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 bool inline_squareToLen();
328 bool inline_mulAdd();
329
330 bool inline_profileBoolean();
331 };
332
333
334 //---------------------------make_vm_intrinsic----------------------------
335 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
336 vmIntrinsics::ID id = m->intrinsic_id();
337 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
338
339 ccstr disable_intr = NULL;
340
341 if ((DisableIntrinsic[0] != '\0'
342 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
343 (method_has_option_value("DisableIntrinsic", disable_intr)
344 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
345 // disabled by a user request on the command line:
346 // example: -XX:DisableIntrinsic=_hashCode,_getClass
347 return NULL;
348 }
349
350 if (!m->is_loaded()) {
351 // do not attempt to inline unloaded methods
352 return NULL;
353 }
354
355 // Only a few intrinsics implement a virtual dispatch.
356 // They are expensive calls which are also frequently overridden.
357 if (is_virtual) {
358 switch (id) {
359 case vmIntrinsics::_hashCode:
360 case vmIntrinsics::_clone:
361 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
362 break;
363 default:
364 return NULL;
365 }
366 }
367
368 // -XX:-InlineNatives disables nearly all intrinsics:
369 if (!InlineNatives) {
370 switch (id) {
371 case vmIntrinsics::_indexOf:
372 case vmIntrinsics::_compareTo:
373 case vmIntrinsics::_equals:
374 case vmIntrinsics::_equalsC:
375 case vmIntrinsics::_getAndAddInt:
376 case vmIntrinsics::_getAndAddLong:
377 case vmIntrinsics::_getAndSetInt:
378 case vmIntrinsics::_getAndSetLong:
379 case vmIntrinsics::_getAndSetObject:
380 case vmIntrinsics::_loadFence:
381 case vmIntrinsics::_storeFence:
382 case vmIntrinsics::_fullFence:
383 break; // InlineNatives does not control String.compareTo
384 case vmIntrinsics::_Reference_get:
385 break; // InlineNatives does not control Reference.get
386 default:
387 return NULL;
388 }
389 }
390
391 int predicates = 0;
392 bool does_virtual_dispatch = false;
393
394 switch (id) {
395 case vmIntrinsics::_compareTo:
396 if (!SpecialStringCompareTo) return NULL;
397 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
398 break;
399 case vmIntrinsics::_indexOf:
400 if (!SpecialStringIndexOf) return NULL;
401 break;
402 case vmIntrinsics::_equals:
403 if (!SpecialStringEquals) return NULL;
404 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
405 break;
406 case vmIntrinsics::_equalsC:
407 if (!SpecialArraysEquals) return NULL;
408 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
409 break;
410 case vmIntrinsics::_arraycopy:
411 if (!InlineArrayCopy) return NULL;
412 break;
413 case vmIntrinsics::_copyMemory:
414 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
415 if (!InlineArrayCopy) return NULL;
416 break;
417 case vmIntrinsics::_hashCode:
418 if (!InlineObjectHash) return NULL;
419 does_virtual_dispatch = true;
420 break;
421 case vmIntrinsics::_clone:
422 does_virtual_dispatch = true;
423 case vmIntrinsics::_copyOf:
424 case vmIntrinsics::_copyOfRange:
425 if (!InlineObjectCopy) return NULL;
426 // These also use the arraycopy intrinsic mechanism:
427 if (!InlineArrayCopy) return NULL;
428 break;
429 case vmIntrinsics::_encodeISOArray:
430 if (!SpecialEncodeISOArray) return NULL;
431 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
432 break;
433 case vmIntrinsics::_checkIndex:
434 // We do not intrinsify this. The optimizer does fine with it.
435 return NULL;
436
437 case vmIntrinsics::_getCallerClass:
438 if (!UseNewReflection) return NULL;
439 if (!InlineReflectionGetCallerClass) return NULL;
440 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
441 break;
442
443 case vmIntrinsics::_bitCount_i:
444 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
445 break;
446
447 case vmIntrinsics::_bitCount_l:
448 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
449 break;
450
451 case vmIntrinsics::_numberOfLeadingZeros_i:
452 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
453 break;
454
455 case vmIntrinsics::_numberOfLeadingZeros_l:
456 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
457 break;
458
459 case vmIntrinsics::_numberOfTrailingZeros_i:
460 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
461 break;
462
463 case vmIntrinsics::_numberOfTrailingZeros_l:
464 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
465 break;
466
467 case vmIntrinsics::_reverseBytes_c:
468 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
469 break;
470 case vmIntrinsics::_reverseBytes_s:
471 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
472 break;
473 case vmIntrinsics::_reverseBytes_i:
474 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
475 break;
476 case vmIntrinsics::_reverseBytes_l:
477 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
478 break;
479
480 case vmIntrinsics::_Reference_get:
481 // Use the intrinsic version of Reference.get() so that the value in
482 // the referent field can be registered by the G1 pre-barrier code.
483 // Also add memory barrier to prevent commoning reads from this field
484 // across safepoint since GC can change it value.
485 break;
486
487 case vmIntrinsics::_compareAndSwapObject:
488 #ifdef _LP64
489 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
490 #endif
491 break;
492
493 case vmIntrinsics::_compareAndSwapLong:
494 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
495 break;
496
497 case vmIntrinsics::_getAndAddInt:
498 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
499 break;
500
501 case vmIntrinsics::_getAndAddLong:
502 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
503 break;
504
505 case vmIntrinsics::_getAndSetInt:
506 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
507 break;
508
509 case vmIntrinsics::_getAndSetLong:
510 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
511 break;
512
513 case vmIntrinsics::_getAndSetObject:
514 #ifdef _LP64
515 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
516 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
517 break;
518 #else
519 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
520 break;
521 #endif
522
523 case vmIntrinsics::_aescrypt_encryptBlock:
524 case vmIntrinsics::_aescrypt_decryptBlock:
525 if (!UseAESIntrinsics) return NULL;
526 break;
527
528 case vmIntrinsics::_multiplyToLen:
529 if (!UseMultiplyToLenIntrinsic) return NULL;
530 break;
531
532 case vmIntrinsics::_squareToLen:
533 if (!UseSquareToLenIntrinsic) return NULL;
534 break;
535
536 case vmIntrinsics::_mulAdd:
537 if (!UseMulAddIntrinsic) return NULL;
538 break;
539
540 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
541 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
542 if (!UseAESIntrinsics) return NULL;
543 // these two require the predicated logic
544 predicates = 1;
545 break;
546
547 case vmIntrinsics::_sha_implCompress:
548 if (!UseSHA1Intrinsics) return NULL;
549 break;
550
551 case vmIntrinsics::_sha2_implCompress:
552 if (!UseSHA256Intrinsics) return NULL;
553 break;
554
555 case vmIntrinsics::_sha5_implCompress:
556 if (!UseSHA512Intrinsics) return NULL;
557 break;
558
559 case vmIntrinsics::_digestBase_implCompressMB:
560 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
561 predicates = 3;
562 break;
563
564 case vmIntrinsics::_updateCRC32:
565 case vmIntrinsics::_updateBytesCRC32:
566 case vmIntrinsics::_updateByteBufferCRC32:
567 if (!UseCRC32Intrinsics) return NULL;
568 break;
569
570 case vmIntrinsics::_incrementExactI:
571 case vmIntrinsics::_addExactI:
572 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
573 break;
574 case vmIntrinsics::_incrementExactL:
575 case vmIntrinsics::_addExactL:
576 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
577 break;
578 case vmIntrinsics::_decrementExactI:
579 case vmIntrinsics::_subtractExactI:
580 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
581 break;
582 case vmIntrinsics::_decrementExactL:
583 case vmIntrinsics::_subtractExactL:
584 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
585 break;
586 case vmIntrinsics::_negateExactI:
587 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
588 break;
589 case vmIntrinsics::_negateExactL:
590 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
591 break;
592 case vmIntrinsics::_multiplyExactI:
593 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
594 break;
595 case vmIntrinsics::_multiplyExactL:
596 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
597 break;
598
599 default:
600 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
601 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
602 break;
603 }
604
605 // -XX:-InlineClassNatives disables natives from the Class class.
606 // The flag applies to all reflective calls, notably Array.newArray
607 // (visible to Java programmers as Array.newInstance).
608 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
609 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
610 if (!InlineClassNatives) return NULL;
611 }
612
613 // -XX:-InlineThreadNatives disables natives from the Thread class.
614 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
615 if (!InlineThreadNatives) return NULL;
616 }
617
618 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
619 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
620 m->holder()->name() == ciSymbol::java_lang_Float() ||
621 m->holder()->name() == ciSymbol::java_lang_Double()) {
622 if (!InlineMathNatives) return NULL;
623 }
624
625 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
626 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
627 if (!InlineUnsafeOps) return NULL;
628 }
629
630 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
631 }
632
633 //----------------------register_library_intrinsics-----------------------
634 // Initialize this file's data structures, for each Compile instance.
635 void Compile::register_library_intrinsics() {
636 // Nothing to do here.
637 }
638
639 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
640 LibraryCallKit kit(jvms, this);
641 Compile* C = kit.C;
642 int nodes = C->unique();
643 #ifndef PRODUCT
644 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
645 char buf[1000];
646 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
647 tty->print_cr("Intrinsic %s", str);
648 }
649 #endif
650 ciMethod* callee = kit.callee();
651 const int bci = kit.bci();
652
653 // Try to inline the intrinsic.
654 if (kit.try_to_inline(_last_predicate)) {
655 if (C->print_intrinsics() || C->print_inlining()) {
656 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
657 }
658 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
659 if (C->log()) {
660 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
661 vmIntrinsics::name_at(intrinsic_id()),
662 (is_virtual() ? " virtual='1'" : ""),
663 C->unique() - nodes);
664 }
665 // Push the result from the inlined method onto the stack.
666 kit.push_result();
667 return kit.transfer_exceptions_into_jvms();
668 }
669
670 // The intrinsic bailed out
671 if (C->print_intrinsics() || C->print_inlining()) {
672 if (jvms->has_method()) {
673 // Not a root compile.
674 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
675 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
676 } else {
677 // Root compile
678 tty->print("Did not generate intrinsic %s%s at bci:%d in",
679 vmIntrinsics::name_at(intrinsic_id()),
680 (is_virtual() ? " (virtual)" : ""), bci);
681 }
682 }
683 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
684 return NULL;
685 }
686
687 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
688 LibraryCallKit kit(jvms, this);
689 Compile* C = kit.C;
690 int nodes = C->unique();
691 _last_predicate = predicate;
692 #ifndef PRODUCT
693 assert(is_predicated() && predicate < predicates_count(), "sanity");
694 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
695 char buf[1000];
696 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
697 tty->print_cr("Predicate for intrinsic %s", str);
698 }
699 #endif
700 ciMethod* callee = kit.callee();
701 const int bci = kit.bci();
702
703 Node* slow_ctl = kit.try_to_predicate(predicate);
704 if (!kit.failing()) {
705 if (C->print_intrinsics() || C->print_inlining()) {
706 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
707 }
708 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
709 if (C->log()) {
710 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
711 vmIntrinsics::name_at(intrinsic_id()),
712 (is_virtual() ? " virtual='1'" : ""),
713 C->unique() - nodes);
714 }
715 return slow_ctl; // Could be NULL if the check folds.
716 }
717
718 // The intrinsic bailed out
719 if (C->print_intrinsics() || C->print_inlining()) {
720 if (jvms->has_method()) {
721 // Not a root compile.
722 const char* msg = "failed to generate predicate for intrinsic";
723 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
724 } else {
725 // Root compile
726 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
727 vmIntrinsics::name_at(intrinsic_id()),
728 (is_virtual() ? " (virtual)" : ""), bci);
729 }
730 }
731 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
732 return NULL;
733 }
734
735 bool LibraryCallKit::try_to_inline(int predicate) {
736 // Handle symbolic names for otherwise undistinguished boolean switches:
737 const bool is_store = true;
738 const bool is_native_ptr = true;
739 const bool is_static = true;
740 const bool is_volatile = true;
741
742 if (!jvms()->has_method()) {
743 // Root JVMState has a null method.
744 assert(map()->memory()->Opcode() == Op_Parm, "");
745 // Insert the memory aliasing node
746 set_all_memory(reset_memory());
747 }
748 assert(merged_memory(), "");
749
750
751 switch (intrinsic_id()) {
752 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
753 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
754 case vmIntrinsics::_getClass: return inline_native_getClass();
755
756 case vmIntrinsics::_dsin:
757 case vmIntrinsics::_dcos:
758 case vmIntrinsics::_dtan:
759 case vmIntrinsics::_dabs:
760 case vmIntrinsics::_datan2:
761 case vmIntrinsics::_dsqrt:
762 case vmIntrinsics::_dexp:
763 case vmIntrinsics::_dlog:
764 case vmIntrinsics::_dlog10:
765 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
766
767 case vmIntrinsics::_min:
768 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
769
770 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
771 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
772 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
773 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
774 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
775 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
776 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
777 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
778 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
779 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
780 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
781 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
782
783 case vmIntrinsics::_arraycopy: return inline_arraycopy();
784
785 case vmIntrinsics::_compareTo: return inline_string_compareTo();
786 case vmIntrinsics::_indexOf: return inline_string_indexOf();
787 case vmIntrinsics::_equals: return inline_string_equals();
788
789 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
790 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
791 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
792 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
793 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
794 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
795 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
796 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
797 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
798
799 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
800 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
801 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
802 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
803 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
804 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
805 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
806 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
807 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
808
809 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
810 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
811 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
812 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
813 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
814 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
815 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
816 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
817
818 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
819 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
820 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
821 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
822 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
823 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
824 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
825 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
826
827 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
828 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
829 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
830 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
831 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
832 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
833 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
834 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
835 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
836
837 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
838 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
839 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
840 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
841 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
842 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
843 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
844 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
845 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
846
847 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
848 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
849 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
850 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
851
852 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
853 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
854 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
855
856 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
857 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
858 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
859
860 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
861 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
862 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
863 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
864 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
865
866 case vmIntrinsics::_loadFence:
867 case vmIntrinsics::_storeFence:
868 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
869
870 case vmIntrinsics::_currentThread: return inline_native_currentThread();
871 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
872
873 #ifdef TRACE_HAVE_INTRINSICS
874 case vmIntrinsics::_classID: return inline_native_classID();
875 case vmIntrinsics::_threadID: return inline_native_threadID();
876 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
877 #endif
878 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
879 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
880 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
881 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
882 case vmIntrinsics::_newArray: return inline_native_newArray();
883 case vmIntrinsics::_getLength: return inline_native_getLength();
884 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
885 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
886 case vmIntrinsics::_equalsC: return inline_array_equals();
887 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
888
889 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
890
891 case vmIntrinsics::_isInstance:
892 case vmIntrinsics::_getModifiers:
893 case vmIntrinsics::_isInterface:
894 case vmIntrinsics::_isArray:
895 case vmIntrinsics::_isPrimitive:
896 case vmIntrinsics::_getSuperclass:
897 case vmIntrinsics::_getComponentType:
898 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
899
900 case vmIntrinsics::_floatToRawIntBits:
901 case vmIntrinsics::_floatToIntBits:
902 case vmIntrinsics::_intBitsToFloat:
903 case vmIntrinsics::_doubleToRawLongBits:
904 case vmIntrinsics::_doubleToLongBits:
905 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
906
907 case vmIntrinsics::_numberOfLeadingZeros_i:
908 case vmIntrinsics::_numberOfLeadingZeros_l:
909 case vmIntrinsics::_numberOfTrailingZeros_i:
910 case vmIntrinsics::_numberOfTrailingZeros_l:
911 case vmIntrinsics::_bitCount_i:
912 case vmIntrinsics::_bitCount_l:
913 case vmIntrinsics::_reverseBytes_i:
914 case vmIntrinsics::_reverseBytes_l:
915 case vmIntrinsics::_reverseBytes_s:
916 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
917
918 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
919
920 case vmIntrinsics::_Reference_get: return inline_reference_get();
921
922 case vmIntrinsics::_aescrypt_encryptBlock:
923 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
924
925 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
926 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
927 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
928
929 case vmIntrinsics::_sha_implCompress:
930 case vmIntrinsics::_sha2_implCompress:
931 case vmIntrinsics::_sha5_implCompress:
932 return inline_sha_implCompress(intrinsic_id());
933
934 case vmIntrinsics::_digestBase_implCompressMB:
935 return inline_digestBase_implCompressMB(predicate);
936
937 case vmIntrinsics::_multiplyToLen:
938 return inline_multiplyToLen();
939
940 case vmIntrinsics::_squareToLen:
941 return inline_squareToLen();
942
943 case vmIntrinsics::_mulAdd:
944 return inline_mulAdd();
945
946 case vmIntrinsics::_encodeISOArray:
947 return inline_encodeISOArray();
948
949 case vmIntrinsics::_updateCRC32:
950 return inline_updateCRC32();
951 case vmIntrinsics::_updateBytesCRC32:
952 return inline_updateBytesCRC32();
953 case vmIntrinsics::_updateByteBufferCRC32:
954 return inline_updateByteBufferCRC32();
955
956 case vmIntrinsics::_profileBoolean:
957 return inline_profileBoolean();
958
959 default:
960 // If you get here, it may be that someone has added a new intrinsic
961 // to the list in vmSymbols.hpp without implementing it here.
962 #ifndef PRODUCT
963 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
964 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
965 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
966 }
967 #endif
968 return false;
969 }
970 }
971
972 Node* LibraryCallKit::try_to_predicate(int predicate) {
973 if (!jvms()->has_method()) {
974 // Root JVMState has a null method.
975 assert(map()->memory()->Opcode() == Op_Parm, "");
976 // Insert the memory aliasing node
977 set_all_memory(reset_memory());
978 }
979 assert(merged_memory(), "");
980
981 switch (intrinsic_id()) {
982 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
983 return inline_cipherBlockChaining_AESCrypt_predicate(false);
984 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
985 return inline_cipherBlockChaining_AESCrypt_predicate(true);
986 case vmIntrinsics::_digestBase_implCompressMB:
987 return inline_digestBase_implCompressMB_predicate(predicate);
988
989 default:
990 // If you get here, it may be that someone has added a new intrinsic
991 // to the list in vmSymbols.hpp without implementing it here.
992 #ifndef PRODUCT
993 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
994 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
995 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
996 }
997 #endif
998 Node* slow_ctl = control();
999 set_control(top()); // No fast path instrinsic
1000 return slow_ctl;
1001 }
1002 }
1003
1004 //------------------------------set_result-------------------------------
1005 // Helper function for finishing intrinsics.
1006 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1007 record_for_igvn(region);
1008 set_control(_gvn.transform(region));
1009 set_result( _gvn.transform(value));
1010 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1011 }
1012
1013 //------------------------------generate_guard---------------------------
1014 // Helper function for generating guarded fast-slow graph structures.
1015 // The given 'test', if true, guards a slow path. If the test fails
1016 // then a fast path can be taken. (We generally hope it fails.)
1017 // In all cases, GraphKit::control() is updated to the fast path.
1018 // The returned value represents the control for the slow path.
1019 // The return value is never 'top'; it is either a valid control
1020 // or NULL if it is obvious that the slow path can never be taken.
1021 // Also, if region and the slow control are not NULL, the slow edge
1022 // is appended to the region.
1023 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1024 if (stopped()) {
1025 // Already short circuited.
1026 return NULL;
1027 }
1028
1029 // Build an if node and its projections.
1030 // If test is true we take the slow path, which we assume is uncommon.
1031 if (_gvn.type(test) == TypeInt::ZERO) {
1032 // The slow branch is never taken. No need to build this guard.
1033 return NULL;
1034 }
1035
1036 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1037
1038 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1039 if (if_slow == top()) {
1040 // The slow branch is never taken. No need to build this guard.
1041 return NULL;
1042 }
1043
1044 if (region != NULL)
1045 region->add_req(if_slow);
1046
1047 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1048 set_control(if_fast);
1049
1050 return if_slow;
1051 }
1052
1053 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1054 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1055 }
1056 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1057 return generate_guard(test, region, PROB_FAIR);
1058 }
1059
1060 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1061 Node* *pos_index) {
1062 if (stopped())
1063 return NULL; // already stopped
1064 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1065 return NULL; // index is already adequately typed
1066 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1067 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1068 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1069 if (is_neg != NULL && pos_index != NULL) {
1070 // Emulate effect of Parse::adjust_map_after_if.
1071 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1072 ccast->set_req(0, control());
1073 (*pos_index) = _gvn.transform(ccast);
1074 }
1075 return is_neg;
1076 }
1077
1078 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1079 Node* *pos_index) {
1080 if (stopped())
1081 return NULL; // already stopped
1082 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1083 return NULL; // index is already adequately typed
1084 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1085 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1086 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1087 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1088 if (is_notp != NULL && pos_index != NULL) {
1089 // Emulate effect of Parse::adjust_map_after_if.
1090 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1091 ccast->set_req(0, control());
1092 (*pos_index) = _gvn.transform(ccast);
1093 }
1094 return is_notp;
1095 }
1096
1097 // Make sure that 'position' is a valid limit index, in [0..length].
1098 // There are two equivalent plans for checking this:
1099 // A. (offset + copyLength) unsigned<= arrayLength
1100 // B. offset <= (arrayLength - copyLength)
1101 // We require that all of the values above, except for the sum and
1102 // difference, are already known to be non-negative.
1103 // Plan A is robust in the face of overflow, if offset and copyLength
1104 // are both hugely positive.
1105 //
1106 // Plan B is less direct and intuitive, but it does not overflow at
1107 // all, since the difference of two non-negatives is always
1108 // representable. Whenever Java methods must perform the equivalent
1109 // check they generally use Plan B instead of Plan A.
1110 // For the moment we use Plan A.
1111 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1112 Node* subseq_length,
1113 Node* array_length,
1114 RegionNode* region) {
1115 if (stopped())
1116 return NULL; // already stopped
1117 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1118 if (zero_offset && subseq_length->eqv_uncast(array_length))
1119 return NULL; // common case of whole-array copy
1120 Node* last = subseq_length;
1121 if (!zero_offset) // last += offset
1122 last = _gvn.transform(new (C) AddINode(last, offset));
1123 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1124 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1125 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1126 return is_over;
1127 }
1128
1129
1130 //--------------------------generate_current_thread--------------------
1131 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1132 ciKlass* thread_klass = env()->Thread_klass();
1133 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1134 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1135 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1136 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1137 tls_output = thread;
1138 return threadObj;
1139 }
1140
1141
1142 //------------------------------make_string_method_node------------------------
1143 // Helper method for String intrinsic functions. This version is called
1144 // with str1 and str2 pointing to String object nodes.
1145 //
1146 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1147 Node* no_ctrl = NULL;
1148
1149 // Get start addr of string
1150 Node* str1_value = load_String_value(no_ctrl, str1);
1151 Node* str1_offset = load_String_offset(no_ctrl, str1);
1152 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1153
1154 // Get length of string 1
1155 Node* str1_len = load_String_length(no_ctrl, str1);
1156
1157 Node* str2_value = load_String_value(no_ctrl, str2);
1158 Node* str2_offset = load_String_offset(no_ctrl, str2);
1159 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1160
1161 Node* str2_len = NULL;
1162 Node* result = NULL;
1163
1164 switch (opcode) {
1165 case Op_StrIndexOf:
1166 // Get length of string 2
1167 str2_len = load_String_length(no_ctrl, str2);
1168
1169 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1170 str1_start, str1_len, str2_start, str2_len);
1171 break;
1172 case Op_StrComp:
1173 // Get length of string 2
1174 str2_len = load_String_length(no_ctrl, str2);
1175
1176 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1177 str1_start, str1_len, str2_start, str2_len);
1178 break;
1179 case Op_StrEquals:
1180 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1181 str1_start, str2_start, str1_len);
1182 break;
1183 default:
1184 ShouldNotReachHere();
1185 return NULL;
1186 }
1187
1188 // All these intrinsics have checks.
1189 C->set_has_split_ifs(true); // Has chance for split-if optimization
1190
1191 return _gvn.transform(result);
1192 }
1193
1194 // Helper method for String intrinsic functions. This version is called
1195 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1196 // to Int nodes containing the lenghts of str1 and str2.
1197 //
1198 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1199 Node* result = NULL;
1200 switch (opcode) {
1201 case Op_StrIndexOf:
1202 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1203 str1_start, cnt1, str2_start, cnt2);
1204 break;
1205 case Op_StrComp:
1206 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1207 str1_start, cnt1, str2_start, cnt2);
1208 break;
1209 case Op_StrEquals:
1210 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1211 str1_start, str2_start, cnt1);
1212 break;
1213 default:
1214 ShouldNotReachHere();
1215 return NULL;
1216 }
1217
1218 // All these intrinsics have checks.
1219 C->set_has_split_ifs(true); // Has chance for split-if optimization
1220
1221 return _gvn.transform(result);
1222 }
1223
1224 //------------------------------inline_string_compareTo------------------------
1225 // public int java.lang.String.compareTo(String anotherString);
1226 bool LibraryCallKit::inline_string_compareTo() {
1227 Node* receiver = null_check(argument(0));
1228 Node* arg = null_check(argument(1));
1229 if (stopped()) {
1230 return true;
1231 }
1232 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1233 return true;
1234 }
1235
1236 //------------------------------inline_string_equals------------------------
1237 bool LibraryCallKit::inline_string_equals() {
1238 Node* receiver = null_check_receiver();
1239 // NOTE: Do not null check argument for String.equals() because spec
1240 // allows to specify NULL as argument.
1241 Node* argument = this->argument(1);
1242 if (stopped()) {
1243 return true;
1244 }
1245
1246 // paths (plus control) merge
1247 RegionNode* region = new (C) RegionNode(5);
1248 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1249
1250 // does source == target string?
1251 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1252 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1253
1254 Node* if_eq = generate_slow_guard(bol, NULL);
1255 if (if_eq != NULL) {
1256 // receiver == argument
1257 phi->init_req(2, intcon(1));
1258 region->init_req(2, if_eq);
1259 }
1260
1261 // get String klass for instanceOf
1262 ciInstanceKlass* klass = env()->String_klass();
1263
1264 if (!stopped()) {
1265 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1266 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1267 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1268
1269 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1270 //instanceOf == true, fallthrough
1271
1272 if (inst_false != NULL) {
1273 phi->init_req(3, intcon(0));
1274 region->init_req(3, inst_false);
1275 }
1276 }
1277
1278 if (!stopped()) {
1279 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1280
1281 // Properly cast the argument to String
1282 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1283 // This path is taken only when argument's type is String:NotNull.
1284 argument = cast_not_null(argument, false);
1285
1286 Node* no_ctrl = NULL;
1287
1288 // Get start addr of receiver
1289 Node* receiver_val = load_String_value(no_ctrl, receiver);
1290 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1291 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1292
1293 // Get length of receiver
1294 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1295
1296 // Get start addr of argument
1297 Node* argument_val = load_String_value(no_ctrl, argument);
1298 Node* argument_offset = load_String_offset(no_ctrl, argument);
1299 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1300
1301 // Get length of argument
1302 Node* argument_cnt = load_String_length(no_ctrl, argument);
1303
1304 // Check for receiver count != argument count
1305 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1306 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1307 Node* if_ne = generate_slow_guard(bol, NULL);
1308 if (if_ne != NULL) {
1309 phi->init_req(4, intcon(0));
1310 region->init_req(4, if_ne);
1311 }
1312
1313 // Check for count == 0 is done by assembler code for StrEquals.
1314
1315 if (!stopped()) {
1316 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1317 phi->init_req(1, equals);
1318 region->init_req(1, control());
1319 }
1320 }
1321
1322 // post merge
1323 set_control(_gvn.transform(region));
1324 record_for_igvn(region);
1325
1326 set_result(_gvn.transform(phi));
1327 return true;
1328 }
1329
1330 //------------------------------inline_array_equals----------------------------
1331 bool LibraryCallKit::inline_array_equals() {
1332 Node* arg1 = argument(0);
1333 Node* arg2 = argument(1);
1334 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1335 return true;
1336 }
1337
1338 // Java version of String.indexOf(constant string)
1339 // class StringDecl {
1340 // StringDecl(char[] ca) {
1341 // offset = 0;
1342 // count = ca.length;
1343 // value = ca;
1344 // }
1345 // int offset;
1346 // int count;
1347 // char[] value;
1348 // }
1349 //
1350 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1351 // int targetOffset, int cache_i, int md2) {
1352 // int cache = cache_i;
1353 // int sourceOffset = string_object.offset;
1354 // int sourceCount = string_object.count;
1355 // int targetCount = target_object.length;
1356 //
1357 // int targetCountLess1 = targetCount - 1;
1358 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1359 //
1360 // char[] source = string_object.value;
1361 // char[] target = target_object;
1362 // int lastChar = target[targetCountLess1];
1363 //
1364 // outer_loop:
1365 // for (int i = sourceOffset; i < sourceEnd; ) {
1366 // int src = source[i + targetCountLess1];
1367 // if (src == lastChar) {
1368 // // With random strings and a 4-character alphabet,
1369 // // reverse matching at this point sets up 0.8% fewer
1370 // // frames, but (paradoxically) makes 0.3% more probes.
1371 // // Since those probes are nearer the lastChar probe,
1372 // // there is may be a net D$ win with reverse matching.
1373 // // But, reversing loop inhibits unroll of inner loop
1374 // // for unknown reason. So, does running outer loop from
1375 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1376 // for (int j = 0; j < targetCountLess1; j++) {
1377 // if (target[targetOffset + j] != source[i+j]) {
1378 // if ((cache & (1 << source[i+j])) == 0) {
1379 // if (md2 < j+1) {
1380 // i += j+1;
1381 // continue outer_loop;
1382 // }
1383 // }
1384 // i += md2;
1385 // continue outer_loop;
1386 // }
1387 // }
1388 // return i - sourceOffset;
1389 // }
1390 // if ((cache & (1 << src)) == 0) {
1391 // i += targetCountLess1;
1392 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1393 // i++;
1394 // }
1395 // return -1;
1396 // }
1397
1398 //------------------------------string_indexOf------------------------
1399 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1400 jint cache_i, jint md2_i) {
1401
1402 Node* no_ctrl = NULL;
1403 float likely = PROB_LIKELY(0.9);
1404 float unlikely = PROB_UNLIKELY(0.9);
1405
1406 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1407
1408 Node* source = load_String_value(no_ctrl, string_object);
1409 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1410 Node* sourceCount = load_String_length(no_ctrl, string_object);
1411
1412 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1413 jint target_length = target_array->length();
1414 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1415 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1416
1417 // String.value field is known to be @Stable.
1418 if (UseImplicitStableValues) {
1419 target = cast_array_to_stable(target, target_type);
1420 }
1421
1422 IdealKit kit(this, false, true);
1423 #define __ kit.
1424 Node* zero = __ ConI(0);
1425 Node* one = __ ConI(1);
1426 Node* cache = __ ConI(cache_i);
1427 Node* md2 = __ ConI(md2_i);
1428 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1429 Node* targetCount = __ ConI(target_length);
1430 Node* targetCountLess1 = __ ConI(target_length - 1);
1431 Node* targetOffset = __ ConI(targetOffset_i);
1432 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1433
1434 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1435 Node* outer_loop = __ make_label(2 /* goto */);
1436 Node* return_ = __ make_label(1);
1437
1438 __ set(rtn,__ ConI(-1));
1439 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1440 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1441 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1442 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1443 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1444 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1445 Node* tpj = __ AddI(targetOffset, __ value(j));
1446 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1447 Node* ipj = __ AddI(__ value(i), __ value(j));
1448 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1449 __ if_then(targ, BoolTest::ne, src2); {
1450 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1451 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1452 __ increment(i, __ AddI(__ value(j), one));
1453 __ goto_(outer_loop);
1454 } __ end_if(); __ dead(j);
1455 }__ end_if(); __ dead(j);
1456 __ increment(i, md2);
1457 __ goto_(outer_loop);
1458 }__ end_if();
1459 __ increment(j, one);
1460 }__ end_loop(); __ dead(j);
1461 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1462 __ goto_(return_);
1463 }__ end_if();
1464 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1465 __ increment(i, targetCountLess1);
1466 }__ end_if();
1467 __ increment(i, one);
1468 __ bind(outer_loop);
1469 }__ end_loop(); __ dead(i);
1470 __ bind(return_);
1471
1472 // Final sync IdealKit and GraphKit.
1473 final_sync(kit);
1474 Node* result = __ value(rtn);
1475 #undef __
1476 C->set_has_loops(true);
1477 return result;
1478 }
1479
1480 //------------------------------inline_string_indexOf------------------------
1481 bool LibraryCallKit::inline_string_indexOf() {
1482 Node* receiver = argument(0);
1483 Node* arg = argument(1);
1484
1485 Node* result;
1486 // Disable the use of pcmpestri until it can be guaranteed that
1487 // the load doesn't cross into the uncommited space.
1488 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1489 UseSSE42Intrinsics) {
1490 // Generate SSE4.2 version of indexOf
1491 // We currently only have match rules that use SSE4.2
1492
1493 receiver = null_check(receiver);
1494 arg = null_check(arg);
1495 if (stopped()) {
1496 return true;
1497 }
1498
1499 ciInstanceKlass* str_klass = env()->String_klass();
1500 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1501
1502 // Make the merge point
1503 RegionNode* result_rgn = new (C) RegionNode(4);
1504 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1505 Node* no_ctrl = NULL;
1506
1507 // Get start addr of source string
1508 Node* source = load_String_value(no_ctrl, receiver);
1509 Node* source_offset = load_String_offset(no_ctrl, receiver);
1510 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1511
1512 // Get length of source string
1513 Node* source_cnt = load_String_length(no_ctrl, receiver);
1514
1515 // Get start addr of substring
1516 Node* substr = load_String_value(no_ctrl, arg);
1517 Node* substr_offset = load_String_offset(no_ctrl, arg);
1518 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1519
1520 // Get length of source string
1521 Node* substr_cnt = load_String_length(no_ctrl, arg);
1522
1523 // Check for substr count > string count
1524 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1525 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1526 Node* if_gt = generate_slow_guard(bol, NULL);
1527 if (if_gt != NULL) {
1528 result_phi->init_req(2, intcon(-1));
1529 result_rgn->init_req(2, if_gt);
1530 }
1531
1532 if (!stopped()) {
1533 // Check for substr count == 0
1534 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1535 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1536 Node* if_zero = generate_slow_guard(bol, NULL);
1537 if (if_zero != NULL) {
1538 result_phi->init_req(3, intcon(0));
1539 result_rgn->init_req(3, if_zero);
1540 }
1541 }
1542
1543 if (!stopped()) {
1544 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1545 result_phi->init_req(1, result);
1546 result_rgn->init_req(1, control());
1547 }
1548 set_control(_gvn.transform(result_rgn));
1549 record_for_igvn(result_rgn);
1550 result = _gvn.transform(result_phi);
1551
1552 } else { // Use LibraryCallKit::string_indexOf
1553 // don't intrinsify if argument isn't a constant string.
1554 if (!arg->is_Con()) {
1555 return false;
1556 }
1557 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1558 if (str_type == NULL) {
1559 return false;
1560 }
1561 ciInstanceKlass* klass = env()->String_klass();
1562 ciObject* str_const = str_type->const_oop();
1563 if (str_const == NULL || str_const->klass() != klass) {
1564 return false;
1565 }
1566 ciInstance* str = str_const->as_instance();
1567 assert(str != NULL, "must be instance");
1568
1569 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1570 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1571
1572 int o;
1573 int c;
1574 if (java_lang_String::has_offset_field()) {
1575 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1576 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1577 } else {
1578 o = 0;
1579 c = pat->length();
1580 }
1581
1582 // constant strings have no offset and count == length which
1583 // simplifies the resulting code somewhat so lets optimize for that.
1584 if (o != 0 || c != pat->length()) {
1585 return false;
1586 }
1587
1588 receiver = null_check(receiver, T_OBJECT);
1589 // NOTE: No null check on the argument is needed since it's a constant String oop.
1590 if (stopped()) {
1591 return true;
1592 }
1593
1594 // The null string as a pattern always returns 0 (match at beginning of string)
1595 if (c == 0) {
1596 set_result(intcon(0));
1597 return true;
1598 }
1599
1600 // Generate default indexOf
1601 jchar lastChar = pat->char_at(o + (c - 1));
1602 int cache = 0;
1603 int i;
1604 for (i = 0; i < c - 1; i++) {
1605 assert(i < pat->length(), "out of range");
1606 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1607 }
1608
1609 int md2 = c;
1610 for (i = 0; i < c - 1; i++) {
1611 assert(i < pat->length(), "out of range");
1612 if (pat->char_at(o + i) == lastChar) {
1613 md2 = (c - 1) - i;
1614 }
1615 }
1616
1617 result = string_indexOf(receiver, pat, o, cache, md2);
1618 }
1619 set_result(result);
1620 return true;
1621 }
1622
1623 //--------------------------round_double_node--------------------------------
1624 // Round a double node if necessary.
1625 Node* LibraryCallKit::round_double_node(Node* n) {
1626 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1627 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1628 return n;
1629 }
1630
1631 //------------------------------inline_math-----------------------------------
1632 // public static double Math.abs(double)
1633 // public static double Math.sqrt(double)
1634 // public static double Math.log(double)
1635 // public static double Math.log10(double)
1636 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1637 Node* arg = round_double_node(argument(0));
1638 Node* n;
1639 switch (id) {
1640 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1641 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1642 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1643 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1644 default: fatal_unexpected_iid(id); break;
1645 }
1646 set_result(_gvn.transform(n));
1647 return true;
1648 }
1649
1650 //------------------------------inline_trig----------------------------------
1651 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1652 // argument reduction which will turn into a fast/slow diamond.
1653 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1654 Node* arg = round_double_node(argument(0));
1655 Node* n = NULL;
1656
1657 switch (id) {
1658 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1659 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1660 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1661 default: fatal_unexpected_iid(id); break;
1662 }
1663 n = _gvn.transform(n);
1664
1665 // Rounding required? Check for argument reduction!
1666 if (Matcher::strict_fp_requires_explicit_rounding) {
1667 static const double pi_4 = 0.7853981633974483;
1668 static const double neg_pi_4 = -0.7853981633974483;
1669 // pi/2 in 80-bit extended precision
1670 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1671 // -pi/2 in 80-bit extended precision
1672 // 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};
1673 // Cutoff value for using this argument reduction technique
1674 //static const double pi_2_minus_epsilon = 1.564660403643354;
1675 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1676
1677 // Pseudocode for sin:
1678 // if (x <= Math.PI / 4.0) {
1679 // if (x >= -Math.PI / 4.0) return fsin(x);
1680 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1681 // } else {
1682 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1683 // }
1684 // return StrictMath.sin(x);
1685
1686 // Pseudocode for cos:
1687 // if (x <= Math.PI / 4.0) {
1688 // if (x >= -Math.PI / 4.0) return fcos(x);
1689 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1690 // } else {
1691 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1692 // }
1693 // return StrictMath.cos(x);
1694
1695 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1696 // requires a special machine instruction to load it. Instead we'll try
1697 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1698 // probably do the math inside the SIN encoding.
1699
1700 // Make the merge point
1701 RegionNode* r = new (C) RegionNode(3);
1702 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1703
1704 // Flatten arg so we need only 1 test
1705 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1706 // Node for PI/4 constant
1707 Node *pi4 = makecon(TypeD::make(pi_4));
1708 // Check PI/4 : abs(arg)
1709 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1710 // Check: If PI/4 < abs(arg) then go slow
1711 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1712 // Branch either way
1713 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1714 set_control(opt_iff(r,iff));
1715
1716 // Set fast path result
1717 phi->init_req(2, n);
1718
1719 // Slow path - non-blocking leaf call
1720 Node* call = NULL;
1721 switch (id) {
1722 case vmIntrinsics::_dsin:
1723 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1724 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1725 "Sin", NULL, arg, top());
1726 break;
1727 case vmIntrinsics::_dcos:
1728 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1729 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1730 "Cos", NULL, arg, top());
1731 break;
1732 case vmIntrinsics::_dtan:
1733 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1734 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1735 "Tan", NULL, arg, top());
1736 break;
1737 }
1738 assert(control()->in(0) == call, "");
1739 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1740 r->init_req(1, control());
1741 phi->init_req(1, slow_result);
1742
1743 // Post-merge
1744 set_control(_gvn.transform(r));
1745 record_for_igvn(r);
1746 n = _gvn.transform(phi);
1747
1748 C->set_has_split_ifs(true); // Has chance for split-if optimization
1749 }
1750 set_result(n);
1751 return true;
1752 }
1753
1754 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1755 //-------------------
1756 //result=(result.isNaN())? funcAddr():result;
1757 // Check: If isNaN() by checking result!=result? then either trap
1758 // or go to runtime
1759 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1760 // Build the boolean node
1761 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1762
1763 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1764 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1765 // The pow or exp intrinsic returned a NaN, which requires a call
1766 // to the runtime. Recompile with the runtime call.
1767 uncommon_trap(Deoptimization::Reason_intrinsic,
1768 Deoptimization::Action_make_not_entrant);
1769 }
1770 return result;
1771 } else {
1772 // If this inlining ever returned NaN in the past, we compile a call
1773 // to the runtime to properly handle corner cases
1774
1775 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1776 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1777 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1778
1779 if (!if_slow->is_top()) {
1780 RegionNode* result_region = new (C) RegionNode(3);
1781 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1782
1783 result_region->init_req(1, if_fast);
1784 result_val->init_req(1, result);
1785
1786 set_control(if_slow);
1787
1788 const TypePtr* no_memory_effects = NULL;
1789 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1790 no_memory_effects,
1791 x, top(), y, y ? top() : NULL);
1792 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1793 #ifdef ASSERT
1794 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1795 assert(value_top == top(), "second value must be top");
1796 #endif
1797
1798 result_region->init_req(2, control());
1799 result_val->init_req(2, value);
1800 set_control(_gvn.transform(result_region));
1801 return _gvn.transform(result_val);
1802 } else {
1803 return result;
1804 }
1805 }
1806 }
1807
1808 //------------------------------inline_exp-------------------------------------
1809 // Inline exp instructions, if possible. The Intel hardware only misses
1810 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1811 bool LibraryCallKit::inline_exp() {
1812 Node* arg = round_double_node(argument(0));
1813 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1814
1815 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1816 set_result(n);
1817
1818 C->set_has_split_ifs(true); // Has chance for split-if optimization
1819 return true;
1820 }
1821
1822 //------------------------------inline_pow-------------------------------------
1823 // Inline power instructions, if possible.
1824 bool LibraryCallKit::inline_pow() {
1825 // Pseudocode for pow
1826 // if (y == 2) {
1827 // return x * x;
1828 // } else {
1829 // if (x <= 0.0) {
1830 // long longy = (long)y;
1831 // if ((double)longy == y) { // if y is long
1832 // if (y + 1 == y) longy = 0; // huge number: even
1833 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1834 // } else {
1835 // result = NaN;
1836 // }
1837 // } else {
1838 // result = DPow(x,y);
1839 // }
1840 // if (result != result)? {
1841 // result = uncommon_trap() or runtime_call();
1842 // }
1843 // return result;
1844 // }
1845
1846 Node* x = round_double_node(argument(0));
1847 Node* y = round_double_node(argument(2));
1848
1849 Node* result = NULL;
1850
1851 Node* const_two_node = makecon(TypeD::make(2.0));
1852 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1853 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1854 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1855 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1856 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1857
1858 RegionNode* region_node = new (C) RegionNode(3);
1859 region_node->init_req(1, if_true);
1860
1861 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1862 // special case for x^y where y == 2, we can convert it to x * x
1863 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1864
1865 // set control to if_false since we will now process the false branch
1866 set_control(if_false);
1867
1868 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1869 // Short form: skip the fancy tests and just check for NaN result.
1870 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1871 } else {
1872 // If this inlining ever returned NaN in the past, include all
1873 // checks + call to the runtime.
1874
1875 // Set the merge point for If node with condition of (x <= 0.0)
1876 // There are four possible paths to region node and phi node
1877 RegionNode *r = new (C) RegionNode(4);
1878 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1879
1880 // Build the first if node: if (x <= 0.0)
1881 // Node for 0 constant
1882 Node *zeronode = makecon(TypeD::ZERO);
1883 // Check x:0
1884 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1885 // Check: If (x<=0) then go complex path
1886 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1887 // Branch either way
1888 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1889 // Fast path taken; set region slot 3
1890 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1891 r->init_req(3,fast_taken); // Capture fast-control
1892
1893 // Fast path not-taken, i.e. slow path
1894 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1895
1896 // Set fast path result
1897 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1898 phi->init_req(3, fast_result);
1899
1900 // Complex path
1901 // Build the second if node (if y is long)
1902 // Node for (long)y
1903 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1904 // Node for (double)((long) y)
1905 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1906 // Check (double)((long) y) : y
1907 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1908 // Check if (y isn't long) then go to slow path
1909
1910 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1911 // Branch either way
1912 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1913 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1914
1915 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1916
1917 // Calculate DPow(abs(x), y)*(1 & (long)y)
1918 // Node for constant 1
1919 Node *conone = longcon(1);
1920 // 1& (long)y
1921 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1922
1923 // A huge number is always even. Detect a huge number by checking
1924 // if y + 1 == y and set integer to be tested for parity to 0.
1925 // Required for corner case:
1926 // (long)9.223372036854776E18 = max_jlong
1927 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1928 // max_jlong is odd but 9.223372036854776E18 is even
1929 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1930 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1931 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1932 Node* correctedsign = NULL;
1933 if (ConditionalMoveLimit != 0) {
1934 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1935 } else {
1936 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1937 RegionNode *r = new (C) RegionNode(3);
1938 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1939 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1940 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1941 phi->init_req(1, signnode);
1942 phi->init_req(2, longcon(0));
1943 correctedsign = _gvn.transform(phi);
1944 ylong_path = _gvn.transform(r);
1945 record_for_igvn(r);
1946 }
1947
1948 // zero node
1949 Node *conzero = longcon(0);
1950 // Check (1&(long)y)==0?
1951 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1952 // Check if (1&(long)y)!=0?, if so the result is negative
1953 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1954 // abs(x)
1955 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1956 // abs(x)^y
1957 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1958 // -abs(x)^y
1959 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1960 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1961 Node *signresult = NULL;
1962 if (ConditionalMoveLimit != 0) {
1963 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1964 } else {
1965 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1966 RegionNode *r = new (C) RegionNode(3);
1967 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1968 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1969 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1970 phi->init_req(1, absxpowy);
1971 phi->init_req(2, negabsxpowy);
1972 signresult = _gvn.transform(phi);
1973 ylong_path = _gvn.transform(r);
1974 record_for_igvn(r);
1975 }
1976 // Set complex path fast result
1977 r->init_req(2, ylong_path);
1978 phi->init_req(2, signresult);
1979
1980 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1981 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1982 r->init_req(1,slow_path);
1983 phi->init_req(1,slow_result);
1984
1985 // Post merge
1986 set_control(_gvn.transform(r));
1987 record_for_igvn(r);
1988 result = _gvn.transform(phi);
1989 }
1990
1991 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1992
1993 // control from finish_pow_exp is now input to the region node
1994 region_node->set_req(2, control());
1995 // the result from finish_pow_exp is now input to the phi node
1996 phi_node->init_req(2, result);
1997 set_control(_gvn.transform(region_node));
1998 record_for_igvn(region_node);
1999 set_result(_gvn.transform(phi_node));
2000
2001 C->set_has_split_ifs(true); // Has chance for split-if optimization
2002 return true;
2003 }
2004
2005 //------------------------------runtime_math-----------------------------
2006 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2007 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2008 "must be (DD)D or (D)D type");
2009
2010 // Inputs
2011 Node* a = round_double_node(argument(0));
2012 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2013
2014 const TypePtr* no_memory_effects = NULL;
2015 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2016 no_memory_effects,
2017 a, top(), b, b ? top() : NULL);
2018 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2019 #ifdef ASSERT
2020 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2021 assert(value_top == top(), "second value must be top");
2022 #endif
2023
2024 set_result(value);
2025 return true;
2026 }
2027
2028 //------------------------------inline_math_native-----------------------------
2029 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2030 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2031 switch (id) {
2032 // These intrinsics are not properly supported on all hardware
2033 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2034 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2035 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2036 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2037 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2038 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2039
2040 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2041 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2042 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2043 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2044
2045 // These intrinsics are supported on all hardware
2046 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2047 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2048
2049 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2050 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2051 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2052 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2053 #undef FN_PTR
2054
2055 // These intrinsics are not yet correctly implemented
2056 case vmIntrinsics::_datan2:
2057 return false;
2058
2059 default:
2060 fatal_unexpected_iid(id);
2061 return false;
2062 }
2063 }
2064
2065 static bool is_simple_name(Node* n) {
2066 return (n->req() == 1 // constant
2067 || (n->is_Type() && n->as_Type()->type()->singleton())
2068 || n->is_Proj() // parameter or return value
2069 || n->is_Phi() // local of some sort
2070 );
2071 }
2072
2073 //----------------------------inline_min_max-----------------------------------
2074 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2075 set_result(generate_min_max(id, argument(0), argument(1)));
2076 return true;
2077 }
2078
2079 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2080 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2081 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2082 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2083 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2084
2085 {
2086 PreserveJVMState pjvms(this);
2087 PreserveReexecuteState preexecs(this);
2088 jvms()->set_should_reexecute(true);
2089
2090 set_control(slow_path);
2091 set_i_o(i_o());
2092
2093 uncommon_trap(Deoptimization::Reason_intrinsic,
2094 Deoptimization::Action_none);
2095 }
2096
2097 set_control(fast_path);
2098 set_result(math);
2099 }
2100
2101 template <typename OverflowOp>
2102 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2103 typedef typename OverflowOp::MathOp MathOp;
2104
2105 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2106 Node* operation = _gvn.transform( mathOp );
2107 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2108 inline_math_mathExact(operation, ofcheck);
2109 return true;
2110 }
2111
2112 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2113 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2114 }
2115
2116 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2117 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2118 }
2119
2120 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2121 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2122 }
2123
2124 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2125 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2126 }
2127
2128 bool LibraryCallKit::inline_math_negateExactI() {
2129 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2130 }
2131
2132 bool LibraryCallKit::inline_math_negateExactL() {
2133 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2134 }
2135
2136 bool LibraryCallKit::inline_math_multiplyExactI() {
2137 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2138 }
2139
2140 bool LibraryCallKit::inline_math_multiplyExactL() {
2141 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2142 }
2143
2144 Node*
2145 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2146 // These are the candidate return value:
2147 Node* xvalue = x0;
2148 Node* yvalue = y0;
2149
2150 if (xvalue == yvalue) {
2151 return xvalue;
2152 }
2153
2154 bool want_max = (id == vmIntrinsics::_max);
2155
2156 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2157 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2158 if (txvalue == NULL || tyvalue == NULL) return top();
2159 // This is not really necessary, but it is consistent with a
2160 // hypothetical MaxINode::Value method:
2161 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2162
2163 // %%% This folding logic should (ideally) be in a different place.
2164 // Some should be inside IfNode, and there to be a more reliable
2165 // transformation of ?: style patterns into cmoves. We also want
2166 // more powerful optimizations around cmove and min/max.
2167
2168 // Try to find a dominating comparison of these guys.
2169 // It can simplify the index computation for Arrays.copyOf
2170 // and similar uses of System.arraycopy.
2171 // First, compute the normalized version of CmpI(x, y).
2172 int cmp_op = Op_CmpI;
2173 Node* xkey = xvalue;
2174 Node* ykey = yvalue;
2175 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2176 if (ideal_cmpxy->is_Cmp()) {
2177 // E.g., if we have CmpI(length - offset, count),
2178 // it might idealize to CmpI(length, count + offset)
2179 cmp_op = ideal_cmpxy->Opcode();
2180 xkey = ideal_cmpxy->in(1);
2181 ykey = ideal_cmpxy->in(2);
2182 }
2183
2184 // Start by locating any relevant comparisons.
2185 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2186 Node* cmpxy = NULL;
2187 Node* cmpyx = NULL;
2188 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2189 Node* cmp = start_from->fast_out(k);
2190 if (cmp->outcnt() > 0 && // must have prior uses
2191 cmp->in(0) == NULL && // must be context-independent
2192 cmp->Opcode() == cmp_op) { // right kind of compare
2193 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2194 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2195 }
2196 }
2197
2198 const int NCMPS = 2;
2199 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2200 int cmpn;
2201 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2202 if (cmps[cmpn] != NULL) break; // find a result
2203 }
2204 if (cmpn < NCMPS) {
2205 // Look for a dominating test that tells us the min and max.
2206 int depth = 0; // Limit search depth for speed
2207 Node* dom = control();
2208 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2209 if (++depth >= 100) break;
2210 Node* ifproj = dom;
2211 if (!ifproj->is_Proj()) continue;
2212 Node* iff = ifproj->in(0);
2213 if (!iff->is_If()) continue;
2214 Node* bol = iff->in(1);
2215 if (!bol->is_Bool()) continue;
2216 Node* cmp = bol->in(1);
2217 if (cmp == NULL) continue;
2218 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2219 if (cmps[cmpn] == cmp) break;
2220 if (cmpn == NCMPS) continue;
2221 BoolTest::mask btest = bol->as_Bool()->_test._test;
2222 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2223 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2224 // At this point, we know that 'x btest y' is true.
2225 switch (btest) {
2226 case BoolTest::eq:
2227 // They are proven equal, so we can collapse the min/max.
2228 // Either value is the answer. Choose the simpler.
2229 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2230 return yvalue;
2231 return xvalue;
2232 case BoolTest::lt: // x < y
2233 case BoolTest::le: // x <= y
2234 return (want_max ? yvalue : xvalue);
2235 case BoolTest::gt: // x > y
2236 case BoolTest::ge: // x >= y
2237 return (want_max ? xvalue : yvalue);
2238 }
2239 }
2240 }
2241
2242 // We failed to find a dominating test.
2243 // Let's pick a test that might GVN with prior tests.
2244 Node* best_bol = NULL;
2245 BoolTest::mask best_btest = BoolTest::illegal;
2246 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2247 Node* cmp = cmps[cmpn];
2248 if (cmp == NULL) continue;
2249 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2250 Node* bol = cmp->fast_out(j);
2251 if (!bol->is_Bool()) continue;
2252 BoolTest::mask btest = bol->as_Bool()->_test._test;
2253 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2254 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2255 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2256 best_bol = bol->as_Bool();
2257 best_btest = btest;
2258 }
2259 }
2260 }
2261
2262 Node* answer_if_true = NULL;
2263 Node* answer_if_false = NULL;
2264 switch (best_btest) {
2265 default:
2266 if (cmpxy == NULL)
2267 cmpxy = ideal_cmpxy;
2268 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2269 // and fall through:
2270 case BoolTest::lt: // x < y
2271 case BoolTest::le: // x <= y
2272 answer_if_true = (want_max ? yvalue : xvalue);
2273 answer_if_false = (want_max ? xvalue : yvalue);
2274 break;
2275 case BoolTest::gt: // x > y
2276 case BoolTest::ge: // x >= y
2277 answer_if_true = (want_max ? xvalue : yvalue);
2278 answer_if_false = (want_max ? yvalue : xvalue);
2279 break;
2280 }
2281
2282 jint hi, lo;
2283 if (want_max) {
2284 // We can sharpen the minimum.
2285 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2286 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2287 } else {
2288 // We can sharpen the maximum.
2289 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2290 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2291 }
2292
2293 // Use a flow-free graph structure, to avoid creating excess control edges
2294 // which could hinder other optimizations.
2295 // Since Math.min/max is often used with arraycopy, we want
2296 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2297 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2298 answer_if_false, answer_if_true,
2299 TypeInt::make(lo, hi, widen));
2300
2301 return _gvn.transform(cmov);
2302
2303 /*
2304 // This is not as desirable as it may seem, since Min and Max
2305 // nodes do not have a full set of optimizations.
2306 // And they would interfere, anyway, with 'if' optimizations
2307 // and with CMoveI canonical forms.
2308 switch (id) {
2309 case vmIntrinsics::_min:
2310 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2311 case vmIntrinsics::_max:
2312 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2313 default:
2314 ShouldNotReachHere();
2315 }
2316 */
2317 }
2318
2319 inline int
2320 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2321 const TypePtr* base_type = TypePtr::NULL_PTR;
2322 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2323 if (base_type == NULL) {
2324 // Unknown type.
2325 return Type::AnyPtr;
2326 } else if (base_type == TypePtr::NULL_PTR) {
2327 // Since this is a NULL+long form, we have to switch to a rawptr.
2328 base = _gvn.transform(new (C) CastX2PNode(offset));
2329 offset = MakeConX(0);
2330 return Type::RawPtr;
2331 } else if (base_type->base() == Type::RawPtr) {
2332 return Type::RawPtr;
2333 } else if (base_type->isa_oopptr()) {
2334 // Base is never null => always a heap address.
2335 if (base_type->ptr() == TypePtr::NotNull) {
2336 return Type::OopPtr;
2337 }
2338 // Offset is small => always a heap address.
2339 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2340 if (offset_type != NULL &&
2341 base_type->offset() == 0 && // (should always be?)
2342 offset_type->_lo >= 0 &&
2343 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2344 return Type::OopPtr;
2345 }
2346 // Otherwise, it might either be oop+off or NULL+addr.
2347 return Type::AnyPtr;
2348 } else {
2349 // No information:
2350 return Type::AnyPtr;
2351 }
2352 }
2353
2354 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2355 int kind = classify_unsafe_addr(base, offset);
2356 if (kind == Type::RawPtr) {
2357 return basic_plus_adr(top(), base, offset);
2358 } else {
2359 return basic_plus_adr(base, offset);
2360 }
2361 }
2362
2363 //--------------------------inline_number_methods-----------------------------
2364 // inline int Integer.numberOfLeadingZeros(int)
2365 // inline int Long.numberOfLeadingZeros(long)
2366 //
2367 // inline int Integer.numberOfTrailingZeros(int)
2368 // inline int Long.numberOfTrailingZeros(long)
2369 //
2370 // inline int Integer.bitCount(int)
2371 // inline int Long.bitCount(long)
2372 //
2373 // inline char Character.reverseBytes(char)
2374 // inline short Short.reverseBytes(short)
2375 // inline int Integer.reverseBytes(int)
2376 // inline long Long.reverseBytes(long)
2377 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2378 Node* arg = argument(0);
2379 Node* n;
2380 switch (id) {
2381 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2382 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2383 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2384 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2385 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2386 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2387 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2388 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2389 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2390 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2391 default: fatal_unexpected_iid(id); break;
2392 }
2393 set_result(_gvn.transform(n));
2394 return true;
2395 }
2396
2397 //----------------------------inline_unsafe_access----------------------------
2398
2399 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2400
2401 // Helper that guards and inserts a pre-barrier.
2402 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2403 Node* pre_val, bool need_mem_bar) {
2404 // We could be accessing the referent field of a reference object. If so, when G1
2405 // is enabled, we need to log the value in the referent field in an SATB buffer.
2406 // This routine performs some compile time filters and generates suitable
2407 // runtime filters that guard the pre-barrier code.
2408 // Also add memory barrier for non volatile load from the referent field
2409 // to prevent commoning of loads across safepoint.
2410 if (!UseG1GC && !need_mem_bar)
2411 return;
2412
2413 // Some compile time checks.
2414
2415 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2416 const TypeX* otype = offset->find_intptr_t_type();
2417 if (otype != NULL && otype->is_con() &&
2418 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2419 // Constant offset but not the reference_offset so just return
2420 return;
2421 }
2422
2423 // We only need to generate the runtime guards for instances.
2424 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2425 if (btype != NULL) {
2426 if (btype->isa_aryptr()) {
2427 // Array type so nothing to do
2428 return;
2429 }
2430
2431 const TypeInstPtr* itype = btype->isa_instptr();
2432 if (itype != NULL) {
2433 // Can the klass of base_oop be statically determined to be
2434 // _not_ a sub-class of Reference and _not_ Object?
2435 ciKlass* klass = itype->klass();
2436 if ( klass->is_loaded() &&
2437 !klass->is_subtype_of(env()->Reference_klass()) &&
2438 !env()->Object_klass()->is_subtype_of(klass)) {
2439 return;
2440 }
2441 }
2442 }
2443
2444 // The compile time filters did not reject base_oop/offset so
2445 // we need to generate the following runtime filters
2446 //
2447 // if (offset == java_lang_ref_Reference::_reference_offset) {
2448 // if (instance_of(base, java.lang.ref.Reference)) {
2449 // pre_barrier(_, pre_val, ...);
2450 // }
2451 // }
2452
2453 float likely = PROB_LIKELY( 0.999);
2454 float unlikely = PROB_UNLIKELY(0.999);
2455
2456 IdealKit ideal(this);
2457 #define __ ideal.
2458
2459 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2460
2461 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2462 // Update graphKit memory and control from IdealKit.
2463 sync_kit(ideal);
2464
2465 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2466 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2467
2468 // Update IdealKit memory and control from graphKit.
2469 __ sync_kit(this);
2470
2471 Node* one = __ ConI(1);
2472 // is_instof == 0 if base_oop == NULL
2473 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2474
2475 // Update graphKit from IdeakKit.
2476 sync_kit(ideal);
2477
2478 // Use the pre-barrier to record the value in the referent field
2479 pre_barrier(false /* do_load */,
2480 __ ctrl(),
2481 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2482 pre_val /* pre_val */,
2483 T_OBJECT);
2484 if (need_mem_bar) {
2485 // Add memory barrier to prevent commoning reads from this field
2486 // across safepoint since GC can change its value.
2487 insert_mem_bar(Op_MemBarCPUOrder);
2488 }
2489 // Update IdealKit from graphKit.
2490 __ sync_kit(this);
2491
2492 } __ end_if(); // _ref_type != ref_none
2493 } __ end_if(); // offset == referent_offset
2494
2495 // Final sync IdealKit and GraphKit.
2496 final_sync(ideal);
2497 #undef __
2498 }
2499
2500
2501 // Interpret Unsafe.fieldOffset cookies correctly:
2502 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2503
2504 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2505 // Attempt to infer a sharper value type from the offset and base type.
2506 ciKlass* sharpened_klass = NULL;
2507
2508 // See if it is an instance field, with an object type.
2509 if (alias_type->field() != NULL) {
2510 assert(!is_native_ptr, "native pointer op cannot use a java address");
2511 if (alias_type->field()->type()->is_klass()) {
2512 sharpened_klass = alias_type->field()->type()->as_klass();
2513 }
2514 }
2515
2516 // See if it is a narrow oop array.
2517 if (adr_type->isa_aryptr()) {
2518 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2519 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2520 if (elem_type != NULL) {
2521 sharpened_klass = elem_type->klass();
2522 }
2523 }
2524 }
2525
2526 // The sharpened class might be unloaded if there is no class loader
2527 // contraint in place.
2528 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2529 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2530
2531 #ifndef PRODUCT
2532 if (C->print_intrinsics() || C->print_inlining()) {
2533 tty->print(" from base type: "); adr_type->dump();
2534 tty->print(" sharpened value: "); tjp->dump();
2535 }
2536 #endif
2537 // Sharpen the value type.
2538 return tjp;
2539 }
2540 return NULL;
2541 }
2542
2543 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2544 if (callee()->is_static()) return false; // caller must have the capability!
2545
2546 #ifndef PRODUCT
2547 {
2548 ResourceMark rm;
2549 // Check the signatures.
2550 ciSignature* sig = callee()->signature();
2551 #ifdef ASSERT
2552 if (!is_store) {
2553 // Object getObject(Object base, int/long offset), etc.
2554 BasicType rtype = sig->return_type()->basic_type();
2555 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2556 rtype = T_ADDRESS; // it is really a C void*
2557 assert(rtype == type, "getter must return the expected value");
2558 if (!is_native_ptr) {
2559 assert(sig->count() == 2, "oop getter has 2 arguments");
2560 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2561 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2562 } else {
2563 assert(sig->count() == 1, "native getter has 1 argument");
2564 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2565 }
2566 } else {
2567 // void putObject(Object base, int/long offset, Object x), etc.
2568 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2569 if (!is_native_ptr) {
2570 assert(sig->count() == 3, "oop putter has 3 arguments");
2571 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2572 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2573 } else {
2574 assert(sig->count() == 2, "native putter has 2 arguments");
2575 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2576 }
2577 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2578 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2579 vtype = T_ADDRESS; // it is really a C void*
2580 assert(vtype == type, "putter must accept the expected value");
2581 }
2582 #endif // ASSERT
2583 }
2584 #endif //PRODUCT
2585
2586 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2587
2588 Node* receiver = argument(0); // type: oop
2589
2590 // Build address expression. See the code in inline_unsafe_prefetch.
2591 Node* adr;
2592 Node* heap_base_oop = top();
2593 Node* offset = top();
2594 Node* val;
2595
2596 if (!is_native_ptr) {
2597 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2598 Node* base = argument(1); // type: oop
2599 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2600 offset = argument(2); // type: long
2601 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2602 // to be plain byte offsets, which are also the same as those accepted
2603 // by oopDesc::field_base.
2604 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2605 "fieldOffset must be byte-scaled");
2606 // 32-bit machines ignore the high half!
2607 offset = ConvL2X(offset);
2608 adr = make_unsafe_address(base, offset);
2609 heap_base_oop = base;
2610 val = is_store ? argument(4) : NULL;
2611 } else {
2612 Node* ptr = argument(1); // type: long
2613 ptr = ConvL2X(ptr); // adjust Java long to machine word
2614 adr = make_unsafe_address(NULL, ptr);
2615 val = is_store ? argument(3) : NULL;
2616 }
2617
2618 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2619
2620 // First guess at the value type.
2621 const Type *value_type = Type::get_const_basic_type(type);
2622
2623 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2624 // there was not enough information to nail it down.
2625 Compile::AliasType* alias_type = C->alias_type(adr_type);
2626 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2627
2628 // We will need memory barriers unless we can determine a unique
2629 // alias category for this reference. (Note: If for some reason
2630 // the barriers get omitted and the unsafe reference begins to "pollute"
2631 // the alias analysis of the rest of the graph, either Compile::can_alias
2632 // or Compile::must_alias will throw a diagnostic assert.)
2633 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2634
2635 // If we are reading the value of the referent field of a Reference
2636 // object (either by using Unsafe directly or through reflection)
2637 // then, if G1 is enabled, we need to record the referent in an
2638 // SATB log buffer using the pre-barrier mechanism.
2639 // Also we need to add memory barrier to prevent commoning reads
2640 // from this field across safepoint since GC can change its value.
2641 bool need_read_barrier = !is_native_ptr && !is_store &&
2642 offset != top() && heap_base_oop != top();
2643
2644 if (!is_store && type == T_OBJECT) {
2645 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2646 if (tjp != NULL) {
2647 value_type = tjp;
2648 }
2649 }
2650
2651 receiver = null_check(receiver);
2652 if (stopped()) {
2653 return true;
2654 }
2655 // Heap pointers get a null-check from the interpreter,
2656 // as a courtesy. However, this is not guaranteed by Unsafe,
2657 // and it is not possible to fully distinguish unintended nulls
2658 // from intended ones in this API.
2659
2660 if (is_volatile) {
2661 // We need to emit leading and trailing CPU membars (see below) in
2662 // addition to memory membars when is_volatile. This is a little
2663 // too strong, but avoids the need to insert per-alias-type
2664 // volatile membars (for stores; compare Parse::do_put_xxx), which
2665 // we cannot do effectively here because we probably only have a
2666 // rough approximation of type.
2667 need_mem_bar = true;
2668 // For Stores, place a memory ordering barrier now.
2669 if (is_store) {
2670 insert_mem_bar(Op_MemBarRelease);
2671 } else {
2672 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2673 insert_mem_bar(Op_MemBarVolatile);
2674 }
2675 }
2676 }
2677
2678 // Memory barrier to prevent normal and 'unsafe' accesses from
2679 // bypassing each other. Happens after null checks, so the
2680 // exception paths do not take memory state from the memory barrier,
2681 // so there's no problems making a strong assert about mixing users
2682 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2683 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2684 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2685
2686 if (!is_store) {
2687 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2688 // To be valid, unsafe loads may depend on other conditions than
2689 // the one that guards them: pin the Load node
2690 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2691 // load value
2692 switch (type) {
2693 case T_BOOLEAN:
2694 case T_CHAR:
2695 case T_BYTE:
2696 case T_SHORT:
2697 case T_INT:
2698 case T_LONG:
2699 case T_FLOAT:
2700 case T_DOUBLE:
2701 break;
2702 case T_OBJECT:
2703 if (need_read_barrier) {
2704 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2705 }
2706 break;
2707 case T_ADDRESS:
2708 // Cast to an int type.
2709 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2710 p = ConvX2UL(p);
2711 break;
2712 default:
2713 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2714 break;
2715 }
2716 // The load node has the control of the preceding MemBarCPUOrder. All
2717 // following nodes will have the control of the MemBarCPUOrder inserted at
2718 // the end of this method. So, pushing the load onto the stack at a later
2719 // point is fine.
2720 set_result(p);
2721 } else {
2722 // place effect of store into memory
2723 switch (type) {
2724 case T_DOUBLE:
2725 val = dstore_rounding(val);
2726 break;
2727 case T_ADDRESS:
2728 // Repackage the long as a pointer.
2729 val = ConvL2X(val);
2730 val = _gvn.transform(new (C) CastX2PNode(val));
2731 break;
2732 }
2733
2734 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2735 if (type != T_OBJECT ) {
2736 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2737 } else {
2738 // Possibly an oop being stored to Java heap or native memory
2739 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2740 // oop to Java heap.
2741 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2742 } else {
2743 // We can't tell at compile time if we are storing in the Java heap or outside
2744 // of it. So we need to emit code to conditionally do the proper type of
2745 // store.
2746
2747 IdealKit ideal(this);
2748 #define __ ideal.
2749 // QQQ who knows what probability is here??
2750 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2751 // Sync IdealKit and graphKit.
2752 sync_kit(ideal);
2753 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2754 // Update IdealKit memory.
2755 __ sync_kit(this);
2756 } __ else_(); {
2757 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2758 } __ end_if();
2759 // Final sync IdealKit and GraphKit.
2760 final_sync(ideal);
2761 #undef __
2762 }
2763 }
2764 }
2765
2766 if (is_volatile) {
2767 if (!is_store) {
2768 insert_mem_bar(Op_MemBarAcquire);
2769 } else {
2770 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2771 insert_mem_bar(Op_MemBarVolatile);
2772 }
2773 }
2774 }
2775
2776 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2777
2778 return true;
2779 }
2780
2781 //----------------------------inline_unsafe_prefetch----------------------------
2782
2783 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2784 #ifndef PRODUCT
2785 {
2786 ResourceMark rm;
2787 // Check the signatures.
2788 ciSignature* sig = callee()->signature();
2789 #ifdef ASSERT
2790 // Object getObject(Object base, int/long offset), etc.
2791 BasicType rtype = sig->return_type()->basic_type();
2792 if (!is_native_ptr) {
2793 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2794 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2795 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2796 } else {
2797 assert(sig->count() == 1, "native prefetch has 1 argument");
2798 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2799 }
2800 #endif // ASSERT
2801 }
2802 #endif // !PRODUCT
2803
2804 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2805
2806 const int idx = is_static ? 0 : 1;
2807 if (!is_static) {
2808 null_check_receiver();
2809 if (stopped()) {
2810 return true;
2811 }
2812 }
2813
2814 // Build address expression. See the code in inline_unsafe_access.
2815 Node *adr;
2816 if (!is_native_ptr) {
2817 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2818 Node* base = argument(idx + 0); // type: oop
2819 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2820 Node* offset = argument(idx + 1); // type: long
2821 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2822 // to be plain byte offsets, which are also the same as those accepted
2823 // by oopDesc::field_base.
2824 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2825 "fieldOffset must be byte-scaled");
2826 // 32-bit machines ignore the high half!
2827 offset = ConvL2X(offset);
2828 adr = make_unsafe_address(base, offset);
2829 } else {
2830 Node* ptr = argument(idx + 0); // type: long
2831 ptr = ConvL2X(ptr); // adjust Java long to machine word
2832 adr = make_unsafe_address(NULL, ptr);
2833 }
2834
2835 // Generate the read or write prefetch
2836 Node *prefetch;
2837 if (is_store) {
2838 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2839 } else {
2840 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2841 }
2842 prefetch->init_req(0, control());
2843 set_i_o(_gvn.transform(prefetch));
2844
2845 return true;
2846 }
2847
2848 //----------------------------inline_unsafe_load_store----------------------------
2849 // This method serves a couple of different customers (depending on LoadStoreKind):
2850 //
2851 // LS_cmpxchg:
2852 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2853 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2854 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2855 //
2856 // LS_xadd:
2857 // public int getAndAddInt( Object o, long offset, int delta)
2858 // public long getAndAddLong(Object o, long offset, long delta)
2859 //
2860 // LS_xchg:
2861 // int getAndSet(Object o, long offset, int newValue)
2862 // long getAndSet(Object o, long offset, long newValue)
2863 // Object getAndSet(Object o, long offset, Object newValue)
2864 //
2865 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2866 // This basic scheme here is the same as inline_unsafe_access, but
2867 // differs in enough details that combining them would make the code
2868 // overly confusing. (This is a true fact! I originally combined
2869 // them, but even I was confused by it!) As much code/comments as
2870 // possible are retained from inline_unsafe_access though to make
2871 // the correspondences clearer. - dl
2872
2873 if (callee()->is_static()) return false; // caller must have the capability!
2874
2875 #ifndef PRODUCT
2876 BasicType rtype;
2877 {
2878 ResourceMark rm;
2879 // Check the signatures.
2880 ciSignature* sig = callee()->signature();
2881 rtype = sig->return_type()->basic_type();
2882 if (kind == LS_xadd || kind == LS_xchg) {
2883 // Check the signatures.
2884 #ifdef ASSERT
2885 assert(rtype == type, "get and set must return the expected type");
2886 assert(sig->count() == 3, "get and set has 3 arguments");
2887 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2888 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2889 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2890 #endif // ASSERT
2891 } else if (kind == LS_cmpxchg) {
2892 // Check the signatures.
2893 #ifdef ASSERT
2894 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2895 assert(sig->count() == 4, "CAS has 4 arguments");
2896 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2897 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2898 #endif // ASSERT
2899 } else {
2900 ShouldNotReachHere();
2901 }
2902 }
2903 #endif //PRODUCT
2904
2905 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2906
2907 // Get arguments:
2908 Node* receiver = NULL;
2909 Node* base = NULL;
2910 Node* offset = NULL;
2911 Node* oldval = NULL;
2912 Node* newval = NULL;
2913 if (kind == LS_cmpxchg) {
2914 const bool two_slot_type = type2size[type] == 2;
2915 receiver = argument(0); // type: oop
2916 base = argument(1); // type: oop
2917 offset = argument(2); // type: long
2918 oldval = argument(4); // type: oop, int, or long
2919 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2920 } else if (kind == LS_xadd || kind == LS_xchg){
2921 receiver = argument(0); // type: oop
2922 base = argument(1); // type: oop
2923 offset = argument(2); // type: long
2924 oldval = NULL;
2925 newval = argument(4); // type: oop, int, or long
2926 }
2927
2928 // Null check receiver.
2929 receiver = null_check(receiver);
2930 if (stopped()) {
2931 return true;
2932 }
2933
2934 // Build field offset expression.
2935 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2936 // to be plain byte offsets, which are also the same as those accepted
2937 // by oopDesc::field_base.
2938 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2939 // 32-bit machines ignore the high half of long offsets
2940 offset = ConvL2X(offset);
2941 Node* adr = make_unsafe_address(base, offset);
2942 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2943
2944 // For CAS, unlike inline_unsafe_access, there seems no point in
2945 // trying to refine types. Just use the coarse types here.
2946 const Type *value_type = Type::get_const_basic_type(type);
2947 Compile::AliasType* alias_type = C->alias_type(adr_type);
2948 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2949
2950 if (kind == LS_xchg && type == T_OBJECT) {
2951 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2952 if (tjp != NULL) {
2953 value_type = tjp;
2954 }
2955 }
2956
2957 int alias_idx = C->get_alias_index(adr_type);
2958
2959 // Memory-model-wise, a LoadStore acts like a little synchronized
2960 // block, so needs barriers on each side. These don't translate
2961 // into actual barriers on most machines, but we still need rest of
2962 // compiler to respect ordering.
2963
2964 insert_mem_bar(Op_MemBarRelease);
2965 insert_mem_bar(Op_MemBarCPUOrder);
2966
2967 // 4984716: MemBars must be inserted before this
2968 // memory node in order to avoid a false
2969 // dependency which will confuse the scheduler.
2970 Node *mem = memory(alias_idx);
2971
2972 // For now, we handle only those cases that actually exist: ints,
2973 // longs, and Object. Adding others should be straightforward.
2974 Node* load_store;
2975 switch(type) {
2976 case T_INT:
2977 if (kind == LS_xadd) {
2978 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2979 } else if (kind == LS_xchg) {
2980 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2981 } else if (kind == LS_cmpxchg) {
2982 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2983 } else {
2984 ShouldNotReachHere();
2985 }
2986 break;
2987 case T_LONG:
2988 if (kind == LS_xadd) {
2989 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2990 } else if (kind == LS_xchg) {
2991 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2992 } else if (kind == LS_cmpxchg) {
2993 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2994 } else {
2995 ShouldNotReachHere();
2996 }
2997 break;
2998 case T_OBJECT:
2999 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3000 // could be delayed during Parse (for example, in adjust_map_after_if()).
3001 // Execute transformation here to avoid barrier generation in such case.
3002 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3003 newval = _gvn.makecon(TypePtr::NULL_PTR);
3004
3005 // Reference stores need a store barrier.
3006 if (kind == LS_xchg) {
3007 // If pre-barrier must execute before the oop store, old value will require do_load here.
3008 if (!can_move_pre_barrier()) {
3009 pre_barrier(true /* do_load*/,
3010 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3011 NULL /* pre_val*/,
3012 T_OBJECT);
3013 } // Else move pre_barrier to use load_store value, see below.
3014 } else if (kind == LS_cmpxchg) {
3015 // Same as for newval above:
3016 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3017 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3018 }
3019 // The only known value which might get overwritten is oldval.
3020 pre_barrier(false /* do_load */,
3021 control(), NULL, NULL, max_juint, NULL, NULL,
3022 oldval /* pre_val */,
3023 T_OBJECT);
3024 } else {
3025 ShouldNotReachHere();
3026 }
3027
3028 #ifdef _LP64
3029 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3030 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3031 if (kind == LS_xchg) {
3032 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3033 newval_enc, adr_type, value_type->make_narrowoop()));
3034 } else {
3035 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3036 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3037 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3038 newval_enc, oldval_enc));
3039 }
3040 } else
3041 #endif
3042 {
3043 if (kind == LS_xchg) {
3044 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3045 } else {
3046 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3047 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3048 }
3049 }
3050 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3051 break;
3052 default:
3053 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3054 break;
3055 }
3056
3057 // SCMemProjNodes represent the memory state of a LoadStore. Their
3058 // main role is to prevent LoadStore nodes from being optimized away
3059 // when their results aren't used.
3060 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3061 set_memory(proj, alias_idx);
3062
3063 if (type == T_OBJECT && kind == LS_xchg) {
3064 #ifdef _LP64
3065 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3066 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3067 }
3068 #endif
3069 if (can_move_pre_barrier()) {
3070 // Don't need to load pre_val. The old value is returned by load_store.
3071 // The pre_barrier can execute after the xchg as long as no safepoint
3072 // gets inserted between them.
3073 pre_barrier(false /* do_load */,
3074 control(), NULL, NULL, max_juint, NULL, NULL,
3075 load_store /* pre_val */,
3076 T_OBJECT);
3077 }
3078 }
3079
3080 // Add the trailing membar surrounding the access
3081 insert_mem_bar(Op_MemBarCPUOrder);
3082 insert_mem_bar(Op_MemBarAcquire);
3083
3084 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3085 set_result(load_store);
3086 return true;
3087 }
3088
3089 //----------------------------inline_unsafe_ordered_store----------------------
3090 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3091 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3092 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3093 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3094 // This is another variant of inline_unsafe_access, differing in
3095 // that it always issues store-store ("release") barrier and ensures
3096 // store-atomicity (which only matters for "long").
3097
3098 if (callee()->is_static()) return false; // caller must have the capability!
3099
3100 #ifndef PRODUCT
3101 {
3102 ResourceMark rm;
3103 // Check the signatures.
3104 ciSignature* sig = callee()->signature();
3105 #ifdef ASSERT
3106 BasicType rtype = sig->return_type()->basic_type();
3107 assert(rtype == T_VOID, "must return void");
3108 assert(sig->count() == 3, "has 3 arguments");
3109 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3110 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3111 #endif // ASSERT
3112 }
3113 #endif //PRODUCT
3114
3115 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3116
3117 // Get arguments:
3118 Node* receiver = argument(0); // type: oop
3119 Node* base = argument(1); // type: oop
3120 Node* offset = argument(2); // type: long
3121 Node* val = argument(4); // type: oop, int, or long
3122
3123 // Null check receiver.
3124 receiver = null_check(receiver);
3125 if (stopped()) {
3126 return true;
3127 }
3128
3129 // Build field offset expression.
3130 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3131 // 32-bit machines ignore the high half of long offsets
3132 offset = ConvL2X(offset);
3133 Node* adr = make_unsafe_address(base, offset);
3134 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3135 const Type *value_type = Type::get_const_basic_type(type);
3136 Compile::AliasType* alias_type = C->alias_type(adr_type);
3137
3138 insert_mem_bar(Op_MemBarRelease);
3139 insert_mem_bar(Op_MemBarCPUOrder);
3140 // Ensure that the store is atomic for longs:
3141 const bool require_atomic_access = true;
3142 Node* store;
3143 if (type == T_OBJECT) // reference stores need a store barrier.
3144 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3145 else {
3146 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3147 }
3148 insert_mem_bar(Op_MemBarCPUOrder);
3149 return true;
3150 }
3151
3152 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3153 // Regardless of form, don't allow previous ld/st to move down,
3154 // then issue acquire, release, or volatile mem_bar.
3155 insert_mem_bar(Op_MemBarCPUOrder);
3156 switch(id) {
3157 case vmIntrinsics::_loadFence:
3158 insert_mem_bar(Op_LoadFence);
3159 return true;
3160 case vmIntrinsics::_storeFence:
3161 insert_mem_bar(Op_StoreFence);
3162 return true;
3163 case vmIntrinsics::_fullFence:
3164 insert_mem_bar(Op_MemBarVolatile);
3165 return true;
3166 default:
3167 fatal_unexpected_iid(id);
3168 return false;
3169 }
3170 }
3171
3172 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3173 if (!kls->is_Con()) {
3174 return true;
3175 }
3176 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3177 if (klsptr == NULL) {
3178 return true;
3179 }
3180 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3181 // don't need a guard for a klass that is already initialized
3182 return !ik->is_initialized();
3183 }
3184
3185 //----------------------------inline_unsafe_allocate---------------------------
3186 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3187 bool LibraryCallKit::inline_unsafe_allocate() {
3188 if (callee()->is_static()) return false; // caller must have the capability!
3189
3190 null_check_receiver(); // null-check, then ignore
3191 Node* cls = null_check(argument(1));
3192 if (stopped()) return true;
3193
3194 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3195 kls = null_check(kls);
3196 if (stopped()) return true; // argument was like int.class
3197
3198 Node* test = NULL;
3199 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3200 // Note: The argument might still be an illegal value like
3201 // Serializable.class or Object[].class. The runtime will handle it.
3202 // But we must make an explicit check for initialization.
3203 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3204 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3205 // can generate code to load it as unsigned byte.
3206 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3207 Node* bits = intcon(InstanceKlass::fully_initialized);
3208 test = _gvn.transform(new (C) SubINode(inst, bits));
3209 // The 'test' is non-zero if we need to take a slow path.
3210 }
3211
3212 Node* obj = new_instance(kls, test);
3213 set_result(obj);
3214 return true;
3215 }
3216
3217 #ifdef TRACE_HAVE_INTRINSICS
3218 /*
3219 * oop -> myklass
3220 * myklass->trace_id |= USED
3221 * return myklass->trace_id & ~0x3
3222 */
3223 bool LibraryCallKit::inline_native_classID() {
3224 null_check_receiver(); // null-check, then ignore
3225 Node* cls = null_check(argument(1), T_OBJECT);
3226 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3227 kls = null_check(kls, T_OBJECT);
3228 ByteSize offset = TRACE_ID_OFFSET;
3229 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3230 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3231 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3232 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3233 Node* clsused = longcon(0x01l); // set the class bit
3234 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3235
3236 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3237 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3238 set_result(andl);
3239 return true;
3240 }
3241
3242 bool LibraryCallKit::inline_native_threadID() {
3243 Node* tls_ptr = NULL;
3244 Node* cur_thr = generate_current_thread(tls_ptr);
3245 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3246 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3247 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3248
3249 Node* threadid = NULL;
3250 size_t thread_id_size = OSThread::thread_id_size();
3251 if (thread_id_size == (size_t) BytesPerLong) {
3252 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3253 } else if (thread_id_size == (size_t) BytesPerInt) {
3254 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3255 } else {
3256 ShouldNotReachHere();
3257 }
3258 set_result(threadid);
3259 return true;
3260 }
3261 #endif
3262
3263 //------------------------inline_native_time_funcs--------------
3264 // inline code for System.currentTimeMillis() and System.nanoTime()
3265 // these have the same type and signature
3266 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3267 const TypeFunc* tf = OptoRuntime::void_long_Type();
3268 const TypePtr* no_memory_effects = NULL;
3269 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3270 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3271 #ifdef ASSERT
3272 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3273 assert(value_top == top(), "second value must be top");
3274 #endif
3275 set_result(value);
3276 return true;
3277 }
3278
3279 //------------------------inline_native_currentThread------------------
3280 bool LibraryCallKit::inline_native_currentThread() {
3281 Node* junk = NULL;
3282 set_result(generate_current_thread(junk));
3283 return true;
3284 }
3285
3286 //------------------------inline_native_isInterrupted------------------
3287 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3288 bool LibraryCallKit::inline_native_isInterrupted() {
3289 // Add a fast path to t.isInterrupted(clear_int):
3290 // (t == Thread.current() &&
3291 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3292 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3293 // So, in the common case that the interrupt bit is false,
3294 // we avoid making a call into the VM. Even if the interrupt bit
3295 // is true, if the clear_int argument is false, we avoid the VM call.
3296 // However, if the receiver is not currentThread, we must call the VM,
3297 // because there must be some locking done around the operation.
3298
3299 // We only go to the fast case code if we pass two guards.
3300 // Paths which do not pass are accumulated in the slow_region.
3301
3302 enum {
3303 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3304 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3305 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3306 PATH_LIMIT
3307 };
3308
3309 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3310 // out of the function.
3311 insert_mem_bar(Op_MemBarCPUOrder);
3312
3313 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3314 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3315
3316 RegionNode* slow_region = new (C) RegionNode(1);
3317 record_for_igvn(slow_region);
3318
3319 // (a) Receiving thread must be the current thread.
3320 Node* rec_thr = argument(0);
3321 Node* tls_ptr = NULL;
3322 Node* cur_thr = generate_current_thread(tls_ptr);
3323 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3324 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3325
3326 generate_slow_guard(bol_thr, slow_region);
3327
3328 // (b) Interrupt bit on TLS must be false.
3329 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3330 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3331 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3332
3333 // Set the control input on the field _interrupted read to prevent it floating up.
3334 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3335 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3336 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3337
3338 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3339
3340 // First fast path: if (!TLS._interrupted) return false;
3341 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3342 result_rgn->init_req(no_int_result_path, false_bit);
3343 result_val->init_req(no_int_result_path, intcon(0));
3344
3345 // drop through to next case
3346 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3347
3348 #ifndef TARGET_OS_FAMILY_windows
3349 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3350 Node* clr_arg = argument(1);
3351 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3352 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3353 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3354
3355 // Second fast path: ... else if (!clear_int) return true;
3356 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3357 result_rgn->init_req(no_clear_result_path, false_arg);
3358 result_val->init_req(no_clear_result_path, intcon(1));
3359
3360 // drop through to next case
3361 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3362 #else
3363 // To return true on Windows you must read the _interrupted field
3364 // and check the the event state i.e. take the slow path.
3365 #endif // TARGET_OS_FAMILY_windows
3366
3367 // (d) Otherwise, go to the slow path.
3368 slow_region->add_req(control());
3369 set_control( _gvn.transform(slow_region));
3370
3371 if (stopped()) {
3372 // There is no slow path.
3373 result_rgn->init_req(slow_result_path, top());
3374 result_val->init_req(slow_result_path, top());
3375 } else {
3376 // non-virtual because it is a private non-static
3377 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3378
3379 Node* slow_val = set_results_for_java_call(slow_call);
3380 // this->control() comes from set_results_for_java_call
3381
3382 Node* fast_io = slow_call->in(TypeFunc::I_O);
3383 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3384
3385 // These two phis are pre-filled with copies of of the fast IO and Memory
3386 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3387 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3388
3389 result_rgn->init_req(slow_result_path, control());
3390 result_io ->init_req(slow_result_path, i_o());
3391 result_mem->init_req(slow_result_path, reset_memory());
3392 result_val->init_req(slow_result_path, slow_val);
3393
3394 set_all_memory(_gvn.transform(result_mem));
3395 set_i_o( _gvn.transform(result_io));
3396 }
3397
3398 C->set_has_split_ifs(true); // Has chance for split-if optimization
3399 set_result(result_rgn, result_val);
3400 return true;
3401 }
3402
3403 //---------------------------load_mirror_from_klass----------------------------
3404 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3405 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3406 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3407 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3408 }
3409
3410 //-----------------------load_klass_from_mirror_common-------------------------
3411 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3412 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3413 // and branch to the given path on the region.
3414 // If never_see_null, take an uncommon trap on null, so we can optimistically
3415 // compile for the non-null case.
3416 // If the region is NULL, force never_see_null = true.
3417 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3418 bool never_see_null,
3419 RegionNode* region,
3420 int null_path,
3421 int offset) {
3422 if (region == NULL) never_see_null = true;
3423 Node* p = basic_plus_adr(mirror, offset);
3424 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3425 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3426 Node* null_ctl = top();
3427 kls = null_check_oop(kls, &null_ctl, never_see_null);
3428 if (region != NULL) {
3429 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3430 region->init_req(null_path, null_ctl);
3431 } else {
3432 assert(null_ctl == top(), "no loose ends");
3433 }
3434 return kls;
3435 }
3436
3437 //--------------------(inline_native_Class_query helpers)---------------------
3438 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3439 // Fall through if (mods & mask) == bits, take the guard otherwise.
3440 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3441 // Branch around if the given klass has the given modifier bit set.
3442 // Like generate_guard, adds a new path onto the region.
3443 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3444 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3445 Node* mask = intcon(modifier_mask);
3446 Node* bits = intcon(modifier_bits);
3447 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3448 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3449 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3450 return generate_fair_guard(bol, region);
3451 }
3452 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3453 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3454 }
3455
3456 //-------------------------inline_native_Class_query-------------------
3457 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3458 const Type* return_type = TypeInt::BOOL;
3459 Node* prim_return_value = top(); // what happens if it's a primitive class?
3460 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3461 bool expect_prim = false; // most of these guys expect to work on refs
3462
3463 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3464
3465 Node* mirror = argument(0);
3466 Node* obj = top();
3467
3468 switch (id) {
3469 case vmIntrinsics::_isInstance:
3470 // nothing is an instance of a primitive type
3471 prim_return_value = intcon(0);
3472 obj = argument(1);
3473 break;
3474 case vmIntrinsics::_getModifiers:
3475 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3476 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3477 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3478 break;
3479 case vmIntrinsics::_isInterface:
3480 prim_return_value = intcon(0);
3481 break;
3482 case vmIntrinsics::_isArray:
3483 prim_return_value = intcon(0);
3484 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3485 break;
3486 case vmIntrinsics::_isPrimitive:
3487 prim_return_value = intcon(1);
3488 expect_prim = true; // obviously
3489 break;
3490 case vmIntrinsics::_getSuperclass:
3491 prim_return_value = null();
3492 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3493 break;
3494 case vmIntrinsics::_getComponentType:
3495 prim_return_value = null();
3496 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3497 break;
3498 case vmIntrinsics::_getClassAccessFlags:
3499 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3500 return_type = TypeInt::INT; // not bool! 6297094
3501 break;
3502 default:
3503 fatal_unexpected_iid(id);
3504 break;
3505 }
3506
3507 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3508 if (mirror_con == NULL) return false; // cannot happen?
3509
3510 #ifndef PRODUCT
3511 if (C->print_intrinsics() || C->print_inlining()) {
3512 ciType* k = mirror_con->java_mirror_type();
3513 if (k) {
3514 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3515 k->print_name();
3516 tty->cr();
3517 }
3518 }
3519 #endif
3520
3521 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3522 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3523 record_for_igvn(region);
3524 PhiNode* phi = new (C) PhiNode(region, return_type);
3525
3526 // The mirror will never be null of Reflection.getClassAccessFlags, however
3527 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3528 // if it is. See bug 4774291.
3529
3530 // For Reflection.getClassAccessFlags(), the null check occurs in
3531 // the wrong place; see inline_unsafe_access(), above, for a similar
3532 // situation.
3533 mirror = null_check(mirror);
3534 // If mirror or obj is dead, only null-path is taken.
3535 if (stopped()) return true;
3536
3537 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3538
3539 // Now load the mirror's klass metaobject, and null-check it.
3540 // Side-effects region with the control path if the klass is null.
3541 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3542 // If kls is null, we have a primitive mirror.
3543 phi->init_req(_prim_path, prim_return_value);
3544 if (stopped()) { set_result(region, phi); return true; }
3545 bool safe_for_replace = (region->in(_prim_path) == top());
3546
3547 Node* p; // handy temp
3548 Node* null_ctl;
3549
3550 // Now that we have the non-null klass, we can perform the real query.
3551 // For constant classes, the query will constant-fold in LoadNode::Value.
3552 Node* query_value = top();
3553 switch (id) {
3554 case vmIntrinsics::_isInstance:
3555 // nothing is an instance of a primitive type
3556 query_value = gen_instanceof(obj, kls, safe_for_replace);
3557 break;
3558
3559 case vmIntrinsics::_getModifiers:
3560 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3561 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3562 break;
3563
3564 case vmIntrinsics::_isInterface:
3565 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3566 if (generate_interface_guard(kls, region) != NULL)
3567 // A guard was added. If the guard is taken, it was an interface.
3568 phi->add_req(intcon(1));
3569 // If we fall through, it's a plain class.
3570 query_value = intcon(0);
3571 break;
3572
3573 case vmIntrinsics::_isArray:
3574 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3575 if (generate_array_guard(kls, region) != NULL)
3576 // A guard was added. If the guard is taken, it was an array.
3577 phi->add_req(intcon(1));
3578 // If we fall through, it's a plain class.
3579 query_value = intcon(0);
3580 break;
3581
3582 case vmIntrinsics::_isPrimitive:
3583 query_value = intcon(0); // "normal" path produces false
3584 break;
3585
3586 case vmIntrinsics::_getSuperclass:
3587 // The rules here are somewhat unfortunate, but we can still do better
3588 // with random logic than with a JNI call.
3589 // Interfaces store null or Object as _super, but must report null.
3590 // Arrays store an intermediate super as _super, but must report Object.
3591 // Other types can report the actual _super.
3592 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3593 if (generate_interface_guard(kls, region) != NULL)
3594 // A guard was added. If the guard is taken, it was an interface.
3595 phi->add_req(null());
3596 if (generate_array_guard(kls, region) != NULL)
3597 // A guard was added. If the guard is taken, it was an array.
3598 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3599 // If we fall through, it's a plain class. Get its _super.
3600 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3601 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3602 null_ctl = top();
3603 kls = null_check_oop(kls, &null_ctl);
3604 if (null_ctl != top()) {
3605 // If the guard is taken, Object.superClass is null (both klass and mirror).
3606 region->add_req(null_ctl);
3607 phi ->add_req(null());
3608 }
3609 if (!stopped()) {
3610 query_value = load_mirror_from_klass(kls);
3611 }
3612 break;
3613
3614 case vmIntrinsics::_getComponentType:
3615 if (generate_array_guard(kls, region) != NULL) {
3616 // Be sure to pin the oop load to the guard edge just created:
3617 Node* is_array_ctrl = region->in(region->req()-1);
3618 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3619 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3620 phi->add_req(cmo);
3621 }
3622 query_value = null(); // non-array case is null
3623 break;
3624
3625 case vmIntrinsics::_getClassAccessFlags:
3626 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3627 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3628 break;
3629
3630 default:
3631 fatal_unexpected_iid(id);
3632 break;
3633 }
3634
3635 // Fall-through is the normal case of a query to a real class.
3636 phi->init_req(1, query_value);
3637 region->init_req(1, control());
3638
3639 C->set_has_split_ifs(true); // Has chance for split-if optimization
3640 set_result(region, phi);
3641 return true;
3642 }
3643
3644 //--------------------------inline_native_subtype_check------------------------
3645 // This intrinsic takes the JNI calls out of the heart of
3646 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3647 bool LibraryCallKit::inline_native_subtype_check() {
3648 // Pull both arguments off the stack.
3649 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3650 args[0] = argument(0);
3651 args[1] = argument(1);
3652 Node* klasses[2]; // corresponding Klasses: superk, subk
3653 klasses[0] = klasses[1] = top();
3654
3655 enum {
3656 // A full decision tree on {superc is prim, subc is prim}:
3657 _prim_0_path = 1, // {P,N} => false
3658 // {P,P} & superc!=subc => false
3659 _prim_same_path, // {P,P} & superc==subc => true
3660 _prim_1_path, // {N,P} => false
3661 _ref_subtype_path, // {N,N} & subtype check wins => true
3662 _both_ref_path, // {N,N} & subtype check loses => false
3663 PATH_LIMIT
3664 };
3665
3666 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3667 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3668 record_for_igvn(region);
3669
3670 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3671 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3672 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3673
3674 // First null-check both mirrors and load each mirror's klass metaobject.
3675 int which_arg;
3676 for (which_arg = 0; which_arg <= 1; which_arg++) {
3677 Node* arg = args[which_arg];
3678 arg = null_check(arg);
3679 if (stopped()) break;
3680 args[which_arg] = arg;
3681
3682 Node* p = basic_plus_adr(arg, class_klass_offset);
3683 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3684 klasses[which_arg] = _gvn.transform(kls);
3685 }
3686
3687 // Having loaded both klasses, test each for null.
3688 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3689 for (which_arg = 0; which_arg <= 1; which_arg++) {
3690 Node* kls = klasses[which_arg];
3691 Node* null_ctl = top();
3692 kls = null_check_oop(kls, &null_ctl, never_see_null);
3693 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3694 region->init_req(prim_path, null_ctl);
3695 if (stopped()) break;
3696 klasses[which_arg] = kls;
3697 }
3698
3699 if (!stopped()) {
3700 // now we have two reference types, in klasses[0..1]
3701 Node* subk = klasses[1]; // the argument to isAssignableFrom
3702 Node* superk = klasses[0]; // the receiver
3703 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3704 // now we have a successful reference subtype check
3705 region->set_req(_ref_subtype_path, control());
3706 }
3707
3708 // If both operands are primitive (both klasses null), then
3709 // we must return true when they are identical primitives.
3710 // It is convenient to test this after the first null klass check.
3711 set_control(region->in(_prim_0_path)); // go back to first null check
3712 if (!stopped()) {
3713 // Since superc is primitive, make a guard for the superc==subc case.
3714 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3715 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3716 generate_guard(bol_eq, region, PROB_FAIR);
3717 if (region->req() == PATH_LIMIT+1) {
3718 // A guard was added. If the added guard is taken, superc==subc.
3719 region->swap_edges(PATH_LIMIT, _prim_same_path);
3720 region->del_req(PATH_LIMIT);
3721 }
3722 region->set_req(_prim_0_path, control()); // Not equal after all.
3723 }
3724
3725 // these are the only paths that produce 'true':
3726 phi->set_req(_prim_same_path, intcon(1));
3727 phi->set_req(_ref_subtype_path, intcon(1));
3728
3729 // pull together the cases:
3730 assert(region->req() == PATH_LIMIT, "sane region");
3731 for (uint i = 1; i < region->req(); i++) {
3732 Node* ctl = region->in(i);
3733 if (ctl == NULL || ctl == top()) {
3734 region->set_req(i, top());
3735 phi ->set_req(i, top());
3736 } else if (phi->in(i) == NULL) {
3737 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3738 }
3739 }
3740
3741 set_control(_gvn.transform(region));
3742 set_result(_gvn.transform(phi));
3743 return true;
3744 }
3745
3746 //---------------------generate_array_guard_common------------------------
3747 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3748 bool obj_array, bool not_array) {
3749 // If obj_array/non_array==false/false:
3750 // Branch around if the given klass is in fact an array (either obj or prim).
3751 // If obj_array/non_array==false/true:
3752 // Branch around if the given klass is not an array klass of any kind.
3753 // If obj_array/non_array==true/true:
3754 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3755 // If obj_array/non_array==true/false:
3756 // Branch around if the kls is an oop array (Object[] or subtype)
3757 //
3758 // Like generate_guard, adds a new path onto the region.
3759 jint layout_con = 0;
3760 Node* layout_val = get_layout_helper(kls, layout_con);
3761 if (layout_val == NULL) {
3762 bool query = (obj_array
3763 ? Klass::layout_helper_is_objArray(layout_con)
3764 : Klass::layout_helper_is_array(layout_con));
3765 if (query == not_array) {
3766 return NULL; // never a branch
3767 } else { // always a branch
3768 Node* always_branch = control();
3769 if (region != NULL)
3770 region->add_req(always_branch);
3771 set_control(top());
3772 return always_branch;
3773 }
3774 }
3775 // Now test the correct condition.
3776 jint nval = (obj_array
3777 ? ((jint)Klass::_lh_array_tag_type_value
3778 << Klass::_lh_array_tag_shift)
3779 : Klass::_lh_neutral_value);
3780 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3781 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3782 // invert the test if we are looking for a non-array
3783 if (not_array) btest = BoolTest(btest).negate();
3784 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3785 return generate_fair_guard(bol, region);
3786 }
3787
3788
3789 //-----------------------inline_native_newArray--------------------------
3790 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3791 bool LibraryCallKit::inline_native_newArray() {
3792 Node* mirror = argument(0);
3793 Node* count_val = argument(1);
3794
3795 mirror = null_check(mirror);
3796 // If mirror or obj is dead, only null-path is taken.
3797 if (stopped()) return true;
3798
3799 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3800 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3801 PhiNode* result_val = new(C) PhiNode(result_reg,
3802 TypeInstPtr::NOTNULL);
3803 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3804 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3805 TypePtr::BOTTOM);
3806
3807 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3808 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3809 result_reg, _slow_path);
3810 Node* normal_ctl = control();
3811 Node* no_array_ctl = result_reg->in(_slow_path);
3812
3813 // Generate code for the slow case. We make a call to newArray().
3814 set_control(no_array_ctl);
3815 if (!stopped()) {
3816 // Either the input type is void.class, or else the
3817 // array klass has not yet been cached. Either the
3818 // ensuing call will throw an exception, or else it
3819 // will cache the array klass for next time.
3820 PreserveJVMState pjvms(this);
3821 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3822 Node* slow_result = set_results_for_java_call(slow_call);
3823 // this->control() comes from set_results_for_java_call
3824 result_reg->set_req(_slow_path, control());
3825 result_val->set_req(_slow_path, slow_result);
3826 result_io ->set_req(_slow_path, i_o());
3827 result_mem->set_req(_slow_path, reset_memory());
3828 }
3829
3830 set_control(normal_ctl);
3831 if (!stopped()) {
3832 // Normal case: The array type has been cached in the java.lang.Class.
3833 // The following call works fine even if the array type is polymorphic.
3834 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3835 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3836 result_reg->init_req(_normal_path, control());
3837 result_val->init_req(_normal_path, obj);
3838 result_io ->init_req(_normal_path, i_o());
3839 result_mem->init_req(_normal_path, reset_memory());
3840 }
3841
3842 // Return the combined state.
3843 set_i_o( _gvn.transform(result_io) );
3844 set_all_memory( _gvn.transform(result_mem));
3845
3846 C->set_has_split_ifs(true); // Has chance for split-if optimization
3847 set_result(result_reg, result_val);
3848 return true;
3849 }
3850
3851 //----------------------inline_native_getLength--------------------------
3852 // public static native int java.lang.reflect.Array.getLength(Object array);
3853 bool LibraryCallKit::inline_native_getLength() {
3854 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3855
3856 Node* array = null_check(argument(0));
3857 // If array is dead, only null-path is taken.
3858 if (stopped()) return true;
3859
3860 // Deoptimize if it is a non-array.
3861 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3862
3863 if (non_array != NULL) {
3864 PreserveJVMState pjvms(this);
3865 set_control(non_array);
3866 uncommon_trap(Deoptimization::Reason_intrinsic,
3867 Deoptimization::Action_maybe_recompile);
3868 }
3869
3870 // If control is dead, only non-array-path is taken.
3871 if (stopped()) return true;
3872
3873 // The works fine even if the array type is polymorphic.
3874 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3875 Node* result = load_array_length(array);
3876
3877 C->set_has_split_ifs(true); // Has chance for split-if optimization
3878 set_result(result);
3879 return true;
3880 }
3881
3882 //------------------------inline_array_copyOf----------------------------
3883 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3884 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3885 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3886 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3887
3888 // Get the arguments.
3889 Node* original = argument(0);
3890 Node* start = is_copyOfRange? argument(1): intcon(0);
3891 Node* end = is_copyOfRange? argument(2): argument(1);
3892 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3893
3894 Node* newcopy;
3895
3896 // Set the original stack and the reexecute bit for the interpreter to reexecute
3897 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3898 { PreserveReexecuteState preexecs(this);
3899 jvms()->set_should_reexecute(true);
3900
3901 array_type_mirror = null_check(array_type_mirror);
3902 original = null_check(original);
3903
3904 // Check if a null path was taken unconditionally.
3905 if (stopped()) return true;
3906
3907 Node* orig_length = load_array_length(original);
3908
3909 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3910 klass_node = null_check(klass_node);
3911
3912 RegionNode* bailout = new (C) RegionNode(1);
3913 record_for_igvn(bailout);
3914
3915 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3916 // Bail out if that is so.
3917 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3918 if (not_objArray != NULL) {
3919 // Improve the klass node's type from the new optimistic assumption:
3920 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3921 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3922 Node* cast = new (C) CastPPNode(klass_node, akls);
3923 cast->init_req(0, control());
3924 klass_node = _gvn.transform(cast);
3925 }
3926
3927 // Bail out if either start or end is negative.
3928 generate_negative_guard(start, bailout, &start);
3929 generate_negative_guard(end, bailout, &end);
3930
3931 Node* length = end;
3932 if (_gvn.type(start) != TypeInt::ZERO) {
3933 length = _gvn.transform(new (C) SubINode(end, start));
3934 }
3935
3936 // Bail out if length is negative.
3937 // Without this the new_array would throw
3938 // NegativeArraySizeException but IllegalArgumentException is what
3939 // should be thrown
3940 generate_negative_guard(length, bailout, &length);
3941
3942 if (bailout->req() > 1) {
3943 PreserveJVMState pjvms(this);
3944 set_control(_gvn.transform(bailout));
3945 uncommon_trap(Deoptimization::Reason_intrinsic,
3946 Deoptimization::Action_maybe_recompile);
3947 }
3948
3949 if (!stopped()) {
3950 // How many elements will we copy from the original?
3951 // The answer is MinI(orig_length - start, length).
3952 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3953 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3954
3955 newcopy = new_array(klass_node, length, 0); // no argments to push
3956
3957 // Generate a direct call to the right arraycopy function(s).
3958 // We know the copy is disjoint but we might not know if the
3959 // oop stores need checking.
3960 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3961 // This will fail a store-check if x contains any non-nulls.
3962 bool disjoint_bases = true;
3963 // if start > orig_length then the length of the copy may be
3964 // negative.
3965 bool length_never_negative = !is_copyOfRange;
3966 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3967 original, start, newcopy, intcon(0), moved,
3968 disjoint_bases, length_never_negative);
3969 }
3970 } // original reexecute is set back here
3971
3972 C->set_has_split_ifs(true); // Has chance for split-if optimization
3973 if (!stopped()) {
3974 set_result(newcopy);
3975 }
3976 return true;
3977 }
3978
3979
3980 //----------------------generate_virtual_guard---------------------------
3981 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3982 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3983 RegionNode* slow_region) {
3984 ciMethod* method = callee();
3985 int vtable_index = method->vtable_index();
3986 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3987 err_msg_res("bad index %d", vtable_index));
3988 // Get the Method* out of the appropriate vtable entry.
3989 int entry_offset = (InstanceKlass::vtable_start_offset() +
3990 vtable_index*vtableEntry::size()) * wordSize +
3991 vtableEntry::method_offset_in_bytes();
3992 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3993 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3994
3995 // Compare the target method with the expected method (e.g., Object.hashCode).
3996 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3997
3998 Node* native_call = makecon(native_call_addr);
3999 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4000 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4001
4002 return generate_slow_guard(test_native, slow_region);
4003 }
4004
4005 //-----------------------generate_method_call----------------------------
4006 // Use generate_method_call to make a slow-call to the real
4007 // method if the fast path fails. An alternative would be to
4008 // use a stub like OptoRuntime::slow_arraycopy_Java.
4009 // This only works for expanding the current library call,
4010 // not another intrinsic. (E.g., don't use this for making an
4011 // arraycopy call inside of the copyOf intrinsic.)
4012 CallJavaNode*
4013 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4014 // When compiling the intrinsic method itself, do not use this technique.
4015 guarantee(callee() != C->method(), "cannot make slow-call to self");
4016
4017 ciMethod* method = callee();
4018 // ensure the JVMS we have will be correct for this call
4019 guarantee(method_id == method->intrinsic_id(), "must match");
4020
4021 const TypeFunc* tf = TypeFunc::make(method);
4022 CallJavaNode* slow_call;
4023 if (is_static) {
4024 assert(!is_virtual, "");
4025 slow_call = new(C) CallStaticJavaNode(C, tf,
4026 SharedRuntime::get_resolve_static_call_stub(),
4027 method, bci());
4028 } else if (is_virtual) {
4029 null_check_receiver();
4030 int vtable_index = Method::invalid_vtable_index;
4031 if (UseInlineCaches) {
4032 // Suppress the vtable call
4033 } else {
4034 // hashCode and clone are not a miranda methods,
4035 // so the vtable index is fixed.
4036 // No need to use the linkResolver to get it.
4037 vtable_index = method->vtable_index();
4038 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4039 err_msg_res("bad index %d", vtable_index));
4040 }
4041 slow_call = new(C) CallDynamicJavaNode(tf,
4042 SharedRuntime::get_resolve_virtual_call_stub(),
4043 method, vtable_index, bci());
4044 } else { // neither virtual nor static: opt_virtual
4045 null_check_receiver();
4046 slow_call = new(C) CallStaticJavaNode(C, tf,
4047 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4048 method, bci());
4049 slow_call->set_optimized_virtual(true);
4050 }
4051 set_arguments_for_java_call(slow_call);
4052 set_edges_for_java_call(slow_call);
4053 return slow_call;
4054 }
4055
4056
4057 /**
4058 * Build special case code for calls to hashCode on an object. This call may
4059 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4060 * slightly different code.
4061 */
4062 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4063 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4064 assert(!(is_virtual && is_static), "either virtual, special, or static");
4065
4066 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4067
4068 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4069 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4070 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4071 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4072 Node* obj = NULL;
4073 if (!is_static) {
4074 // Check for hashing null object
4075 obj = null_check_receiver();
4076 if (stopped()) return true; // unconditionally null
4077 result_reg->init_req(_null_path, top());
4078 result_val->init_req(_null_path, top());
4079 } else {
4080 // Do a null check, and return zero if null.
4081 // System.identityHashCode(null) == 0
4082 obj = argument(0);
4083 Node* null_ctl = top();
4084 obj = null_check_oop(obj, &null_ctl);
4085 result_reg->init_req(_null_path, null_ctl);
4086 result_val->init_req(_null_path, _gvn.intcon(0));
4087 }
4088
4089 // Unconditionally null? Then return right away.
4090 if (stopped()) {
4091 set_control( result_reg->in(_null_path));
4092 if (!stopped())
4093 set_result(result_val->in(_null_path));
4094 return true;
4095 }
4096
4097 // We only go to the fast case code if we pass a number of guards. The
4098 // paths which do not pass are accumulated in the slow_region.
4099 RegionNode* slow_region = new (C) RegionNode(1);
4100 record_for_igvn(slow_region);
4101
4102 // If this is a virtual call, we generate a funny guard. We pull out
4103 // the vtable entry corresponding to hashCode() from the target object.
4104 // If the target method which we are calling happens to be the native
4105 // Object hashCode() method, we pass the guard. We do not need this
4106 // guard for non-virtual calls -- the caller is known to be the native
4107 // Object hashCode().
4108 if (is_virtual) {
4109 // After null check, get the object's klass.
4110 Node* obj_klass = load_object_klass(obj);
4111 generate_virtual_guard(obj_klass, slow_region);
4112 }
4113
4114 // Get the header out of the object, use LoadMarkNode when available
4115 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4116 // The control of the load must be NULL. Otherwise, the load can move before
4117 // the null check after castPP removal.
4118 Node* no_ctrl = NULL;
4119 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4120
4121 // Test the header to see if it is unlocked.
4122 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4123 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4124 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4125 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4126 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4127
4128 generate_slow_guard(test_unlocked, slow_region);
4129
4130 // Get the hash value and check to see that it has been properly assigned.
4131 // We depend on hash_mask being at most 32 bits and avoid the use of
4132 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4133 // vm: see markOop.hpp.
4134 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4135 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4136 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4137 // This hack lets the hash bits live anywhere in the mark object now, as long
4138 // as the shift drops the relevant bits into the low 32 bits. Note that
4139 // Java spec says that HashCode is an int so there's no point in capturing
4140 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4141 hshifted_header = ConvX2I(hshifted_header);
4142 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4143
4144 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4145 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4146 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4147
4148 generate_slow_guard(test_assigned, slow_region);
4149
4150 Node* init_mem = reset_memory();
4151 // fill in the rest of the null path:
4152 result_io ->init_req(_null_path, i_o());
4153 result_mem->init_req(_null_path, init_mem);
4154
4155 result_val->init_req(_fast_path, hash_val);
4156 result_reg->init_req(_fast_path, control());
4157 result_io ->init_req(_fast_path, i_o());
4158 result_mem->init_req(_fast_path, init_mem);
4159
4160 // Generate code for the slow case. We make a call to hashCode().
4161 set_control(_gvn.transform(slow_region));
4162 if (!stopped()) {
4163 // No need for PreserveJVMState, because we're using up the present state.
4164 set_all_memory(init_mem);
4165 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4166 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4167 Node* slow_result = set_results_for_java_call(slow_call);
4168 // this->control() comes from set_results_for_java_call
4169 result_reg->init_req(_slow_path, control());
4170 result_val->init_req(_slow_path, slow_result);
4171 result_io ->set_req(_slow_path, i_o());
4172 result_mem ->set_req(_slow_path, reset_memory());
4173 }
4174
4175 // Return the combined state.
4176 set_i_o( _gvn.transform(result_io) );
4177 set_all_memory( _gvn.transform(result_mem));
4178
4179 set_result(result_reg, result_val);
4180 return true;
4181 }
4182
4183 //---------------------------inline_native_getClass----------------------------
4184 // public final native Class<?> java.lang.Object.getClass();
4185 //
4186 // Build special case code for calls to getClass on an object.
4187 bool LibraryCallKit::inline_native_getClass() {
4188 Node* obj = null_check_receiver();
4189 if (stopped()) return true;
4190 set_result(load_mirror_from_klass(load_object_klass(obj)));
4191 return true;
4192 }
4193
4194 //-----------------inline_native_Reflection_getCallerClass---------------------
4195 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4196 //
4197 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4198 //
4199 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4200 // in that it must skip particular security frames and checks for
4201 // caller sensitive methods.
4202 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4203 #ifndef PRODUCT
4204 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4205 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4206 }
4207 #endif
4208
4209 if (!jvms()->has_method()) {
4210 #ifndef PRODUCT
4211 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4212 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4213 }
4214 #endif
4215 return false;
4216 }
4217
4218 // Walk back up the JVM state to find the caller at the required
4219 // depth.
4220 JVMState* caller_jvms = jvms();
4221
4222 // Cf. JVM_GetCallerClass
4223 // NOTE: Start the loop at depth 1 because the current JVM state does
4224 // not include the Reflection.getCallerClass() frame.
4225 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4226 ciMethod* m = caller_jvms->method();
4227 switch (n) {
4228 case 0:
4229 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4230 break;
4231 case 1:
4232 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4233 if (!m->caller_sensitive()) {
4234 #ifndef PRODUCT
4235 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4236 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4237 }
4238 #endif
4239 return false; // bail-out; let JVM_GetCallerClass do the work
4240 }
4241 break;
4242 default:
4243 if (!m->is_ignored_by_security_stack_walk()) {
4244 // We have reached the desired frame; return the holder class.
4245 // Acquire method holder as java.lang.Class and push as constant.
4246 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4247 ciInstance* caller_mirror = caller_klass->java_mirror();
4248 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4249
4250 #ifndef PRODUCT
4251 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4252 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());
4253 tty->print_cr(" JVM state at this point:");
4254 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4255 ciMethod* m = jvms()->of_depth(i)->method();
4256 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4257 }
4258 }
4259 #endif
4260 return true;
4261 }
4262 break;
4263 }
4264 }
4265
4266 #ifndef PRODUCT
4267 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4268 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4269 tty->print_cr(" JVM state at this point:");
4270 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4271 ciMethod* m = jvms()->of_depth(i)->method();
4272 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4273 }
4274 }
4275 #endif
4276
4277 return false; // bail-out; let JVM_GetCallerClass do the work
4278 }
4279
4280 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4281 Node* arg = argument(0);
4282 Node* result;
4283
4284 switch (id) {
4285 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4286 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4287 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4288 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4289
4290 case vmIntrinsics::_doubleToLongBits: {
4291 // two paths (plus control) merge in a wood
4292 RegionNode *r = new (C) RegionNode(3);
4293 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4294
4295 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4296 // Build the boolean node
4297 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4298
4299 // Branch either way.
4300 // NaN case is less traveled, which makes all the difference.
4301 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4302 Node *opt_isnan = _gvn.transform(ifisnan);
4303 assert( opt_isnan->is_If(), "Expect an IfNode");
4304 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4305 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4306
4307 set_control(iftrue);
4308
4309 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4310 Node *slow_result = longcon(nan_bits); // return NaN
4311 phi->init_req(1, _gvn.transform( slow_result ));
4312 r->init_req(1, iftrue);
4313
4314 // Else fall through
4315 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4316 set_control(iffalse);
4317
4318 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4319 r->init_req(2, iffalse);
4320
4321 // Post merge
4322 set_control(_gvn.transform(r));
4323 record_for_igvn(r);
4324
4325 C->set_has_split_ifs(true); // Has chance for split-if optimization
4326 result = phi;
4327 assert(result->bottom_type()->isa_long(), "must be");
4328 break;
4329 }
4330
4331 case vmIntrinsics::_floatToIntBits: {
4332 // two paths (plus control) merge in a wood
4333 RegionNode *r = new (C) RegionNode(3);
4334 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4335
4336 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4337 // Build the boolean node
4338 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4339
4340 // Branch either way.
4341 // NaN case is less traveled, which makes all the difference.
4342 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4343 Node *opt_isnan = _gvn.transform(ifisnan);
4344 assert( opt_isnan->is_If(), "Expect an IfNode");
4345 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4346 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4347
4348 set_control(iftrue);
4349
4350 static const jint nan_bits = 0x7fc00000;
4351 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4352 phi->init_req(1, _gvn.transform( slow_result ));
4353 r->init_req(1, iftrue);
4354
4355 // Else fall through
4356 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4357 set_control(iffalse);
4358
4359 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4360 r->init_req(2, iffalse);
4361
4362 // Post merge
4363 set_control(_gvn.transform(r));
4364 record_for_igvn(r);
4365
4366 C->set_has_split_ifs(true); // Has chance for split-if optimization
4367 result = phi;
4368 assert(result->bottom_type()->isa_int(), "must be");
4369 break;
4370 }
4371
4372 default:
4373 fatal_unexpected_iid(id);
4374 break;
4375 }
4376 set_result(_gvn.transform(result));
4377 return true;
4378 }
4379
4380 #ifdef _LP64
4381 #define XTOP ,top() /*additional argument*/
4382 #else //_LP64
4383 #define XTOP /*no additional argument*/
4384 #endif //_LP64
4385
4386 //----------------------inline_unsafe_copyMemory-------------------------
4387 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4388 bool LibraryCallKit::inline_unsafe_copyMemory() {
4389 if (callee()->is_static()) return false; // caller must have the capability!
4390 null_check_receiver(); // null-check receiver
4391 if (stopped()) return true;
4392
4393 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4394
4395 Node* src_ptr = argument(1); // type: oop
4396 Node* src_off = ConvL2X(argument(2)); // type: long
4397 Node* dst_ptr = argument(4); // type: oop
4398 Node* dst_off = ConvL2X(argument(5)); // type: long
4399 Node* size = ConvL2X(argument(7)); // type: long
4400
4401 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4402 "fieldOffset must be byte-scaled");
4403
4404 Node* src = make_unsafe_address(src_ptr, src_off);
4405 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4406
4407 // Conservatively insert a memory barrier on all memory slices.
4408 // Do not let writes of the copy source or destination float below the copy.
4409 insert_mem_bar(Op_MemBarCPUOrder);
4410
4411 // Call it. Note that the length argument is not scaled.
4412 make_runtime_call(RC_LEAF|RC_NO_FP,
4413 OptoRuntime::fast_arraycopy_Type(),
4414 StubRoutines::unsafe_arraycopy(),
4415 "unsafe_arraycopy",
4416 TypeRawPtr::BOTTOM,
4417 src, dst, size XTOP);
4418
4419 // Do not let reads of the copy destination float above the copy.
4420 insert_mem_bar(Op_MemBarCPUOrder);
4421
4422 return true;
4423 }
4424
4425 //------------------------clone_coping-----------------------------------
4426 // Helper function for inline_native_clone.
4427 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4428 assert(obj_size != NULL, "");
4429 Node* raw_obj = alloc_obj->in(1);
4430 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4431
4432 AllocateNode* alloc = NULL;
4433 if (ReduceBulkZeroing) {
4434 // We will be completely responsible for initializing this object -
4435 // mark Initialize node as complete.
4436 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4437 // The object was just allocated - there should be no any stores!
4438 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4439 // Mark as complete_with_arraycopy so that on AllocateNode
4440 // expansion, we know this AllocateNode is initialized by an array
4441 // copy and a StoreStore barrier exists after the array copy.
4442 alloc->initialization()->set_complete_with_arraycopy();
4443 }
4444
4445 // Copy the fastest available way.
4446 // TODO: generate fields copies for small objects instead.
4447 Node* src = obj;
4448 Node* dest = alloc_obj;
4449 Node* size = _gvn.transform(obj_size);
4450
4451 // Exclude the header but include array length to copy by 8 bytes words.
4452 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4453 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4454 instanceOopDesc::base_offset_in_bytes();
4455 // base_off:
4456 // 8 - 32-bit VM
4457 // 12 - 64-bit VM, compressed klass
4458 // 16 - 64-bit VM, normal klass
4459 if (base_off % BytesPerLong != 0) {
4460 assert(UseCompressedClassPointers, "");
4461 if (is_array) {
4462 // Exclude length to copy by 8 bytes words.
4463 base_off += sizeof(int);
4464 } else {
4465 // Include klass to copy by 8 bytes words.
4466 base_off = instanceOopDesc::klass_offset_in_bytes();
4467 }
4468 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4469 }
4470 src = basic_plus_adr(src, base_off);
4471 dest = basic_plus_adr(dest, base_off);
4472
4473 // Compute the length also, if needed:
4474 Node* countx = size;
4475 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4476 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4477
4478 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4479 bool disjoint_bases = true;
4480 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4481 src, NULL, dest, NULL, countx,
4482 /*dest_uninitialized*/true);
4483
4484 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4485 if (card_mark) {
4486 assert(!is_array, "");
4487 // Put in store barrier for any and all oops we are sticking
4488 // into this object. (We could avoid this if we could prove
4489 // that the object type contains no oop fields at all.)
4490 Node* no_particular_value = NULL;
4491 Node* no_particular_field = NULL;
4492 int raw_adr_idx = Compile::AliasIdxRaw;
4493 post_barrier(control(),
4494 memory(raw_adr_type),
4495 alloc_obj,
4496 no_particular_field,
4497 raw_adr_idx,
4498 no_particular_value,
4499 T_OBJECT,
4500 false);
4501 }
4502
4503 // Do not let reads from the cloned object float above the arraycopy.
4504 if (alloc != NULL) {
4505 // Do not let stores that initialize this object be reordered with
4506 // a subsequent store that would make this object accessible by
4507 // other threads.
4508 // Record what AllocateNode this StoreStore protects so that
4509 // escape analysis can go from the MemBarStoreStoreNode to the
4510 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4511 // based on the escape status of the AllocateNode.
4512 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4513 } else {
4514 insert_mem_bar(Op_MemBarCPUOrder);
4515 }
4516 }
4517
4518 //------------------------inline_native_clone----------------------------
4519 // protected native Object java.lang.Object.clone();
4520 //
4521 // Here are the simple edge cases:
4522 // null receiver => normal trap
4523 // virtual and clone was overridden => slow path to out-of-line clone
4524 // not cloneable or finalizer => slow path to out-of-line Object.clone
4525 //
4526 // The general case has two steps, allocation and copying.
4527 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4528 //
4529 // Copying also has two cases, oop arrays and everything else.
4530 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4531 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4532 //
4533 // These steps fold up nicely if and when the cloned object's klass
4534 // can be sharply typed as an object array, a type array, or an instance.
4535 //
4536 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4537 PhiNode* result_val;
4538
4539 // Set the reexecute bit for the interpreter to reexecute
4540 // the bytecode that invokes Object.clone if deoptimization happens.
4541 { PreserveReexecuteState preexecs(this);
4542 jvms()->set_should_reexecute(true);
4543
4544 Node* obj = null_check_receiver();
4545 if (stopped()) return true;
4546
4547 Node* obj_klass = load_object_klass(obj);
4548 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4549 const TypeOopPtr* toop = ((tklass != NULL)
4550 ? tklass->as_instance_type()
4551 : TypeInstPtr::NOTNULL);
4552
4553 // Conservatively insert a memory barrier on all memory slices.
4554 // Do not let writes into the original float below the clone.
4555 insert_mem_bar(Op_MemBarCPUOrder);
4556
4557 // paths into result_reg:
4558 enum {
4559 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4560 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4561 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4562 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4563 PATH_LIMIT
4564 };
4565 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4566 result_val = new(C) PhiNode(result_reg,
4567 TypeInstPtr::NOTNULL);
4568 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4569 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4570 TypePtr::BOTTOM);
4571 record_for_igvn(result_reg);
4572
4573 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4574 int raw_adr_idx = Compile::AliasIdxRaw;
4575
4576 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4577 if (array_ctl != NULL) {
4578 // It's an array.
4579 PreserveJVMState pjvms(this);
4580 set_control(array_ctl);
4581 Node* obj_length = load_array_length(obj);
4582 Node* obj_size = NULL;
4583 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4584
4585 if (!use_ReduceInitialCardMarks()) {
4586 // If it is an oop array, it requires very special treatment,
4587 // because card marking is required on each card of the array.
4588 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4589 if (is_obja != NULL) {
4590 PreserveJVMState pjvms2(this);
4591 set_control(is_obja);
4592 // Generate a direct call to the right arraycopy function(s).
4593 bool disjoint_bases = true;
4594 bool length_never_negative = true;
4595 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4596 obj, intcon(0), alloc_obj, intcon(0),
4597 obj_length,
4598 disjoint_bases, length_never_negative);
4599 result_reg->init_req(_objArray_path, control());
4600 result_val->init_req(_objArray_path, alloc_obj);
4601 result_i_o ->set_req(_objArray_path, i_o());
4602 result_mem ->set_req(_objArray_path, reset_memory());
4603 }
4604 }
4605 // Otherwise, there are no card marks to worry about.
4606 // (We can dispense with card marks if we know the allocation
4607 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4608 // causes the non-eden paths to take compensating steps to
4609 // simulate a fresh allocation, so that no further
4610 // card marks are required in compiled code to initialize
4611 // the object.)
4612
4613 if (!stopped()) {
4614 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4615
4616 // Present the results of the copy.
4617 result_reg->init_req(_array_path, control());
4618 result_val->init_req(_array_path, alloc_obj);
4619 result_i_o ->set_req(_array_path, i_o());
4620 result_mem ->set_req(_array_path, reset_memory());
4621 }
4622 }
4623
4624 // We only go to the instance fast case code if we pass a number of guards.
4625 // The paths which do not pass are accumulated in the slow_region.
4626 RegionNode* slow_region = new (C) RegionNode(1);
4627 record_for_igvn(slow_region);
4628 if (!stopped()) {
4629 // It's an instance (we did array above). Make the slow-path tests.
4630 // If this is a virtual call, we generate a funny guard. We grab
4631 // the vtable entry corresponding to clone() from the target object.
4632 // If the target method which we are calling happens to be the
4633 // Object clone() method, we pass the guard. We do not need this
4634 // guard for non-virtual calls; the caller is known to be the native
4635 // Object clone().
4636 if (is_virtual) {
4637 generate_virtual_guard(obj_klass, slow_region);
4638 }
4639
4640 // The object must be cloneable and must not have a finalizer.
4641 // Both of these conditions may be checked in a single test.
4642 // We could optimize the cloneable test further, but we don't care.
4643 generate_access_flags_guard(obj_klass,
4644 // Test both conditions:
4645 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4646 // Must be cloneable but not finalizer:
4647 JVM_ACC_IS_CLONEABLE,
4648 slow_region);
4649 }
4650
4651 if (!stopped()) {
4652 // It's an instance, and it passed the slow-path tests.
4653 PreserveJVMState pjvms(this);
4654 Node* obj_size = NULL;
4655 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4656 // is reexecuted if deoptimization occurs and there could be problems when merging
4657 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4658 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4659
4660 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4661
4662 // Present the results of the slow call.
4663 result_reg->init_req(_instance_path, control());
4664 result_val->init_req(_instance_path, alloc_obj);
4665 result_i_o ->set_req(_instance_path, i_o());
4666 result_mem ->set_req(_instance_path, reset_memory());
4667 }
4668
4669 // Generate code for the slow case. We make a call to clone().
4670 set_control(_gvn.transform(slow_region));
4671 if (!stopped()) {
4672 PreserveJVMState pjvms(this);
4673 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4674 Node* slow_result = set_results_for_java_call(slow_call);
4675 // this->control() comes from set_results_for_java_call
4676 result_reg->init_req(_slow_path, control());
4677 result_val->init_req(_slow_path, slow_result);
4678 result_i_o ->set_req(_slow_path, i_o());
4679 result_mem ->set_req(_slow_path, reset_memory());
4680 }
4681
4682 // Return the combined state.
4683 set_control( _gvn.transform(result_reg));
4684 set_i_o( _gvn.transform(result_i_o));
4685 set_all_memory( _gvn.transform(result_mem));
4686 } // original reexecute is set back here
4687
4688 set_result(_gvn.transform(result_val));
4689 return true;
4690 }
4691
4692 //------------------------------basictype2arraycopy----------------------------
4693 address LibraryCallKit::basictype2arraycopy(BasicType t,
4694 Node* src_offset,
4695 Node* dest_offset,
4696 bool disjoint_bases,
4697 const char* &name,
4698 bool dest_uninitialized) {
4699 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4700 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4701
4702 bool aligned = false;
4703 bool disjoint = disjoint_bases;
4704
4705 // if the offsets are the same, we can treat the memory regions as
4706 // disjoint, because either the memory regions are in different arrays,
4707 // or they are identical (which we can treat as disjoint.) We can also
4708 // treat a copy with a destination index less that the source index
4709 // as disjoint since a low->high copy will work correctly in this case.
4710 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4711 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4712 // both indices are constants
4713 int s_offs = src_offset_inttype->get_con();
4714 int d_offs = dest_offset_inttype->get_con();
4715 int element_size = type2aelembytes(t);
4716 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4717 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4718 if (s_offs >= d_offs) disjoint = true;
4719 } else if (src_offset == dest_offset && src_offset != NULL) {
4720 // This can occur if the offsets are identical non-constants.
4721 disjoint = true;
4722 }
4723
4724 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4725 }
4726
4727
4728 //------------------------------inline_arraycopy-----------------------
4729 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4730 // Object dest, int destPos,
4731 // int length);
4732 bool LibraryCallKit::inline_arraycopy() {
4733 // Get the arguments.
4734 Node* src = argument(0); // type: oop
4735 Node* src_offset = argument(1); // type: int
4736 Node* dest = argument(2); // type: oop
4737 Node* dest_offset = argument(3); // type: int
4738 Node* length = argument(4); // type: int
4739
4740 // Compile time checks. If any of these checks cannot be verified at compile time,
4741 // we do not make a fast path for this call. Instead, we let the call remain as it
4742 // is. The checks we choose to mandate at compile time are:
4743 //
4744 // (1) src and dest are arrays.
4745 const Type* src_type = src->Value(&_gvn);
4746 const Type* dest_type = dest->Value(&_gvn);
4747 const TypeAryPtr* top_src = src_type->isa_aryptr();
4748 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4749
4750 // Do we have the type of src?
4751 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4752 // Do we have the type of dest?
4753 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4754 // Is the type for src from speculation?
4755 bool src_spec = false;
4756 // Is the type for dest from speculation?
4757 bool dest_spec = false;
4758
4759 if (!has_src || !has_dest) {
4760 // We don't have sufficient type information, let's see if
4761 // speculative types can help. We need to have types for both src
4762 // and dest so that it pays off.
4763
4764 // Do we already have or could we have type information for src
4765 bool could_have_src = has_src;
4766 // Do we already have or could we have type information for dest
4767 bool could_have_dest = has_dest;
4768
4769 ciKlass* src_k = NULL;
4770 if (!has_src) {
4771 src_k = src_type->speculative_type();
4772 if (src_k != NULL && src_k->is_array_klass()) {
4773 could_have_src = true;
4774 }
4775 }
4776
4777 ciKlass* dest_k = NULL;
4778 if (!has_dest) {
4779 dest_k = dest_type->speculative_type();
4780 if (dest_k != NULL && dest_k->is_array_klass()) {
4781 could_have_dest = true;
4782 }
4783 }
4784
4785 if (could_have_src && could_have_dest) {
4786 // This is going to pay off so emit the required guards
4787 if (!has_src) {
4788 src = maybe_cast_profiled_obj(src, src_k);
4789 src_type = _gvn.type(src);
4790 top_src = src_type->isa_aryptr();
4791 has_src = (top_src != NULL && top_src->klass() != NULL);
4792 src_spec = true;
4793 }
4794 if (!has_dest) {
4795 dest = maybe_cast_profiled_obj(dest, dest_k);
4796 dest_type = _gvn.type(dest);
4797 top_dest = dest_type->isa_aryptr();
4798 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4799 dest_spec = true;
4800 }
4801 }
4802 }
4803
4804 if (!has_src || !has_dest) {
4805 // Conservatively insert a memory barrier on all memory slices.
4806 // Do not let writes into the source float below the arraycopy.
4807 insert_mem_bar(Op_MemBarCPUOrder);
4808
4809 // Call StubRoutines::generic_arraycopy stub.
4810 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4811 src, src_offset, dest, dest_offset, length);
4812
4813 // Do not let reads from the destination float above the arraycopy.
4814 // Since we cannot type the arrays, we don't know which slices
4815 // might be affected. We could restrict this barrier only to those
4816 // memory slices which pertain to array elements--but don't bother.
4817 if (!InsertMemBarAfterArraycopy)
4818 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4819 insert_mem_bar(Op_MemBarCPUOrder);
4820 return true;
4821 }
4822
4823 // (2) src and dest arrays must have elements of the same BasicType
4824 // Figure out the size and type of the elements we will be copying.
4825 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4826 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4827 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4828 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4829
4830 if (src_elem != dest_elem || dest_elem == T_VOID) {
4831 // The component types are not the same or are not recognized. Punt.
4832 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4833 generate_slow_arraycopy(TypePtr::BOTTOM,
4834 src, src_offset, dest, dest_offset, length,
4835 /*dest_uninitialized*/false);
4836 return true;
4837 }
4838
4839 if (src_elem == T_OBJECT) {
4840 // If both arrays are object arrays then having the exact types
4841 // for both will remove the need for a subtype check at runtime
4842 // before the call and may make it possible to pick a faster copy
4843 // routine (without a subtype check on every element)
4844 // Do we have the exact type of src?
4845 bool could_have_src = src_spec;
4846 // Do we have the exact type of dest?
4847 bool could_have_dest = dest_spec;
4848 ciKlass* src_k = top_src->klass();
4849 ciKlass* dest_k = top_dest->klass();
4850 if (!src_spec) {
4851 src_k = src_type->speculative_type();
4852 if (src_k != NULL && src_k->is_array_klass()) {
4853 could_have_src = true;
4854 }
4855 }
4856 if (!dest_spec) {
4857 dest_k = dest_type->speculative_type();
4858 if (dest_k != NULL && dest_k->is_array_klass()) {
4859 could_have_dest = true;
4860 }
4861 }
4862 if (could_have_src && could_have_dest) {
4863 // If we can have both exact types, emit the missing guards
4864 if (could_have_src && !src_spec) {
4865 src = maybe_cast_profiled_obj(src, src_k);
4866 }
4867 if (could_have_dest && !dest_spec) {
4868 dest = maybe_cast_profiled_obj(dest, dest_k);
4869 }
4870 }
4871 }
4872
4873 //---------------------------------------------------------------------------
4874 // We will make a fast path for this call to arraycopy.
4875
4876 // We have the following tests left to perform:
4877 //
4878 // (3) src and dest must not be null.
4879 // (4) src_offset must not be negative.
4880 // (5) dest_offset must not be negative.
4881 // (6) length must not be negative.
4882 // (7) src_offset + length must not exceed length of src.
4883 // (8) dest_offset + length must not exceed length of dest.
4884 // (9) each element of an oop array must be assignable
4885
4886 RegionNode* slow_region = new (C) RegionNode(1);
4887 record_for_igvn(slow_region);
4888
4889 // (3) operands must not be null
4890 // We currently perform our null checks with the null_check routine.
4891 // This means that the null exceptions will be reported in the caller
4892 // rather than (correctly) reported inside of the native arraycopy call.
4893 // This should be corrected, given time. We do our null check with the
4894 // stack pointer restored.
4895 src = null_check(src, T_ARRAY);
4896 dest = null_check(dest, T_ARRAY);
4897
4898 // (4) src_offset must not be negative.
4899 generate_negative_guard(src_offset, slow_region);
4900
4901 // (5) dest_offset must not be negative.
4902 generate_negative_guard(dest_offset, slow_region);
4903
4904 // (6) length must not be negative (moved to generate_arraycopy()).
4905 // generate_negative_guard(length, slow_region);
4906
4907 // (7) src_offset + length must not exceed length of src.
4908 generate_limit_guard(src_offset, length,
4909 load_array_length(src),
4910 slow_region);
4911
4912 // (8) dest_offset + length must not exceed length of dest.
4913 generate_limit_guard(dest_offset, length,
4914 load_array_length(dest),
4915 slow_region);
4916
4917 // (9) each element of an oop array must be assignable
4918 // The generate_arraycopy subroutine checks this.
4919
4920 // This is where the memory effects are placed:
4921 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4922 generate_arraycopy(adr_type, dest_elem,
4923 src, src_offset, dest, dest_offset, length,
4924 false, false, slow_region);
4925
4926 return true;
4927 }
4928
4929 //-----------------------------generate_arraycopy----------------------
4930 // Generate an optimized call to arraycopy.
4931 // Caller must guard against non-arrays.
4932 // Caller must determine a common array basic-type for both arrays.
4933 // Caller must validate offsets against array bounds.
4934 // The slow_region has already collected guard failure paths
4935 // (such as out of bounds length or non-conformable array types).
4936 // The generated code has this shape, in general:
4937 //
4938 // if (length == 0) return // via zero_path
4939 // slowval = -1
4940 // if (types unknown) {
4941 // slowval = call generic copy loop
4942 // if (slowval == 0) return // via checked_path
4943 // } else if (indexes in bounds) {
4944 // if ((is object array) && !(array type check)) {
4945 // slowval = call checked copy loop
4946 // if (slowval == 0) return // via checked_path
4947 // } else {
4948 // call bulk copy loop
4949 // return // via fast_path
4950 // }
4951 // }
4952 // // adjust params for remaining work:
4953 // if (slowval != -1) {
4954 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4955 // }
4956 // slow_region:
4957 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4958 // return // via slow_call_path
4959 //
4960 // This routine is used from several intrinsics: System.arraycopy,
4961 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4962 //
4963 void
4964 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4965 BasicType basic_elem_type,
4966 Node* src, Node* src_offset,
4967 Node* dest, Node* dest_offset,
4968 Node* copy_length,
4969 bool disjoint_bases,
4970 bool length_never_negative,
4971 RegionNode* slow_region) {
4972
4973 if (slow_region == NULL) {
4974 slow_region = new(C) RegionNode(1);
4975 record_for_igvn(slow_region);
4976 }
4977
4978 Node* original_dest = dest;
4979 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4980 bool dest_uninitialized = false;
4981
4982 // See if this is the initialization of a newly-allocated array.
4983 // If so, we will take responsibility here for initializing it to zero.
4984 // (Note: Because tightly_coupled_allocation performs checks on the
4985 // out-edges of the dest, we need to avoid making derived pointers
4986 // from it until we have checked its uses.)
4987 if (ReduceBulkZeroing
4988 && !ZeroTLAB // pointless if already zeroed
4989 && basic_elem_type != T_CONFLICT // avoid corner case
4990 && !src->eqv_uncast(dest)
4991 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4992 != NULL)
4993 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4994 && alloc->maybe_set_complete(&_gvn)) {
4995 // "You break it, you buy it."
4996 InitializeNode* init = alloc->initialization();
4997 assert(init->is_complete(), "we just did this");
4998 init->set_complete_with_arraycopy();
4999 assert(dest->is_CheckCastPP(), "sanity");
5000 assert(dest->in(0)->in(0) == init, "dest pinned");
5001 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5002 // From this point on, every exit path is responsible for
5003 // initializing any non-copied parts of the object to zero.
5004 // Also, if this flag is set we make sure that arraycopy interacts properly
5005 // with G1, eliding pre-barriers. See CR 6627983.
5006 dest_uninitialized = true;
5007 } else {
5008 // No zeroing elimination here.
5009 alloc = NULL;
5010 //original_dest = dest;
5011 //dest_uninitialized = false;
5012 }
5013
5014 // Results are placed here:
5015 enum { fast_path = 1, // normal void-returning assembly stub
5016 checked_path = 2, // special assembly stub with cleanup
5017 slow_call_path = 3, // something went wrong; call the VM
5018 zero_path = 4, // bypass when length of copy is zero
5019 bcopy_path = 5, // copy primitive array by 64-bit blocks
5020 PATH_LIMIT = 6
5021 };
5022 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5023 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5024 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5025 record_for_igvn(result_region);
5026 _gvn.set_type_bottom(result_i_o);
5027 _gvn.set_type_bottom(result_memory);
5028 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5029
5030 // The slow_control path:
5031 Node* slow_control;
5032 Node* slow_i_o = i_o();
5033 Node* slow_mem = memory(adr_type);
5034 debug_only(slow_control = (Node*) badAddress);
5035
5036 // Checked control path:
5037 Node* checked_control = top();
5038 Node* checked_mem = NULL;
5039 Node* checked_i_o = NULL;
5040 Node* checked_value = NULL;
5041
5042 if (basic_elem_type == T_CONFLICT) {
5043 assert(!dest_uninitialized, "");
5044 Node* cv = generate_generic_arraycopy(adr_type,
5045 src, src_offset, dest, dest_offset,
5046 copy_length, dest_uninitialized);
5047 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5048 checked_control = control();
5049 checked_i_o = i_o();
5050 checked_mem = memory(adr_type);
5051 checked_value = cv;
5052 set_control(top()); // no fast path
5053 }
5054
5055 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5056 if (not_pos != NULL) {
5057 PreserveJVMState pjvms(this);
5058 set_control(not_pos);
5059
5060 // (6) length must not be negative.
5061 if (!length_never_negative) {
5062 generate_negative_guard(copy_length, slow_region);
5063 }
5064
5065 // copy_length is 0.
5066 if (!stopped() && dest_uninitialized) {
5067 Node* dest_length = alloc->in(AllocateNode::ALength);
5068 if (copy_length->eqv_uncast(dest_length)
5069 || _gvn.find_int_con(dest_length, 1) <= 0) {
5070 // There is no zeroing to do. No need for a secondary raw memory barrier.
5071 } else {
5072 // Clear the whole thing since there are no source elements to copy.
5073 generate_clear_array(adr_type, dest, basic_elem_type,
5074 intcon(0), NULL,
5075 alloc->in(AllocateNode::AllocSize));
5076 // Use a secondary InitializeNode as raw memory barrier.
5077 // Currently it is needed only on this path since other
5078 // paths have stub or runtime calls as raw memory barriers.
5079 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5080 Compile::AliasIdxRaw,
5081 top())->as_Initialize();
5082 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5083 }
5084 }
5085
5086 // Present the results of the fast call.
5087 result_region->init_req(zero_path, control());
5088 result_i_o ->init_req(zero_path, i_o());
5089 result_memory->init_req(zero_path, memory(adr_type));
5090 }
5091
5092 if (!stopped() && dest_uninitialized) {
5093 // We have to initialize the *uncopied* part of the array to zero.
5094 // The copy destination is the slice dest[off..off+len]. The other slices
5095 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5096 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5097 Node* dest_length = alloc->in(AllocateNode::ALength);
5098 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5099 copy_length));
5100
5101 // If there is a head section that needs zeroing, do it now.
5102 if (find_int_con(dest_offset, -1) != 0) {
5103 generate_clear_array(adr_type, dest, basic_elem_type,
5104 intcon(0), dest_offset,
5105 NULL);
5106 }
5107
5108 // Next, perform a dynamic check on the tail length.
5109 // It is often zero, and we can win big if we prove this.
5110 // There are two wins: Avoid generating the ClearArray
5111 // with its attendant messy index arithmetic, and upgrade
5112 // the copy to a more hardware-friendly word size of 64 bits.
5113 Node* tail_ctl = NULL;
5114 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5115 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5116 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5117 tail_ctl = generate_slow_guard(bol_lt, NULL);
5118 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5119 }
5120
5121 // At this point, let's assume there is no tail.
5122 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5123 // There is no tail. Try an upgrade to a 64-bit copy.
5124 bool didit = false;
5125 { PreserveJVMState pjvms(this);
5126 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5127 src, src_offset, dest, dest_offset,
5128 dest_size, dest_uninitialized);
5129 if (didit) {
5130 // Present the results of the block-copying fast call.
5131 result_region->init_req(bcopy_path, control());
5132 result_i_o ->init_req(bcopy_path, i_o());
5133 result_memory->init_req(bcopy_path, memory(adr_type));
5134 }
5135 }
5136 if (didit)
5137 set_control(top()); // no regular fast path
5138 }
5139
5140 // Clear the tail, if any.
5141 if (tail_ctl != NULL) {
5142 Node* notail_ctl = stopped() ? NULL : control();
5143 set_control(tail_ctl);
5144 if (notail_ctl == NULL) {
5145 generate_clear_array(adr_type, dest, basic_elem_type,
5146 dest_tail, NULL,
5147 dest_size);
5148 } else {
5149 // Make a local merge.
5150 Node* done_ctl = new(C) RegionNode(3);
5151 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5152 done_ctl->init_req(1, notail_ctl);
5153 done_mem->init_req(1, memory(adr_type));
5154 generate_clear_array(adr_type, dest, basic_elem_type,
5155 dest_tail, NULL,
5156 dest_size);
5157 done_ctl->init_req(2, control());
5158 done_mem->init_req(2, memory(adr_type));
5159 set_control( _gvn.transform(done_ctl));
5160 set_memory( _gvn.transform(done_mem), adr_type );
5161 }
5162 }
5163 }
5164
5165 BasicType copy_type = basic_elem_type;
5166 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5167 if (!stopped() && copy_type == T_OBJECT) {
5168 // If src and dest have compatible element types, we can copy bits.
5169 // Types S[] and D[] are compatible if D is a supertype of S.
5170 //
5171 // If they are not, we will use checked_oop_disjoint_arraycopy,
5172 // which performs a fast optimistic per-oop check, and backs off
5173 // further to JVM_ArrayCopy on the first per-oop check that fails.
5174 // (Actually, we don't move raw bits only; the GC requires card marks.)
5175
5176 // Get the Klass* for both src and dest
5177 Node* src_klass = load_object_klass(src);
5178 Node* dest_klass = load_object_klass(dest);
5179
5180 // Generate the subtype check.
5181 // This might fold up statically, or then again it might not.
5182 //
5183 // Non-static example: Copying List<String>.elements to a new String[].
5184 // The backing store for a List<String> is always an Object[],
5185 // but its elements are always type String, if the generic types
5186 // are correct at the source level.
5187 //
5188 // Test S[] against D[], not S against D, because (probably)
5189 // the secondary supertype cache is less busy for S[] than S.
5190 // This usually only matters when D is an interface.
5191 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5192 // Plug failing path into checked_oop_disjoint_arraycopy
5193 if (not_subtype_ctrl != top()) {
5194 PreserveJVMState pjvms(this);
5195 set_control(not_subtype_ctrl);
5196 // (At this point we can assume disjoint_bases, since types differ.)
5197 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5198 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5199 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5200 Node* dest_elem_klass = _gvn.transform(n1);
5201 Node* cv = generate_checkcast_arraycopy(adr_type,
5202 dest_elem_klass,
5203 src, src_offset, dest, dest_offset,
5204 ConvI2X(copy_length), dest_uninitialized);
5205 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5206 checked_control = control();
5207 checked_i_o = i_o();
5208 checked_mem = memory(adr_type);
5209 checked_value = cv;
5210 }
5211 // At this point we know we do not need type checks on oop stores.
5212
5213 // Let's see if we need card marks:
5214 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5215 // If we do not need card marks, copy using the jint or jlong stub.
5216 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5217 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5218 "sizes agree");
5219 }
5220 }
5221
5222 if (!stopped()) {
5223 // Generate the fast path, if possible.
5224 PreserveJVMState pjvms(this);
5225 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5226 src, src_offset, dest, dest_offset,
5227 ConvI2X(copy_length), dest_uninitialized);
5228
5229 // Present the results of the fast call.
5230 result_region->init_req(fast_path, control());
5231 result_i_o ->init_req(fast_path, i_o());
5232 result_memory->init_req(fast_path, memory(adr_type));
5233 }
5234
5235 // Here are all the slow paths up to this point, in one bundle:
5236 slow_control = top();
5237 if (slow_region != NULL)
5238 slow_control = _gvn.transform(slow_region);
5239 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5240
5241 set_control(checked_control);
5242 if (!stopped()) {
5243 // Clean up after the checked call.
5244 // The returned value is either 0 or -1^K,
5245 // where K = number of partially transferred array elements.
5246 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5247 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5248 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5249
5250 // If it is 0, we are done, so transfer to the end.
5251 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5252 result_region->init_req(checked_path, checks_done);
5253 result_i_o ->init_req(checked_path, checked_i_o);
5254 result_memory->init_req(checked_path, checked_mem);
5255
5256 // If it is not zero, merge into the slow call.
5257 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5258 RegionNode* slow_reg2 = new(C) RegionNode(3);
5259 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5260 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5261 record_for_igvn(slow_reg2);
5262 slow_reg2 ->init_req(1, slow_control);
5263 slow_i_o2 ->init_req(1, slow_i_o);
5264 slow_mem2 ->init_req(1, slow_mem);
5265 slow_reg2 ->init_req(2, control());
5266 slow_i_o2 ->init_req(2, checked_i_o);
5267 slow_mem2 ->init_req(2, checked_mem);
5268
5269 slow_control = _gvn.transform(slow_reg2);
5270 slow_i_o = _gvn.transform(slow_i_o2);
5271 slow_mem = _gvn.transform(slow_mem2);
5272
5273 if (alloc != NULL) {
5274 // We'll restart from the very beginning, after zeroing the whole thing.
5275 // This can cause double writes, but that's OK since dest is brand new.
5276 // So we ignore the low 31 bits of the value returned from the stub.
5277 } else {
5278 // We must continue the copy exactly where it failed, or else
5279 // another thread might see the wrong number of writes to dest.
5280 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5281 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5282 slow_offset->init_req(1, intcon(0));
5283 slow_offset->init_req(2, checked_offset);
5284 slow_offset = _gvn.transform(slow_offset);
5285
5286 // Adjust the arguments by the conditionally incoming offset.
5287 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5288 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5289 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5290
5291 // Tweak the node variables to adjust the code produced below:
5292 src_offset = src_off_plus;
5293 dest_offset = dest_off_plus;
5294 copy_length = length_minus;
5295 }
5296 }
5297
5298 set_control(slow_control);
5299 if (!stopped()) {
5300 // Generate the slow path, if needed.
5301 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5302
5303 set_memory(slow_mem, adr_type);
5304 set_i_o(slow_i_o);
5305
5306 if (dest_uninitialized) {
5307 generate_clear_array(adr_type, dest, basic_elem_type,
5308 intcon(0), NULL,
5309 alloc->in(AllocateNode::AllocSize));
5310 }
5311
5312 generate_slow_arraycopy(adr_type,
5313 src, src_offset, dest, dest_offset,
5314 copy_length, /*dest_uninitialized*/false);
5315
5316 result_region->init_req(slow_call_path, control());
5317 result_i_o ->init_req(slow_call_path, i_o());
5318 result_memory->init_req(slow_call_path, memory(adr_type));
5319 }
5320
5321 // Remove unused edges.
5322 for (uint i = 1; i < result_region->req(); i++) {
5323 if (result_region->in(i) == NULL)
5324 result_region->init_req(i, top());
5325 }
5326
5327 // Finished; return the combined state.
5328 set_control( _gvn.transform(result_region));
5329 set_i_o( _gvn.transform(result_i_o) );
5330 set_memory( _gvn.transform(result_memory), adr_type );
5331
5332 // The memory edges above are precise in order to model effects around
5333 // array copies accurately to allow value numbering of field loads around
5334 // arraycopy. Such field loads, both before and after, are common in Java
5335 // collections and similar classes involving header/array data structures.
5336 //
5337 // But with low number of register or when some registers are used or killed
5338 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5339 // The next memory barrier is added to avoid it. If the arraycopy can be
5340 // optimized away (which it can, sometimes) then we can manually remove
5341 // the membar also.
5342 //
5343 // Do not let reads from the cloned object float above the arraycopy.
5344 if (alloc != NULL) {
5345 // Do not let stores that initialize this object be reordered with
5346 // a subsequent store that would make this object accessible by
5347 // other threads.
5348 // Record what AllocateNode this StoreStore protects so that
5349 // escape analysis can go from the MemBarStoreStoreNode to the
5350 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5351 // based on the escape status of the AllocateNode.
5352 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5353 } else if (InsertMemBarAfterArraycopy)
5354 insert_mem_bar(Op_MemBarCPUOrder);
5355 }
5356
5357
5358 // Helper function which determines if an arraycopy immediately follows
5359 // an allocation, with no intervening tests or other escapes for the object.
5360 AllocateArrayNode*
5361 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5362 RegionNode* slow_region) {
5363 if (stopped()) return NULL; // no fast path
5364 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5365
5366 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5367 if (alloc == NULL) return NULL;
5368
5369 Node* rawmem = memory(Compile::AliasIdxRaw);
5370 // Is the allocation's memory state untouched?
5371 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5372 // Bail out if there have been raw-memory effects since the allocation.
5373 // (Example: There might have been a call or safepoint.)
5374 return NULL;
5375 }
5376 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5377 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5378 return NULL;
5379 }
5380
5381 // There must be no unexpected observers of this allocation.
5382 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5383 Node* obs = ptr->fast_out(i);
5384 if (obs != this->map()) {
5385 return NULL;
5386 }
5387 }
5388
5389 // This arraycopy must unconditionally follow the allocation of the ptr.
5390 Node* alloc_ctl = ptr->in(0);
5391 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5392
5393 Node* ctl = control();
5394 while (ctl != alloc_ctl) {
5395 // There may be guards which feed into the slow_region.
5396 // Any other control flow means that we might not get a chance
5397 // to finish initializing the allocated object.
5398 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5399 IfNode* iff = ctl->in(0)->as_If();
5400 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5401 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5402 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5403 ctl = iff->in(0); // This test feeds the known slow_region.
5404 continue;
5405 }
5406 // One more try: Various low-level checks bottom out in
5407 // uncommon traps. If the debug-info of the trap omits
5408 // any reference to the allocation, as we've already
5409 // observed, then there can be no objection to the trap.
5410 bool found_trap = false;
5411 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5412 Node* obs = not_ctl->fast_out(j);
5413 if (obs->in(0) == not_ctl && obs->is_Call() &&
5414 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5415 found_trap = true; break;
5416 }
5417 }
5418 if (found_trap) {
5419 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5420 continue;
5421 }
5422 }
5423 return NULL;
5424 }
5425
5426 // If we get this far, we have an allocation which immediately
5427 // precedes the arraycopy, and we can take over zeroing the new object.
5428 // The arraycopy will finish the initialization, and provide
5429 // a new control state to which we will anchor the destination pointer.
5430
5431 return alloc;
5432 }
5433
5434 // Helper for initialization of arrays, creating a ClearArray.
5435 // It writes zero bits in [start..end), within the body of an array object.
5436 // The memory effects are all chained onto the 'adr_type' alias category.
5437 //
5438 // Since the object is otherwise uninitialized, we are free
5439 // to put a little "slop" around the edges of the cleared area,
5440 // as long as it does not go back into the array's header,
5441 // or beyond the array end within the heap.
5442 //
5443 // The lower edge can be rounded down to the nearest jint and the
5444 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5445 //
5446 // Arguments:
5447 // adr_type memory slice where writes are generated
5448 // dest oop of the destination array
5449 // basic_elem_type element type of the destination
5450 // slice_idx array index of first element to store
5451 // slice_len number of elements to store (or NULL)
5452 // dest_size total size in bytes of the array object
5453 //
5454 // Exactly one of slice_len or dest_size must be non-NULL.
5455 // If dest_size is non-NULL, zeroing extends to the end of the object.
5456 // If slice_len is non-NULL, the slice_idx value must be a constant.
5457 void
5458 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5459 Node* dest,
5460 BasicType basic_elem_type,
5461 Node* slice_idx,
5462 Node* slice_len,
5463 Node* dest_size) {
5464 // one or the other but not both of slice_len and dest_size:
5465 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5466 if (slice_len == NULL) slice_len = top();
5467 if (dest_size == NULL) dest_size = top();
5468
5469 // operate on this memory slice:
5470 Node* mem = memory(adr_type); // memory slice to operate on
5471
5472 // scaling and rounding of indexes:
5473 int scale = exact_log2(type2aelembytes(basic_elem_type));
5474 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5475 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5476 int bump_bit = (-1 << scale) & BytesPerInt;
5477
5478 // determine constant starts and ends
5479 const intptr_t BIG_NEG = -128;
5480 assert(BIG_NEG + 2*abase < 0, "neg enough");
5481 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5482 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5483 if (slice_len_con == 0) {
5484 return; // nothing to do here
5485 }
5486 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5487 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5488 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5489 assert(end_con < 0, "not two cons");
5490 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5491 BytesPerLong);
5492 }
5493
5494 if (start_con >= 0 && end_con >= 0) {
5495 // Constant start and end. Simple.
5496 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5497 start_con, end_con, &_gvn);
5498 } else if (start_con >= 0 && dest_size != top()) {
5499 // Constant start, pre-rounded end after the tail of the array.
5500 Node* end = dest_size;
5501 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5502 start_con, end, &_gvn);
5503 } else if (start_con >= 0 && slice_len != top()) {
5504 // Constant start, non-constant end. End needs rounding up.
5505 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5506 intptr_t end_base = abase + (slice_idx_con << scale);
5507 int end_round = (-1 << scale) & (BytesPerLong - 1);
5508 Node* end = ConvI2X(slice_len);
5509 if (scale != 0)
5510 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5511 end_base += end_round;
5512 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5513 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5514 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5515 start_con, end, &_gvn);
5516 } else if (start_con < 0 && dest_size != top()) {
5517 // Non-constant start, pre-rounded end after the tail of the array.
5518 // This is almost certainly a "round-to-end" operation.
5519 Node* start = slice_idx;
5520 start = ConvI2X(start);
5521 if (scale != 0)
5522 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5523 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5524 if ((bump_bit | clear_low) != 0) {
5525 int to_clear = (bump_bit | clear_low);
5526 // Align up mod 8, then store a jint zero unconditionally
5527 // just before the mod-8 boundary.
5528 if (((abase + bump_bit) & ~to_clear) - bump_bit
5529 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5530 bump_bit = 0;
5531 assert((abase & to_clear) == 0, "array base must be long-aligned");
5532 } else {
5533 // Bump 'start' up to (or past) the next jint boundary:
5534 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5535 assert((abase & clear_low) == 0, "array base must be int-aligned");
5536 }
5537 // Round bumped 'start' down to jlong boundary in body of array.
5538 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5539 if (bump_bit != 0) {
5540 // Store a zero to the immediately preceding jint:
5541 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5542 Node* p1 = basic_plus_adr(dest, x1);
5543 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5544 mem = _gvn.transform(mem);
5545 }
5546 }
5547 Node* end = dest_size; // pre-rounded
5548 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5549 start, end, &_gvn);
5550 } else {
5551 // Non-constant start, unrounded non-constant end.
5552 // (Nobody zeroes a random midsection of an array using this routine.)
5553 ShouldNotReachHere(); // fix caller
5554 }
5555
5556 // Done.
5557 set_memory(mem, adr_type);
5558 }
5559
5560
5561 bool
5562 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5563 BasicType basic_elem_type,
5564 AllocateNode* alloc,
5565 Node* src, Node* src_offset,
5566 Node* dest, Node* dest_offset,
5567 Node* dest_size, bool dest_uninitialized) {
5568 // See if there is an advantage from block transfer.
5569 int scale = exact_log2(type2aelembytes(basic_elem_type));
5570 if (scale >= LogBytesPerLong)
5571 return false; // it is already a block transfer
5572
5573 // Look at the alignment of the starting offsets.
5574 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5575
5576 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5577 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5578 if (src_off_con < 0 || dest_off_con < 0)
5579 // At present, we can only understand constants.
5580 return false;
5581
5582 intptr_t src_off = abase + (src_off_con << scale);
5583 intptr_t dest_off = abase + (dest_off_con << scale);
5584
5585 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5586 // Non-aligned; too bad.
5587 // One more chance: Pick off an initial 32-bit word.
5588 // This is a common case, since abase can be odd mod 8.
5589 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5590 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5591 Node* sptr = basic_plus_adr(src, src_off);
5592 Node* dptr = basic_plus_adr(dest, dest_off);
5593 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5594 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5595 src_off += BytesPerInt;
5596 dest_off += BytesPerInt;
5597 } else {
5598 return false;
5599 }
5600 }
5601 assert(src_off % BytesPerLong == 0, "");
5602 assert(dest_off % BytesPerLong == 0, "");
5603
5604 // Do this copy by giant steps.
5605 Node* sptr = basic_plus_adr(src, src_off);
5606 Node* dptr = basic_plus_adr(dest, dest_off);
5607 Node* countx = dest_size;
5608 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5609 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5610
5611 bool disjoint_bases = true; // since alloc != NULL
5612 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5613 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5614
5615 return true;
5616 }
5617
5618
5619 // Helper function; generates code for the slow case.
5620 // We make a call to a runtime method which emulates the native method,
5621 // but without the native wrapper overhead.
5622 void
5623 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5624 Node* src, Node* src_offset,
5625 Node* dest, Node* dest_offset,
5626 Node* copy_length, bool dest_uninitialized) {
5627 assert(!dest_uninitialized, "Invariant");
5628 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5629 OptoRuntime::slow_arraycopy_Type(),
5630 OptoRuntime::slow_arraycopy_Java(),
5631 "slow_arraycopy", adr_type,
5632 src, src_offset, dest, dest_offset,
5633 copy_length);
5634
5635 // Handle exceptions thrown by this fellow:
5636 make_slow_call_ex(call, env()->Throwable_klass(), false);
5637 }
5638
5639 // Helper function; generates code for cases requiring runtime checks.
5640 Node*
5641 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5642 Node* dest_elem_klass,
5643 Node* src, Node* src_offset,
5644 Node* dest, Node* dest_offset,
5645 Node* copy_length, bool dest_uninitialized) {
5646 if (stopped()) return NULL;
5647
5648 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5649 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5650 return NULL;
5651 }
5652
5653 // Pick out the parameters required to perform a store-check
5654 // for the target array. This is an optimistic check. It will
5655 // look in each non-null element's class, at the desired klass's
5656 // super_check_offset, for the desired klass.
5657 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5658 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5659 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5660 Node* check_offset = ConvI2X(_gvn.transform(n3));
5661 Node* check_value = dest_elem_klass;
5662
5663 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5664 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5665
5666 // (We know the arrays are never conjoint, because their types differ.)
5667 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5668 OptoRuntime::checkcast_arraycopy_Type(),
5669 copyfunc_addr, "checkcast_arraycopy", adr_type,
5670 // five arguments, of which two are
5671 // intptr_t (jlong in LP64)
5672 src_start, dest_start,
5673 copy_length XTOP,
5674 check_offset XTOP,
5675 check_value);
5676
5677 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5678 }
5679
5680
5681 // Helper function; generates code for cases requiring runtime checks.
5682 Node*
5683 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5684 Node* src, Node* src_offset,
5685 Node* dest, Node* dest_offset,
5686 Node* copy_length, bool dest_uninitialized) {
5687 assert(!dest_uninitialized, "Invariant");
5688 if (stopped()) return NULL;
5689 address copyfunc_addr = StubRoutines::generic_arraycopy();
5690 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5691 return NULL;
5692 }
5693
5694 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5695 OptoRuntime::generic_arraycopy_Type(),
5696 copyfunc_addr, "generic_arraycopy", adr_type,
5697 src, src_offset, dest, dest_offset, copy_length);
5698
5699 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5700 }
5701
5702 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5703 void
5704 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5705 BasicType basic_elem_type,
5706 bool disjoint_bases,
5707 Node* src, Node* src_offset,
5708 Node* dest, Node* dest_offset,
5709 Node* copy_length, bool dest_uninitialized) {
5710 if (stopped()) return; // nothing to do
5711
5712 Node* src_start = src;
5713 Node* dest_start = dest;
5714 if (src_offset != NULL || dest_offset != NULL) {
5715 assert(src_offset != NULL && dest_offset != NULL, "");
5716 src_start = array_element_address(src, src_offset, basic_elem_type);
5717 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5718 }
5719
5720 // Figure out which arraycopy runtime method to call.
5721 const char* copyfunc_name = "arraycopy";
5722 address copyfunc_addr =
5723 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5724 disjoint_bases, copyfunc_name, dest_uninitialized);
5725
5726 // Call it. Note that the count_ix value is not scaled to a byte-size.
5727 make_runtime_call(RC_LEAF|RC_NO_FP,
5728 OptoRuntime::fast_arraycopy_Type(),
5729 copyfunc_addr, copyfunc_name, adr_type,
5730 src_start, dest_start, copy_length XTOP);
5731 }
5732
5733 //-------------inline_encodeISOArray-----------------------------------
5734 // encode char[] to byte[] in ISO_8859_1
5735 bool LibraryCallKit::inline_encodeISOArray() {
5736 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5737 // no receiver since it is static method
5738 Node *src = argument(0);
5739 Node *src_offset = argument(1);
5740 Node *dst = argument(2);
5741 Node *dst_offset = argument(3);
5742 Node *length = argument(4);
5743
5744 const Type* src_type = src->Value(&_gvn);
5745 const Type* dst_type = dst->Value(&_gvn);
5746 const TypeAryPtr* top_src = src_type->isa_aryptr();
5747 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5748 if (top_src == NULL || top_src->klass() == NULL ||
5749 top_dest == NULL || top_dest->klass() == NULL) {
5750 // failed array check
5751 return false;
5752 }
5753
5754 // Figure out the size and type of the elements we will be copying.
5755 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5756 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5757 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5758 return false;
5759 }
5760 Node* src_start = array_element_address(src, src_offset, src_elem);
5761 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5762 // 'src_start' points to src array + scaled offset
5763 // 'dst_start' points to dst array + scaled offset
5764
5765 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5766 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5767 enc = _gvn.transform(enc);
5768 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5769 set_memory(res_mem, mtype);
5770 set_result(enc);
5771 return true;
5772 }
5773
5774 //-------------inline_multiplyToLen-----------------------------------
5775 bool LibraryCallKit::inline_multiplyToLen() {
5776 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5777
5778 address stubAddr = StubRoutines::multiplyToLen();
5779 if (stubAddr == NULL) {
5780 return false; // Intrinsic's stub is not implemented on this platform
5781 }
5782 const char* stubName = "multiplyToLen";
5783
5784 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5785
5786 Node* x = argument(1);
5787 Node* xlen = argument(2);
5788 Node* y = argument(3);
5789 Node* ylen = argument(4);
5790 Node* z = argument(5);
5791
5792 const Type* x_type = x->Value(&_gvn);
5793 const Type* y_type = y->Value(&_gvn);
5794 const TypeAryPtr* top_x = x_type->isa_aryptr();
5795 const TypeAryPtr* top_y = y_type->isa_aryptr();
5796 if (top_x == NULL || top_x->klass() == NULL ||
5797 top_y == NULL || top_y->klass() == NULL) {
5798 // failed array check
5799 return false;
5800 }
5801
5802 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5803 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5804 if (x_elem != T_INT || y_elem != T_INT) {
5805 return false;
5806 }
5807
5808 // Set the original stack and the reexecute bit for the interpreter to reexecute
5809 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5810 // on the return from z array allocation in runtime.
5811 { PreserveReexecuteState preexecs(this);
5812 jvms()->set_should_reexecute(true);
5813
5814 Node* x_start = array_element_address(x, intcon(0), x_elem);
5815 Node* y_start = array_element_address(y, intcon(0), y_elem);
5816 // 'x_start' points to x array + scaled xlen
5817 // 'y_start' points to y array + scaled ylen
5818
5819 // Allocate the result array
5820 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5821 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5822 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5823
5824 IdealKit ideal(this);
5825
5826 #define __ ideal.
5827 Node* one = __ ConI(1);
5828 Node* zero = __ ConI(0);
5829 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5830 __ set(need_alloc, zero);
5831 __ set(z_alloc, z);
5832 __ if_then(z, BoolTest::eq, null()); {
5833 __ increment (need_alloc, one);
5834 } __ else_(); {
5835 // Update graphKit memory and control from IdealKit.
5836 sync_kit(ideal);
5837 Node* zlen_arg = load_array_length(z);
5838 // Update IdealKit memory and control from graphKit.
5839 __ sync_kit(this);
5840 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5841 __ increment (need_alloc, one);
5842 } __ end_if();
5843 } __ end_if();
5844
5845 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5846 // Update graphKit memory and control from IdealKit.
5847 sync_kit(ideal);
5848 Node * narr = new_array(klass_node, zlen, 1);
5849 // Update IdealKit memory and control from graphKit.
5850 __ sync_kit(this);
5851 __ set(z_alloc, narr);
5852 } __ end_if();
5853
5854 sync_kit(ideal);
5855 z = __ value(z_alloc);
5856 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5857 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5858 // Final sync IdealKit and GraphKit.
5859 final_sync(ideal);
5860 #undef __
5861
5862 Node* z_start = array_element_address(z, intcon(0), T_INT);
5863
5864 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5865 OptoRuntime::multiplyToLen_Type(),
5866 stubAddr, stubName, TypePtr::BOTTOM,
5867 x_start, xlen, y_start, ylen, z_start, zlen);
5868 } // original reexecute is set back here
5869
5870 C->set_has_split_ifs(true); // Has chance for split-if optimization
5871 set_result(z);
5872 return true;
5873 }
5874
5875 //-------------inline_squareToLen------------------------------------
5876 bool LibraryCallKit::inline_squareToLen() {
5877 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5878
5879 address stubAddr = StubRoutines::squareToLen();
5880 if (stubAddr == NULL) {
5881 return false; // Intrinsic's stub is not implemented on this platform
5882 }
5883 const char* stubName = "squareToLen";
5884
5885 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5886
5887 Node* x = argument(0);
5888 Node* len = argument(1);
5889 Node* z = argument(2);
5890 Node* zlen = argument(3);
5891
5892 const Type* x_type = x->Value(&_gvn);
5893 const Type* z_type = z->Value(&_gvn);
5894 const TypeAryPtr* top_x = x_type->isa_aryptr();
5895 const TypeAryPtr* top_z = z_type->isa_aryptr();
5896 if (top_x == NULL || top_x->klass() == NULL ||
5897 top_z == NULL || top_z->klass() == NULL) {
5898 // failed array check
5899 return false;
5900 }
5901
5902 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5903 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5904 if (x_elem != T_INT || z_elem != T_INT) {
5905 return false;
5906 }
5907
5908
5909 Node* x_start = array_element_address(x, intcon(0), x_elem);
5910 Node* z_start = array_element_address(z, intcon(0), z_elem);
5911
5912 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5913 OptoRuntime::squareToLen_Type(),
5914 stubAddr, stubName, TypePtr::BOTTOM,
5915 x_start, len, z_start, zlen);
5916
5917 set_result(z);
5918 return true;
5919 }
5920
5921 //-------------inline_mulAdd------------------------------------------
5922 bool LibraryCallKit::inline_mulAdd() {
5923 assert(UseMulAddIntrinsic, "not implementated on this platform");
5924
5925 address stubAddr = StubRoutines::mulAdd();
5926 if (stubAddr == NULL) {
5927 return false; // Intrinsic's stub is not implemented on this platform
5928 }
5929 const char* stubName = "mulAdd";
5930
5931 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5932
5933 Node* out = argument(0);
5934 Node* in = argument(1);
5935 Node* offset = argument(2);
5936 Node* len = argument(3);
5937 Node* k = argument(4);
5938
5939 const Type* out_type = out->Value(&_gvn);
5940 const Type* in_type = in->Value(&_gvn);
5941 const TypeAryPtr* top_out = out_type->isa_aryptr();
5942 const TypeAryPtr* top_in = in_type->isa_aryptr();
5943 if (top_out == NULL || top_out->klass() == NULL ||
5944 top_in == NULL || top_in->klass() == NULL) {
5945 // failed array check
5946 return false;
5947 }
5948
5949 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5950 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5951 if (out_elem != T_INT || in_elem != T_INT) {
5952 return false;
5953 }
5954
5955 Node* outlen = load_array_length(out);
5956 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
5957 Node* out_start = array_element_address(out, intcon(0), out_elem);
5958 Node* in_start = array_element_address(in, intcon(0), in_elem);
5959
5960 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5961 OptoRuntime::mulAdd_Type(),
5962 stubAddr, stubName, TypePtr::BOTTOM,
5963 out_start,in_start, new_offset, len, k);
5964 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5965 set_result(result);
5966 return true;
5967 }
5968
5969
5970 /**
5971 * Calculate CRC32 for byte.
5972 * int java.util.zip.CRC32.update(int crc, int b)
5973 */
5974 bool LibraryCallKit::inline_updateCRC32() {
5975 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5976 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5977 // no receiver since it is static method
5978 Node* crc = argument(0); // type: int
5979 Node* b = argument(1); // type: int
5980
5981 /*
5982 * int c = ~ crc;
5983 * b = timesXtoThe32[(b ^ c) & 0xFF];
5984 * b = b ^ (c >>> 8);
5985 * crc = ~b;
5986 */
5987
5988 Node* M1 = intcon(-1);
5989 crc = _gvn.transform(new (C) XorINode(crc, M1));
5990 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5991 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5992
5993 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5994 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5995 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5996 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5997
5998 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5999 result = _gvn.transform(new (C) XorINode(crc, result));
6000 result = _gvn.transform(new (C) XorINode(result, M1));
6001 set_result(result);
6002 return true;
6003 }
6004
6005 /**
6006 * Calculate CRC32 for byte[] array.
6007 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6008 */
6009 bool LibraryCallKit::inline_updateBytesCRC32() {
6010 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6011 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6012 // no receiver since it is static method
6013 Node* crc = argument(0); // type: int
6014 Node* src = argument(1); // type: oop
6015 Node* offset = argument(2); // type: int
6016 Node* length = argument(3); // type: int
6017
6018 const Type* src_type = src->Value(&_gvn);
6019 const TypeAryPtr* top_src = src_type->isa_aryptr();
6020 if (top_src == NULL || top_src->klass() == NULL) {
6021 // failed array check
6022 return false;
6023 }
6024
6025 // Figure out the size and type of the elements we will be copying.
6026 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6027 if (src_elem != T_BYTE) {
6028 return false;
6029 }
6030
6031 // 'src_start' points to src array + scaled offset
6032 Node* src_start = array_element_address(src, offset, src_elem);
6033
6034 // We assume that range check is done by caller.
6035 // TODO: generate range check (offset+length < src.length) in debug VM.
6036
6037 // Call the stub.
6038 address stubAddr = StubRoutines::updateBytesCRC32();
6039 const char *stubName = "updateBytesCRC32";
6040
6041 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6042 stubAddr, stubName, TypePtr::BOTTOM,
6043 crc, src_start, length);
6044 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6045 set_result(result);
6046 return true;
6047 }
6048
6049 /**
6050 * Calculate CRC32 for ByteBuffer.
6051 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6052 */
6053 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6054 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6055 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6056 // no receiver since it is static method
6057 Node* crc = argument(0); // type: int
6058 Node* src = argument(1); // type: long
6059 Node* offset = argument(3); // type: int
6060 Node* length = argument(4); // type: int
6061
6062 src = ConvL2X(src); // adjust Java long to machine word
6063 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6064 offset = ConvI2X(offset);
6065
6066 // 'src_start' points to src array + scaled offset
6067 Node* src_start = basic_plus_adr(top(), base, offset);
6068
6069 // Call the stub.
6070 address stubAddr = StubRoutines::updateBytesCRC32();
6071 const char *stubName = "updateBytesCRC32";
6072
6073 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6074 stubAddr, stubName, TypePtr::BOTTOM,
6075 crc, src_start, length);
6076 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6077 set_result(result);
6078 return true;
6079 }
6080
6081 //----------------------------inline_reference_get----------------------------
6082 // public T java.lang.ref.Reference.get();
6083 bool LibraryCallKit::inline_reference_get() {
6084 const int referent_offset = java_lang_ref_Reference::referent_offset;
6085 guarantee(referent_offset > 0, "should have already been set");
6086
6087 // Get the argument:
6088 Node* reference_obj = null_check_receiver();
6089 if (stopped()) return true;
6090
6091 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6092
6093 ciInstanceKlass* klass = env()->Object_klass();
6094 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6095
6096 Node* no_ctrl = NULL;
6097 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6098
6099 // Use the pre-barrier to record the value in the referent field
6100 pre_barrier(false /* do_load */,
6101 control(),
6102 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6103 result /* pre_val */,
6104 T_OBJECT);
6105
6106 // Add memory barrier to prevent commoning reads from this field
6107 // across safepoint since GC can change its value.
6108 insert_mem_bar(Op_MemBarCPUOrder);
6109
6110 set_result(result);
6111 return true;
6112 }
6113
6114
6115 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6116 bool is_exact=true, bool is_static=false) {
6117
6118 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6119 assert(tinst != NULL, "obj is null");
6120 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6121 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6122
6123 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6124 ciSymbol::make(fieldTypeString),
6125 is_static);
6126 if (field == NULL) return (Node *) NULL;
6127 assert (field != NULL, "undefined field");
6128
6129 // Next code copied from Parse::do_get_xxx():
6130
6131 // Compute address and memory type.
6132 int offset = field->offset_in_bytes();
6133 bool is_vol = field->is_volatile();
6134 ciType* field_klass = field->type();
6135 assert(field_klass->is_loaded(), "should be loaded");
6136 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6137 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6138 BasicType bt = field->layout_type();
6139
6140 // Build the resultant type of the load
6141 const Type *type;
6142 if (bt == T_OBJECT) {
6143 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6144 } else {
6145 type = Type::get_const_basic_type(bt);
6146 }
6147
6148 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6149 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6150 }
6151 // Build the load.
6152 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6153 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6154 // If reference is volatile, prevent following memory ops from
6155 // floating up past the volatile read. Also prevents commoning
6156 // another volatile read.
6157 if (is_vol) {
6158 // Memory barrier includes bogus read of value to force load BEFORE membar
6159 insert_mem_bar(Op_MemBarAcquire, loadedField);
6160 }
6161 return loadedField;
6162 }
6163
6164
6165 //------------------------------inline_aescrypt_Block-----------------------
6166 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6167 address stubAddr;
6168 const char *stubName;
6169 assert(UseAES, "need AES instruction support");
6170
6171 switch(id) {
6172 case vmIntrinsics::_aescrypt_encryptBlock:
6173 stubAddr = StubRoutines::aescrypt_encryptBlock();
6174 stubName = "aescrypt_encryptBlock";
6175 break;
6176 case vmIntrinsics::_aescrypt_decryptBlock:
6177 stubAddr = StubRoutines::aescrypt_decryptBlock();
6178 stubName = "aescrypt_decryptBlock";
6179 break;
6180 }
6181 if (stubAddr == NULL) return false;
6182
6183 Node* aescrypt_object = argument(0);
6184 Node* src = argument(1);
6185 Node* src_offset = argument(2);
6186 Node* dest = argument(3);
6187 Node* dest_offset = argument(4);
6188
6189 // (1) src and dest are arrays.
6190 const Type* src_type = src->Value(&_gvn);
6191 const Type* dest_type = dest->Value(&_gvn);
6192 const TypeAryPtr* top_src = src_type->isa_aryptr();
6193 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6194 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6195
6196 // for the quick and dirty code we will skip all the checks.
6197 // we are just trying to get the call to be generated.
6198 Node* src_start = src;
6199 Node* dest_start = dest;
6200 if (src_offset != NULL || dest_offset != NULL) {
6201 assert(src_offset != NULL && dest_offset != NULL, "");
6202 src_start = array_element_address(src, src_offset, T_BYTE);
6203 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6204 }
6205
6206 // now need to get the start of its expanded key array
6207 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6208 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6209 if (k_start == NULL) return false;
6210
6211 if (Matcher::pass_original_key_for_aes()) {
6212 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6213 // compatibility issues between Java key expansion and SPARC crypto instructions
6214 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6215 if (original_k_start == NULL) return false;
6216
6217 // Call the stub.
6218 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6219 stubAddr, stubName, TypePtr::BOTTOM,
6220 src_start, dest_start, k_start, original_k_start);
6221 } else {
6222 // Call the stub.
6223 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6224 stubAddr, stubName, TypePtr::BOTTOM,
6225 src_start, dest_start, k_start);
6226 }
6227
6228 return true;
6229 }
6230
6231 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6232 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6233 address stubAddr;
6234 const char *stubName;
6235
6236 assert(UseAES, "need AES instruction support");
6237
6238 switch(id) {
6239 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6240 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6241 stubName = "cipherBlockChaining_encryptAESCrypt";
6242 break;
6243 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6244 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6245 stubName = "cipherBlockChaining_decryptAESCrypt";
6246 break;
6247 }
6248 if (stubAddr == NULL) return false;
6249
6250 Node* cipherBlockChaining_object = argument(0);
6251 Node* src = argument(1);
6252 Node* src_offset = argument(2);
6253 Node* len = argument(3);
6254 Node* dest = argument(4);
6255 Node* dest_offset = argument(5);
6256
6257 // (1) src and dest are arrays.
6258 const Type* src_type = src->Value(&_gvn);
6259 const Type* dest_type = dest->Value(&_gvn);
6260 const TypeAryPtr* top_src = src_type->isa_aryptr();
6261 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6262 assert (top_src != NULL && top_src->klass() != NULL
6263 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6264
6265 // checks are the responsibility of the caller
6266 Node* src_start = src;
6267 Node* dest_start = dest;
6268 if (src_offset != NULL || dest_offset != NULL) {
6269 assert(src_offset != NULL && dest_offset != NULL, "");
6270 src_start = array_element_address(src, src_offset, T_BYTE);
6271 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6272 }
6273
6274 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6275 // (because of the predicated logic executed earlier).
6276 // so we cast it here safely.
6277 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6278
6279 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6280 if (embeddedCipherObj == NULL) return false;
6281
6282 // cast it to what we know it will be at runtime
6283 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6284 assert(tinst != NULL, "CBC obj is null");
6285 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6286 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6287 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6288
6289 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6290 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6291 const TypeOopPtr* xtype = aklass->as_instance_type();
6292 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6293 aescrypt_object = _gvn.transform(aescrypt_object);
6294
6295 // we need to get the start of the aescrypt_object's expanded key array
6296 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6297 if (k_start == NULL) return false;
6298
6299 // similarly, get the start address of the r vector
6300 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6301 if (objRvec == NULL) return false;
6302 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6303
6304 Node* cbcCrypt;
6305 if (Matcher::pass_original_key_for_aes()) {
6306 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6307 // compatibility issues between Java key expansion and SPARC crypto instructions
6308 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6309 if (original_k_start == NULL) return false;
6310
6311 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6312 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6313 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6314 stubAddr, stubName, TypePtr::BOTTOM,
6315 src_start, dest_start, k_start, r_start, len, original_k_start);
6316 } else {
6317 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6318 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6319 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6320 stubAddr, stubName, TypePtr::BOTTOM,
6321 src_start, dest_start, k_start, r_start, len);
6322 }
6323
6324 // return cipher length (int)
6325 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6326 set_result(retvalue);
6327 return true;
6328 }
6329
6330 //------------------------------get_key_start_from_aescrypt_object-----------------------
6331 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6332 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6333 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6334 if (objAESCryptKey == NULL) return (Node *) NULL;
6335
6336 // now have the array, need to get the start address of the K array
6337 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6338 return k_start;
6339 }
6340
6341 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6342 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6343 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6344 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6345 if (objAESCryptKey == NULL) return (Node *) NULL;
6346
6347 // now have the array, need to get the start address of the lastKey array
6348 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6349 return original_k_start;
6350 }
6351
6352 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6353 // Return node representing slow path of predicate check.
6354 // the pseudo code we want to emulate with this predicate is:
6355 // for encryption:
6356 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6357 // for decryption:
6358 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6359 // note cipher==plain is more conservative than the original java code but that's OK
6360 //
6361 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6362 // The receiver was checked for NULL already.
6363 Node* objCBC = argument(0);
6364
6365 // Load embeddedCipher field of CipherBlockChaining object.
6366 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6367
6368 // get AESCrypt klass for instanceOf check
6369 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6370 // will have same classloader as CipherBlockChaining object
6371 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6372 assert(tinst != NULL, "CBCobj is null");
6373 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6374
6375 // we want to do an instanceof comparison against the AESCrypt class
6376 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6377 if (!klass_AESCrypt->is_loaded()) {
6378 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6379 Node* ctrl = control();
6380 set_control(top()); // no regular fast path
6381 return ctrl;
6382 }
6383 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6384
6385 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6386 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6387 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6388
6389 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6390
6391 // for encryption, we are done
6392 if (!decrypting)
6393 return instof_false; // even if it is NULL
6394
6395 // for decryption, we need to add a further check to avoid
6396 // taking the intrinsic path when cipher and plain are the same
6397 // see the original java code for why.
6398 RegionNode* region = new(C) RegionNode(3);
6399 region->init_req(1, instof_false);
6400 Node* src = argument(1);
6401 Node* dest = argument(4);
6402 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6403 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6404 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6405 region->init_req(2, src_dest_conjoint);
6406
6407 record_for_igvn(region);
6408 return _gvn.transform(region);
6409 }
6410
6411 //------------------------------inline_sha_implCompress-----------------------
6412 //
6413 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6414 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6415 //
6416 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6417 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6418 //
6419 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6420 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6421 //
6422 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6423 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6424
6425 Node* sha_obj = argument(0);
6426 Node* src = argument(1); // type oop
6427 Node* ofs = argument(2); // type int
6428
6429 const Type* src_type = src->Value(&_gvn);
6430 const TypeAryPtr* top_src = src_type->isa_aryptr();
6431 if (top_src == NULL || top_src->klass() == NULL) {
6432 // failed array check
6433 return false;
6434 }
6435 // Figure out the size and type of the elements we will be copying.
6436 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6437 if (src_elem != T_BYTE) {
6438 return false;
6439 }
6440 // 'src_start' points to src array + offset
6441 Node* src_start = array_element_address(src, ofs, src_elem);
6442 Node* state = NULL;
6443 address stubAddr;
6444 const char *stubName;
6445
6446 switch(id) {
6447 case vmIntrinsics::_sha_implCompress:
6448 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6449 state = get_state_from_sha_object(sha_obj);
6450 stubAddr = StubRoutines::sha1_implCompress();
6451 stubName = "sha1_implCompress";
6452 break;
6453 case vmIntrinsics::_sha2_implCompress:
6454 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6455 state = get_state_from_sha_object(sha_obj);
6456 stubAddr = StubRoutines::sha256_implCompress();
6457 stubName = "sha256_implCompress";
6458 break;
6459 case vmIntrinsics::_sha5_implCompress:
6460 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6461 state = get_state_from_sha5_object(sha_obj);
6462 stubAddr = StubRoutines::sha512_implCompress();
6463 stubName = "sha512_implCompress";
6464 break;
6465 default:
6466 fatal_unexpected_iid(id);
6467 return false;
6468 }
6469 if (state == NULL) return false;
6470
6471 // Call the stub.
6472 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6473 stubAddr, stubName, TypePtr::BOTTOM,
6474 src_start, state);
6475
6476 return true;
6477 }
6478
6479 //------------------------------inline_digestBase_implCompressMB-----------------------
6480 //
6481 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6482 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6483 //
6484 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6485 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6486 "need SHA1/SHA256/SHA512 instruction support");
6487 assert((uint)predicate < 3, "sanity");
6488 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6489
6490 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6491 Node* src = argument(1); // byte[] array
6492 Node* ofs = argument(2); // type int
6493 Node* limit = argument(3); // type int
6494
6495 const Type* src_type = src->Value(&_gvn);
6496 const TypeAryPtr* top_src = src_type->isa_aryptr();
6497 if (top_src == NULL || top_src->klass() == NULL) {
6498 // failed array check
6499 return false;
6500 }
6501 // Figure out the size and type of the elements we will be copying.
6502 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6503 if (src_elem != T_BYTE) {
6504 return false;
6505 }
6506 // 'src_start' points to src array + offset
6507 Node* src_start = array_element_address(src, ofs, src_elem);
6508
6509 const char* klass_SHA_name = NULL;
6510 const char* stub_name = NULL;
6511 address stub_addr = NULL;
6512 bool long_state = false;
6513
6514 switch (predicate) {
6515 case 0:
6516 if (UseSHA1Intrinsics) {
6517 klass_SHA_name = "sun/security/provider/SHA";
6518 stub_name = "sha1_implCompressMB";
6519 stub_addr = StubRoutines::sha1_implCompressMB();
6520 }
6521 break;
6522 case 1:
6523 if (UseSHA256Intrinsics) {
6524 klass_SHA_name = "sun/security/provider/SHA2";
6525 stub_name = "sha256_implCompressMB";
6526 stub_addr = StubRoutines::sha256_implCompressMB();
6527 }
6528 break;
6529 case 2:
6530 if (UseSHA512Intrinsics) {
6531 klass_SHA_name = "sun/security/provider/SHA5";
6532 stub_name = "sha512_implCompressMB";
6533 stub_addr = StubRoutines::sha512_implCompressMB();
6534 long_state = true;
6535 }
6536 break;
6537 default:
6538 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6539 }
6540 if (klass_SHA_name != NULL) {
6541 // get DigestBase klass to lookup for SHA klass
6542 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6543 assert(tinst != NULL, "digestBase_obj is not instance???");
6544 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6545
6546 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6547 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6548 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6549 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6550 }
6551 return false;
6552 }
6553 //------------------------------inline_sha_implCompressMB-----------------------
6554 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6555 bool long_state, address stubAddr, const char *stubName,
6556 Node* src_start, Node* ofs, Node* limit) {
6557 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6558 const TypeOopPtr* xtype = aklass->as_instance_type();
6559 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6560 sha_obj = _gvn.transform(sha_obj);
6561
6562 Node* state;
6563 if (long_state) {
6564 state = get_state_from_sha5_object(sha_obj);
6565 } else {
6566 state = get_state_from_sha_object(sha_obj);
6567 }
6568 if (state == NULL) return false;
6569
6570 // Call the stub.
6571 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6572 OptoRuntime::digestBase_implCompressMB_Type(),
6573 stubAddr, stubName, TypePtr::BOTTOM,
6574 src_start, state, ofs, limit);
6575 // return ofs (int)
6576 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6577 set_result(result);
6578
6579 return true;
6580 }
6581
6582 //------------------------------get_state_from_sha_object-----------------------
6583 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6584 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6585 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6586 if (sha_state == NULL) return (Node *) NULL;
6587
6588 // now have the array, need to get the start address of the state array
6589 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6590 return state;
6591 }
6592
6593 //------------------------------get_state_from_sha5_object-----------------------
6594 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6595 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6596 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6597 if (sha_state == NULL) return (Node *) NULL;
6598
6599 // now have the array, need to get the start address of the state array
6600 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6601 return state;
6602 }
6603
6604 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6605 // Return node representing slow path of predicate check.
6606 // the pseudo code we want to emulate with this predicate is:
6607 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6608 //
6609 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6610 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6611 "need SHA1/SHA256/SHA512 instruction support");
6612 assert((uint)predicate < 3, "sanity");
6613
6614 // The receiver was checked for NULL already.
6615 Node* digestBaseObj = argument(0);
6616
6617 // get DigestBase klass for instanceOf check
6618 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6619 assert(tinst != NULL, "digestBaseObj is null");
6620 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6621
6622 const char* klass_SHA_name = NULL;
6623 switch (predicate) {
6624 case 0:
6625 if (UseSHA1Intrinsics) {
6626 // we want to do an instanceof comparison against the SHA class
6627 klass_SHA_name = "sun/security/provider/SHA";
6628 }
6629 break;
6630 case 1:
6631 if (UseSHA256Intrinsics) {
6632 // we want to do an instanceof comparison against the SHA2 class
6633 klass_SHA_name = "sun/security/provider/SHA2";
6634 }
6635 break;
6636 case 2:
6637 if (UseSHA512Intrinsics) {
6638 // we want to do an instanceof comparison against the SHA5 class
6639 klass_SHA_name = "sun/security/provider/SHA5";
6640 }
6641 break;
6642 default:
6643 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6644 }
6645
6646 ciKlass* klass_SHA = NULL;
6647 if (klass_SHA_name != NULL) {
6648 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6649 }
6650 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6651 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6652 Node* ctrl = control();
6653 set_control(top()); // no intrinsic path
6654 return ctrl;
6655 }
6656 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6657
6658 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6659 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6660 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6661 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6662
6663 return instof_false; // even if it is NULL
6664 }
6665
6666 bool LibraryCallKit::inline_profileBoolean() {
6667 Node* counts = argument(1);
6668 const TypeAryPtr* ary = NULL;
6669 ciArray* aobj = NULL;
6670 if (counts->is_Con()
6671 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6672 && (aobj = ary->const_oop()->as_array()) != NULL
6673 && (aobj->length() == 2)) {
6674 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6675 jint false_cnt = aobj->element_value(0).as_int();
6676 jint true_cnt = aobj->element_value(1).as_int();
6677
6678 if (C->log() != NULL) {
6679 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6680 false_cnt, true_cnt);
6681 }
6682
6683 if (false_cnt + true_cnt == 0) {
6684 // According to profile, never executed.
6685 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6686 Deoptimization::Action_reinterpret);
6687 return true;
6688 }
6689
6690 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6691 // is a number of each value occurrences.
6692 Node* result = argument(0);
6693 if (false_cnt == 0 || true_cnt == 0) {
6694 // According to profile, one value has been never seen.
6695 int expected_val = (false_cnt == 0) ? 1 : 0;
6696
6697 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6698 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6699
6700 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6701 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6702 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6703
6704 { // Slow path: uncommon trap for never seen value and then reexecute
6705 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6706 // the value has been seen at least once.
6707 PreserveJVMState pjvms(this);
6708 PreserveReexecuteState preexecs(this);
6709 jvms()->set_should_reexecute(true);
6710
6711 set_control(slow_path);
6712 set_i_o(i_o());
6713
6714 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6715 Deoptimization::Action_reinterpret);
6716 }
6717 // The guard for never seen value enables sharpening of the result and
6718 // returning a constant. It allows to eliminate branches on the same value
6719 // later on.
6720 set_control(fast_path);
6721 result = intcon(expected_val);
6722 }
6723 // Stop profiling.
6724 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6725 // By replacing method body with profile data (represented as ProfileBooleanNode
6726 // on IR level) we effectively disable profiling.
6727 // It enables full speed execution once optimized code is generated.
6728 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6729 C->record_for_igvn(profile);
6730 set_result(profile);
6731 return true;
6732 } else {
6733 // Continue profiling.
6734 // Profile data isn't available at the moment. So, execute method's bytecode version.
6735 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6736 // is compiled and counters aren't available since corresponding MethodHandle
6737 // isn't a compile-time constant.
6738 return false;
6739 }
6740 }