr = argument(1); // type: long
2597 ptr = ConvL2X(ptr); // adjust Java long to machine word
2598 adr = make_unsafe_address(NULL, ptr);
2599 val = is_store ? argument(3) : NULL;
2600 }
2601
2602 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2603
2604 // First guess at the value type.
2605 const Type *value_type = Type::get_const_basic_type(type);
2606
2607 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2608 // there was not enough information to nail it down.
2609 Compile::AliasType* alias_type = C->alias_type(adr_type);
2610 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2611
2612 // We will need memory barriers unless we can determine a unique
2613 // alias category for this reference. (Note: If for some reason
2614 // the barriers get omitted and the unsafe reference begins to "pollute"
2615 // the alias analysis of the rest of the graph, either Compile::can_alias
2616 // or Compile::must_alias will throw a diagnostic assert.)
2617 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2618
2619 // If we are reading the value of the referent field of a Reference
2620 // object (either by using Unsafe directly or through reflection)
2621 // then, if G1 is enabled, we need to record the referent in an
2622 // SATB log buffer using the pre-barrier mechanism.
2623 // Also we need to add memory barrier to prevent commoning reads
2624 // from this field across safepoint since GC can change its value.
2625 bool need_read_barrier = !is_native_ptr && !is_store &&
2626 offset != top() && heap_base_oop != top();
2627
2628 if (!is_store && type == T_OBJECT) {
2629 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2630 if (tjp != NULL) {
2631 value_type = tjp;
2632 }
2633 }
2634
2635 receiver = null_check(receiver);
2636 if (stopped()) {
2637 return true;
2638 }
2639 // Heap pointers get a null-check from the interpreter,
2640 // as a courtesy. However, this is not guaranteed by Unsafe,
2641 // and it is not possible to fully distinguish unintended nulls
2642 // from intended ones in this API.
2643
2644 if (is_volatile) {
2645 // We need to emit leading and trailing CPU membars (see below) in
2646 // addition to memory membars when is_volatile. This is a little
2647 // too strong, but avoids the need to insert per-alias-type
2648 // volatile membars (for stores; compare Parse::do_put_xxx), which
2649 // we cannot do effectively here because we probably only have a
2650 // rough approximation of type.
2651 need_mem_bar = true;
2652 // For Stores, place a memory ordering barrier now.
2653 if (is_store) {
2654 insert_mem_bar(Op_MemBarRelease);
2655 } else {
2656 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2657 insert_mem_bar(Op_MemBarVolatile);
2658 }
2659 }
2660 }
2661
2662 // Memory barrier to prevent normal and 'unsafe' accesses from
2663 // bypassing each other. Happens after null checks, so the
2664 // exception paths do not take memory state from the memory barrier,
2665 // so there's no problems making a strong assert about mixing users
2666 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2667 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2668 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2669
2670 if (!is_store) {
2671 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2672 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2673 // load value
2674 switch (type) {
2675 case T_BOOLEAN:
2676 case T_CHAR:
2677 case T_BYTE:
2678 case T_SHORT:
2679 case T_INT:
2680 case T_LONG:
2681 case T_FLOAT:
2682 case T_DOUBLE:
2683 break;
2684 case T_OBJECT:
2685 if (need_read_barrier) {
2686 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2687 }
2688 break;
2689 case T_ADDRESS:
2690 // Cast to an int type.
2691 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2692 p = ConvX2UL(p);
2693 break;
2694 default:
2695 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2696 break;
2697 }
2698 // The load node has the control of the preceding MemBarCPUOrder. All
2699 // following nodes will have the control of the MemBarCPUOrder inserted at
2700 // the end of this method. So, pushing the load onto the stack at a later
2701 // point is fine.
2702 set_result(p);
2703 } else {
2704 // place effect of store into memory
2705 switch (type) {
2706 case T_DOUBLE:
2707 val = dstore_rounding(val);
2708 break;
2709 case T_ADDRESS:
2710 // Repackage the long as a pointer.
2711 val = ConvL2X(val);
2712 val = _gvn.transform(new (C) CastX2PNode(val));
2713 break;
2714 }
2715
2716 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2717 if (type != T_OBJECT ) {
2718 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2719 } else {
2720 // Possibly an oop being stored to Java heap or native memory
2721 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2722 // oop to Java heap.
2723 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2724 } else {
2725 // We can't tell at compile time if we are storing in the Java heap or outside
2726 // of it. So we need to emit code to conditionally do the proper type of
2727 // store.
2728
2729 IdealKit ideal(this);
2730 #define __ ideal.
2731 // QQQ who knows what probability is here??
2732 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2733 // Sync IdealKit and graphKit.
2734 sync_kit(ideal);
2735 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2736 // Update IdealKit memory.
2737 __ sync_kit(this);
2738 } __ else_(); {
2739 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2740 } __ end_if();
2741 // Final sync IdealKit and GraphKit.
2742 final_sync(ideal);
2743 #undef __
2744 }
2745 }
2746 }
2747
2748 if (is_volatile) {
2749 if (!is_store) {
2750 insert_mem_bar(Op_MemBarAcquire);
2751 } else {
2752 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2753 insert_mem_bar(Op_MemBarVolatile);
2754 }
2755 }
2756 }
2757
2758 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2759
2760 return true;
2761 }
2762
2763 //----------------------------inline_unsafe_prefetch----------------------------
2764
2765 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2766 #ifndef PRODUCT
2767 {
2768 ResourceMark rm;
2769 // Check the signatures.
2770 ciSignature* sig = callee()->signature();
2771 #ifdef ASSERT
2772 // Object getObject(Object base, int/long offset), etc.
2773 BasicType rtype = sig->return_type()->basic_type();
2774 if (!is_native_ptr) {
2775 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2776 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2777 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2778 } else {
2779 assert(sig->count() == 1, "native prefetch has 1 argument");
2780 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2781 }
2782 #endif // ASSERT
2783 }
2784 #endif // !PRODUCT
2785
2786 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2787
2788 const int idx = is_static ? 0 : 1;
2789 if (!is_static) {
2790 null_check_receiver();
2791 if (stopped()) {
2792 return true;
2793 }
2794 }
2795
2796 // Build address expression. See the code in inline_unsafe_access.
2797 Node *adr;
2798 if (!is_native_ptr) {
2799 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2800 Node* base = argument(idx + 0); // type: oop
2801 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2802 Node* offset = argument(idx + 1); // type: long
2803 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2804 // to be plain byte offsets, which are also the same as those accepted
2805 // by oopDesc::field_base.
2806 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2807 "fieldOffset must be byte-scaled");
2808 // 32-bit machines ignore the high half!
2809 offset = ConvL2X(offset);
2810 adr = make_unsafe_address(base, offset);
2811 } else {
2812 Node* ptr = argument(idx + 0); // type: long
2813 ptr = ConvL2X(ptr); // adjust Java long to machine word
2814 adr = make_unsafe_address(NULL, ptr);
2815 }
2816
2817 // Generate the read or write prefetch
2818 Node *prefetch;
2819 if (is_store) {
2820 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2821 } else {
2822 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2823 }
2824 prefetch->init_req(0, control());
2825 set_i_o(_gvn.transform(prefetch));
2826
2827 return true;
2828 }
2829
2830 //----------------------------inline_unsafe_load_store----------------------------
2831 // This method serves a couple of different customers (depending on LoadStoreKind):
2832 //
2833 // LS_cmpxchg:
2834 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2835 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2836 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2837 //
2838 // LS_xadd:
2839 // public int getAndAddInt( Object o, long offset, int delta)
2840 // public long getAndAddLong(Object o, long offset, long delta)
2841 //
2842 // LS_xchg:
2843 // int getAndSet(Object o, long offset, int newValue)
2844 // long getAndSet(Object o, long offset, long newValue)
2845 // Object getAndSet(Object o, long offset, Object newValue)
2846 //
2847 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2848 // This basic scheme here is the same as inline_unsafe_access, but
2849 // differs in enough details that combining them would make the code
2850 // overly confusing. (This is a true fact! I originally combined
2851 // them, but even I was confused by it!) As much code/comments as
2852 // possible are retained from inline_unsafe_access though to make
2853 // the correspondences clearer. - dl
2854
2855 if (callee()->is_static()) return false; // caller must have the capability!
2856
2857 #ifndef PRODUCT
2858 BasicType rtype;
2859 {
2860 ResourceMark rm;
2861 // Check the signatures.
2862 ciSignature* sig = callee()->signature();
2863 rtype = sig->return_type()->basic_type();
2864 if (kind == LS_xadd || kind == LS_xchg) {
2865 // Check the signatures.
2866 #ifdef ASSERT
2867 assert(rtype == type, "get and set must return the expected type");
2868 assert(sig->count() == 3, "get and set has 3 arguments");
2869 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2870 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2871 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2872 #endif // ASSERT
2873 } else if (kind == LS_cmpxchg) {
2874 // Check the signatures.
2875 #ifdef ASSERT
2876 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2877 assert(sig->count() == 4, "CAS has 4 arguments");
2878 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2879 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2880 #endif // ASSERT
2881 } else {
2882 ShouldNotReachHere();
2883 }
2884 }
2885 #endif //PRODUCT
2886
2887 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2888
2889 // Get arguments:
2890 Node* receiver = NULL;
2891 Node* base = NULL;
2892 Node* offset = NULL;
2893 Node* oldval = NULL;
2894 Node* newval = NULL;
2895 if (kind == LS_cmpxchg) {
2896 const bool two_slot_type = type2size[type] == 2;
2897 receiver = argument(0); // type: oop
2898 base = argument(1); // type: oop
2899 offset = argument(2); // type: long
2900 oldval = argument(4); // type: oop, int, or long
2901 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2902 } else if (kind == LS_xadd || kind == LS_xchg){
2903 receiver = argument(0); // type: oop
2904 base = argument(1); // type: oop
2905 offset = argument(2); // type: long
2906 oldval = NULL;
2907 newval = argument(4); // type: oop, int, or long
2908 }
2909
2910 // Null check receiver.
2911 receiver = null_check(receiver);
2912 if (stopped()) {
2913 return true;
2914 }
2915
2916 // Build field offset expression.
2917 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2918 // to be plain byte offsets, which are also the same as those accepted
2919 // by oopDesc::field_base.
2920 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2921 // 32-bit machines ignore the high half of long offsets
2922 offset = ConvL2X(offset);
2923 Node* adr = make_unsafe_address(base, offset);
2924 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2925
2926 // For CAS, unlike inline_unsafe_access, there seems no point in
2927 // trying to refine types. Just use the coarse types here.
2928 const Type *value_type = Type::get_const_basic_type(type);
2929 Compile::AliasType* alias_type = C->alias_type(adr_type);
2930 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2931
2932 if (kind == LS_xchg && type == T_OBJECT) {
2933 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2934 if (tjp != NULL) {
2935 value_type = tjp;
2936 }
2937 }
2938
2939 int alias_idx = C->get_alias_index(adr_type);
2940
2941 // Memory-model-wise, a LoadStore acts like a little synchronized
2942 // block, so needs barriers on each side. These don't translate
2943 // into actual barriers on most machines, but we still need rest of
2944 // compiler to respect ordering.
2945
2946 insert_mem_bar(Op_MemBarRelease);
2947 insert_mem_bar(Op_MemBarCPUOrder);
2948
2949 // 4984716: MemBars must be inserted before this
2950 // memory node in order to avoid a false
2951 // dependency which will confuse the scheduler.
2952 Node *mem = memory(alias_idx);
2953
2954 // For now, we handle only those cases that actually exist: ints,
2955 // longs, and Object. Adding others should be straightforward.
2956 Node* load_store;
2957 switch(type) {
2958 case T_INT:
2959 if (kind == LS_xadd) {
2960 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2961 } else if (kind == LS_xchg) {
2962 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2963 } else if (kind == LS_cmpxchg) {
2964 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2965 } else {
2966 ShouldNotReachHere();
2967 }
2968 break;
2969 case T_LONG:
2970 if (kind == LS_xadd) {
2971 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2972 } else if (kind == LS_xchg) {
2973 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2974 } else if (kind == LS_cmpxchg) {
2975 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2976 } else {
2977 ShouldNotReachHere();
2978 }
2979 break;
2980 case T_OBJECT:
2981 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2982 // could be delayed during Parse (for example, in adjust_map_after_if()).
2983 // Execute transformation here to avoid barrier generation in such case.
2984 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2985 newval = _gvn.makecon(TypePtr::NULL_PTR);
2986
2987 // Reference stores need a store barrier.
2988 if (kind == LS_xchg) {
2989 // If pre-barrier must execute before the oop store, old value will require do_load here.
2990 if (!can_move_pre_barrier()) {
2991 pre_barrier(true /* do_load*/,
2992 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2993 NULL /* pre_val*/,
2994 T_OBJECT);
2995 } // Else move pre_barrier to use load_store value, see below.
2996 } else if (kind == LS_cmpxchg) {
2997 // Same as for newval above:
2998 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2999 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3000 }
3001 // The only known value which might get overwritten is oldval.
3002 pre_barrier(false /* do_load */,
3003 control(), NULL, NULL, max_juint, NULL, NULL,
3004 oldval /* pre_val */,
3005 T_OBJECT);
3006 } else {
3007 ShouldNotReachHere();
3008 }
3009
3010 #ifdef _LP64
3011 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3012 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3013 if (kind == LS_xchg) {
3014 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3015 newval_enc, adr_type, value_type->make_narrowoop()));
3016 } else {
3017 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3018 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3019 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3020 newval_enc, oldval_enc));
3021 }
3022 } else
3023 #endif
3024 {
3025 if (kind == LS_xchg) {
3026 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3027 } else {
3028 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3029 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3030 }
3031 }
3032 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3033 break;
3034 default:
3035 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3036 break;
3037 }
3038
3039 // SCMemProjNodes represent the memory state of a LoadStore. Their
3040 // main role is to prevent LoadStore nodes from being optimized away
3041 // when their results aren't used.
3042 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3043 set_memory(proj, alias_idx);
3044
3045 if (type == T_OBJECT && kind == LS_xchg) {
3046 #ifdef _LP64
3047 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3048 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3049 }
3050 #endif
3051 if (can_move_pre_barrier()) {
3052 // Don't need to load pre_val. The old value is returned by load_store.
3053 // The pre_barrier can execute after the xchg as long as no safepoint
3054 // gets inserted between them.
3055 pre_barrier(false /* do_load */,
3056 control(), NULL, NULL, max_juint, NULL, NULL,
3057 load_store /* pre_val */,
3058 T_OBJECT);
3059 }
3060 }
3061
3062 // Add the trailing membar surrounding the access
3063 insert_mem_bar(Op_MemBarCPUOrder);
3064 insert_mem_bar(Op_MemBarAcquire);
3065
3066 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3067 set_result(load_store);
3068 return true;
3069 }
3070
3071 //----------------------------inline_unsafe_ordered_store----------------------
3072 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3073 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3074 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3075 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3076 // This is another variant of inline_unsafe_access, differing in
3077 // that it always issues store-store ("release") barrier and ensures
3078 // store-atomicity (which only matters for "long").
3079
3080 if (callee()->is_static()) return false; // caller must have the capability!
3081
3082 #ifndef PRODUCT
3083 {
3084 ResourceMark rm;
3085 // Check the signatures.
3086 ciSignature* sig = callee()->signature();
3087 #ifdef ASSERT
3088 BasicType rtype = sig->return_type()->basic_type();
3089 assert(rtype == T_VOID, "must return void");
3090 assert(sig->count() == 3, "has 3 arguments");
3091 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3092 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3093 #endif // ASSERT
3094 }
3095 #endif //PRODUCT
3096
3097 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3098
3099 // Get arguments:
3100 Node* receiver = argument(0); // type: oop
3101 Node* base = argument(1); // type: oop
3102 Node* offset = argument(2); // type: long
3103 Node* val = argument(4); // type: oop, int, or long
3104
3105 // Null check receiver.
3106 receiver = null_check(receiver);
3107 if (stopped()) {
3108 return true;
3109 }
3110
3111 // Build field offset expression.
3112 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3113 // 32-bit machines ignore the high half of long offsets
3114 offset = ConvL2X(offset);
3115 Node* adr = make_unsafe_address(base, offset);
3116 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3117 const Type *value_type = Type::get_const_basic_type(type);
3118 Compile::AliasType* alias_type = C->alias_type(adr_type);
3119
3120 insert_mem_bar(Op_MemBarRelease);
3121 insert_mem_bar(Op_MemBarCPUOrder);
3122 // Ensure that the store is atomic for longs:
3123 const bool require_atomic_access = true;
3124 Node* store;
3125 if (type == T_OBJECT) // reference stores need a store barrier.
3126 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3127 else {
3128 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3129 }
3130 insert_mem_bar(Op_MemBarCPUOrder);
3131 return true;
3132 }
3133
3134 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3135 // Regardless of form, don't allow previous ld/st to move down,
3136 // then issue acquire, release, or volatile mem_bar.
3137 insert_mem_bar(Op_MemBarCPUOrder);
3138 switch(id) {
3139 case vmIntrinsics::_loadFence:
3140 insert_mem_bar(Op_LoadFence);
3141 return true;
3142 case vmIntrinsics::_storeFence:
3143 insert_mem_bar(Op_StoreFence);
3144 return true;
3145 case vmIntrinsics::_fullFence:
3146 insert_mem_bar(Op_MemBarVolatile);
3147 return true;
3148 default:
3149 fatal_unexpected_iid(id);
3150 return false;
3151 }
3152 }
3153
3154 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3155 if (!kls->is_Con()) {
3156 return true;
3157 }
3158 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3159 if (klsptr == NULL) {
3160 return true;
3161 }
3162 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3163 // don't need a guard for a klass that is already initialized
3164 return !ik->is_initialized();
3165 }
3166
3167 //----------------------------inline_unsafe_allocate---------------------------
3168 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3169 bool LibraryCallKit::inline_unsafe_allocate() {
3170 if (callee()->is_static()) return false; // caller must have the capability!
3171
3172 null_check_receiver(); // null-check, then ignore
3173 Node* cls = null_check(argument(1));
3174 if (stopped()) return true;
3175
3176 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3177 kls = null_check(kls);
3178 if (stopped()) return true; // argument was like int.class
3179
3180 Node* test = NULL;
3181 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3182 // Note: The argument might still be an illegal value like
3183 // Serializable.class or Object[].class. The runtime will handle it.
3184 // But we must make an explicit check for initialization.
3185 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3186 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3187 // can generate code to load it as unsigned byte.
3188 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3189 Node* bits = intcon(InstanceKlass::fully_initialized);
3190 test = _gvn.transform(new (C) SubINode(inst, bits));
3191 // The 'test' is non-zero if we need to take a slow path.
3192 }
3193
3194 Node* obj = new_instance(kls, test);
3195 set_result(obj);
3196 return true;
3197 }
3198
3199 #ifdef TRACE_HAVE_INTRINSICS
3200 /*
3201 * oop -> myklass
3202 * myklass->trace_id |= USED
3203 * return myklass->trace_id & ~0x3
3204 */
3205 bool LibraryCallKit::inline_native_classID() {
3206 null_check_receiver(); // null-check, then ignore
3207 Node* cls = null_check(argument(1), T_OBJECT);
3208 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3209 kls = null_check(kls, T_OBJECT);
3210 ByteSize offset = TRACE_ID_OFFSET;
3211 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3212 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3213 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3214 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3215 Node* clsused = longcon(0x01l); // set the class bit
3216 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3217
3218 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3219 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3220 set_result(andl);
3221 return true;
3222 }
3223
3224 bool LibraryCallKit::inline_native_threadID() {
3225 Node* tls_ptr = NULL;
3226 Node* cur_thr = generate_current_thread(tls_ptr);
3227 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3228 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3229 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3230
3231 Node* threadid = NULL;
3232 size_t thread_id_size = OSThread::thread_id_size();
3233 if (thread_id_size == (size_t) BytesPerLong) {
3234 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3235 } else if (thread_id_size == (size_t) BytesPerInt) {
3236 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3237 } else {
3238 ShouldNotReachHere();
3239 }
3240 set_result(threadid);
3241 return true;
3242 }
3243 #endif
3244
3245 //------------------------inline_native_time_funcs--------------
3246 // inline code for System.currentTimeMillis() and System.nanoTime()
3247 // these have the same type and signature
3248 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3249 const TypeFunc* tf = OptoRuntime::void_long_Type();
3250 const TypePtr* no_memory_effects = NULL;
3251 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3252 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3253 #ifdef ASSERT
3254 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3255 assert(value_top == top(), "second value must be top");
3256 #endif
3257 set_result(value);
3258 return true;
3259 }
3260
3261 //------------------------inline_native_currentThread------------------
3262 bool LibraryCallKit::inline_native_currentThread() {
3263 Node* junk = NULL;
3264 set_result(generate_current_thread(junk));
3265 return true;
3266 }
3267
3268 //------------------------inline_native_isInterrupted------------------
3269 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3270 bool LibraryCallKit::inline_native_isInterrupted() {
3271 // Add a fast path to t.isInterrupted(clear_int):
3272 // (t == Thread.current() &&
3273 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3274 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3275 // So, in the common case that the interrupt bit is false,
3276 // we avoid making a call into the VM. Even if the interrupt bit
3277 // is true, if the clear_int argument is false, we avoid the VM call.
3278 // However, if the receiver is not currentThread, we must call the VM,
3279 // because there must be some locking done around the operation.
3280
3281 // We only go to the fast case code if we pass two guards.
3282 // Paths which do not pass are accumulated in the slow_region.
3283
3284 enum {
3285 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3286 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3287 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3288 PATH_LIMIT
3289 };
3290
3291 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3292 // out of the function.
3293 insert_mem_bar(Op_MemBarCPUOrder);
3294
3295 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3296 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3297
3298 RegionNode* slow_region = new (C) RegionNode(1);
3299 record_for_igvn(slow_region);
3300
3301 // (a) Receiving thread must be the current thread.
3302 Node* rec_thr = argument(0);
3303 Node* tls_ptr = NULL;
3304 Node* cur_thr = generate_current_thread(tls_ptr);
3305 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3306 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3307
3308 generate_slow_guard(bol_thr, slow_region);
3309
3310 // (b) Interrupt bit on TLS must be false.
3311 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3312 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3313 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3314
3315 // Set the control input on the field _interrupted read to prevent it floating up.
3316 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3317 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3318 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3319
3320 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3321
3322 // First fast path: if (!TLS._interrupted) return false;
3323 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3324 result_rgn->init_req(no_int_result_path, false_bit);
3325 result_val->init_req(no_int_result_path, intcon(0));
3326
3327 // drop through to next case
3328 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3329
3330 #ifndef TARGET_OS_FAMILY_windows
3331 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3332 Node* clr_arg = argument(1);
3333 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3334 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3335 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3336
3337 // Second fast path: ... else if (!clear_int) return true;
3338 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3339 result_rgn->init_req(no_clear_result_path, false_arg);
3340 result_val->init_req(no_clear_result_path, intcon(1));
3341
3342 // drop through to next case
3343 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3344 #else
3345 // To return true on Windows you must read the _interrupted field
3346 // and check the the event state i.e. take the slow path.
3347 #endif // TARGET_OS_FAMILY_windows
3348
3349 // (d) Otherwise, go to the slow path.
3350 slow_region->add_req(control());
3351 set_control( _gvn.transform(slow_region));
3352
3353 if (stopped()) {
3354 // There is no slow path.
3355 result_rgn->init_req(slow_result_path, top());
3356 result_val->init_req(slow_result_path, top());
3357 } else {
3358 // non-virtual because it is a private non-static
3359 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3360
3361 Node* slow_val = set_results_for_java_call(slow_call);
3362 // this->control() comes from set_results_for_java_call
3363
3364 Node* fast_io = slow_call->in(TypeFunc::I_O);
3365 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3366
3367 // These two phis are pre-filled with copies of of the fast IO and Memory
3368 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3369 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3370
3371 result_rgn->init_req(slow_result_path, control());
3372 result_io ->init_req(slow_result_path, i_o());
3373 result_mem->init_req(slow_result_path, reset_memory());
3374 result_val->init_req(slow_result_path, slow_val);
3375
3376 set_all_memory(_gvn.transform(result_mem));
3377 set_i_o( _gvn.transform(result_io));
3378 }
3379
3380 C->set_has_split_ifs(true); // Has chance for split-if optimization
3381 set_result(result_rgn, result_val);
3382 return true;
3383 }
3384
3385 //---------------------------load_mirror_from_klass----------------------------
3386 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3387 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3388 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3389 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3390 }
3391
3392 //-----------------------load_klass_from_mirror_common-------------------------
3393 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3394 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3395 // and branch to the given path on the region.
3396 // If never_see_null, take an uncommon trap on null, so we can optimistically
3397 // compile for the non-null case.
3398 // If the region is NULL, force never_see_null = true.
3399 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3400 bool never_see_null,
3401 RegionNode* region,
3402 int null_path,
3403 int offset) {
3404 if (region == NULL) never_see_null = true;
3405 Node* p = basic_plus_adr(mirror, offset);
3406 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3407 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3408 Node* null_ctl = top();
3409 kls = null_check_oop(kls, &null_ctl, never_see_null);
3410 if (region != NULL) {
3411 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3412 region->init_req(null_path, null_ctl);
3413 } else {
3414 assert(null_ctl == top(), "no loose ends");
3415 }
3416 return kls;
3417 }
3418
3419 //--------------------(inline_native_Class_query helpers)---------------------
3420 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3421 // Fall through if (mods & mask) == bits, take the guard otherwise.
3422 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3423 // Branch around if the given klass has the given modifier bit set.
3424 // Like generate_guard, adds a new path onto the region.
3425 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3426 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3427 Node* mask = intcon(modifier_mask);
3428 Node* bits = intcon(modifier_bits);
3429 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3430 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3431 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3432 return generate_fair_guard(bol, region);
3433 }
3434 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3435 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3436 }
3437
3438 //-------------------------inline_native_Class_query-------------------
3439 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3440 const Type* return_type = TypeInt::BOOL;
3441 Node* prim_return_value = top(); // what happens if it's a primitive class?
3442 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3443 bool expect_prim = false; // most of these guys expect to work on refs
3444
3445 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3446
3447 Node* mirror = argument(0);
3448 Node* obj = top();
3449
3450 switch (id) {
3451 case vmIntrinsics::_isInstance:
3452 // nothing is an instance of a primitive type
3453 prim_return_value = intcon(0);
3454 obj = argument(1);
3455 break;
3456 case vmIntrinsics::_getModifiers:
3457 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3458 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3459 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3460 break;
3461 case vmIntrinsics::_isInterface:
3462 prim_return_value = intcon(0);
3463 break;
3464 case vmIntrinsics::_isArray:
3465 prim_return_value = intcon(0);
3466 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3467 break;
3468 case vmIntrinsics::_isPrimitive:
3469 prim_return_value = intcon(1);
3470 expect_prim = true; // obviously
3471 break;
3472 case vmIntrinsics::_getSuperclass:
3473 prim_return_value = null();
3474 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3475 break;
3476 case vmIntrinsics::_getComponentType:
3477 prim_return_value = null();
3478 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3479 break;
3480 case vmIntrinsics::_getClassAccessFlags:
3481 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3482 return_type = TypeInt::INT; // not bool! 6297094
3483 break;
3484 default:
3485 fatal_unexpected_iid(id);
3486 break;
3487 }
3488
3489 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3490 if (mirror_con == NULL) return false; // cannot happen?
3491
3492 #ifndef PRODUCT
3493 if (C->print_intrinsics() || C->print_inlining()) {
3494 ciType* k = mirror_con->java_mirror_type();
3495 if (k) {
3496 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3497 k->print_name();
3498 tty->cr();
3499 }
3500 }
3501 #endif
3502
3503 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3504 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3505 record_for_igvn(region);
3506 PhiNode* phi = new (C) PhiNode(region, return_type);
3507
3508 // The mirror will never be null of Reflection.getClassAccessFlags, however
3509 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3510 // if it is. See bug 4774291.
3511
3512 // For Reflection.getClassAccessFlags(), the null check occurs in
3513 // the wrong place; see inline_unsafe_access(), above, for a similar
3514 // situation.
3515 mirror = null_check(mirror);
3516 // If mirror or obj is dead, only null-path is taken.
3517 if (stopped()) return true;
3518
3519 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3520
3521 // Now load the mirror's klass metaobject, and null-check it.
3522 // Side-effects region with the control path if the klass is null.
3523 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3524 // If kls is null, we have a primitive mirror.
3525 phi->init_req(_prim_path, prim_return_value);
3526 if (stopped()) { set_result(region, phi); return true; }
3527 bool safe_for_replace = (region->in(_prim_path) == top());
3528
3529 Node* p; // handy temp
3530 Node* null_ctl;
3531
3532 // Now that we have the non-null klass, we can perform the real query.
3533 // For constant classes, the query will constant-fold in LoadNode::Value.
3534 Node* query_value = top();
3535 switch (id) {
3536 case vmIntrinsics::_isInstance:
3537 // nothing is an instance of a primitive type
3538 query_value = gen_instanceof(obj, kls, safe_for_replace);
3539 break;
3540
3541 case vmIntrinsics::_getModifiers:
3542 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3543 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3544 break;
3545
3546 case vmIntrinsics::_isInterface:
3547 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3548 if (generate_interface_guard(kls, region) != NULL)
3549 // A guard was added. If the guard is taken, it was an interface.
3550 phi->add_req(intcon(1));
3551 // If we fall through, it's a plain class.
3552 query_value = intcon(0);
3553 break;
3554
3555 case vmIntrinsics::_isArray:
3556 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3557 if (generate_array_guard(kls, region) != NULL)
3558 // A guard was added. If the guard is taken, it was an array.
3559 phi->add_req(intcon(1));
3560 // If we fall through, it's a plain class.
3561 query_value = intcon(0);
3562 break;
3563
3564 case vmIntrinsics::_isPrimitive:
3565 query_value = intcon(0); // "normal" path produces false
3566 break;
3567
3568 case vmIntrinsics::_getSuperclass:
3569 // The rules here are somewhat unfortunate, but we can still do better
3570 // with random logic than with a JNI call.
3571 // Interfaces store null or Object as _super, but must report null.
3572 // Arrays store an intermediate super as _super, but must report Object.
3573 // Other types can report the actual _super.
3574 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3575 if (generate_interface_guard(kls, region) != NULL)
3576 // A guard was added. If the guard is taken, it was an interface.
3577 phi->add_req(null());
3578 if (generate_array_guard(kls, region) != NULL)
3579 // A guard was added. If the guard is taken, it was an array.
3580 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3581 // If we fall through, it's a plain class. Get its _super.
3582 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3583 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3584 null_ctl = top();
3585 kls = null_check_oop(kls, &null_ctl);
3586 if (null_ctl != top()) {
3587 // If the guard is taken, Object.superClass is null (both klass and mirror).
3588 region->add_req(null_ctl);
3589 phi ->add_req(null());
3590 }
3591 if (!stopped()) {
3592 query_value = load_mirror_from_klass(kls);
3593 }
3594 break;
3595
3596 case vmIntrinsics::_getComponentType:
3597 if (generate_array_guard(kls, region) != NULL) {
3598 // Be sure to pin the oop load to the guard edge just created:
3599 Node* is_array_ctrl = region->in(region->req()-1);
3600 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3601 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3602 phi->add_req(cmo);
3603 }
3604 query_value = null(); // non-array case is null
3605 break;
3606
3607 case vmIntrinsics::_getClassAccessFlags:
3608 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3609 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3610 break;
3611
3612 default:
3613 fatal_unexpected_iid(id);
3614 break;
3615 }
3616
3617 // Fall-through is the normal case of a query to a real class.
3618 phi->init_req(1, query_value);
3619 region->init_req(1, control());
3620
3621 C->set_has_split_ifs(true); // Has chance for split-if optimization
3622 set_result(region, phi);
3623 return true;
3624 }
3625
3626 //--------------------------inline_native_subtype_check------------------------
3627 // This intrinsic takes the JNI calls out of the heart of
3628 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3629 bool LibraryCallKit::inline_native_subtype_check() {
3630 // Pull both arguments off the stack.
3631 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3632 args[0] = argument(0);
3633 args[1] = argument(1);
3634 Node* klasses[2]; // corresponding Klasses: superk, subk
3635 klasses[0] = klasses[1] = top();
3636
3637 enum {
3638 // A full decision tree on {superc is prim, subc is prim}:
3639 _prim_0_path = 1, // {P,N} => false
3640 // {P,P} & superc!=subc => false
3641 _prim_same_path, // {P,P} & superc==subc => true
3642 _prim_1_path, // {N,P} => false
3643 _ref_subtype_path, // {N,N} & subtype check wins => true
3644 _both_ref_path, // {N,N} & subtype check loses => false
3645 PATH_LIMIT
3646 };
3647
3648 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3649 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3650 record_for_igvn(region);
3651
3652 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3653 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3654 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3655
3656 // First null-check both mirrors and load each mirror's klass metaobject.
3657 int which_arg;
3658 for (which_arg = 0; which_arg <= 1; which_arg++) {
3659 Node* arg = args[which_arg];
3660 arg = null_check(arg);
3661 if (stopped()) break;
3662 args[which_arg] = arg;
3663
3664 Node* p = basic_plus_adr(arg, class_klass_offset);
3665 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3666 klasses[which_arg] = _gvn.transform(kls);
3667 }
3668
3669 // Having loaded both klasses, test each for null.
3670 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3671 for (which_arg = 0; which_arg <= 1; which_arg++) {
3672 Node* kls = klasses[which_arg];
3673 Node* null_ctl = top();
3674 kls = null_check_oop(kls, &null_ctl, never_see_null);
3675 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3676 region->init_req(prim_path, null_ctl);
3677 if (stopped()) break;
3678 klasses[which_arg] = kls;
3679 }
3680
3681 if (!stopped()) {
3682 // now we have two reference types, in klasses[0..1]
3683 Node* subk = klasses[1]; // the argument to isAssignableFrom
3684 Node* superk = klasses[0]; // the receiver
3685 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3686 // now we have a successful reference subtype check
3687 region->set_req(_ref_subtype_path, control());
3688 }
3689
3690 // If both operands are primitive (both klasses null), then
3691 // we must return true when they are identical primitives.
3692 // It is convenient to test this after the first null klass check.
3693 set_control(region->in(_prim_0_path)); // go back to first null check
3694 if (!stopped()) {
3695 // Since superc is primitive, make a guard for the superc==subc case.
3696 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3697 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3698 generate_guard(bol_eq, region, PROB_FAIR);
3699 if (region->req() == PATH_LIMIT+1) {
3700 // A guard was added. If the added guard is taken, superc==subc.
3701 region->swap_edges(PATH_LIMIT, _prim_same_path);
3702 region->del_req(PATH_LIMIT);
3703 }
3704 region->set_req(_prim_0_path, control()); // Not equal after all.
3705 }
3706
3707 // these are the only paths that produce 'true':
3708 phi->set_req(_prim_same_path, intcon(1));
3709 phi->set_req(_ref_subtype_path, intcon(1));
3710
3711 // pull together the cases:
3712 assert(region->req() == PATH_LIMIT, "sane region");
3713 for (uint i = 1; i < region->req(); i++) {
3714 Node* ctl = region->in(i);
3715 if (ctl == NULL || ctl == top()) {
3716 region->set_req(i, top());
3717 phi ->set_req(i, top());
3718 } else if (phi->in(i) == NULL) {
3719 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3720 }
3721 }
3722
3723 set_control(_gvn.transform(region));
3724 set_result(_gvn.transform(phi));
3725 return true;
3726 }
3727
3728 //---------------------generate_array_guard_common------------------------
3729 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3730 bool obj_array, bool not_array) {
3731 // If obj_array/non_array==false/false:
3732 // Branch around if the given klass is in fact an array (either obj or prim).
3733 // If obj_array/non_array==false/true:
3734 // Branch around if the given klass is not an array klass of any kind.
3735 // If obj_array/non_array==true/true:
3736 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3737 // If obj_array/non_array==true/false:
3738 // Branch around if the kls is an oop array (Object[] or subtype)
3739 //
3740 // Like generate_guard, adds a new path onto the region.
3741 jint layout_con = 0;
3742 Node* layout_val = get_layout_helper(kls, layout_con);
3743 if (layout_val == NULL) {
3744 bool query = (obj_array
3745 ? Klass::layout_helper_is_objArray(layout_con)
3746 : Klass::layout_helper_is_array(layout_con));
3747 if (query == not_array) {
3748 return NULL; // never a branch
3749 } else { // always a branch
3750 Node* always_branch = control();
3751 if (region != NULL)
3752 region->add_req(always_branch);
3753 set_control(top());
3754 return always_branch;
3755 }
3756 }
3757 // Now test the correct condition.
3758 jint nval = (obj_array
3759 ? ((jint)Klass::_lh_array_tag_type_value
3760 << Klass::_lh_array_tag_shift)
3761 : Klass::_lh_neutral_value);
3762 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3763 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3764 // invert the test if we are looking for a non-array
3765 if (not_array) btest = BoolTest(btest).negate();
3766 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3767 return generate_fair_guard(bol, region);
3768 }
3769
3770
3771 //-----------------------inline_native_newArray--------------------------
3772 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3773 bool LibraryCallKit::inline_native_newArray() {
3774 Node* mirror = argument(0);
3775 Node* count_val = argument(1);
3776
3777 mirror = null_check(mirror);
3778 // If mirror or obj is dead, only null-path is taken.
3779 if (stopped()) return true;
3780
3781 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3782 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3783 PhiNode* result_val = new(C) PhiNode(result_reg,
3784 TypeInstPtr::NOTNULL);
3785 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3786 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3787 TypePtr::BOTTOM);
3788
3789 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3790 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3791 result_reg, _slow_path);
3792 Node* normal_ctl = control();
3793 Node* no_array_ctl = result_reg->in(_slow_path);
3794
3795 // Generate code for the slow case. We make a call to newArray().
3796 set_control(no_array_ctl);
3797 if (!stopped()) {
3798 // Either the input type is void.class, or else the
3799 // array klass has not yet been cached. Either the
3800 // ensuing call will throw an exception, or else it
3801 // will cache the array klass for next time.
3802 PreserveJVMState pjvms(this);
3803 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3804 Node* slow_result = set_results_for_java_call(slow_call);
3805 // this->control() comes from set_results_for_java_call
3806 result_reg->set_req(_slow_path, control());
3807 result_val->set_req(_slow_path, slow_result);
3808 result_io ->set_req(_slow_path, i_o());
3809 result_mem->set_req(_slow_path, reset_memory());
3810 }
3811
3812 set_control(normal_ctl);
3813 if (!stopped()) {
3814 // Normal case: The array type has been cached in the java.lang.Class.
3815 // The following call works fine even if the array type is polymorphic.
3816 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3817 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3818 result_reg->init_req(_normal_path, control());
3819 result_val->init_req(_normal_path, obj);
3820 result_io ->init_req(_normal_path, i_o());
3821 result_mem->init_req(_normal_path, reset_memory());
3822 }
3823
3824 // Return the combined state.
3825 set_i_o( _gvn.transform(result_io) );
3826 set_all_memory( _gvn.transform(result_mem));
3827
3828 C->set_has_split_ifs(true); // Has chance for split-if optimization
3829 set_result(result_reg, result_val);
3830 return true;
3831 }
3832
3833 //----------------------inline_native_getLength--------------------------
3834 // public static native int java.lang.reflect.Array.getLength(Object array);
3835 bool LibraryCallKit::inline_native_getLength() {
3836 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3837
3838 Node* array = null_check(argument(0));
3839 // If array is dead, only null-path is taken.
3840 if (stopped()) return true;
3841
3842 // Deoptimize if it is a non-array.
3843 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3844
3845 if (non_array != NULL) {
3846 PreserveJVMState pjvms(this);
3847 set_control(non_array);
3848 uncommon_trap(Deoptimization::Reason_intrinsic,
3849 Deoptimization::Action_maybe_recompile);
3850 }
3851
3852 // If control is dead, only non-array-path is taken.
3853 if (stopped()) return true;
3854
3855 // The works fine even if the array type is polymorphic.
3856 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3857 Node* result = load_array_length(array);
3858
3859 C->set_has_split_ifs(true); // Has chance for split-if optimization
3860 set_result(result);
3861 return true;
3862 }
3863
3864 //------------------------inline_array_copyOf----------------------------
3865 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3866 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3867 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3868 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3869
3870 // Get the arguments.
3871 Node* original = argument(0);
3872 Node* start = is_copyOfRange? argument(1): intcon(0);
3873 Node* end = is_copyOfRange? argument(2): argument(1);
3874 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3875
3876 Node* newcopy;
3877
3878 // Set the original stack and the reexecute bit for the interpreter to reexecute
3879 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3880 { PreserveReexecuteState preexecs(this);
3881 jvms()->set_should_reexecute(true);
3882
3883 array_type_mirror = null_check(array_type_mirror);
3884 original = null_check(original);
3885
3886 // Check if a null path was taken unconditionally.
3887 if (stopped()) return true;
3888
3889 Node* orig_length = load_array_length(original);
3890
3891 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3892 klass_node = null_check(klass_node);
3893
3894 RegionNode* bailout = new (C) RegionNode(1);
3895 record_for_igvn(bailout);
3896
3897 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3898 // Bail out if that is so.
3899 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3900 if (not_objArray != NULL) {
3901 // Improve the klass node's type from the new optimistic assumption:
3902 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3903 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3904 Node* cast = new (C) CastPPNode(klass_node, akls);
3905 cast->init_req(0, control());
3906 klass_node = _gvn.transform(cast);
3907 }
3908
3909 // Bail out if either start or end is negative.
3910 generate_negative_guard(start, bailout, &start);
3911 generate_negative_guard(end, bailout, &end);
3912
3913 Node* length = end;
3914 if (_gvn.type(start) != TypeInt::ZERO) {
3915 length = _gvn.transform(new (C) SubINode(end, start));
3916 }
3917
3918 // Bail out if length is negative.
3919 // Without this the new_array would throw
3920 // NegativeArraySizeException but IllegalArgumentException is what
3921 // should be thrown
3922 generate_negative_guard(length, bailout, &length);
3923
3924 if (bailout->req() > 1) {
3925 PreserveJVMState pjvms(this);
3926 set_control(_gvn.transform(bailout));
3927 uncommon_trap(Deoptimization::Reason_intrinsic,
3928 Deoptimization::Action_maybe_recompile);
3929 }
3930
3931 if (!stopped()) {
3932 // How many elements will we copy from the original?
3933 // The answer is MinI(orig_length - start, length).
3934 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3935 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3936
3937 newcopy = new_array(klass_node, length, 0); // no argments to push
3938
3939 // Generate a direct call to the right arraycopy function(s).
3940 // We know the copy is disjoint but we might not know if the
3941 // oop stores need checking.
3942 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3943 // This will fail a store-check if x contains any non-nulls.
3944 bool disjoint_bases = true;
3945 // if start > orig_length then the length of the copy may be
3946 // negative.
3947 bool length_never_negative = !is_copyOfRange;
3948 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3949 original, start, newcopy, intcon(0), moved,
3950 disjoint_bases, length_never_negative);
3951 }
3952 } // original reexecute is set back here
3953
3954 C->set_has_split_ifs(true); // Has chance for split-if optimization
3955 if (!stopped()) {
3956 set_result(newcopy);
3957 }
3958 return true;
3959 }
3960
3961
3962 //----------------------generate_virtual_guard---------------------------
3963 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3964 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3965 RegionNode* slow_region) {
3966 ciMethod* method = callee();
3967 int vtable_index = method->vtable_index();
3968 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3969 err_msg_res("bad index %d", vtable_index));
3970 // Get the Method* out of the appropriate vtable entry.
3971 int entry_offset = (InstanceKlass::vtable_start_offset() +
3972 vtable_index*vtableEntry::size()) * wordSize +
3973 vtableEntry::method_offset_in_bytes();
3974 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3975 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3976
3977 // Compare the target method with the expected method (e.g., Object.hashCode).
3978 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3979
3980 Node* native_call = makecon(native_call_addr);
3981 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3982 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3983
3984 return generate_slow_guard(test_native, slow_region);
3985 }
3986
3987 //-----------------------generate_method_call----------------------------
3988 // Use generate_method_call to make a slow-call to the real
3989 // method if the fast path fails. An alternative would be to
3990 // use a stub like OptoRuntime::slow_arraycopy_Java.
3991 // This only works for expanding the current library call,
3992 // not another intrinsic. (E.g., don't use this for making an
3993 // arraycopy call inside of the copyOf intrinsic.)
3994 CallJavaNode*
3995 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3996 // When compiling the intrinsic method itself, do not use this technique.
3997 guarantee(callee() != C->method(), "cannot make slow-call to self");
3998
3999 ciMethod* method = callee();
4000 // ensure the JVMS we have will be correct for this call
4001 guarantee(method_id == method->intrinsic_id(), "must match");
4002
4003 const TypeFunc* tf = TypeFunc::make(method);
4004 CallJavaNode* slow_call;
4005 if (is_static) {
4006 assert(!is_virtual, "");
4007 slow_call = new(C) CallStaticJavaNode(C, tf,
4008 SharedRuntime::get_resolve_static_call_stub(),
4009 method, bci());
4010 } else if (is_virtual) {
4011 null_check_receiver();
4012 int vtable_index = Method::invalid_vtable_index;
4013 if (UseInlineCaches) {
4014 // Suppress the vtable call
4015 } else {
4016 // hashCode and clone are not a miranda methods,
4017 // so the vtable index is fixed.
4018 // No need to use the linkResolver to get it.
4019 vtable_index = method->vtable_index();
4020 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4021 err_msg_res("bad index %d", vtable_index));
4022 }
4023 slow_call = new(C) CallDynamicJavaNode(tf,
4024 SharedRuntime::get_resolve_virtual_call_stub(),
4025 method, vtable_index, bci());
4026 } else { // neither virtual nor static: opt_virtual
4027 null_check_receiver();
4028 slow_call = new(C) CallStaticJavaNode(C, tf,
4029 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4030 method, bci());
4031 slow_call->set_optimized_virtual(true);
4032 }
4033 set_arguments_for_java_call(slow_call);
4034 set_edges_for_java_call(slow_call);
4035 return slow_call;
4036 }
4037
4038
4039 /**
4040 * Build special case code for calls to hashCode on an object. This call may
4041 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4042 * slightly different code.
4043 */
4044 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4045 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4046 assert(!(is_virtual && is_static), "either virtual, special, or static");
4047
4048 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4049
4050 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4051 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4052 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4053 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4054 Node* obj = NULL;
4055 if (!is_static) {
4056 // Check for hashing null object
4057 obj = null_check_receiver();
4058 if (stopped()) return true; // unconditionally null
4059 result_reg->init_req(_null_path, top());
4060 result_val->init_req(_null_path, top());
4061 } else {
4062 // Do a null check, and return zero if null.
4063 // System.identityHashCode(null) == 0
4064 obj = argument(0);
4065 Node* null_ctl = top();
4066 obj = null_check_oop(obj, &null_ctl);
4067 result_reg->init_req(_null_path, null_ctl);
4068 result_val->init_req(_null_path, _gvn.intcon(0));
4069 }
4070
4071 // Unconditionally null? Then return right away.
4072 if (stopped()) {
4073 set_control( result_reg->in(_null_path));
4074 if (!stopped())
4075 set_result(result_val->in(_null_path));
4076 return true;
4077 }
4078
4079 // We only go to the fast case code if we pass a number of guards. The
4080 // paths which do not pass are accumulated in the slow_region.
4081 RegionNode* slow_region = new (C) RegionNode(1);
4082 record_for_igvn(slow_region);
4083
4084 // If this is a virtual call, we generate a funny guard. We pull out
4085 // the vtable entry corresponding to hashCode() from the target object.
4086 // If the target method which we are calling happens to be the native
4087 // Object hashCode() method, we pass the guard. We do not need this
4088 // guard for non-virtual calls -- the caller is known to be the native
4089 // Object hashCode().
4090 if (is_virtual) {
4091 // After null check, get the object's klass.
4092 Node* obj_klass = load_object_klass(obj);
4093 generate_virtual_guard(obj_klass, slow_region);
4094 }
4095
4096 // Get the header out of the object, use LoadMarkNode when available
4097 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4098 // The control of the load must be NULL. Otherwise, the load can move before
4099 // the null check after castPP removal.
4100 Node* no_ctrl = NULL;
4101 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4102
4103 // Test the header to see if it is unlocked.
4104 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4105 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4106 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4107 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4108 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4109
4110 generate_slow_guard(test_unlocked, slow_region);
4111
4112 // Get the hash value and check to see that it has been properly assigned.
4113 // We depend on hash_mask being at most 32 bits and avoid the use of
4114 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4115 // vm: see markOop.hpp.
4116 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4117 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4118 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4119 // This hack lets the hash bits live anywhere in the mark object now, as long
4120 // as the shift drops the relevant bits into the low 32 bits. Note that
4121 // Java spec says that HashCode is an int so there's no point in capturing
4122 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4123 hshifted_header = ConvX2I(hshifted_header);
4124 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4125
4126 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4127 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4128 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4129
4130 generate_slow_guard(test_assigned, slow_region);
4131
4132 Node* init_mem = reset_memory();
4133 // fill in the rest of the null path:
4134 result_io ->init_req(_null_path, i_o());
4135 result_mem->init_req(_null_path, init_mem);
4136
4137 result_val->init_req(_fast_path, hash_val);
4138 result_reg->init_req(_fast_path, control());
4139 result_io ->init_req(_fast_path, i_o());
4140 result_mem->init_req(_fast_path, init_mem);
4141
4142 // Generate code for the slow case. We make a call to hashCode().
4143 set_control(_gvn.transform(slow_region));
4144 if (!stopped()) {
4145 // No need for PreserveJVMState, because we're using up the present state.
4146 set_all_memory(init_mem);
4147 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4148 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4149 Node* slow_result = set_results_for_java_call(slow_call);
4150 // this->control() comes from set_results_for_java_call
4151 result_reg->init_req(_slow_path, control());
4152 result_val->init_req(_slow_path, slow_result);
4153 result_io ->set_req(_slow_path, i_o());
4154 result_mem ->set_req(_slow_path, reset_memory());
4155 }
4156
4157 // Return the combined state.
4158 set_i_o( _gvn.transform(result_io) );
4159 set_all_memory( _gvn.transform(result_mem));
4160
4161 set_result(result_reg, result_val);
4162 return true;
4163 }
4164
4165 //---------------------------inline_native_getClass----------------------------
4166 // public final native Class<?> java.lang.Object.getClass();
4167 //
4168 // Build special case code for calls to getClass on an object.
4169 bool LibraryCallKit::inline_native_getClass() {
4170 Node* obj = null_check_receiver();
4171 if (stopped()) return true;
4172 set_result(load_mirror_from_klass(load_object_klass(obj)));
4173 return true;
4174 }
4175
4176 //-----------------inline_native_Reflection_getCallerClass---------------------
4177 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4178 //
4179 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4180 //
4181 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4182 // in that it must skip particular security frames and checks for
4183 // caller sensitive methods.
4184 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4185 #ifndef PRODUCT
4186 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4187 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4188 }
4189 #endif
4190
4191 if (!jvms()->has_method()) {
4192 #ifndef PRODUCT
4193 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4194 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4195 }
4196 #endif
4197 return false;
4198 }
4199
4200 // Walk back up the JVM state to find the caller at the required
4201 // depth.
4202 JVMState* caller_jvms = jvms();
4203
4204 // Cf. JVM_GetCallerClass
4205 // NOTE: Start the loop at depth 1 because the current JVM state does
4206 // not include the Reflection.getCallerClass() frame.
4207 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4208 ciMethod* m = caller_jvms->method();
4209 switch (n) {
4210 case 0:
4211 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4212 break;
4213 case 1:
4214 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4215 if (!m->caller_sensitive()) {
4216 #ifndef PRODUCT
4217 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4218 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4219 }
4220 #endif
4221 return false; // bail-out; let JVM_GetCallerClass do the work
4222 }
4223 break;
4224 default:
4225 if (!m->is_ignored_by_security_stack_walk()) {
4226 // We have reached the desired frame; return the holder class.
4227 // Acquire method holder as java.lang.Class and push as constant.
4228 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4229 ciInstance* caller_mirror = caller_klass->java_mirror();
4230 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4231
4232 #ifndef PRODUCT
4233 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4234 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());
4235 tty->print_cr(" JVM state at this point:");
4236 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4237 ciMethod* m = jvms()->of_depth(i)->method();
4238 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4239 }
4240 }
4241 #endif
4242 return true;
4243 }
4244 break;
4245 }
4246 }
4247
4248 #ifndef PRODUCT
4249 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4250 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4251 tty->print_cr(" JVM state at this point:");
4252 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4253 ciMethod* m = jvms()->of_depth(i)->method();
4254 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4255 }
4256 }
4257 #endif
4258
4259 return false; // bail-out; let JVM_GetCallerClass do the work
4260 }
4261
4262 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4263 Node* arg = argument(0);
4264 Node* result;
4265
4266 switch (id) {
4267 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4268 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4269 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4270 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4271
4272 case vmIntrinsics::_doubleToLongBits: {
4273 // two paths (plus control) merge in a wood
4274 RegionNode *r = new (C) RegionNode(3);
4275 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4276
4277 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4278 // Build the boolean node
4279 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4280
4281 // Branch either way.
4282 // NaN case is less traveled, which makes all the difference.
4283 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4284 Node *opt_isnan = _gvn.transform(ifisnan);
4285 assert( opt_isnan->is_If(), "Expect an IfNode");
4286 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4287 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4288
4289 set_control(iftrue);
4290
4291 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4292 Node *slow_result = longcon(nan_bits); // return NaN
4293 phi->init_req(1, _gvn.transform( slow_result ));
4294 r->init_req(1, iftrue);
4295
4296 // Else fall through
4297 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4298 set_control(iffalse);
4299
4300 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4301 r->init_req(2, iffalse);
4302
4303 // Post merge
4304 set_control(_gvn.transform(r));
4305 record_for_igvn(r);
4306
4307 C->set_has_split_ifs(true); // Has chance for split-if optimization
4308 result = phi;
4309 assert(result->bottom_type()->isa_long(), "must be");
4310 break;
4311 }
4312
4313 case vmIntrinsics::_floatToIntBits: {
4314 // two paths (plus control) merge in a wood
4315 RegionNode *r = new (C) RegionNode(3);
4316 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4317
4318 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4319 // Build the boolean node
4320 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4321
4322 // Branch either way.
4323 // NaN case is less traveled, which makes all the difference.
4324 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4325 Node *opt_isnan = _gvn.transform(ifisnan);
4326 assert( opt_isnan->is_If(), "Expect an IfNode");
4327 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4328 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4329
4330 set_control(iftrue);
4331
4332 static const jint nan_bits = 0x7fc00000;
4333 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4334 phi->init_req(1, _gvn.transform( slow_result ));
4335 r->init_req(1, iftrue);
4336
4337 // Else fall through
4338 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4339 set_control(iffalse);
4340
4341 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4342 r->init_req(2, iffalse);
4343
4344 // Post merge
4345 set_control(_gvn.transform(r));
4346 record_for_igvn(r);
4347
4348 C->set_has_split_ifs(true); // Has chance for split-if optimization
4349 result = phi;
4350 assert(result->bottom_type()->isa_int(), "must be");
4351 break;
4352 }
4353
4354 default:
4355 fatal_unexpected_iid(id);
4356 break;
4357 }
4358 set_result(_gvn.transform(result));
4359 return true;
4360 }
4361
4362 #ifdef _LP64
4363 #define XTOP ,top() /*additional argument*/
4364 #else //_LP64
4365 #define XTOP /*no additional argument*/
4366 #endif //_LP64
4367
4368 //----------------------inline_unsafe_copyMemory-------------------------
4369 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4370 bool LibraryCallKit::inline_unsafe_copyMemory() {
4371 if (callee()->is_static()) return false; // caller must have the capability!
4372 null_check_receiver(); // null-check receiver
4373 if (stopped()) return true;
4374
4375 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4376
4377 Node* src_ptr = argument(1); // type: oop
4378 Node* src_off = ConvL2X(argument(2)); // type: long
4379 Node* dst_ptr = argument(4); // type: oop
4380 Node* dst_off = ConvL2X(argument(5)); // type: long
4381 Node* size = ConvL2X(argument(7)); // type: long
4382
4383 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4384 "fieldOffset must be byte-scaled");
4385
4386 Node* src = make_unsafe_address(src_ptr, src_off);
4387 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4388
4389 // Conservatively insert a memory barrier on all memory slices.
4390 // Do not let writes of the copy source or destination float below the copy.
4391 insert_mem_bar(Op_MemBarCPUOrder);
4392
4393 // Call it. Note that the length argument is not scaled.
4394 make_runtime_call(RC_LEAF|RC_NO_FP,
4395 OptoRuntime::fast_arraycopy_Type(),
4396 StubRoutines::unsafe_arraycopy(),
4397 "unsafe_arraycopy",
4398 TypeRawPtr::BOTTOM,
4399 src, dst, size XTOP);
4400
4401 // Do not let reads of the copy destination float above the copy.
4402 insert_mem_bar(Op_MemBarCPUOrder);
4403
4404 return true;
4405 }
4406
4407 //------------------------clone_coping-----------------------------------
4408 // Helper function for inline_native_clone.
4409 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4410 assert(obj_size != NULL, "");
4411 Node* raw_obj = alloc_obj->in(1);
4412 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4413
4414 AllocateNode* alloc = NULL;
4415 if (ReduceBulkZeroing) {
4416 // We will be completely responsible for initializing this object -
4417 // mark Initialize node as complete.
4418 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4419 // The object was just allocated - there should be no any stores!
4420 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4421 // Mark as complete_with_arraycopy so that on AllocateNode
4422 // expansion, we know this AllocateNode is initialized by an array
4423 // copy and a StoreStore barrier exists after the array copy.
4424 alloc->initialization()->set_complete_with_arraycopy();
4425 }
4426
4427 // Copy the fastest available way.
4428 // TODO: generate fields copies for small objects instead.
4429 Node* src = obj;
4430 Node* dest = alloc_obj;
4431 Node* size = _gvn.transform(obj_size);
4432
4433 // Exclude the header but include array length to copy by 8 bytes words.
4434 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4435 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4436 instanceOopDesc::base_offset_in_bytes();
4437 // base_off:
4438 // 8 - 32-bit VM
4439 // 12 - 64-bit VM, compressed klass
4440 // 16 - 64-bit VM, normal klass
4441 if (base_off % BytesPerLong != 0) {
4442 assert(UseCompressedClassPointers, "");
4443 if (is_array) {
4444 // Exclude length to copy by 8 bytes words.
4445 base_off += sizeof(int);
4446 } else {
4447 // Include klass to copy by 8 bytes words.
4448 base_off = instanceOopDesc::klass_offset_in_bytes();
4449 }
4450 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4451 }
4452 src = basic_plus_adr(src, base_off);
4453 dest = basic_plus_adr(dest, base_off);
4454
4455 // Compute the length also, if needed:
4456 Node* countx = size;
4457 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4458 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4459
4460 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4461 bool disjoint_bases = true;
4462 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4463 src, NULL, dest, NULL, countx,
4464 /*dest_uninitialized*/true);
4465
4466 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4467 if (card_mark) {
4468 assert(!is_array, "");
4469 // Put in store barrier for any and all oops we are sticking
4470 // into this object. (We could avoid this if we could prove
4471 // that the object type contains no oop fields at all.)
4472 Node* no_particular_value = NULL;
4473 Node* no_particular_field = NULL;
4474 int raw_adr_idx = Compile::AliasIdxRaw;
4475 post_barrier(control(),
4476 memory(raw_adr_type),
4477 alloc_obj,
4478 no_particular_field,
4479 raw_adr_idx,
4480 no_particular_value,
4481 T_OBJECT,
4482 false);
4483 }
4484
4485 // Do not let reads from the cloned object float above the arraycopy.
4486 if (alloc != NULL) {
4487 // Do not let stores that initialize this object be reordered with
4488 // a subsequent store that would make this object accessible by
4489 // other threads.
4490 // Record what AllocateNode this StoreStore protects so that
4491 // escape analysis can go from the MemBarStoreStoreNode to the
4492 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4493 // based on the escape status of the AllocateNode.
4494 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4495 } else {
4496 insert_mem_bar(Op_MemBarCPUOrder);
4497 }
4498 }
4499
4500 //------------------------inline_native_clone----------------------------
4501 // protected native Object java.lang.Object.clone();
4502 //
4503 // Here are the simple edge cases:
4504 // null receiver => normal trap
4505 // virtual and clone was overridden => slow path to out-of-line clone
4506 // not cloneable or finalizer => slow path to out-of-line Object.clone
4507 //
4508 // The general case has two steps, allocation and copying.
4509 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4510 //
4511 // Copying also has two cases, oop arrays and everything else.
4512 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4513 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4514 //
4515 // These steps fold up nicely if and when the cloned object's klass
4516 // can be sharply typed as an object array, a type array, or an instance.
4517 //
4518 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4519 PhiNode* result_val;
4520
4521 // Set the reexecute bit for the interpreter to reexecute
4522 // the bytecode that invokes Object.clone if deoptimization happens.
4523 { PreserveReexecuteState preexecs(this);
4524 jvms()->set_should_reexecute(true);
4525
4526 Node* obj = null_check_receiver();
4527 if (stopped()) return true;
4528
4529 Node* obj_klass = load_object_klass(obj);
4530 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4531 const TypeOopPtr* toop = ((tklass != NULL)
4532 ? tklass->as_instance_type()
4533 : TypeInstPtr::NOTNULL);
4534
4535 // Conservatively insert a memory barrier on all memory slices.
4536 // Do not let writes into the original float below the clone.
4537 insert_mem_bar(Op_MemBarCPUOrder);
4538
4539 // paths into result_reg:
4540 enum {
4541 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4542 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4543 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4544 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4545 PATH_LIMIT
4546 };
4547 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4548 result_val = new(C) PhiNode(result_reg,
4549 TypeInstPtr::NOTNULL);
4550 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4551 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4552 TypePtr::BOTTOM);
4553 record_for_igvn(result_reg);
4554
4555 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4556 int raw_adr_idx = Compile::AliasIdxRaw;
4557
4558 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4559 if (array_ctl != NULL) {
4560 // It's an array.
4561 PreserveJVMState pjvms(this);
4562 set_control(array_ctl);
4563 Node* obj_length = load_array_length(obj);
4564 Node* obj_size = NULL;
4565 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4566
4567 if (!use_ReduceInitialCardMarks()) {
4568 // If it is an oop array, it requires very special treatment,
4569 // because card marking is required on each card of the array.
4570 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4571 if (is_obja != NULL) {
4572 PreserveJVMState pjvms2(this);
4573 set_control(is_obja);
4574 // Generate a direct call to the right arraycopy function(s).
4575 bool disjoint_bases = true;
4576 bool length_never_negative = true;
4577 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4578 obj, intcon(0), alloc_obj, intcon(0),
4579 obj_length,
4580 disjoint_bases, length_never_negative);
4581 result_reg->init_req(_objArray_path, control());
4582 result_val->init_req(_objArray_path, alloc_obj);
4583 result_i_o ->set_req(_objArray_path, i_o());
4584 result_mem ->set_req(_objArray_path, reset_memory());
4585 }
4586 }
4587 // Otherwise, there are no card marks to worry about.
4588 // (We can dispense with card marks if we know the allocation
4589 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4590 // causes the non-eden paths to take compensating steps to
4591 // simulate a fresh allocation, so that no further
4592 // card marks are required in compiled code to initialize
4593 // the object.)
4594
4595 if (!stopped()) {
4596 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4597
4598 // Present the results of the copy.
4599 result_reg->init_req(_array_path, control());
4600 result_val->init_req(_array_path, alloc_obj);
4601 result_i_o ->set_req(_array_path, i_o());
4602 result_mem ->set_req(_array_path, reset_memory());
4603 }
4604 }
4605
4606 // We only go to the instance fast case code if we pass a number of guards.
4607 // The paths which do not pass are accumulated in the slow_region.
4608 RegionNode* slow_region = new (C) RegionNode(1);
4609 record_for_igvn(slow_region);
4610 if (!stopped()) {
4611 // It's an instance (we did array above). Make the slow-path tests.
4612 // If this is a virtual call, we generate a funny guard. We grab
4613 // the vtable entry corresponding to clone() from the target object.
4614 // If the target method which we are calling happens to be the
4615 // Object clone() method, we pass the guard. We do not need this
4616 // guard for non-virtual calls; the caller is known to be the native
4617 // Object clone().
4618 if (is_virtual) {
4619 generate_virtual_guard(obj_klass, slow_region);
4620 }
4621
4622 // The object must be cloneable and must not have a finalizer.
4623 // Both of these conditions may be checked in a single test.
4624 // We could optimize the cloneable test further, but we don't care.
4625 generate_access_flags_guard(obj_klass,
4626 // Test both conditions:
4627 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4628 // Must be cloneable but not finalizer:
4629 JVM_ACC_IS_CLONEABLE,
4630 slow_region);
4631 }
4632
4633 if (!stopped()) {
4634 // It's an instance, and it passed the slow-path tests.
4635 PreserveJVMState pjvms(this);
4636 Node* obj_size = NULL;
4637 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4638 // is reexecuted if deoptimization occurs and there could be problems when merging
4639 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4640 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4641
4642 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4643
4644 // Present the results of the slow call.
4645 result_reg->init_req(_instance_path, control());
4646 result_val->init_req(_instance_path, alloc_obj);
4647 result_i_o ->set_req(_instance_path, i_o());
4648 result_mem ->set_req(_instance_path, reset_memory());
4649 }
4650
4651 // Generate code for the slow case. We make a call to clone().
4652 set_control(_gvn.transform(slow_region));
4653 if (!stopped()) {
4654 PreserveJVMState pjvms(this);
4655 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4656 Node* slow_result = set_results_for_java_call(slow_call);
4657 // this->control() comes from set_results_for_java_call
4658 result_reg->init_req(_slow_path, control());
4659 result_val->init_req(_slow_path, slow_result);
4660 result_i_o ->set_req(_slow_path, i_o());
4661 result_mem ->set_req(_slow_path, reset_memory());
4662 }
4663
4664 // Return the combined state.
4665 set_control( _gvn.transform(result_reg));
4666 set_i_o( _gvn.transform(result_i_o));
4667 set_all_memory( _gvn.transform(result_mem));
4668 } // original reexecute is set back here
4669
4670 set_result(_gvn.transform(result_val));
4671 return true;
4672 }
4673
4674 //------------------------------basictype2arraycopy----------------------------
4675 address LibraryCallKit::basictype2arraycopy(BasicType t,
4676 Node* src_offset,
4677 Node* dest_offset,
4678 bool disjoint_bases,
4679 const char* &name,
4680 bool dest_uninitialized) {
4681 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4682 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4683
4684 bool aligned = false;
4685 bool disjoint = disjoint_bases;
4686
4687 // if the offsets are the same, we can treat the memory regions as
4688 // disjoint, because either the memory regions are in different arrays,
4689 // or they are identical (which we can treat as disjoint.) We can also
4690 // treat a copy with a destination index less that the source index
4691 // as disjoint since a low->high copy will work correctly in this case.
4692 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4693 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4694 // both indices are constants
4695 int s_offs = src_offset_inttype->get_con();
4696 int d_offs = dest_offset_inttype->get_con();
4697 int element_size = type2aelembytes(t);
4698 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4699 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4700 if (s_offs >= d_offs) disjoint = true;
4701 } else if (src_offset == dest_offset && src_offset != NULL) {
4702 // This can occur if the offsets are identical non-constants.
4703 disjoint = true;
4704 }
4705
4706 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4707 }
4708
4709
4710 //------------------------------inline_arraycopy-----------------------
4711 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4712 // Object dest, int destPos,
4713 // int length);
4714 bool LibraryCallKit::inline_arraycopy() {
4715 // Get the arguments.
4716 Node* src = argument(0); // type: oop
4717 Node* src_offset = argument(1); // type: int
4718 Node* dest = argument(2); // type: oop
4719 Node* dest_offset = argument(3); // type: int
4720 Node* length = argument(4); // type: int
4721
4722 // Compile time checks. If any of these checks cannot be verified at compile time,
4723 // we do not make a fast path for this call. Instead, we let the call remain as it
4724 // is. The checks we choose to mandate at compile time are:
4725 //
4726 // (1) src and dest are arrays.
4727 const Type* src_type = src->Value(&_gvn);
4728 const Type* dest_type = dest->Value(&_gvn);
4729 const TypeAryPtr* top_src = src_type->isa_aryptr();
4730 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4731
4732 // Do we have the type of src?
4733 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4734 // Do we have the type of dest?
4735 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4736 // Is the type for src from speculation?
4737 bool src_spec = false;
4738 // Is the type for dest from speculation?
4739 bool dest_spec = false;
4740
4741 if (!has_src || !has_dest) {
4742 // We don't have sufficient type information, let's see if
4743 // speculative types can help. We need to have types for both src
4744 // and dest so that it pays off.
4745
4746 // Do we already have or could we have type information for src
4747 bool could_have_src = has_src;
4748 // Do we already have or could we have type information for dest
4749 bool could_have_dest = has_dest;
4750
4751 ciKlass* src_k = NULL;
4752 if (!has_src) {
4753 src_k = src_type->speculative_type();
4754 if (src_k != NULL && src_k->is_array_klass()) {
4755 could_have_src = true;
4756 }
4757 }
4758
4759 ciKlass* dest_k = NULL;
4760 if (!has_dest) {
4761 dest_k = dest_type->speculative_type();
4762 if (dest_k != NULL && dest_k->is_array_klass()) {
4763 could_have_dest = true;
4764 }
4765 }
4766
4767 if (could_have_src && could_have_dest) {
4768 // This is going to pay off so emit the required guards
4769 if (!has_src) {
4770 src = maybe_cast_profiled_obj(src, src_k);
4771 src_type = _gvn.type(src);
4772 top_src = src_type->isa_aryptr();
4773 has_src = (top_src != NULL && top_src->klass() != NULL);
4774 src_spec = true;
4775 }
4776 if (!has_dest) {
4777 dest = maybe_cast_profiled_obj(dest, dest_k);
4778 dest_type = _gvn.type(dest);
4779 top_dest = dest_type->isa_aryptr();
4780 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4781 dest_spec = true;
4782 }
4783 }
4784 }
4785
4786 if (!has_src || !has_dest) {
4787 // Conservatively insert a memory barrier on all memory slices.
4788 // Do not let writes into the source float below the arraycopy.
4789 insert_mem_bar(Op_MemBarCPUOrder);
4790
4791 // Call StubRoutines::generic_arraycopy stub.
4792 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4793 src, src_offset, dest, dest_offset, length);
4794
4795 // Do not let reads from the destination float above the arraycopy.
4796 // Since we cannot type the arrays, we don't know which slices
4797 // might be affected. We could restrict this barrier only to those
4798 // memory slices which pertain to array elements--but don't bother.
4799 if (!InsertMemBarAfterArraycopy)
4800 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4801 insert_mem_bar(Op_MemBarCPUOrder);
4802 return true;
4803 }
4804
4805 // (2) src and dest arrays must have elements of the same BasicType
4806 // Figure out the size and type of the elements we will be copying.
4807 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4808 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4809 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4810 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4811
4812 if (src_elem != dest_elem || dest_elem == T_VOID) {
4813 // The component types are not the same or are not recognized. Punt.
4814 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4815 generate_slow_arraycopy(TypePtr::BOTTOM,
4816 src, src_offset, dest, dest_offset, length,
4817 /*dest_uninitialized*/false);
4818 return true;
4819 }
4820
4821 if (src_elem == T_OBJECT) {
4822 // If both arrays are object arrays then having the exact types
4823 // for both will remove the need for a subtype check at runtime
4824 // before the call and may make it possible to pick a faster copy
4825 // routine (without a subtype check on every element)
4826 // Do we have the exact type of src?
4827 bool could_have_src = src_spec;
4828 // Do we have the exact type of dest?
4829 bool could_have_dest = dest_spec;
4830 ciKlass* src_k = top_src->klass();
4831 ciKlass* dest_k = top_dest->klass();
4832 if (!src_spec) {
4833 src_k = src_type->speculative_type();
4834 if (src_k != NULL && src_k->is_array_klass()) {
4835 could_have_src = true;
4836 }
4837 }
4838 if (!dest_spec) {
4839 dest_k = dest_type->speculative_type();
4840 if (dest_k != NULL && dest_k->is_array_klass()) {
4841 could_have_dest = true;
4842 }
4843 }
4844 if (could_have_src && could_have_dest) {
4845 // If we can have both exact types, emit the missing guards
4846 if (could_have_src && !src_spec) {
4847 src = maybe_cast_profiled_obj(src, src_k);
4848 }
4849 if (could_have_dest && !dest_spec) {
4850 dest = maybe_cast_profiled_obj(dest, dest_k);
4851 }
4852 }
4853 }
4854
4855 //---------------------------------------------------------------------------
4856 // We will make a fast path for this call to arraycopy.
4857
4858 // We have the following tests left to perform:
4859 //
4860 // (3) src and dest must not be null.
4861 // (4) src_offset must not be negative.
4862 // (5) dest_offset must not be negative.
4863 // (6) length must not be negative.
4864 // (7) src_offset + length must not exceed length of src.
4865 // (8) dest_offset + length must not exceed length of dest.
4866 // (9) each element of an oop array must be assignable
4867
4868 RegionNode* slow_region = new (C) RegionNode(1);
4869 record_for_igvn(slow_region);
4870
4871 // (3) operands must not be null
4872 // We currently perform our null checks with the null_check routine.
4873 // This means that the null exceptions will be reported in the caller
4874 // rather than (correctly) reported inside of the native arraycopy call.
4875 // This should be corrected, given time. We do our null check with the
4876 // stack pointer restored.
4877 src = null_check(src, T_ARRAY);
4878 dest = null_check(dest, T_ARRAY);
4879
4880 // (4) src_offset must not be negative.
4881 generate_negative_guard(src_offset, slow_region);
4882
4883 // (5) dest_offset must not be negative.
4884 generate_negative_guard(dest_offset, slow_region);
4885
4886 // (6) length must not be negative (moved to generate_arraycopy()).
4887 // generate_negative_guard(length, slow_region);
4888
4889 // (7) src_offset + length must not exceed length of src.
4890 generate_limit_guard(src_offset, length,
4891 load_array_length(src),
4892 slow_region);
4893
4894 // (8) dest_offset + length must not exceed length of dest.
4895 generate_limit_guard(dest_offset, length,
4896 load_array_length(dest),
4897 slow_region);
4898
4899 // (9) each element of an oop array must be assignable
4900 // The generate_arraycopy subroutine checks this.
4901
4902 // This is where the memory effects are placed:
4903 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4904 generate_arraycopy(adr_type, dest_elem,
4905 src, src_offset, dest, dest_offset, length,
4906 false, false, slow_region);
4907
4908 return true;
4909 }
4910
4911 //-----------------------------generate_arraycopy----------------------
4912 // Generate an optimized call to arraycopy.
4913 // Caller must guard against non-arrays.
4914 // Caller must determine a common array basic-type for both arrays.
4915 // Caller must validate offsets against array bounds.
4916 // The slow_region has already collected guard failure paths
4917 // (such as out of bounds length or non-conformable array types).
4918 // The generated code has this shape, in general:
4919 //
4920 // if (length == 0) return // via zero_path
4921 // slowval = -1
4922 // if (types unknown) {
4923 // slowval = call generic copy loop
4924 // if (slowval == 0) return // via checked_path
4925 // } else if (indexes in bounds) {
4926 // if ((is object array) && !(array type check)) {
4927 // slowval = call checked copy loop
4928 // if (slowval == 0) return // via checked_path
4929 // } else {
4930 // call bulk copy loop
4931 // return // via fast_path
4932 // }
4933 // }
4934 // // adjust params for remaining work:
4935 // if (slowval != -1) {
4936 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4937 // }
4938 // slow_region:
4939 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4940 // return // via slow_call_path
4941 //
4942 // This routine is used from several intrinsics: System.arraycopy,
4943 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4944 //
4945 void
4946 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4947 BasicType basic_elem_type,
4948 Node* src, Node* src_offset,
4949 Node* dest, Node* dest_offset,
4950 Node* copy_length,
4951 bool disjoint_bases,
4952 bool length_never_negative,
4953 RegionNode* slow_region) {
4954
4955 if (slow_region == NULL) {
4956 slow_region = new(C) RegionNode(1);
4957 record_for_igvn(slow_region);
4958 }
4959
4960 Node* original_dest = dest;
4961 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4962 bool dest_uninitialized = false;
4963
4964 // See if this is the initialization of a newly-allocated array.
4965 // If so, we will take responsibility here for initializing it to zero.
4966 // (Note: Because tightly_coupled_allocation performs checks on the
4967 // out-edges of the dest, we need to avoid making derived pointers
4968 // from it until we have checked its uses.)
4969 if (ReduceBulkZeroing
4970 && !ZeroTLAB // pointless if already zeroed
4971 && basic_elem_type != T_CONFLICT // avoid corner case
4972 && !src->eqv_uncast(dest)
4973 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4974 != NULL)
4975 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4976 && alloc->maybe_set_complete(&_gvn)) {
4977 // "You break it, you buy it."
4978 InitializeNode* init = alloc->initialization();
4979 assert(init->is_complete(), "we just did this");
4980 init->set_complete_with_arraycopy();
4981 assert(dest->is_CheckCastPP(), "sanity");
4982 assert(dest->in(0)->in(0) == init, "dest pinned");
4983 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4984 // From this point on, every exit path is responsible for
4985 // initializing any non-copied parts of the object to zero.
4986 // Also, if this flag is set we make sure that arraycopy interacts properly
4987 // with G1, eliding pre-barriers. See CR 6627983.
4988 dest_uninitialized = true;
4989 } else {
4990 // No zeroing elimination here.
4991 alloc = NULL;
4992 //original_dest = dest;
4993 //dest_uninitialized = false;
4994 }
4995
4996 // Results are placed here:
4997 enum { fast_path = 1, // normal void-returning assembly stub
4998 checked_path = 2, // special assembly stub with cleanup
4999 slow_call_path = 3, // something went wrong; call the VM
5000 zero_path = 4, // bypass when length of copy is zero
5001 bcopy_path = 5, // copy primitive array by 64-bit blocks
5002 PATH_LIMIT = 6
5003 };
5004 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5005 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5006 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5007 record_for_igvn(result_region);
5008 _gvn.set_type_bottom(result_i_o);
5009 _gvn.set_type_bottom(result_memory);
5010 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5011
5012 // The slow_control path:
5013 Node* slow_control;
5014 Node* slow_i_o = i_o();
5015 Node* slow_mem = memory(adr_type);
5016 debug_only(slow_control = (Node*) badAddress);
5017
5018 // Checked control path:
5019 Node* checked_control = top();
5020 Node* checked_mem = NULL;
5021 Node* checked_i_o = NULL;
5022 Node* checked_value = NULL;
5023
5024 if (basic_elem_type == T_CONFLICT) {
5025 assert(!dest_uninitialized, "");
5026 Node* cv = generate_generic_arraycopy(adr_type,
5027 src, src_offset, dest, dest_offset,
5028 copy_length, dest_uninitialized);
5029 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5030 checked_control = control();
5031 checked_i_o = i_o();
5032 checked_mem = memory(adr_type);
5033 checked_value = cv;
5034 set_control(top()); // no fast path
5035 }
5036
5037 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5038 if (not_pos != NULL) {
5039 PreserveJVMState pjvms(this);
5040 set_control(not_pos);
5041
5042 // (6) length must not be negative.
5043 if (!length_never_negative) {
5044 generate_negative_guard(copy_length, slow_region);
5045 }
5046
5047 // copy_length is 0.
5048 if (!stopped() && dest_uninitialized) {
5049 Node* dest_length = alloc->in(AllocateNode::ALength);
5050 if (copy_length->eqv_uncast(dest_length)
5051 || _gvn.find_int_con(dest_length, 1) <= 0) {
5052 // There is no zeroing to do. No need for a secondary raw memory barrier.
5053 } else {
5054 // Clear the whole thing since there are no source elements to copy.
5055 generate_clear_array(adr_type, dest, basic_elem_type,
5056 intcon(0), NULL,
5057 alloc->in(AllocateNode::AllocSize));
5058 // Use a secondary InitializeNode as raw memory barrier.
5059 // Currently it is needed only on this path since other
5060 // paths have stub or runtime calls as raw memory barriers.
5061 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5062 Compile::AliasIdxRaw,
5063 top())->as_Initialize();
5064 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5065 }
5066 }
5067
5068 // Present the results of the fast call.
5069 result_region->init_req(zero_path, control());
5070 result_i_o ->init_req(zero_path, i_o());
5071 result_memory->init_req(zero_path, memory(adr_type));
5072 }
5073
5074 if (!stopped() && dest_uninitialized) {
5075 // We have to initialize the *uncopied* part of the array to zero.
5076 // The copy destination is the slice dest[off..off+len]. The other slices
5077 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5078 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5079 Node* dest_length = alloc->in(AllocateNode::ALength);
5080 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5081 copy_length));
5082
5083 // If there is a head section that needs zeroing, do it now.
5084 if (find_int_con(dest_offset, -1) != 0) {
5085 generate_clear_array(adr_type, dest, basic_elem_type,
5086 intcon(0), dest_offset,
5087 NULL);
5088 }
5089
5090 // Next, perform a dynamic check on the tail length.
5091 // It is often zero, and we can win big if we prove this.
5092 // There are two wins: Avoid generating the ClearArray
5093 // with its attendant messy index arithmetic, and upgrade
5094 // the copy to a more hardware-friendly word size of 64 bits.
5095 Node* tail_ctl = NULL;
5096 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5097 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5098 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5099 tail_ctl = generate_slow_guard(bol_lt, NULL);
5100 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5101 }
5102
5103 // At this point, let's assume there is no tail.
5104 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5105 // There is no tail. Try an upgrade to a 64-bit copy.
5106 bool didit = false;
5107 { PreserveJVMState pjvms(this);
5108 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5109 src, src_offset, dest, dest_offset,
5110 dest_size, dest_uninitialized);
5111 if (didit) {
5112 // Present the results of the block-copying fast call.
5113 result_region->init_req(bcopy_path, control());
5114 result_i_o ->init_req(bcopy_path, i_o());
5115 result_memory->init_req(bcopy_path, memory(adr_type));
5116 }
5117 }
5118 if (didit)
5119 set_control(top()); // no regular fast path
5120 }
5121
5122 // Clear the tail, if any.
5123 if (tail_ctl != NULL) {
5124 Node* notail_ctl = stopped() ? NULL : control();
5125 set_control(tail_ctl);
5126 if (notail_ctl == NULL) {
5127 generate_clear_array(adr_type, dest, basic_elem_type,
5128 dest_tail, NULL,
5129 dest_size);
5130 } else {
5131 // Make a local merge.
5132 Node* done_ctl = new(C) RegionNode(3);
5133 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5134 done_ctl->init_req(1, notail_ctl);
5135 done_mem->init_req(1, memory(adr_type));
5136 generate_clear_array(adr_type, dest, basic_elem_type,
5137 dest_tail, NULL,
5138 dest_size);
5139 done_ctl->init_req(2, control());
5140 done_mem->init_req(2, memory(adr_type));
5141 set_control( _gvn.transform(done_ctl));
5142 set_memory( _gvn.transform(done_mem), adr_type );
5143 }
5144 }
5145 }
5146
5147 BasicType copy_type = basic_elem_type;
5148 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5149 if (!stopped() && copy_type == T_OBJECT) {
5150 // If src and dest have compatible element types, we can copy bits.
5151 // Types S[] and D[] are compatible if D is a supertype of S.
5152 //
5153 // If they are not, we will use checked_oop_disjoint_arraycopy,
5154 // which performs a fast optimistic per-oop check, and backs off
5155 // further to JVM_ArrayCopy on the first per-oop check that fails.
5156 // (Actually, we don't move raw bits only; the GC requires card marks.)
5157
5158 // Get the Klass* for both src and dest
5159 Node* src_klass = load_object_klass(src);
5160 Node* dest_klass = load_object_klass(dest);
5161
5162 // Generate the subtype check.
5163 // This might fold up statically, or then again it might not.
5164 //
5165 // Non-static example: Copying List<String>.elements to a new String[].
5166 // The backing store for a List<String> is always an Object[],
5167 // but its elements are always type String, if the generic types
5168 // are correct at the source level.
5169 //
5170 // Test S[] against D[], not S against D, because (probably)
5171 // the secondary supertype cache is less busy for S[] than S.
5172 // This usually only matters when D is an interface.
5173 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5174 // Plug failing path into checked_oop_disjoint_arraycopy
5175 if (not_subtype_ctrl != top()) {
5176 PreserveJVMState pjvms(this);
5177 set_control(not_subtype_ctrl);
5178 // (At this point we can assume disjoint_bases, since types differ.)
5179 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5180 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5181 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5182 Node* dest_elem_klass = _gvn.transform(n1);
5183 Node* cv = generate_checkcast_arraycopy(adr_type,
5184 dest_elem_klass,
5185 src, src_offset, dest, dest_offset,
5186 ConvI2X(copy_length), dest_uninitialized);
5187 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5188 checked_control = control();
5189 checked_i_o = i_o();
5190 checked_mem = memory(adr_type);
5191 checked_value = cv;
5192 }
5193 // At this point we know we do not need type checks on oop stores.
5194
5195 // Let's see if we need card marks:
5196 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5197 // If we do not need card marks, copy using the jint or jlong stub.
5198 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5199 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5200 "sizes agree");
5201 }
5202 }
5203
5204 if (!stopped()) {
5205 // Generate the fast path, if possible.
5206 PreserveJVMState pjvms(this);
5207 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5208 src, src_offset, dest, dest_offset,
5209 ConvI2X(copy_length), dest_uninitialized);
5210
5211 // Present the results of the fast call.
5212 result_region->init_req(fast_path, control());
5213 result_i_o ->init_req(fast_path, i_o());
5214 result_memory->init_req(fast_path, memory(adr_type));
5215 }
5216
5217 // Here are all the slow paths up to this point, in one bundle:
5218 slow_control = top();
5219 if (slow_region != NULL)
5220 slow_control = _gvn.transform(slow_region);
5221 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5222
5223 set_control(checked_control);
5224 if (!stopped()) {
5225 // Clean up after the checked call.
5226 // The returned value is either 0 or -1^K,
5227 // where K = number of partially transferred array elements.
5228 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5229 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5230 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5231
5232 // If it is 0, we are done, so transfer to the end.
5233 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5234 result_region->init_req(checked_path, checks_done);
5235 result_i_o ->init_req(checked_path, checked_i_o);
5236 result_memory->init_req(checked_path, checked_mem);
5237
5238 // If it is not zero, merge into the slow call.
5239 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5240 RegionNode* slow_reg2 = new(C) RegionNode(3);
5241 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5242 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5243 record_for_igvn(slow_reg2);
5244 slow_reg2 ->init_req(1, slow_control);
5245 slow_i_o2 ->init_req(1, slow_i_o);
5246 slow_mem2 ->init_req(1, slow_mem);
5247 slow_reg2 ->init_req(2, control());
5248 slow_i_o2 ->init_req(2, checked_i_o);
5249 slow_mem2 ->init_req(2, checked_mem);
5250
5251 slow_control = _gvn.transform(slow_reg2);
5252 slow_i_o = _gvn.transform(slow_i_o2);
5253 slow_mem = _gvn.transform(slow_mem2);
5254
5255 if (alloc != NULL) {
5256 // We'll restart from the very beginning, after zeroing the whole thing.
5257 // This can cause double writes, but that's OK since dest is brand new.
5258 // So we ignore the low 31 bits of the value returned from the stub.
5259 } else {
5260 // We must continue the copy exactly where it failed, or else
5261 // another thread might see the wrong number of writes to dest.
5262 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5263 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5264 slow_offset->init_req(1, intcon(0));
5265 slow_offset->init_req(2, checked_offset);
5266 slow_offset = _gvn.transform(slow_offset);
5267
5268 // Adjust the arguments by the conditionally incoming offset.
5269 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5270 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5271 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5272
5273 // Tweak the node variables to adjust the code produced below:
5274 src_offset = src_off_plus;
5275 dest_offset = dest_off_plus;
5276 copy_length = length_minus;
5277 }
5278 }
5279
5280 set_control(slow_control);
5281 if (!stopped()) {
5282 // Generate the slow path, if needed.
5283 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5284
5285 set_memory(slow_mem, adr_type);
5286 set_i_o(slow_i_o);
5287
5288 if (dest_uninitialized) {
5289 generate_clear_array(adr_type, dest, basic_elem_type,
5290 intcon(0), NULL,
5291 alloc->in(AllocateNode::AllocSize));
5292 }
5293
5294 generate_slow_arraycopy(adr_type,
5295 src, src_offset, dest, dest_offset,
5296 copy_length, /*dest_uninitialized*/false);
5297
5298 result_region->init_req(slow_call_path, control());
5299 result_i_o ->init_req(slow_call_path, i_o());
5300 result_memory->init_req(slow_call_path, memory(adr_type));
5301 }
5302
5303 // Remove unused edges.
5304 for (uint i = 1; i < result_region->req(); i++) {
5305 if (result_region->in(i) == NULL)
5306 result_region->init_req(i, top());
5307 }
5308
5309 // Finished; return the combined state.
5310 set_control( _gvn.transform(result_region));
5311 set_i_o( _gvn.transform(result_i_o) );
5312 set_memory( _gvn.transform(result_memory), adr_type );
5313
5314 // The memory edges above are precise in order to model effects around
5315 // array copies accurately to allow value numbering of field loads around
5316 // arraycopy. Such field loads, both before and after, are common in Java
5317 // collections and similar classes involving header/array data structures.
5318 //
5319 // But with low number of register or when some registers are used or killed
5320 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5321 // The next memory barrier is added to avoid it. If the arraycopy can be
5322 // optimized away (which it can, sometimes) then we can manually remove
5323 // the membar also.
5324 //
5325 // Do not let reads from the cloned object float above the arraycopy.
5326 if (alloc != NULL) {
5327 // Do not let stores that initialize this object be reordered with
5328 // a subsequent store that would make this object accessible by
5329 // other threads.
5330 // Record what AllocateNode this StoreStore protects so that
5331 // escape analysis can go from the MemBarStoreStoreNode to the
5332 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5333 // based on the escape status of the AllocateNode.
5334 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5335 } else if (InsertMemBarAfterArraycopy)
5336 insert_mem_bar(Op_MemBarCPUOrder);
5337 }
5338
5339
5340 // Helper function which determines if an arraycopy immediately follows
5341 // an allocation, with no intervening tests or other escapes for the object.
5342 AllocateArrayNode*
5343 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5344 RegionNode* slow_region) {
5345 if (stopped()) return NULL; // no fast path
5346 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5347
5348 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5349 if (alloc == NULL) return NULL;
5350
5351 Node* rawmem = memory(Compile::AliasIdxRaw);
5352 // Is the allocation's memory state untouched?
5353 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5354 // Bail out if there have been raw-memory effects since the allocation.
5355 // (Example: There might have been a call or safepoint.)
5356 return NULL;
5357 }
5358 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5359 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5360 return NULL;
5361 }
5362
5363 // There must be no unexpected observers of this allocation.
5364 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5365 Node* obs = ptr->fast_out(i);
5366 if (obs != this->map()) {
5367 return NULL;
5368 }
5369 }
5370
5371 // This arraycopy must unconditionally follow the allocation of the ptr.
5372 Node* alloc_ctl = ptr->in(0);
5373 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5374
5375 Node* ctl = control();
5376 while (ctl != alloc_ctl) {
5377 // There may be guards which feed into the slow_region.
5378 // Any other control flow means that we might not get a chance
5379 // to finish initializing the allocated object.
5380 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5381 IfNode* iff = ctl->in(0)->as_If();
5382 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5383 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5384 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5385 ctl = iff->in(0); // This test feeds the known slow_region.
5386 continue;
5387 }
5388 // One more try: Various low-level checks bottom out in
5389 // uncommon traps. If the debug-info of the trap omits
5390 // any reference to the allocation, as we've already
5391 // observed, then there can be no objection to the trap.
5392 bool found_trap = false;
5393 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5394 Node* obs = not_ctl->fast_out(j);
5395 if (obs->in(0) == not_ctl && obs->is_Call() &&
5396 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5397 found_trap = true; break;
5398 }
5399 }
5400 if (found_trap) {
5401 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5402 continue;
5403 }
5404 }
5405 return NULL;
5406 }
5407
5408 // If we get this far, we have an allocation which immediately
5409 // precedes the arraycopy, and we can take over zeroing the new object.
5410 // The arraycopy will finish the initialization, and provide
5411 // a new control state to which we will anchor the destination pointer.
5412
5413 return alloc;
5414 }
5415
5416 // Helper for initialization of arrays, creating a ClearArray.
5417 // It writes zero bits in [start..end), within the body of an array object.
5418 // The memory effects are all chained onto the 'adr_type' alias category.
5419 //
5420 // Since the object is otherwise uninitialized, we are free
5421 // to put a little "slop" around the edges of the cleared area,
5422 // as long as it does not go back into the array's header,
5423 // or beyond the array end within the heap.
5424 //
5425 // The lower edge can be rounded down to the nearest jint and the
5426 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5427 //
5428 // Arguments:
5429 // adr_type memory slice where writes are generated
5430 // dest oop of the destination array
5431 // basic_elem_type element type of the destination
5432 // slice_idx array index of first element to store
5433 // slice_len number of elements to store (or NULL)
5434 // dest_size total size in bytes of the array object
5435 //
5436 // Exactly one of slice_len or dest_size must be non-NULL.
5437 // If dest_size is non-NULL, zeroing extends to the end of the object.
5438 // If slice_len is non-NULL, the slice_idx value must be a constant.
5439 void
5440 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5441 Node* dest,
5442 BasicType basic_elem_type,
5443 Node* slice_idx,
5444 Node* slice_len,
5445 Node* dest_size) {
5446 // one or the other but not both of slice_len and dest_size:
5447 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5448 if (slice_len == NULL) slice_len = top();
5449 if (dest_size == NULL) dest_size = top();
5450
5451 // operate on this memory slice:
5452 Node* mem = memory(adr_type); // memory slice to operate on
5453
5454 // scaling and rounding of indexes:
5455 int scale = exact_log2(type2aelembytes(basic_elem_type));
5456 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5457 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5458 int bump_bit = (-1 << scale) & BytesPerInt;
5459
5460 // determine constant starts and ends
5461 const intptr_t BIG_NEG = -128;
5462 assert(BIG_NEG + 2*abase < 0, "neg enough");
5463 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5464 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5465 if (slice_len_con == 0) {
5466 return; // nothing to do here
5467 }
5468 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5469 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5470 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5471 assert(end_con < 0, "not two cons");
5472 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5473 BytesPerLong);
5474 }
5475
5476 if (start_con >= 0 && end_con >= 0) {
5477 // Constant start and end. Simple.
5478 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5479 start_con, end_con, &_gvn);
5480 } else if (start_con >= 0 && dest_size != top()) {
5481 // Constant start, pre-rounded end after the tail of the array.
5482 Node* end = dest_size;
5483 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5484 start_con, end, &_gvn);
5485 } else if (start_con >= 0 && slice_len != top()) {
5486 // Constant start, non-constant end. End needs rounding up.
5487 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5488 intptr_t end_base = abase + (slice_idx_con << scale);
5489 int end_round = (-1 << scale) & (BytesPerLong - 1);
5490 Node* end = ConvI2X(slice_len);
5491 if (scale != 0)
5492 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5493 end_base += end_round;
5494 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5495 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5496 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5497 start_con, end, &_gvn);
5498 } else if (start_con < 0 && dest_size != top()) {
5499 // Non-constant start, pre-rounded end after the tail of the array.
5500 // This is almost certainly a "round-to-end" operation.
5501 Node* start = slice_idx;
5502 start = ConvI2X(start);
5503 if (scale != 0)
5504 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5505 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5506 if ((bump_bit | clear_low) != 0) {
5507 int to_clear = (bump_bit | clear_low);
5508 // Align up mod 8, then store a jint zero unconditionally
5509 // just before the mod-8 boundary.
5510 if (((abase + bump_bit) & ~to_clear) - bump_bit
5511 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5512 bump_bit = 0;
5513 assert((abase & to_clear) == 0, "array base must be long-aligned");
5514 } else {
5515 // Bump 'start' up to (or past) the next jint boundary:
5516 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5517 assert((abase & clear_low) == 0, "array base must be int-aligned");
5518 }
5519 // Round bumped 'start' down to jlong boundary in body of array.
5520 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5521 if (bump_bit != 0) {
5522 // Store a zero to the immediately preceding jint:
5523 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5524 Node* p1 = basic_plus_adr(dest, x1);
5525 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5526 mem = _gvn.transform(mem);
5527 }
5528 }
5529 Node* end = dest_size; // pre-rounded
5530 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5531 start, end, &_gvn);
5532 } else {
5533 // Non-constant start, unrounded non-constant end.
5534 // (Nobody zeroes a random midsection of an array using this routine.)
5535 ShouldNotReachHere(); // fix caller
5536 }
5537
5538 // Done.
5539 set_memory(mem, adr_type);
5540 }
5541
5542
5543 bool
5544 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5545 BasicType basic_elem_type,
5546 AllocateNode* alloc,
5547 Node* src, Node* src_offset,
5548 Node* dest, Node* dest_offset,
5549 Node* dest_size, bool dest_uninitialized) {
5550 // See if there is an advantage from block transfer.
5551 int scale = exact_log2(type2aelembytes(basic_elem_type));
5552 if (scale >= LogBytesPerLong)
5553 return false; // it is already a block transfer
5554
5555 // Look at the alignment of the starting offsets.
5556 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5557
5558 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5559 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5560 if (src_off_con < 0 || dest_off_con < 0)
5561 // At present, we can only understand constants.
5562 return false;
5563
5564 intptr_t src_off = abase + (src_off_con << scale);
5565 intptr_t dest_off = abase + (dest_off_con << scale);
5566
5567 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5568 // Non-aligned; too bad.
5569 // One more chance: Pick off an initial 32-bit word.
5570 // This is a common case, since abase can be odd mod 8.
5571 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5572 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5573 Node* sptr = basic_plus_adr(src, src_off);
5574 Node* dptr = basic_plus_adr(dest, dest_off);
5575 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5576 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5577 src_off += BytesPerInt;
5578 dest_off += BytesPerInt;
5579 } else {
5580 return false;
5581 }
5582 }
5583 assert(src_off % BytesPerLong == 0, "");
5584 assert(dest_off % BytesPerLong == 0, "");
5585
5586 // Do this copy by giant steps.
5587 Node* sptr = basic_plus_adr(src, src_off);
5588 Node* dptr = basic_plus_adr(dest, dest_off);
5589 Node* countx = dest_size;
5590 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5591 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5592
5593 bool disjoint_bases = true; // since alloc != NULL
5594 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5595 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5596
5597 return true;
5598 }
5599
5600
5601 // Helper function; generates code for the slow case.
5602 // We make a call to a runtime method which emulates the native method,
5603 // but without the native wrapper overhead.
5604 void
5605 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5606 Node* src, Node* src_offset,
5607 Node* dest, Node* dest_offset,
5608 Node* copy_length, bool dest_uninitialized) {
5609 assert(!dest_uninitialized, "Invariant");
5610 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5611 OptoRuntime::slow_arraycopy_Type(),
5612 OptoRuntime::slow_arraycopy_Java(),
5613 "slow_arraycopy", adr_type,
5614 src, src_offset, dest, dest_offset,
5615 copy_length);
5616
5617 // Handle exceptions thrown by this fellow:
5618 make_slow_call_ex(call, env()->Throwable_klass(), false);
5619 }
5620
5621 // Helper function; generates code for cases requiring runtime checks.
5622 Node*
5623 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5624 Node* dest_elem_klass,
5625 Node* src, Node* src_offset,
5626 Node* dest, Node* dest_offset,
5627 Node* copy_length, bool dest_uninitialized) {
5628 if (stopped()) return NULL;
5629
5630 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5631 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5632 return NULL;
5633 }
5634
5635 // Pick out the parameters required to perform a store-check
5636 // for the target array. This is an optimistic check. It will
5637 // look in each non-null element's class, at the desired klass's
5638 // super_check_offset, for the desired klass.
5639 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5640 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5641 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5642 Node* check_offset = ConvI2X(_gvn.transform(n3));
5643 Node* check_value = dest_elem_klass;
5644
5645 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5646 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5647
5648 // (We know the arrays are never conjoint, because their types differ.)
5649 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5650 OptoRuntime::checkcast_arraycopy_Type(),
5651 copyfunc_addr, "checkcast_arraycopy", adr_type,
5652 // five arguments, of which two are
5653 // intptr_t (jlong in LP64)
5654 src_start, dest_start,
5655 copy_length XTOP,
5656 check_offset XTOP,
5657 check_value);
5658
5659 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5660 }
5661
5662
5663 // Helper function; generates code for cases requiring runtime checks.
5664 Node*
5665 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5666 Node* src, Node* src_offset,
5667 Node* dest, Node* dest_offset,
5668 Node* copy_length, bool dest_uninitialized) {
5669 assert(!dest_uninitialized, "Invariant");
5670 if (stopped()) return NULL;
5671 address copyfunc_addr = StubRoutines::generic_arraycopy();
5672 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5673 return NULL;
5674 }
5675
5676 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5677 OptoRuntime::generic_arraycopy_Type(),
5678 copyfunc_addr, "generic_arraycopy", adr_type,
5679 src, src_offset, dest, dest_offset, copy_length);
5680
5681 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5682 }
5683
5684 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5685 void
5686 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5687 BasicType basic_elem_type,
5688 bool disjoint_bases,
5689 Node* src, Node* src_offset,
5690 Node* dest, Node* dest_offset,
5691 Node* copy_length, bool dest_uninitialized) {
5692 if (stopped()) return; // nothing to do
5693
5694 Node* src_start = src;
5695 Node* dest_start = dest;
5696 if (src_offset != NULL || dest_offset != NULL) {
5697 assert(src_offset != NULL && dest_offset != NULL, "");
5698 src_start = array_element_address(src, src_offset, basic_elem_type);
5699 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5700 }
5701
5702 // Figure out which arraycopy runtime method to call.
5703 const char* copyfunc_name = "arraycopy";
5704 address copyfunc_addr =
5705 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5706 disjoint_bases, copyfunc_name, dest_uninitialized);
5707
5708 // Call it. Note that the count_ix value is not scaled to a byte-size.
5709 make_runtime_call(RC_LEAF|RC_NO_FP,
5710 OptoRuntime::fast_arraycopy_Type(),
5711 copyfunc_addr, copyfunc_name, adr_type,
5712 src_start, dest_start, copy_length XTOP);
5713 }
5714
5715 //-------------inline_encodeISOArray-----------------------------------
5716 // encode char[] to byte[] in ISO_8859_1
5717 bool LibraryCallKit::inline_encodeISOArray() {
5718 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5719 // no receiver since it is static method
5720 Node *src = argument(0);
5721 Node *src_offset = argument(1);
5722 Node *dst = argument(2);
5723 Node *dst_offset = argument(3);
5724 Node *length = argument(4);
5725
5726 const Type* src_type = src->Value(&_gvn);
5727 const Type* dst_type = dst->Value(&_gvn);
5728 const TypeAryPtr* top_src = src_type->isa_aryptr();
5729 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5730 if (top_src == NULL || top_src->klass() == NULL ||
5731 top_dest == NULL || top_dest->klass() == NULL) {
5732 // failed array check
5733 return false;
5734 }
5735
5736 // Figure out the size and type of the elements we will be copying.
5737 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5738 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5739 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5740 return false;
5741 }
5742 Node* src_start = array_element_address(src, src_offset, src_elem);
5743 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5744 // 'src_start' points to src array + scaled offset
5745 // 'dst_start' points to dst array + scaled offset
5746
5747 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5748 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5749 enc = _gvn.transform(enc);
5750 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5751 set_memory(res_mem, mtype);
5752 set_result(enc);
5753 return true;
5754 }
5755
5756 //-------------inline_multiplyToLen-----------------------------------
5757 bool LibraryCallKit::inline_multiplyToLen() {
5758 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5759
5760 address stubAddr = StubRoutines::multiplyToLen();
5761 if (stubAddr == NULL) {
5762 return false; // Intrinsic's stub is not implemented on this platform
5763 }
5764 const char* stubName = "multiplyToLen";
5765
5766 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5767
5768 Node* x = argument(1);
5769 Node* xlen = argument(2);
5770 Node* y = argument(3);
5771 Node* ylen = argument(4);
5772 Node* z = argument(5);
5773
5774 const Type* x_type = x->Value(&_gvn);
5775 const Type* y_type = y->Value(&_gvn);
5776 const TypeAryPtr* top_x = x_type->isa_aryptr();
5777 const TypeAryPtr* top_y = y_type->isa_aryptr();
5778 if (top_x == NULL || top_x->klass() == NULL ||
5779 top_y == NULL || top_y->klass() == NULL) {
5780 // failed array check
5781 return false;
5782 }
5783
5784 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5785 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5786 if (x_elem != T_INT || y_elem != T_INT) {
5787 return false;
5788 }
5789
5790 // Set the original stack and the reexecute bit for the interpreter to reexecute
5791 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5792 // on the return from z array allocation in runtime.
5793 { PreserveReexecuteState preexecs(this);
5794 jvms()->set_should_reexecute(true);
5795
5796 Node* x_start = array_element_address(x, intcon(0), x_elem);
5797 Node* y_start = array_element_address(y, intcon(0), y_elem);
5798 // 'x_start' points to x array + scaled xlen
5799 // 'y_start' points to y array + scaled ylen
5800
5801 // Allocate the result array
5802 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5803 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5804 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5805
5806 IdealKit ideal(this);
5807
5808 #define __ ideal.
5809 Node* one = __ ConI(1);
5810 Node* zero = __ ConI(0);
5811 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5812 __ set(need_alloc, zero);
5813 __ set(z_alloc, z);
5814 __ if_then(z, BoolTest::eq, null()); {
5815 __ increment (need_alloc, one);
5816 } __ else_(); {
5817 // Update graphKit memory and control from IdealKit.
5818 sync_kit(ideal);
5819 Node* zlen_arg = load_array_length(z);
5820 // Update IdealKit memory and control from graphKit.
5821 __ sync_kit(this);
5822 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5823 __ increment (need_alloc, one);
5824 } __ end_if();
5825 } __ end_if();
5826
5827 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5828 // Update graphKit memory and control from IdealKit.
5829 sync_kit(ideal);
5830 Node * narr = new_array(klass_node, zlen, 1);
5831 // Update IdealKit memory and control from graphKit.
5832 __ sync_kit(this);
5833 __ set(z_alloc, narr);
5834 } __ end_if();
5835
5836 sync_kit(ideal);
5837 z = __ value(z_alloc);
5838 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5839 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5840 // Final sync IdealKit and GraphKit.
5841 final_sync(ideal);
5842 #undef __
5843
5844 Node* z_start = array_element_address(z, intcon(0), T_INT);
5845
5846 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5847 OptoRuntime::multiplyToLen_Type(),
5848 stubAddr, stubName, TypePtr::BOTTOM,
5849 x_start, xlen, y_start, ylen, z_start, zlen);
5850 } // original reexecute is set back here
5851
5852 C->set_has_split_ifs(true); // Has chance for split-if optimization
5853 set_result(z);
5854 return true;
5855 }
5856
5857
5858 /**
5859 * Calculate CRC32 for byte.
5860 * int java.util.zip.CRC32.update(int crc, int b)
5861 */
5862 bool LibraryCallKit::inline_updateCRC32() {
5863 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5864 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5865 // no receiver since it is static method
5866 Node* crc = argument(0); // type: int
5867 Node* b = argument(1); // type: int
5868
5869 /*
5870 * int c = ~ crc;
5871 * b = timesXtoThe32[(b ^ c) & 0xFF];
5872 * b = b ^ (c >>> 8);
5873 * crc = ~b;
5874 */
5875
5876 Node* M1 = intcon(-1);
5877 crc = _gvn.transform(new (C) XorINode(crc, M1));
5878 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5879 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5880
5881 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5882 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5883 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5884 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5885
5886 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5887 result = _gvn.transform(new (C) XorINode(crc, result));
5888 result = _gvn.transform(new (C) XorINode(result, M1));
5889 set_result(result);
5890 return true;
5891 }
5892
5893 /**
5894 * Calculate CRC32 for byte[] array.
5895 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5896 */
5897 bool LibraryCallKit::inline_updateBytesCRC32() {
5898 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5899 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5900 // no receiver since it is static method
5901 Node* crc = argument(0); // type: int
5902 Node* src = argument(1); // type: oop
5903 Node* offset = argument(2); // type: int
5904 Node* length = argument(3); // type: int
5905
5906 const Type* src_type = src->Value(&_gvn);
5907 const TypeAryPtr* top_src = src_type->isa_aryptr();
5908 if (top_src == NULL || top_src->klass() == NULL) {
5909 // failed array check
5910 return false;
5911 }
5912
5913 // Figure out the size and type of the elements we will be copying.
5914 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5915 if (src_elem != T_BYTE) {
5916 return false;
5917 }
5918
5919 // 'src_start' points to src array + scaled offset
5920 Node* src_start = array_element_address(src, offset, src_elem);
5921
5922 // We assume that range check is done by caller.
5923 // TODO: generate range check (offset+length < src.length) in debug VM.
5924
5925 // Call the stub.
5926 address stubAddr = StubRoutines::updateBytesCRC32();
5927 const char *stubName = "updateBytesCRC32";
5928
5929 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5930 stubAddr, stubName, TypePtr::BOTTOM,
5931 crc, src_start, length);
5932 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5933 set_result(result);
5934 return true;
5935 }
5936
5937 /**
5938 * Calculate CRC32 for ByteBuffer.
5939 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5940 */
5941 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5942 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5943 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5944 // no receiver since it is static method
5945 Node* crc = argument(0); // type: int
5946 Node* src = argument(1); // type: long
5947 Node* offset = argument(3); // type: int
5948 Node* length = argument(4); // type: int
5949
5950 src = ConvL2X(src); // adjust Java long to machine word
5951 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5952 offset = ConvI2X(offset);
5953
5954 // 'src_start' points to src array + scaled offset
5955 Node* src_start = basic_plus_adr(top(), base, offset);
5956
5957 // Call the stub.
5958 address stubAddr = StubRoutines::updateBytesCRC32();
5959 const char *stubName = "updateBytesCRC32";
5960
5961 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5962 stubAddr, stubName, TypePtr::BOTTOM,
5963 crc, src_start, length);
5964 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5965 set_result(result);
5966 return true;
5967 }
5968
5969 //----------------------------inline_reference_get----------------------------
5970 // public T java.lang.ref.Reference.get();
5971 bool LibraryCallKit::inline_reference_get() {
5972 const int referent_offset = java_lang_ref_Reference::referent_offset;
5973 guarantee(referent_offset > 0, "should have already been set");
5974
5975 // Get the argument:
5976 Node* reference_obj = null_check_receiver();
5977 if (stopped()) return true;
5978
5979 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5980
5981 ciInstanceKlass* klass = env()->Object_klass();
5982 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5983
5984 Node* no_ctrl = NULL;
5985 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5986
5987 // Use the pre-barrier to record the value in the referent field
5988 pre_barrier(false /* do_load */,
5989 control(),
5990 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5991 result /* pre_val */,
5992 T_OBJECT);
5993
5994 // Add memory barrier to prevent commoning reads from this field
5995 // across safepoint since GC can change its value.
5996 insert_mem_bar(Op_MemBarCPUOrder);
5997
5998 set_result(result);
5999 return true;
6000 }
6001
6002
6003 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6004 bool is_exact=true, bool is_static=false) {
6005
6006 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6007 assert(tinst != NULL, "obj is null");
6008 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6009 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6010
6011 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6012 ciSymbol::make(fieldTypeString),
6013 is_static);
6014 if (field == NULL) return (Node *) NULL;
6015 assert (field != NULL, "undefined field");
6016
6017 // Next code copied from Parse::do_get_xxx():
6018
6019 // Compute address and memory type.
6020 int offset = field->offset_in_bytes();
6021 bool is_vol = field->is_volatile();
6022 ciType* field_klass = field->type();
6023 assert(field_klass->is_loaded(), "should be loaded");
6024 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6025 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6026 BasicType bt = field->layout_type();
6027
6028 // Build the resultant type of the load
6029 const Type *type;
6030 if (bt == T_OBJECT) {
6031 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6032 } else {
6033 type = Type::get_const_basic_type(bt);
6034 }
6035
6036 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6037 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6038 }
6039 // Build the load.
6040 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6041 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
6042 // If reference is volatile, prevent following memory ops from
6043 // floating up past the volatile read. Also prevents commoning
6044 // another volatile read.
6045 if (is_vol) {
6046 // Memory barrier includes bogus read of value to force load BEFORE membar
6047 insert_mem_bar(Op_MemBarAcquire, loadedField);
6048 }
6049 return loadedField;
6050 }
6051
6052
6053 //------------------------------inline_aescrypt_Block-----------------------
6054 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6055 address stubAddr;
6056 const char *stubName;
6057 assert(UseAES, "need AES instruction support");
6058
6059 switch(id) {
6060 case vmIntrinsics::_aescrypt_encryptBlock:
6061 stubAddr = StubRoutines::aescrypt_encryptBlock();
6062 stubName = "aescrypt_encryptBlock";
6063 break;
6064 case vmIntrinsics::_aescrypt_decryptBlock:
6065 stubAddr = StubRoutines::aescrypt_decryptBlock();
6066 stubName = "aescrypt_decryptBlock";
6067 break;
6068 }
6069 if (stubAddr == NULL) return false;
6070
6071 Node* aescrypt_object = argument(0);
6072 Node* src = argument(1);
6073 Node* src_offset = argument(2);
6074 Node* dest = argument(3);
6075 Node* dest_offset = argument(4);
6076
6077 // (1) src and dest are arrays.
6078 const Type* src_type = src->Value(&_gvn);
6079 const Type* dest_type = dest->Value(&_gvn);
6080 const TypeAryPtr* top_src = src_type->isa_aryptr();
6081 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6082 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6083
6084 // for the quick and dirty code we will skip all the checks.
6085 // we are just trying to get the call to be generated.
6086 Node* src_start = src;
6087 Node* dest_start = dest;
6088 if (src_offset != NULL || dest_offset != NULL) {
6089 assert(src_offset != NULL && dest_offset != NULL, "");
6090 src_start = array_element_address(src, src_offset, T_BYTE);
6091 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6092 }
6093
6094 // now need to get the start of its expanded key array
6095 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6096 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6097 if (k_start == NULL) return false;
6098
6099 if (Matcher::pass_original_key_for_aes()) {
6100 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6101 // compatibility issues between Java key expansion and SPARC crypto instructions
6102 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6103 if (original_k_start == NULL) return false;
6104
6105 // Call the stub.
6106 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6107 stubAddr, stubName, TypePtr::BOTTOM,
6108 src_start, dest_start, k_start, original_k_start);
6109 } else {
6110 // Call the stub.
6111 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6112 stubAddr, stubName, TypePtr::BOTTOM,
6113 src_start, dest_start, k_start);
6114 }
6115
6116 return true;
6117 }
6118
6119 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6120 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6121 address stubAddr;
6122 const char *stubName;
6123
6124 assert(UseAES, "need AES instruction support");
6125
6126 switch(id) {
6127 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6128 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6129 stubName = "cipherBlockChaining_encryptAESCrypt";
6130 break;
6131 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6132 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6133 stubName = "cipherBlockChaining_decryptAESCrypt";
6134 break;
6135 }
6136 if (stubAddr == NULL) return false;
6137
6138 Node* cipherBlockChaining_object = argument(0);
6139 Node* src = argument(1);
6140 Node* src_offset = argument(2);
6141 Node* len = argument(3);
6142 Node* dest = argument(4);
6143 Node* dest_offset = argument(5);
6144
6145 // (1) src and dest are arrays.
6146 const Type* src_type = src->Value(&_gvn);
6147 const Type* dest_type = dest->Value(&_gvn);
6148 const TypeAryPtr* top_src = src_type->isa_aryptr();
6149 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6150 assert (top_src != NULL && top_src->klass() != NULL
6151 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6152
6153 // checks are the responsibility of the caller
6154 Node* src_start = src;
6155 Node* dest_start = dest;
6156 if (src_offset != NULL || dest_offset != NULL) {
6157 assert(src_offset != NULL && dest_offset != NULL, "");
6158 src_start = array_element_address(src, src_offset, T_BYTE);
6159 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6160 }
6161
6162 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6163 // (because of the predicated logic executed earlier).
6164 // so we cast it here safely.
6165 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6166
6167 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6168 if (embeddedCipherObj == NULL) return false;
6169
6170 // cast it to what we know it will be at runtime
6171 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6172 assert(tinst != NULL, "CBC obj is null");
6173 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6174 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6175 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6176
6177 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6178 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6179 const TypeOopPtr* xtype = aklass->as_instance_type();
6180 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6181 aescrypt_object = _gvn.transform(aescrypt_object);
6182
6183 // we need to get the start of the aescrypt_object's expanded key array
6184 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6185 if (k_start == NULL) return false;
6186
6187 // similarly, get the start address of the r vector
6188 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6189 if (objRvec == NULL) return false;
6190 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6191
6192 Node* cbcCrypt;
6193 if (Matcher::pass_original_key_for_aes()) {
6194 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6195 // compatibility issues between Java key expansion and SPARC crypto instructions
6196 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6197 if (original_k_start == NULL) return false;
6198
6199 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6200 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6201 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6202 stubAddr, stubName, TypePtr::BOTTOM,
6203 src_start, dest_start, k_start, r_start, len, original_k_start);
6204 } else {
6205 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6206 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6207 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6208 stubAddr, stubName, TypePtr::BOTTOM,
6209 src_start, dest_start, k_start, r_start, len);
6210 }
6211
6212 // return cipher length (int)
6213 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6214 set_result(retvalue);
6215 return true;
6216 }
6217
6218 //------------------------------get_key_start_from_aescrypt_object-----------------------
6219 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6220 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6221 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6222 if (objAESCryptKey == NULL) return (Node *) NULL;
6223
6224 // now have the array, need to get the start address of the K array
6225 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6226 return k_start;
6227 }
6228
6229 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6230 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6231 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6232 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6233 if (objAESCryptKey == NULL) return (Node *) NULL;
6234
6235 // now have the array, need to get the start address of the lastKey array
6236 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6237 return original_k_start;
6238 }
6239
6240 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6241 // Return node representing slow path of predicate check.
6242 // the pseudo code we want to emulate with this predicate is:
6243 // for encryption:
6244 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6245 // for decryption:
6246 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6247 // note cipher==plain is more conservative than the original java code but that's OK
6248 //
6249 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6250 // The receiver was checked for NULL already.
6251 Node* objCBC = argument(0);
6252
6253 // Load embeddedCipher field of CipherBlockChaining object.
6254 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6255
6256 // get AESCrypt klass for instanceOf check
6257 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6258 // will have same classloader as CipherBlockChaining object
6259 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6260 assert(tinst != NULL, "CBCobj is null");
6261 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6262
6263 // we want to do an instanceof comparison against the AESCrypt class
6264 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6265 if (!klass_AESCrypt->is_loaded()) {
6266 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6267 Node* ctrl = control();
6268 set_control(top()); // no regular fast path
6269 return ctrl;
6270 }
6271 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6272
6273 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6274 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6275 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6276
6277 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6278
6279 // for encryption, we are done
6280 if (!decrypting)
6281 return instof_false; // even if it is NULL
6282
6283 // for decryption, we need to add a further check to avoid
6284 // taking the intrinsic path when cipher and plain are the same
6285 // see the original java code for why.
6286 RegionNode* region = new(C) RegionNode(3);
6287 region->init_req(1, instof_false);
6288 Node* src = argument(1);
6289 Node* dest = argument(4);
6290 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6291 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6292 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6293 region->init_req(2, src_dest_conjoint);
6294
6295 record_for_igvn(region);
6296 return _gvn.transform(region);
6297 }
6298
6299 //------------------------------inline_sha_implCompress-----------------------
6300 //
6301 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6302 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6303 //
6304 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6305 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6306 //
6307 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6308 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6309 //
6310 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6311 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6312
6313 Node* sha_obj = argument(0);
6314 Node* src = argument(1); // type oop
6315 Node* ofs = argument(2); // type int
6316
6317 const Type* src_type = src->Value(&_gvn);
6318 const TypeAryPtr* top_src = src_type->isa_aryptr();
6319 if (top_src == NULL || top_src->klass() == NULL) {
6320 // failed array check
6321 return false;
6322 }
6323 // Figure out the size and type of the elements we will be copying.
6324 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6325 if (src_elem != T_BYTE) {
6326 return false;
6327 }
6328 // 'src_start' points to src array + offset
6329 Node* src_start = array_element_address(src, ofs, src_elem);
6330 Node* state = NULL;
6331 address stubAddr;
6332 const char *stubName;
6333
6334 switch(id) {
6335 case vmIntrinsics::_sha_implCompress:
6336 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6337 state = get_state_from_sha_object(sha_obj);
6338 stubAddr = StubRoutines::sha1_implCompress();
6339 stubName = "sha1_implCompress";
6340 break;
6341 case vmIntrinsics::_sha2_implCompress:
6342 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6343 state = get_state_from_sha_object(sha_obj);
6344 stubAddr = StubRoutines::sha256_implCompress();
6345 stubName = "sha256_implCompress";
6346 break;
6347 case vmIntrinsics::_sha5_implCompress:
6348 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6349 state = get_state_from_sha5_object(sha_obj);
6350 stubAddr = StubRoutines::sha512_implCompress();
6351 stubName = "sha512_implCompress";
6352 break;
6353 default:
6354 fatal_unexpected_iid(id);
6355 return false;
6356 }
6357 if (state == NULL) return false;
6358
6359 // Call the stub.
6360 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6361 stubAddr, stubName, TypePtr::BOTTOM,
6362 src_start, state);
6363
6364 return true;
6365 }
6366
6367 //------------------------------inline_digestBase_implCompressMB-----------------------
6368 //
6369 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6370 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6371 //
6372 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6373 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6374 "need SHA1/SHA256/SHA512 instruction support");
6375 assert((uint)predicate < 3, "sanity");
6376 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6377
6378 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6379 Node* src = argument(1); // byte[] array
6380 Node* ofs = argument(2); // type int
6381 Node* limit = argument(3); // type int
6382
6383 const Type* src_type = src->Value(&_gvn);
6384 const TypeAryPtr* top_src = src_type->isa_aryptr();
6385 if (top_src == NULL || top_src->klass() == NULL) {
6386 // failed array check
6387 return false;
6388 }
6389 // Figure out the size and type of the elements we will be copying.
6390 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6391 if (src_elem != T_BYTE) {
6392 return false;
6393 }
6394 // 'src_start' points to src array + offset
6395 Node* src_start = array_element_address(src, ofs, src_elem);
6396
6397 const char* klass_SHA_name = NULL;
6398 const char* stub_name = NULL;
6399 address stub_addr = NULL;
6400 bool long_state = false;
6401
6402 switch (predicate) {
6403 case 0:
6404 if (UseSHA1Intrinsics) {
6405 klass_SHA_name = "sun/security/provider/SHA";
6406 stub_name = "sha1_implCompressMB";
6407 stub_addr = StubRoutines::sha1_implCompressMB();
6408 }
6409 break;
6410 case 1:
6411 if (UseSHA256Intrinsics) {
6412 klass_SHA_name = "sun/security/provider/SHA2";
6413 stub_name = "sha256_implCompressMB";
6414 stub_addr = StubRoutines::sha256_implCompressMB();
6415 }
6416 break;
6417 case 2:
6418 if (UseSHA512Intrinsics) {
6419 klass_SHA_name = "sun/security/provider/SHA5";
6420 stub_name = "sha512_implCompressMB";
6421 stub_addr = StubRoutines::sha512_implCompressMB();
6422 long_state = true;
6423 }
6424 break;
6425 default:
6426 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6427 }
6428 if (klass_SHA_name != NULL) {
6429 // get DigestBase klass to lookup for SHA klass
6430 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6431 assert(tinst != NULL, "digestBase_obj is not instance???");
6432 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6433
6434 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6435 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6436 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6437 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6438 }
6439 return false;
6440 }
6441 //------------------------------inline_sha_implCompressMB-----------------------
6442 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6443 bool long_state, address stubAddr, const char *stubName,
6444 Node* src_start, Node* ofs, Node* limit) {
6445 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6446 const TypeOopPtr* xtype = aklass->as_instance_type();
6447 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6448 sha_obj = _gvn.transform(sha_obj);
6449
6450 Node* state;
6451 if (long_state) {
6452 state = get_state_from_sha5_object(sha_obj);
6453 } else {
6454 state = get_state_from_sha_object(sha_obj);
6455 }
6456 if (state == NULL) return false;
6457
6458 // Call the stub.
6459 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6460 OptoRuntime::digestBase_implCompressMB_Type(),
6461 stubAddr, stubName, TypePtr::BOTTOM,
6462 src_start, state, ofs, limit);
6463 // return ofs (int)
6464 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6465 set_result(result);
6466
6467 return true;
6468 }
6469
6470 //------------------------------get_state_from_sha_object-----------------------
6471 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6472 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6473 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6474 if (sha_state == NULL) return (Node *) NULL;
6475
6476 // now have the array, need to get the start address of the state array
6477 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6478 return state;
6479 }
6480
6481 //------------------------------get_state_from_sha5_object-----------------------
6482 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6483 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6484 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6485 if (sha_state == NULL) return (Node *) NULL;
6486
6487 // now have the array, need to get the start address of the state array
6488 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6489 return state;
6490 }
6491
6492 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6493 // Return node representing slow path of predicate check.
6494 // the pseudo code we want to emulate with this predicate is:
6495 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6496 //
6497 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6498 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6499 "need SHA1/SHA256/SHA512 instruction support");
6500 assert((uint)predicate < 3, "sanity");
6501
6502 // The receiver was checked for NULL already.
6503 Node* digestBaseObj = argument(0);
6504
6505 // get DigestBase klass for instanceOf check
6506 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6507 assert(tinst != NULL, "digestBaseObj is null");
6508 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6509
6510 const char* klass_SHA_name = NULL;
6511 switch (predicate) {
6512 case 0:
6513 if (UseSHA1Intrinsics) {
6514 // we want to do an instanceof comparison against the SHA class
6515 klass_SHA_name = "sun/security/provider/SHA";
6516 }
6517 break;
6518 case 1:
6519 if (UseSHA256Intrinsics) {
6520 // we want to do an instanceof comparison against the SHA2 class
6521 klass_SHA_name = "sun/security/provider/SHA2";
6522 }
6523 break;
6524 case 2:
6525 if (UseSHA512Intrinsics) {
6526 // we want to do an instanceof comparison against the SHA5 class
6527 klass_SHA_name = "sun/security/provider/SHA5";
6528 }
6529 break;
6530 default:
6531 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6532 }
6533
6534 ciKlass* klass_SHA = NULL;
6535 if (klass_SHA_name != NULL) {
6536 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6537 }
6538 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6539 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6540 Node* ctrl = control();
6541 set_control(top()); // no intrinsic path
6542 return ctrl;
6543 }
6544 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6545
6546 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6547 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6548 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6549 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6550
6551 return instof_false; // even if it is NULL
6552 }
6553
6554 bool LibraryCallKit::inline_profileBoolean() {
6555 Node* counts = argument(1);
6556 const TypeAryPtr* ary = NULL;
6557 ciArray* aobj = NULL;
6558 if (counts->is_Con()
6559 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6560 && (aobj = ary->const_oop()->as_array()) != NULL
6561 && (aobj->length() == 2)) {
6562 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6563 jint false_cnt = aobj->element_value(0).as_int();
6564 jint true_cnt = aobj->element_value(1).as_int();
6565
6566 method()->set_injected_profile(true);
6567
6568 if (C->log() != NULL) {
6569 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6570 false_cnt, true_cnt);
6571 }
6572
6573 if (false_cnt + true_cnt == 0) {
6574 // According to profile, never executed.
6575 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6576 Deoptimization::Action_reinterpret);
6577 return true;
6578 }
6579
6580 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6581 // is a number of each value occurrences.
6582 Node* result = argument(0);
6583 if (false_cnt == 0 || true_cnt == 0) {
6584 // According to profile, one value has been never seen.
6585 int expected_val = (false_cnt == 0) ? 1 : 0;
6586
6587 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6588 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6589
6590 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6591 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6592 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6593
6594 { // Slow path: uncommon trap for never seen value and then reexecute
6595 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6596 // the value has been seen at least once.
6597 PreserveJVMState pjvms(this);
6598 PreserveReexecuteState preexecs(this);
6599 jvms()->set_should_reexecute(true);
6600
6601 set_control(slow_path);
6602 set_i_o(i_o());
6603
6604 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6605 Deoptimization::Action_reinterpret);
6606 }
6607 // The guard for never seen value enables sharpening of the result and
6608 // returning a constant. It allows to eliminate branches on the same value
6609 // later on.
6610 set_control(fast_path);
6611 result = intcon(expected_val);
6612 }
6613 // Stop profiling.
6614 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6615 // By replacing method body with profile data (represented as ProfileBooleanNode
6616 // on IR level) we effectively disable profiling.
6617 // It enables full speed execution once optimized code is generated.
6618 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6619 C->record_for_igvn(profile);
6620 set_result(profile);
6621 return true;
6622 } else {
6623 // Continue profiling.
6624 // Profile data isn't available at the moment. So, execute method's bytecode version.
6625 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6626 // is compiled and counters aren't available since corresponding MethodHandle
6627 // isn't a compile-time constant.
6628 return false;
6629 }
6630 }