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, bool is_unaligned);
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 bool inline_montgomeryMultiply();
330 bool inline_montgomerySquare();
331
332 bool inline_profileBoolean();
333 };
334
335
336 //---------------------------make_vm_intrinsic----------------------------
337 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
338 vmIntrinsics::ID id = m->intrinsic_id();
339 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
340
341 ccstr disable_intr = NULL;
342
343 if ((DisableIntrinsic[0] != '\0'
344 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
345 (method_has_option_value("DisableIntrinsic", disable_intr)
346 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
347 // disabled by a user request on the command line:
348 // example: -XX:DisableIntrinsic=_hashCode,_getClass
349 return NULL;
350 }
351
352 if (!m->is_loaded()) {
353 // do not attempt to inline unloaded methods
354 return NULL;
355 }
356
357 // Only a few intrinsics implement a virtual dispatch.
358 // They are expensive calls which are also frequently overridden.
359 if (is_virtual) {
360 switch (id) {
361 case vmIntrinsics::_hashCode:
362 case vmIntrinsics::_clone:
363 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
364 break;
365 default:
366 return NULL;
367 }
368 }
369
370 // -XX:-InlineNatives disables nearly all intrinsics:
371 if (!InlineNatives) {
372 switch (id) {
373 case vmIntrinsics::_indexOf:
374 case vmIntrinsics::_compareTo:
375 case vmIntrinsics::_equals:
376 case vmIntrinsics::_equalsC:
377 case vmIntrinsics::_getAndAddInt:
378 case vmIntrinsics::_getAndAddLong:
379 case vmIntrinsics::_getAndSetInt:
380 case vmIntrinsics::_getAndSetLong:
381 case vmIntrinsics::_getAndSetObject:
382 case vmIntrinsics::_loadFence:
383 case vmIntrinsics::_storeFence:
384 case vmIntrinsics::_fullFence:
385 break; // InlineNatives does not control String.compareTo
386 case vmIntrinsics::_Reference_get:
387 break; // InlineNatives does not control Reference.get
388 default:
389 return NULL;
390 }
391 }
392
393 int predicates = 0;
394 bool does_virtual_dispatch = false;
395
396 switch (id) {
397 case vmIntrinsics::_compareTo:
398 if (!SpecialStringCompareTo) return NULL;
399 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
400 break;
401 case vmIntrinsics::_indexOf:
402 if (!SpecialStringIndexOf) return NULL;
403 break;
404 case vmIntrinsics::_equals:
405 if (!SpecialStringEquals) return NULL;
406 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
407 break;
408 case vmIntrinsics::_equalsC:
409 if (!SpecialArraysEquals) return NULL;
410 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
411 break;
412 case vmIntrinsics::_arraycopy:
413 if (!InlineArrayCopy) return NULL;
414 break;
415 case vmIntrinsics::_copyMemory:
416 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
417 if (!InlineArrayCopy) return NULL;
418 break;
419 case vmIntrinsics::_hashCode:
420 if (!InlineObjectHash) return NULL;
421 does_virtual_dispatch = true;
422 break;
423 case vmIntrinsics::_clone:
424 does_virtual_dispatch = true;
425 case vmIntrinsics::_copyOf:
426 case vmIntrinsics::_copyOfRange:
427 if (!InlineObjectCopy) return NULL;
428 // These also use the arraycopy intrinsic mechanism:
429 if (!InlineArrayCopy) return NULL;
430 break;
431 case vmIntrinsics::_encodeISOArray:
432 if (!SpecialEncodeISOArray) return NULL;
433 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
434 break;
435 case vmIntrinsics::_checkIndex:
436 // We do not intrinsify this. The optimizer does fine with it.
437 return NULL;
438
439 case vmIntrinsics::_getCallerClass:
440 if (!UseNewReflection) return NULL;
441 if (!InlineReflectionGetCallerClass) return NULL;
442 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
443 break;
444
445 case vmIntrinsics::_bitCount_i:
446 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
447 break;
448
449 case vmIntrinsics::_bitCount_l:
450 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
451 break;
452
453 case vmIntrinsics::_numberOfLeadingZeros_i:
454 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
455 break;
456
457 case vmIntrinsics::_numberOfLeadingZeros_l:
458 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
459 break;
460
461 case vmIntrinsics::_numberOfTrailingZeros_i:
462 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
463 break;
464
465 case vmIntrinsics::_numberOfTrailingZeros_l:
466 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
467 break;
468
469 case vmIntrinsics::_reverseBytes_c:
470 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
471 break;
472 case vmIntrinsics::_reverseBytes_s:
473 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
474 break;
475 case vmIntrinsics::_reverseBytes_i:
476 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
477 break;
478 case vmIntrinsics::_reverseBytes_l:
479 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
480 break;
481
482 case vmIntrinsics::_Reference_get:
483 // Use the intrinsic version of Reference.get() so that the value in
484 // the referent field can be registered by the G1 pre-barrier code.
485 // Also add memory barrier to prevent commoning reads from this field
486 // across safepoint since GC can change it value.
487 break;
488
489 case vmIntrinsics::_compareAndSwapObject:
490 #ifdef _LP64
491 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
492 #endif
493 break;
494
495 case vmIntrinsics::_compareAndSwapLong:
496 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
497 break;
498
499 case vmIntrinsics::_getAndAddInt:
500 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
501 break;
502
503 case vmIntrinsics::_getAndAddLong:
504 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
505 break;
506
507 case vmIntrinsics::_getAndSetInt:
508 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
509 break;
510
511 case vmIntrinsics::_getAndSetLong:
512 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
513 break;
514
515 case vmIntrinsics::_getAndSetObject:
516 #ifdef _LP64
517 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
518 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
519 break;
520 #else
521 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
522 break;
523 #endif
524
525 case vmIntrinsics::_aescrypt_encryptBlock:
526 case vmIntrinsics::_aescrypt_decryptBlock:
527 if (!UseAESIntrinsics) return NULL;
528 break;
529
530 case vmIntrinsics::_multiplyToLen:
531 if (!UseMultiplyToLenIntrinsic) return NULL;
532 break;
533
534 case vmIntrinsics::_squareToLen:
535 if (!UseSquareToLenIntrinsic) return NULL;
536 break;
537
538 case vmIntrinsics::_mulAdd:
539 if (!UseMulAddIntrinsic) return NULL;
540 break;
541
542 case vmIntrinsics::_montgomeryMultiply:
543 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
544 break;
545 case vmIntrinsics::_montgomerySquare:
546 if (!UseMontgomerySquareIntrinsic) return NULL;
547 break;
548
549 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
550 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
551 if (!UseAESIntrinsics) return NULL;
552 // these two require the predicated logic
553 predicates = 1;
554 break;
555
556 case vmIntrinsics::_sha_implCompress:
557 if (!UseSHA1Intrinsics) return NULL;
558 break;
559
560 case vmIntrinsics::_sha2_implCompress:
561 if (!UseSHA256Intrinsics) return NULL;
562 break;
563
564 case vmIntrinsics::_sha5_implCompress:
565 if (!UseSHA512Intrinsics) return NULL;
566 break;
567
568 case vmIntrinsics::_digestBase_implCompressMB:
569 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
570 predicates = 3;
571 break;
572
573 case vmIntrinsics::_updateCRC32:
574 case vmIntrinsics::_updateBytesCRC32:
575 case vmIntrinsics::_updateByteBufferCRC32:
576 if (!UseCRC32Intrinsics) return NULL;
577 break;
578
579 case vmIntrinsics::_incrementExactI:
580 case vmIntrinsics::_addExactI:
581 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
582 break;
583 case vmIntrinsics::_incrementExactL:
584 case vmIntrinsics::_addExactL:
585 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
586 break;
587 case vmIntrinsics::_decrementExactI:
588 case vmIntrinsics::_subtractExactI:
589 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
590 break;
591 case vmIntrinsics::_decrementExactL:
592 case vmIntrinsics::_subtractExactL:
593 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
594 break;
595 case vmIntrinsics::_negateExactI:
596 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
597 break;
598 case vmIntrinsics::_negateExactL:
599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
600 break;
601 case vmIntrinsics::_multiplyExactI:
602 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
603 break;
604 case vmIntrinsics::_multiplyExactL:
605 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
606 break;
607
608 default:
609 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
610 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
611 break;
612 }
613
614 // -XX:-InlineClassNatives disables natives from the Class class.
615 // The flag applies to all reflective calls, notably Array.newArray
616 // (visible to Java programmers as Array.newInstance).
617 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
618 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
619 if (!InlineClassNatives) return NULL;
620 }
621
622 // -XX:-InlineThreadNatives disables natives from the Thread class.
623 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
624 if (!InlineThreadNatives) return NULL;
625 }
626
627 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
628 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
629 m->holder()->name() == ciSymbol::java_lang_Float() ||
630 m->holder()->name() == ciSymbol::java_lang_Double()) {
631 if (!InlineMathNatives) return NULL;
632 }
633
634 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
635 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
636 if (!InlineUnsafeOps) return NULL;
637 }
638
639 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
640 }
641
642 //----------------------register_library_intrinsics-----------------------
643 // Initialize this file's data structures, for each Compile instance.
644 void Compile::register_library_intrinsics() {
645 // Nothing to do here.
646 }
647
648 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
649 LibraryCallKit kit(jvms, this);
650 Compile* C = kit.C;
651 int nodes = C->unique();
652 #ifndef PRODUCT
653 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
654 char buf[1000];
655 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
656 tty->print_cr("Intrinsic %s", str);
657 }
658 #endif
659 ciMethod* callee = kit.callee();
660 const int bci = kit.bci();
661
662 // Try to inline the intrinsic.
663 if (kit.try_to_inline(_last_predicate)) {
664 if (C->print_intrinsics() || C->print_inlining()) {
665 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
666 }
667 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
668 if (C->log()) {
669 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
670 vmIntrinsics::name_at(intrinsic_id()),
671 (is_virtual() ? " virtual='1'" : ""),
672 C->unique() - nodes);
673 }
674 // Push the result from the inlined method onto the stack.
675 kit.push_result();
676 return kit.transfer_exceptions_into_jvms();
677 }
678
679 // The intrinsic bailed out
680 if (C->print_intrinsics() || C->print_inlining()) {
681 if (jvms->has_method()) {
682 // Not a root compile.
683 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
684 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
685 } else {
686 // Root compile
687 tty->print("Did not generate intrinsic %s%s at bci:%d in",
688 vmIntrinsics::name_at(intrinsic_id()),
689 (is_virtual() ? " (virtual)" : ""), bci);
690 }
691 }
692 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
693 return NULL;
694 }
695
696 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
697 LibraryCallKit kit(jvms, this);
698 Compile* C = kit.C;
699 int nodes = C->unique();
700 _last_predicate = predicate;
701 #ifndef PRODUCT
702 assert(is_predicated() && predicate < predicates_count(), "sanity");
703 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
704 char buf[1000];
705 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
706 tty->print_cr("Predicate for intrinsic %s", str);
707 }
708 #endif
709 ciMethod* callee = kit.callee();
710 const int bci = kit.bci();
711
712 Node* slow_ctl = kit.try_to_predicate(predicate);
713 if (!kit.failing()) {
714 if (C->print_intrinsics() || C->print_inlining()) {
715 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
716 }
717 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
718 if (C->log()) {
719 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
720 vmIntrinsics::name_at(intrinsic_id()),
721 (is_virtual() ? " virtual='1'" : ""),
722 C->unique() - nodes);
723 }
724 return slow_ctl; // Could be NULL if the check folds.
725 }
726
727 // The intrinsic bailed out
728 if (C->print_intrinsics() || C->print_inlining()) {
729 if (jvms->has_method()) {
730 // Not a root compile.
731 const char* msg = "failed to generate predicate for intrinsic";
732 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
733 } else {
734 // Root compile
735 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
736 vmIntrinsics::name_at(intrinsic_id()),
737 (is_virtual() ? " (virtual)" : ""), bci);
738 }
739 }
740 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
741 return NULL;
742 }
743
744 bool LibraryCallKit::try_to_inline(int predicate) {
745 // Handle symbolic names for otherwise undistinguished boolean switches:
746 const bool is_store = true;
747 const bool is_native_ptr = true;
748 const bool is_static = true;
749 const bool is_volatile = true;
750
751 if (!jvms()->has_method()) {
752 // Root JVMState has a null method.
753 assert(map()->memory()->Opcode() == Op_Parm, "");
754 // Insert the memory aliasing node
755 set_all_memory(reset_memory());
756 }
757 assert(merged_memory(), "");
758
759
760 switch (intrinsic_id()) {
761 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
762 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
763 case vmIntrinsics::_getClass: return inline_native_getClass();
764
765 case vmIntrinsics::_dsin:
766 case vmIntrinsics::_dcos:
767 case vmIntrinsics::_dtan:
768 case vmIntrinsics::_dabs:
769 case vmIntrinsics::_datan2:
770 case vmIntrinsics::_dsqrt:
771 case vmIntrinsics::_dexp:
772 case vmIntrinsics::_dlog:
773 case vmIntrinsics::_dlog10:
774 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
775
776 case vmIntrinsics::_min:
777 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
778
779 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
780 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
781 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
782 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
783 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
784 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
785 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
786 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
787 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
788 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
789 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
790 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
791
792 case vmIntrinsics::_arraycopy: return inline_arraycopy();
793
794 case vmIntrinsics::_compareTo: return inline_string_compareTo();
795 case vmIntrinsics::_indexOf: return inline_string_indexOf();
796 case vmIntrinsics::_equals: return inline_string_equals();
797
798 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
799 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
800 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
801 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
802 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
803 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
804 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
805 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
806 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
807
808 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
809 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
810 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
811 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
812 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
813 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
814 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
815 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
816 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
817
818 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
819 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
820 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
821 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
822 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
823 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
824 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
825 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
826
827 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
828 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
829 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
830 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
831 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
832 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
833 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
834 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
835
836 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
837 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
838 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
839 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
840 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
841 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
842 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
843 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
844 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
845
846 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
847 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
848 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
849 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
850 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
851 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
852 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
853 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
854 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
855
856 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
857 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
858 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
859 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
860
861 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
862 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
863 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
864
865 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
866 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
867 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
868
869 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
870 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
871 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
872 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
873 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
874
875 case vmIntrinsics::_loadFence:
876 case vmIntrinsics::_storeFence:
877 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
878
879 case vmIntrinsics::_currentThread: return inline_native_currentThread();
880 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
881
882 #ifdef TRACE_HAVE_INTRINSICS
883 case vmIntrinsics::_classID: return inline_native_classID();
884 case vmIntrinsics::_threadID: return inline_native_threadID();
885 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
886 #endif
887 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
888 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
889 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
890 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
891 case vmIntrinsics::_newArray: return inline_native_newArray();
892 case vmIntrinsics::_getLength: return inline_native_getLength();
893 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
894 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
895 case vmIntrinsics::_equalsC: return inline_array_equals();
896 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
897
898 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
899
900 case vmIntrinsics::_isInstance:
901 case vmIntrinsics::_getModifiers:
902 case vmIntrinsics::_isInterface:
903 case vmIntrinsics::_isArray:
904 case vmIntrinsics::_isPrimitive:
905 case vmIntrinsics::_getSuperclass:
906 case vmIntrinsics::_getComponentType:
907 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
908
909 case vmIntrinsics::_floatToRawIntBits:
910 case vmIntrinsics::_floatToIntBits:
911 case vmIntrinsics::_intBitsToFloat:
912 case vmIntrinsics::_doubleToRawLongBits:
913 case vmIntrinsics::_doubleToLongBits:
914 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
915
916 case vmIntrinsics::_numberOfLeadingZeros_i:
917 case vmIntrinsics::_numberOfLeadingZeros_l:
918 case vmIntrinsics::_numberOfTrailingZeros_i:
919 case vmIntrinsics::_numberOfTrailingZeros_l:
920 case vmIntrinsics::_bitCount_i:
921 case vmIntrinsics::_bitCount_l:
922 case vmIntrinsics::_reverseBytes_i:
923 case vmIntrinsics::_reverseBytes_l:
924 case vmIntrinsics::_reverseBytes_s:
925 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
926
927 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
928
929 case vmIntrinsics::_Reference_get: return inline_reference_get();
930
931 case vmIntrinsics::_aescrypt_encryptBlock:
932 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
933
934 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
935 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
936 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
937
938 case vmIntrinsics::_sha_implCompress:
939 case vmIntrinsics::_sha2_implCompress:
940 case vmIntrinsics::_sha5_implCompress:
941 return inline_sha_implCompress(intrinsic_id());
942
943 case vmIntrinsics::_digestBase_implCompressMB:
944 return inline_digestBase_implCompressMB(predicate);
945
946 case vmIntrinsics::_multiplyToLen:
947 return inline_multiplyToLen();
948
949 case vmIntrinsics::_squareToLen:
950 return inline_squareToLen();
951
952 case vmIntrinsics::_mulAdd:
953 return inline_mulAdd();
954
955 case vmIntrinsics::_montgomeryMultiply:
956 return inline_montgomeryMultiply();
957 case vmIntrinsics::_montgomerySquare:
958 return inline_montgomerySquare();
959
960 case vmIntrinsics::_encodeISOArray:
961 return inline_encodeISOArray();
962
963 case vmIntrinsics::_updateCRC32:
964 return inline_updateCRC32();
965 case vmIntrinsics::_updateBytesCRC32:
966 return inline_updateBytesCRC32();
967 case vmIntrinsics::_updateByteBufferCRC32:
968 return inline_updateByteBufferCRC32();
969
970 case vmIntrinsics::_profileBoolean:
971 return inline_profileBoolean();
972
973 default:
974 // If you get here, it may be that someone has added a new intrinsic
975 // to the list in vmSymbols.hpp without implementing it here.
976 #ifndef PRODUCT
977 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
978 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
979 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
980 }
981 #endif
982 return false;
983 }
984 }
985
986 Node* LibraryCallKit::try_to_predicate(int predicate) {
987 if (!jvms()->has_method()) {
988 // Root JVMState has a null method.
989 assert(map()->memory()->Opcode() == Op_Parm, "");
990 // Insert the memory aliasing node
991 set_all_memory(reset_memory());
992 }
993 assert(merged_memory(), "");
994
995 switch (intrinsic_id()) {
996 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
997 return inline_cipherBlockChaining_AESCrypt_predicate(false);
998 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
999 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1000 case vmIntrinsics::_digestBase_implCompressMB:
1001 return inline_digestBase_implCompressMB_predicate(predicate);
1002
1003 default:
1004 // If you get here, it may be that someone has added a new intrinsic
1005 // to the list in vmSymbols.hpp without implementing it here.
1006 #ifndef PRODUCT
1007 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1008 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1009 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1010 }
1011 #endif
1012 Node* slow_ctl = control();
1013 set_control(top()); // No fast path instrinsic
1014 return slow_ctl;
1015 }
1016 }
1017
1018 //------------------------------set_result-------------------------------
1019 // Helper function for finishing intrinsics.
1020 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1021 record_for_igvn(region);
1022 set_control(_gvn.transform(region));
1023 set_result( _gvn.transform(value));
1024 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1025 }
1026
1027 //------------------------------generate_guard---------------------------
1028 // Helper function for generating guarded fast-slow graph structures.
1029 // The given 'test', if true, guards a slow path. If the test fails
1030 // then a fast path can be taken. (We generally hope it fails.)
1031 // In all cases, GraphKit::control() is updated to the fast path.
1032 // The returned value represents the control for the slow path.
1033 // The return value is never 'top'; it is either a valid control
1034 // or NULL if it is obvious that the slow path can never be taken.
1035 // Also, if region and the slow control are not NULL, the slow edge
1036 // is appended to the region.
1037 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1038 if (stopped()) {
1039 // Already short circuited.
1040 return NULL;
1041 }
1042
1043 // Build an if node and its projections.
1044 // If test is true we take the slow path, which we assume is uncommon.
1045 if (_gvn.type(test) == TypeInt::ZERO) {
1046 // The slow branch is never taken. No need to build this guard.
1047 return NULL;
1048 }
1049
1050 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1051
1052 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1053 if (if_slow == top()) {
1054 // The slow branch is never taken. No need to build this guard.
1055 return NULL;
1056 }
1057
1058 if (region != NULL)
1059 region->add_req(if_slow);
1060
1061 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1062 set_control(if_fast);
1063
1064 return if_slow;
1065 }
1066
1067 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1068 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1069 }
1070 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1071 return generate_guard(test, region, PROB_FAIR);
1072 }
1073
1074 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1075 Node* *pos_index) {
1076 if (stopped())
1077 return NULL; // already stopped
1078 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1079 return NULL; // index is already adequately typed
1080 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1081 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1082 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1083 if (is_neg != NULL && pos_index != NULL) {
1084 // Emulate effect of Parse::adjust_map_after_if.
1085 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1086 ccast->set_req(0, control());
1087 (*pos_index) = _gvn.transform(ccast);
1088 }
1089 return is_neg;
1090 }
1091
1092 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1093 Node* *pos_index) {
1094 if (stopped())
1095 return NULL; // already stopped
1096 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1097 return NULL; // index is already adequately typed
1098 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1099 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1100 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1101 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1102 if (is_notp != NULL && pos_index != NULL) {
1103 // Emulate effect of Parse::adjust_map_after_if.
1104 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1105 ccast->set_req(0, control());
1106 (*pos_index) = _gvn.transform(ccast);
1107 }
1108 return is_notp;
1109 }
1110
1111 // Make sure that 'position' is a valid limit index, in [0..length].
1112 // There are two equivalent plans for checking this:
1113 // A. (offset + copyLength) unsigned<= arrayLength
1114 // B. offset <= (arrayLength - copyLength)
1115 // We require that all of the values above, except for the sum and
1116 // difference, are already known to be non-negative.
1117 // Plan A is robust in the face of overflow, if offset and copyLength
1118 // are both hugely positive.
1119 //
1120 // Plan B is less direct and intuitive, but it does not overflow at
1121 // all, since the difference of two non-negatives is always
1122 // representable. Whenever Java methods must perform the equivalent
1123 // check they generally use Plan B instead of Plan A.
1124 // For the moment we use Plan A.
1125 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1126 Node* subseq_length,
1127 Node* array_length,
1128 RegionNode* region) {
1129 if (stopped())
1130 return NULL; // already stopped
1131 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1132 if (zero_offset && subseq_length->eqv_uncast(array_length))
1133 return NULL; // common case of whole-array copy
1134 Node* last = subseq_length;
1135 if (!zero_offset) // last += offset
1136 last = _gvn.transform(new (C) AddINode(last, offset));
1137 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1138 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1139 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1140 return is_over;
1141 }
1142
1143
1144 //--------------------------generate_current_thread--------------------
1145 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1146 ciKlass* thread_klass = env()->Thread_klass();
1147 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1148 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1149 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1150 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1151 tls_output = thread;
1152 return threadObj;
1153 }
1154
1155
1156 //------------------------------make_string_method_node------------------------
1157 // Helper method for String intrinsic functions. This version is called
1158 // with str1 and str2 pointing to String object nodes.
1159 //
1160 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1161 Node* no_ctrl = NULL;
1162
1163 // Get start addr of string
1164 Node* str1_value = load_String_value(no_ctrl, str1);
1165 Node* str1_offset = load_String_offset(no_ctrl, str1);
1166 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1167
1168 // Get length of string 1
1169 Node* str1_len = load_String_length(no_ctrl, str1);
1170
1171 Node* str2_value = load_String_value(no_ctrl, str2);
1172 Node* str2_offset = load_String_offset(no_ctrl, str2);
1173 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1174
1175 Node* str2_len = NULL;
1176 Node* result = NULL;
1177
1178 switch (opcode) {
1179 case Op_StrIndexOf:
1180 // Get length of string 2
1181 str2_len = load_String_length(no_ctrl, str2);
1182
1183 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1184 str1_start, str1_len, str2_start, str2_len);
1185 break;
1186 case Op_StrComp:
1187 // Get length of string 2
1188 str2_len = load_String_length(no_ctrl, str2);
1189
1190 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1191 str1_start, str1_len, str2_start, str2_len);
1192 break;
1193 case Op_StrEquals:
1194 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1195 str1_start, str2_start, str1_len);
1196 break;
1197 default:
1198 ShouldNotReachHere();
1199 return NULL;
1200 }
1201
1202 // All these intrinsics have checks.
1203 C->set_has_split_ifs(true); // Has chance for split-if optimization
1204
1205 return _gvn.transform(result);
1206 }
1207
1208 // Helper method for String intrinsic functions. This version is called
1209 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1210 // to Int nodes containing the lenghts of str1 and str2.
1211 //
1212 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1213 Node* result = NULL;
1214 switch (opcode) {
1215 case Op_StrIndexOf:
1216 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1217 str1_start, cnt1, str2_start, cnt2);
1218 break;
1219 case Op_StrComp:
1220 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1221 str1_start, cnt1, str2_start, cnt2);
1222 break;
1223 case Op_StrEquals:
1224 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1225 str1_start, str2_start, cnt1);
1226 break;
1227 default:
1228 ShouldNotReachHere();
1229 return NULL;
1230 }
1231
1232 // All these intrinsics have checks.
1233 C->set_has_split_ifs(true); // Has chance for split-if optimization
1234
1235 return _gvn.transform(result);
1236 }
1237
1238 //------------------------------inline_string_compareTo------------------------
1239 // public int java.lang.String.compareTo(String anotherString);
1240 bool LibraryCallKit::inline_string_compareTo() {
1241 Node* receiver = null_check(argument(0));
1242 Node* arg = null_check(argument(1));
1243 if (stopped()) {
1244 return true;
1245 }
1246 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1247 return true;
1248 }
1249
1250 //------------------------------inline_string_equals------------------------
1251 bool LibraryCallKit::inline_string_equals() {
1252 Node* receiver = null_check_receiver();
1253 // NOTE: Do not null check argument for String.equals() because spec
1254 // allows to specify NULL as argument.
1255 Node* argument = this->argument(1);
1256 if (stopped()) {
1257 return true;
1258 }
1259
1260 // paths (plus control) merge
1261 RegionNode* region = new (C) RegionNode(5);
1262 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1263
1264 // does source == target string?
1265 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1266 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1267
1268 Node* if_eq = generate_slow_guard(bol, NULL);
1269 if (if_eq != NULL) {
1270 // receiver == argument
1271 phi->init_req(2, intcon(1));
1272 region->init_req(2, if_eq);
1273 }
1274
1275 // get String klass for instanceOf
1276 ciInstanceKlass* klass = env()->String_klass();
1277
1278 if (!stopped()) {
1279 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1280 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1281 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1282
1283 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1284 //instanceOf == true, fallthrough
1285
1286 if (inst_false != NULL) {
1287 phi->init_req(3, intcon(0));
1288 region->init_req(3, inst_false);
1289 }
1290 }
1291
1292 if (!stopped()) {
1293 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1294
1295 // Properly cast the argument to String
1296 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1297 // This path is taken only when argument's type is String:NotNull.
1298 argument = cast_not_null(argument, false);
1299
1300 Node* no_ctrl = NULL;
1301
1302 // Get start addr of receiver
1303 Node* receiver_val = load_String_value(no_ctrl, receiver);
1304 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1305 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1306
1307 // Get length of receiver
1308 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1309
1310 // Get start addr of argument
1311 Node* argument_val = load_String_value(no_ctrl, argument);
1312 Node* argument_offset = load_String_offset(no_ctrl, argument);
1313 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1314
1315 // Get length of argument
1316 Node* argument_cnt = load_String_length(no_ctrl, argument);
1317
1318 // Check for receiver count != argument count
1319 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1320 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1321 Node* if_ne = generate_slow_guard(bol, NULL);
1322 if (if_ne != NULL) {
1323 phi->init_req(4, intcon(0));
1324 region->init_req(4, if_ne);
1325 }
1326
1327 // Check for count == 0 is done by assembler code for StrEquals.
1328
1329 if (!stopped()) {
1330 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1331 phi->init_req(1, equals);
1332 region->init_req(1, control());
1333 }
1334 }
1335
1336 // post merge
1337 set_control(_gvn.transform(region));
1338 record_for_igvn(region);
1339
1340 set_result(_gvn.transform(phi));
1341 return true;
1342 }
1343
1344 //------------------------------inline_array_equals----------------------------
1345 bool LibraryCallKit::inline_array_equals() {
1346 Node* arg1 = argument(0);
1347 Node* arg2 = argument(1);
1348 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1349 return true;
1350 }
1351
1352 // Java version of String.indexOf(constant string)
1353 // class StringDecl {
1354 // StringDecl(char[] ca) {
1355 // offset = 0;
1356 // count = ca.length;
1357 // value = ca;
1358 // }
1359 // int offset;
1360 // int count;
1361 // char[] value;
1362 // }
1363 //
1364 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1365 // int targetOffset, int cache_i, int md2) {
1366 // int cache = cache_i;
1367 // int sourceOffset = string_object.offset;
1368 // int sourceCount = string_object.count;
1369 // int targetCount = target_object.length;
1370 //
1371 // int targetCountLess1 = targetCount - 1;
1372 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1373 //
1374 // char[] source = string_object.value;
1375 // char[] target = target_object;
1376 // int lastChar = target[targetCountLess1];
1377 //
1378 // outer_loop:
1379 // for (int i = sourceOffset; i < sourceEnd; ) {
1380 // int src = source[i + targetCountLess1];
1381 // if (src == lastChar) {
1382 // // With random strings and a 4-character alphabet,
1383 // // reverse matching at this point sets up 0.8% fewer
1384 // // frames, but (paradoxically) makes 0.3% more probes.
1385 // // Since those probes are nearer the lastChar probe,
1386 // // there is may be a net D$ win with reverse matching.
1387 // // But, reversing loop inhibits unroll of inner loop
1388 // // for unknown reason. So, does running outer loop from
1389 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1390 // for (int j = 0; j < targetCountLess1; j++) {
1391 // if (target[targetOffset + j] != source[i+j]) {
1392 // if ((cache & (1 << source[i+j])) == 0) {
1393 // if (md2 < j+1) {
1394 // i += j+1;
1395 // continue outer_loop;
1396 // }
1397 // }
1398 // i += md2;
1399 // continue outer_loop;
1400 // }
1401 // }
1402 // return i - sourceOffset;
1403 // }
1404 // if ((cache & (1 << src)) == 0) {
1405 // i += targetCountLess1;
1406 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1407 // i++;
1408 // }
1409 // return -1;
1410 // }
1411
1412 //------------------------------string_indexOf------------------------
1413 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1414 jint cache_i, jint md2_i) {
1415
1416 Node* no_ctrl = NULL;
1417 float likely = PROB_LIKELY(0.9);
1418 float unlikely = PROB_UNLIKELY(0.9);
1419
1420 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1421
1422 Node* source = load_String_value(no_ctrl, string_object);
1423 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1424 Node* sourceCount = load_String_length(no_ctrl, string_object);
1425
1426 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1427 jint target_length = target_array->length();
1428 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1429 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1430
1431 // String.value field is known to be @Stable.
1432 if (UseImplicitStableValues) {
1433 target = cast_array_to_stable(target, target_type);
1434 }
1435
1436 IdealKit kit(this, false, true);
1437 #define __ kit.
1438 Node* zero = __ ConI(0);
1439 Node* one = __ ConI(1);
1440 Node* cache = __ ConI(cache_i);
1441 Node* md2 = __ ConI(md2_i);
1442 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1443 Node* targetCount = __ ConI(target_length);
1444 Node* targetCountLess1 = __ ConI(target_length - 1);
1445 Node* targetOffset = __ ConI(targetOffset_i);
1446 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1447
1448 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1449 Node* outer_loop = __ make_label(2 /* goto */);
1450 Node* return_ = __ make_label(1);
1451
1452 __ set(rtn,__ ConI(-1));
1453 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1454 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1455 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1456 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1457 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1458 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1459 Node* tpj = __ AddI(targetOffset, __ value(j));
1460 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1461 Node* ipj = __ AddI(__ value(i), __ value(j));
1462 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1463 __ if_then(targ, BoolTest::ne, src2); {
1464 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1465 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1466 __ increment(i, __ AddI(__ value(j), one));
1467 __ goto_(outer_loop);
1468 } __ end_if(); __ dead(j);
1469 }__ end_if(); __ dead(j);
1470 __ increment(i, md2);
1471 __ goto_(outer_loop);
1472 }__ end_if();
1473 __ increment(j, one);
1474 }__ end_loop(); __ dead(j);
1475 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1476 __ goto_(return_);
1477 }__ end_if();
1478 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1479 __ increment(i, targetCountLess1);
1480 }__ end_if();
1481 __ increment(i, one);
1482 __ bind(outer_loop);
1483 }__ end_loop(); __ dead(i);
1484 __ bind(return_);
1485
1486 // Final sync IdealKit and GraphKit.
1487 final_sync(kit);
1488 Node* result = __ value(rtn);
1489 #undef __
1490 C->set_has_loops(true);
1491 return result;
1492 }
1493
1494 //------------------------------inline_string_indexOf------------------------
1495 bool LibraryCallKit::inline_string_indexOf() {
1496 Node* receiver = argument(0);
1497 Node* arg = argument(1);
1498
1499 Node* result;
1500 // Disable the use of pcmpestri until it can be guaranteed that
1501 // the load doesn't cross into the uncommited space.
1502 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1503 UseSSE42Intrinsics) {
1504 // Generate SSE4.2 version of indexOf
1505 // We currently only have match rules that use SSE4.2
1506
1507 receiver = null_check(receiver);
1508 arg = null_check(arg);
1509 if (stopped()) {
1510 return true;
1511 }
1512
1513 ciInstanceKlass* str_klass = env()->String_klass();
1514 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1515
1516 // Make the merge point
1517 RegionNode* result_rgn = new (C) RegionNode(4);
1518 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1519 Node* no_ctrl = NULL;
1520
1521 // Get start addr of source string
1522 Node* source = load_String_value(no_ctrl, receiver);
1523 Node* source_offset = load_String_offset(no_ctrl, receiver);
1524 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1525
1526 // Get length of source string
1527 Node* source_cnt = load_String_length(no_ctrl, receiver);
1528
1529 // Get start addr of substring
1530 Node* substr = load_String_value(no_ctrl, arg);
1531 Node* substr_offset = load_String_offset(no_ctrl, arg);
1532 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1533
1534 // Get length of source string
1535 Node* substr_cnt = load_String_length(no_ctrl, arg);
1536
1537 // Check for substr count > string count
1538 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1539 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1540 Node* if_gt = generate_slow_guard(bol, NULL);
1541 if (if_gt != NULL) {
1542 result_phi->init_req(2, intcon(-1));
1543 result_rgn->init_req(2, if_gt);
1544 }
1545
1546 if (!stopped()) {
1547 // Check for substr count == 0
1548 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1549 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1550 Node* if_zero = generate_slow_guard(bol, NULL);
1551 if (if_zero != NULL) {
1552 result_phi->init_req(3, intcon(0));
1553 result_rgn->init_req(3, if_zero);
1554 }
1555 }
1556
1557 if (!stopped()) {
1558 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1559 result_phi->init_req(1, result);
1560 result_rgn->init_req(1, control());
1561 }
1562 set_control(_gvn.transform(result_rgn));
1563 record_for_igvn(result_rgn);
1564 result = _gvn.transform(result_phi);
1565
1566 } else { // Use LibraryCallKit::string_indexOf
1567 // don't intrinsify if argument isn't a constant string.
1568 if (!arg->is_Con()) {
1569 return false;
1570 }
1571 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1572 if (str_type == NULL) {
1573 return false;
1574 }
1575 ciInstanceKlass* klass = env()->String_klass();
1576 ciObject* str_const = str_type->const_oop();
1577 if (str_const == NULL || str_const->klass() != klass) {
1578 return false;
1579 }
1580 ciInstance* str = str_const->as_instance();
1581 assert(str != NULL, "must be instance");
1582
1583 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1584 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1585
1586 int o;
1587 int c;
1588 if (java_lang_String::has_offset_field()) {
1589 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1590 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1591 } else {
1592 o = 0;
1593 c = pat->length();
1594 }
1595
1596 // constant strings have no offset and count == length which
1597 // simplifies the resulting code somewhat so lets optimize for that.
1598 if (o != 0 || c != pat->length()) {
1599 return false;
1600 }
1601
1602 receiver = null_check(receiver, T_OBJECT);
1603 // NOTE: No null check on the argument is needed since it's a constant String oop.
1604 if (stopped()) {
1605 return true;
1606 }
1607
1608 // The null string as a pattern always returns 0 (match at beginning of string)
1609 if (c == 0) {
1610 set_result(intcon(0));
1611 return true;
1612 }
1613
1614 // Generate default indexOf
1615 jchar lastChar = pat->char_at(o + (c - 1));
1616 int cache = 0;
1617 int i;
1618 for (i = 0; i < c - 1; i++) {
1619 assert(i < pat->length(), "out of range");
1620 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1621 }
1622
1623 int md2 = c;
1624 for (i = 0; i < c - 1; i++) {
1625 assert(i < pat->length(), "out of range");
1626 if (pat->char_at(o + i) == lastChar) {
1627 md2 = (c - 1) - i;
1628 }
1629 }
1630
1631 result = string_indexOf(receiver, pat, o, cache, md2);
1632 }
1633 set_result(result);
1634 return true;
1635 }
1636
1637 //--------------------------round_double_node--------------------------------
1638 // Round a double node if necessary.
1639 Node* LibraryCallKit::round_double_node(Node* n) {
1640 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1641 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1642 return n;
1643 }
1644
1645 //------------------------------inline_math-----------------------------------
1646 // public static double Math.abs(double)
1647 // public static double Math.sqrt(double)
1648 // public static double Math.log(double)
1649 // public static double Math.log10(double)
1650 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1651 Node* arg = round_double_node(argument(0));
1652 Node* n = NULL;
1653 switch (id) {
1654 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1655 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1656 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1657 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1658 default: fatal_unexpected_iid(id); break;
1659 }
1660 set_result(_gvn.transform(n));
1661 return true;
1662 }
1663
1664 //------------------------------inline_trig----------------------------------
1665 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1666 // argument reduction which will turn into a fast/slow diamond.
1667 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1668 Node* arg = round_double_node(argument(0));
1669 Node* n = NULL;
1670
1671 switch (id) {
1672 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1673 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1674 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1675 default: fatal_unexpected_iid(id); break;
1676 }
1677 n = _gvn.transform(n);
1678
1679 // Rounding required? Check for argument reduction!
1680 if (Matcher::strict_fp_requires_explicit_rounding) {
1681 static const double pi_4 = 0.7853981633974483;
1682 static const double neg_pi_4 = -0.7853981633974483;
1683 // pi/2 in 80-bit extended precision
1684 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1685 // -pi/2 in 80-bit extended precision
1686 // 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};
1687 // Cutoff value for using this argument reduction technique
1688 //static const double pi_2_minus_epsilon = 1.564660403643354;
1689 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1690
1691 // Pseudocode for sin:
1692 // if (x <= Math.PI / 4.0) {
1693 // if (x >= -Math.PI / 4.0) return fsin(x);
1694 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1695 // } else {
1696 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1697 // }
1698 // return StrictMath.sin(x);
1699
1700 // Pseudocode for cos:
1701 // if (x <= Math.PI / 4.0) {
1702 // if (x >= -Math.PI / 4.0) return fcos(x);
1703 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1704 // } else {
1705 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1706 // }
1707 // return StrictMath.cos(x);
1708
1709 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1710 // requires a special machine instruction to load it. Instead we'll try
1711 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1712 // probably do the math inside the SIN encoding.
1713
1714 // Make the merge point
1715 RegionNode* r = new (C) RegionNode(3);
1716 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1717
1718 // Flatten arg so we need only 1 test
1719 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1720 // Node for PI/4 constant
1721 Node *pi4 = makecon(TypeD::make(pi_4));
1722 // Check PI/4 : abs(arg)
1723 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1724 // Check: If PI/4 < abs(arg) then go slow
1725 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1726 // Branch either way
1727 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1728 set_control(opt_iff(r,iff));
1729
1730 // Set fast path result
1731 phi->init_req(2, n);
1732
1733 // Slow path - non-blocking leaf call
1734 Node* call = NULL;
1735 switch (id) {
1736 case vmIntrinsics::_dsin:
1737 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1738 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1739 "Sin", NULL, arg, top());
1740 break;
1741 case vmIntrinsics::_dcos:
1742 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1743 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1744 "Cos", NULL, arg, top());
1745 break;
1746 case vmIntrinsics::_dtan:
1747 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1748 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1749 "Tan", NULL, arg, top());
1750 break;
1751 }
1752 assert(control()->in(0) == call, "");
1753 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1754 r->init_req(1, control());
1755 phi->init_req(1, slow_result);
1756
1757 // Post-merge
1758 set_control(_gvn.transform(r));
1759 record_for_igvn(r);
1760 n = _gvn.transform(phi);
1761
1762 C->set_has_split_ifs(true); // Has chance for split-if optimization
1763 }
1764 set_result(n);
1765 return true;
1766 }
1767
1768 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1769 //-------------------
1770 //result=(result.isNaN())? funcAddr():result;
1771 // Check: If isNaN() by checking result!=result? then either trap
1772 // or go to runtime
1773 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1774 // Build the boolean node
1775 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1776
1777 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1778 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1779 // The pow or exp intrinsic returned a NaN, which requires a call
1780 // to the runtime. Recompile with the runtime call.
1781 uncommon_trap(Deoptimization::Reason_intrinsic,
1782 Deoptimization::Action_make_not_entrant);
1783 }
1784 return result;
1785 } else {
1786 // If this inlining ever returned NaN in the past, we compile a call
1787 // to the runtime to properly handle corner cases
1788
1789 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1790 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1791 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1792
1793 if (!if_slow->is_top()) {
1794 RegionNode* result_region = new (C) RegionNode(3);
1795 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1796
1797 result_region->init_req(1, if_fast);
1798 result_val->init_req(1, result);
1799
1800 set_control(if_slow);
1801
1802 const TypePtr* no_memory_effects = NULL;
1803 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1804 no_memory_effects,
1805 x, top(), y, y ? top() : NULL);
1806 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1807 #ifdef ASSERT
1808 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1809 assert(value_top == top(), "second value must be top");
1810 #endif
1811
1812 result_region->init_req(2, control());
1813 result_val->init_req(2, value);
1814 set_control(_gvn.transform(result_region));
1815 return _gvn.transform(result_val);
1816 } else {
1817 return result;
1818 }
1819 }
1820 }
1821
1822 //------------------------------inline_exp-------------------------------------
1823 // Inline exp instructions, if possible. The Intel hardware only misses
1824 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1825 bool LibraryCallKit::inline_exp() {
1826 Node* arg = round_double_node(argument(0));
1827 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1828
1829 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1830 set_result(n);
1831
1832 C->set_has_split_ifs(true); // Has chance for split-if optimization
1833 return true;
1834 }
1835
1836 //------------------------------inline_pow-------------------------------------
1837 // Inline power instructions, if possible.
1838 bool LibraryCallKit::inline_pow() {
1839 // Pseudocode for pow
1840 // if (y == 2) {
1841 // return x * x;
1842 // } else {
1843 // if (x <= 0.0) {
1844 // long longy = (long)y;
1845 // if ((double)longy == y) { // if y is long
1846 // if (y + 1 == y) longy = 0; // huge number: even
1847 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1848 // } else {
1849 // result = NaN;
1850 // }
1851 // } else {
1852 // result = DPow(x,y);
1853 // }
1854 // if (result != result)? {
1855 // result = uncommon_trap() or runtime_call();
1856 // }
1857 // return result;
1858 // }
1859
1860 Node* x = round_double_node(argument(0));
1861 Node* y = round_double_node(argument(2));
1862
1863 Node* result = NULL;
1864
1865 Node* const_two_node = makecon(TypeD::make(2.0));
1866 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1867 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1868 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1869 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1870 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1871
1872 RegionNode* region_node = new (C) RegionNode(3);
1873 region_node->init_req(1, if_true);
1874
1875 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1876 // special case for x^y where y == 2, we can convert it to x * x
1877 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1878
1879 // set control to if_false since we will now process the false branch
1880 set_control(if_false);
1881
1882 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1883 // Short form: skip the fancy tests and just check for NaN result.
1884 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1885 } else {
1886 // If this inlining ever returned NaN in the past, include all
1887 // checks + call to the runtime.
1888
1889 // Set the merge point for If node with condition of (x <= 0.0)
1890 // There are four possible paths to region node and phi node
1891 RegionNode *r = new (C) RegionNode(4);
1892 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1893
1894 // Build the first if node: if (x <= 0.0)
1895 // Node for 0 constant
1896 Node *zeronode = makecon(TypeD::ZERO);
1897 // Check x:0
1898 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1899 // Check: If (x<=0) then go complex path
1900 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1901 // Branch either way
1902 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1903 // Fast path taken; set region slot 3
1904 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1905 r->init_req(3,fast_taken); // Capture fast-control
1906
1907 // Fast path not-taken, i.e. slow path
1908 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1909
1910 // Set fast path result
1911 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1912 phi->init_req(3, fast_result);
1913
1914 // Complex path
1915 // Build the second if node (if y is long)
1916 // Node for (long)y
1917 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1918 // Node for (double)((long) y)
1919 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1920 // Check (double)((long) y) : y
1921 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1922 // Check if (y isn't long) then go to slow path
1923
1924 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1925 // Branch either way
1926 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1927 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1928
1929 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1930
1931 // Calculate DPow(abs(x), y)*(1 & (long)y)
1932 // Node for constant 1
1933 Node *conone = longcon(1);
1934 // 1& (long)y
1935 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1936
1937 // A huge number is always even. Detect a huge number by checking
1938 // if y + 1 == y and set integer to be tested for parity to 0.
1939 // Required for corner case:
1940 // (long)9.223372036854776E18 = max_jlong
1941 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1942 // max_jlong is odd but 9.223372036854776E18 is even
1943 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1944 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1945 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1946 Node* correctedsign = NULL;
1947 if (ConditionalMoveLimit != 0) {
1948 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1949 } else {
1950 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1951 RegionNode *r = new (C) RegionNode(3);
1952 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1953 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1954 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1955 phi->init_req(1, signnode);
1956 phi->init_req(2, longcon(0));
1957 correctedsign = _gvn.transform(phi);
1958 ylong_path = _gvn.transform(r);
1959 record_for_igvn(r);
1960 }
1961
1962 // zero node
1963 Node *conzero = longcon(0);
1964 // Check (1&(long)y)==0?
1965 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1966 // Check if (1&(long)y)!=0?, if so the result is negative
1967 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1968 // abs(x)
1969 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1970 // abs(x)^y
1971 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1972 // -abs(x)^y
1973 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1974 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1975 Node *signresult = NULL;
1976 if (ConditionalMoveLimit != 0) {
1977 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1978 } else {
1979 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1980 RegionNode *r = new (C) RegionNode(3);
1981 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1982 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1983 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1984 phi->init_req(1, absxpowy);
1985 phi->init_req(2, negabsxpowy);
1986 signresult = _gvn.transform(phi);
1987 ylong_path = _gvn.transform(r);
1988 record_for_igvn(r);
1989 }
1990 // Set complex path fast result
1991 r->init_req(2, ylong_path);
1992 phi->init_req(2, signresult);
1993
1994 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1995 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1996 r->init_req(1,slow_path);
1997 phi->init_req(1,slow_result);
1998
1999 // Post merge
2000 set_control(_gvn.transform(r));
2001 record_for_igvn(r);
2002 result = _gvn.transform(phi);
2003 }
2004
2005 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2006
2007 // control from finish_pow_exp is now input to the region node
2008 region_node->set_req(2, control());
2009 // the result from finish_pow_exp is now input to the phi node
2010 phi_node->init_req(2, result);
2011 set_control(_gvn.transform(region_node));
2012 record_for_igvn(region_node);
2013 set_result(_gvn.transform(phi_node));
2014
2015 C->set_has_split_ifs(true); // Has chance for split-if optimization
2016 return true;
2017 }
2018
2019 //------------------------------runtime_math-----------------------------
2020 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2021 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2022 "must be (DD)D or (D)D type");
2023
2024 // Inputs
2025 Node* a = round_double_node(argument(0));
2026 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2027
2028 const TypePtr* no_memory_effects = NULL;
2029 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2030 no_memory_effects,
2031 a, top(), b, b ? top() : NULL);
2032 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2033 #ifdef ASSERT
2034 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2035 assert(value_top == top(), "second value must be top");
2036 #endif
2037
2038 set_result(value);
2039 return true;
2040 }
2041
2042 //------------------------------inline_math_native-----------------------------
2043 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2044 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2045 switch (id) {
2046 // These intrinsics are not properly supported on all hardware
2047 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2048 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2049 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2050 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2051 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2052 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2053
2054 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2055 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2056 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2057 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2058
2059 // These intrinsics are supported on all hardware
2060 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2061 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2062
2063 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2064 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2065 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2066 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2067 #undef FN_PTR
2068
2069 // These intrinsics are not yet correctly implemented
2070 case vmIntrinsics::_datan2:
2071 return false;
2072
2073 default:
2074 fatal_unexpected_iid(id);
2075 return false;
2076 }
2077 }
2078
2079 static bool is_simple_name(Node* n) {
2080 return (n->req() == 1 // constant
2081 || (n->is_Type() && n->as_Type()->type()->singleton())
2082 || n->is_Proj() // parameter or return value
2083 || n->is_Phi() // local of some sort
2084 );
2085 }
2086
2087 //----------------------------inline_min_max-----------------------------------
2088 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2089 set_result(generate_min_max(id, argument(0), argument(1)));
2090 return true;
2091 }
2092
2093 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2094 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2095 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2096 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2097 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2098
2099 {
2100 PreserveJVMState pjvms(this);
2101 PreserveReexecuteState preexecs(this);
2102 jvms()->set_should_reexecute(true);
2103
2104 set_control(slow_path);
2105 set_i_o(i_o());
2106
2107 uncommon_trap(Deoptimization::Reason_intrinsic,
2108 Deoptimization::Action_none);
2109 }
2110
2111 set_control(fast_path);
2112 set_result(math);
2113 }
2114
2115 template <typename OverflowOp>
2116 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2117 typedef typename OverflowOp::MathOp MathOp;
2118
2119 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2120 Node* operation = _gvn.transform( mathOp );
2121 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2122 inline_math_mathExact(operation, ofcheck);
2123 return true;
2124 }
2125
2126 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2127 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2128 }
2129
2130 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2131 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2132 }
2133
2134 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2135 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2136 }
2137
2138 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2139 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2140 }
2141
2142 bool LibraryCallKit::inline_math_negateExactI() {
2143 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2144 }
2145
2146 bool LibraryCallKit::inline_math_negateExactL() {
2147 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2148 }
2149
2150 bool LibraryCallKit::inline_math_multiplyExactI() {
2151 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2152 }
2153
2154 bool LibraryCallKit::inline_math_multiplyExactL() {
2155 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2156 }
2157
2158 Node*
2159 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2160 // These are the candidate return value:
2161 Node* xvalue = x0;
2162 Node* yvalue = y0;
2163
2164 if (xvalue == yvalue) {
2165 return xvalue;
2166 }
2167
2168 bool want_max = (id == vmIntrinsics::_max);
2169
2170 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2171 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2172 if (txvalue == NULL || tyvalue == NULL) return top();
2173 // This is not really necessary, but it is consistent with a
2174 // hypothetical MaxINode::Value method:
2175 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2176
2177 // %%% This folding logic should (ideally) be in a different place.
2178 // Some should be inside IfNode, and there to be a more reliable
2179 // transformation of ?: style patterns into cmoves. We also want
2180 // more powerful optimizations around cmove and min/max.
2181
2182 // Try to find a dominating comparison of these guys.
2183 // It can simplify the index computation for Arrays.copyOf
2184 // and similar uses of System.arraycopy.
2185 // First, compute the normalized version of CmpI(x, y).
2186 int cmp_op = Op_CmpI;
2187 Node* xkey = xvalue;
2188 Node* ykey = yvalue;
2189 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2190 if (ideal_cmpxy->is_Cmp()) {
2191 // E.g., if we have CmpI(length - offset, count),
2192 // it might idealize to CmpI(length, count + offset)
2193 cmp_op = ideal_cmpxy->Opcode();
2194 xkey = ideal_cmpxy->in(1);
2195 ykey = ideal_cmpxy->in(2);
2196 }
2197
2198 // Start by locating any relevant comparisons.
2199 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2200 Node* cmpxy = NULL;
2201 Node* cmpyx = NULL;
2202 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2203 Node* cmp = start_from->fast_out(k);
2204 if (cmp->outcnt() > 0 && // must have prior uses
2205 cmp->in(0) == NULL && // must be context-independent
2206 cmp->Opcode() == cmp_op) { // right kind of compare
2207 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2208 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2209 }
2210 }
2211
2212 const int NCMPS = 2;
2213 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2214 int cmpn;
2215 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2216 if (cmps[cmpn] != NULL) break; // find a result
2217 }
2218 if (cmpn < NCMPS) {
2219 // Look for a dominating test that tells us the min and max.
2220 int depth = 0; // Limit search depth for speed
2221 Node* dom = control();
2222 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2223 if (++depth >= 100) break;
2224 Node* ifproj = dom;
2225 if (!ifproj->is_Proj()) continue;
2226 Node* iff = ifproj->in(0);
2227 if (!iff->is_If()) continue;
2228 Node* bol = iff->in(1);
2229 if (!bol->is_Bool()) continue;
2230 Node* cmp = bol->in(1);
2231 if (cmp == NULL) continue;
2232 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2233 if (cmps[cmpn] == cmp) break;
2234 if (cmpn == NCMPS) continue;
2235 BoolTest::mask btest = bol->as_Bool()->_test._test;
2236 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2237 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2238 // At this point, we know that 'x btest y' is true.
2239 switch (btest) {
2240 case BoolTest::eq:
2241 // They are proven equal, so we can collapse the min/max.
2242 // Either value is the answer. Choose the simpler.
2243 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2244 return yvalue;
2245 return xvalue;
2246 case BoolTest::lt: // x < y
2247 case BoolTest::le: // x <= y
2248 return (want_max ? yvalue : xvalue);
2249 case BoolTest::gt: // x > y
2250 case BoolTest::ge: // x >= y
2251 return (want_max ? xvalue : yvalue);
2252 }
2253 }
2254 }
2255
2256 // We failed to find a dominating test.
2257 // Let's pick a test that might GVN with prior tests.
2258 Node* best_bol = NULL;
2259 BoolTest::mask best_btest = BoolTest::illegal;
2260 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2261 Node* cmp = cmps[cmpn];
2262 if (cmp == NULL) continue;
2263 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2264 Node* bol = cmp->fast_out(j);
2265 if (!bol->is_Bool()) continue;
2266 BoolTest::mask btest = bol->as_Bool()->_test._test;
2267 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2268 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2269 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2270 best_bol = bol->as_Bool();
2271 best_btest = btest;
2272 }
2273 }
2274 }
2275
2276 Node* answer_if_true = NULL;
2277 Node* answer_if_false = NULL;
2278 switch (best_btest) {
2279 default:
2280 if (cmpxy == NULL)
2281 cmpxy = ideal_cmpxy;
2282 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2283 // and fall through:
2284 case BoolTest::lt: // x < y
2285 case BoolTest::le: // x <= y
2286 answer_if_true = (want_max ? yvalue : xvalue);
2287 answer_if_false = (want_max ? xvalue : yvalue);
2288 break;
2289 case BoolTest::gt: // x > y
2290 case BoolTest::ge: // x >= y
2291 answer_if_true = (want_max ? xvalue : yvalue);
2292 answer_if_false = (want_max ? yvalue : xvalue);
2293 break;
2294 }
2295
2296 jint hi, lo;
2297 if (want_max) {
2298 // We can sharpen the minimum.
2299 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2300 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2301 } else {
2302 // We can sharpen the maximum.
2303 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2304 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2305 }
2306
2307 // Use a flow-free graph structure, to avoid creating excess control edges
2308 // which could hinder other optimizations.
2309 // Since Math.min/max is often used with arraycopy, we want
2310 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2311 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2312 answer_if_false, answer_if_true,
2313 TypeInt::make(lo, hi, widen));
2314
2315 return _gvn.transform(cmov);
2316
2317 /*
2318 // This is not as desirable as it may seem, since Min and Max
2319 // nodes do not have a full set of optimizations.
2320 // And they would interfere, anyway, with 'if' optimizations
2321 // and with CMoveI canonical forms.
2322 switch (id) {
2323 case vmIntrinsics::_min:
2324 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2325 case vmIntrinsics::_max:
2326 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2327 default:
2328 ShouldNotReachHere();
2329 }
2330 */
2331 }
2332
2333 inline int
2334 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2335 const TypePtr* base_type = TypePtr::NULL_PTR;
2336 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2337 if (base_type == NULL) {
2338 // Unknown type.
2339 return Type::AnyPtr;
2340 } else if (base_type == TypePtr::NULL_PTR) {
2341 // Since this is a NULL+long form, we have to switch to a rawptr.
2342 base = _gvn.transform(new (C) CastX2PNode(offset));
2343 offset = MakeConX(0);
2344 return Type::RawPtr;
2345 } else if (base_type->base() == Type::RawPtr) {
2346 return Type::RawPtr;
2347 } else if (base_type->isa_oopptr()) {
2348 // Base is never null => always a heap address.
2349 if (base_type->ptr() == TypePtr::NotNull) {
2350 return Type::OopPtr;
2351 }
2352 // Offset is small => always a heap address.
2353 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2354 if (offset_type != NULL &&
2355 base_type->offset() == 0 && // (should always be?)
2356 offset_type->_lo >= 0 &&
2357 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2358 return Type::OopPtr;
2359 }
2360 // Otherwise, it might either be oop+off or NULL+addr.
2361 return Type::AnyPtr;
2362 } else {
2363 // No information:
2364 return Type::AnyPtr;
2365 }
2366 }
2367
2368 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2369 int kind = classify_unsafe_addr(base, offset);
2370 if (kind == Type::RawPtr) {
2371 return basic_plus_adr(top(), base, offset);
2372 } else {
2373 return basic_plus_adr(base, offset);
2374 }
2375 }
2376
2377 //--------------------------inline_number_methods-----------------------------
2378 // inline int Integer.numberOfLeadingZeros(int)
2379 // inline int Long.numberOfLeadingZeros(long)
2380 //
2381 // inline int Integer.numberOfTrailingZeros(int)
2382 // inline int Long.numberOfTrailingZeros(long)
2383 //
2384 // inline int Integer.bitCount(int)
2385 // inline int Long.bitCount(long)
2386 //
2387 // inline char Character.reverseBytes(char)
2388 // inline short Short.reverseBytes(short)
2389 // inline int Integer.reverseBytes(int)
2390 // inline long Long.reverseBytes(long)
2391 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2392 Node* arg = argument(0);
2393 Node* n = NULL;
2394 switch (id) {
2395 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2396 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2397 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2398 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2399 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2400 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2401 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2402 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2403 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2404 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2405 default: fatal_unexpected_iid(id); break;
2406 }
2407 set_result(_gvn.transform(n));
2408 return true;
2409 }
2410
2411 //----------------------------inline_unsafe_access----------------------------
2412
2413 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2414
2415 // Helper that guards and inserts a pre-barrier.
2416 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2417 Node* pre_val, bool need_mem_bar) {
2418 // We could be accessing the referent field of a reference object. If so, when G1
2419 // is enabled, we need to log the value in the referent field in an SATB buffer.
2420 // This routine performs some compile time filters and generates suitable
2421 // runtime filters that guard the pre-barrier code.
2422 // Also add memory barrier for non volatile load from the referent field
2423 // to prevent commoning of loads across safepoint.
2424 if (!UseG1GC && !need_mem_bar)
2425 return;
2426
2427 // Some compile time checks.
2428
2429 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2430 const TypeX* otype = offset->find_intptr_t_type();
2431 if (otype != NULL && otype->is_con() &&
2432 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2433 // Constant offset but not the reference_offset so just return
2434 return;
2435 }
2436
2437 // We only need to generate the runtime guards for instances.
2438 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2439 if (btype != NULL) {
2440 if (btype->isa_aryptr()) {
2441 // Array type so nothing to do
2442 return;
2443 }
2444
2445 const TypeInstPtr* itype = btype->isa_instptr();
2446 if (itype != NULL) {
2447 // Can the klass of base_oop be statically determined to be
2448 // _not_ a sub-class of Reference and _not_ Object?
2449 ciKlass* klass = itype->klass();
2450 if ( klass->is_loaded() &&
2451 !klass->is_subtype_of(env()->Reference_klass()) &&
2452 !env()->Object_klass()->is_subtype_of(klass)) {
2453 return;
2454 }
2455 }
2456 }
2457
2458 // The compile time filters did not reject base_oop/offset so
2459 // we need to generate the following runtime filters
2460 //
2461 // if (offset == java_lang_ref_Reference::_reference_offset) {
2462 // if (instance_of(base, java.lang.ref.Reference)) {
2463 // pre_barrier(_, pre_val, ...);
2464 // }
2465 // }
2466
2467 float likely = PROB_LIKELY( 0.999);
2468 float unlikely = PROB_UNLIKELY(0.999);
2469
2470 IdealKit ideal(this);
2471 #define __ ideal.
2472
2473 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2474
2475 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2476 // Update graphKit memory and control from IdealKit.
2477 sync_kit(ideal);
2478
2479 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2480 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2481
2482 // Update IdealKit memory and control from graphKit.
2483 __ sync_kit(this);
2484
2485 Node* one = __ ConI(1);
2486 // is_instof == 0 if base_oop == NULL
2487 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2488
2489 // Update graphKit from IdeakKit.
2490 sync_kit(ideal);
2491
2492 // Use the pre-barrier to record the value in the referent field
2493 pre_barrier(false /* do_load */,
2494 __ ctrl(),
2495 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2496 pre_val /* pre_val */,
2497 T_OBJECT);
2498 if (need_mem_bar) {
2499 // Add memory barrier to prevent commoning reads from this field
2500 // across safepoint since GC can change its value.
2501 insert_mem_bar(Op_MemBarCPUOrder);
2502 }
2503 // Update IdealKit from graphKit.
2504 __ sync_kit(this);
2505
2506 } __ end_if(); // _ref_type != ref_none
2507 } __ end_if(); // offset == referent_offset
2508
2509 // Final sync IdealKit and GraphKit.
2510 final_sync(ideal);
2511 #undef __
2512 }
2513
2514
2515 // Interpret Unsafe.fieldOffset cookies correctly:
2516 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2517
2518 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2519 // Attempt to infer a sharper value type from the offset and base type.
2520 ciKlass* sharpened_klass = NULL;
2521
2522 // See if it is an instance field, with an object type.
2523 if (alias_type->field() != NULL) {
2524 assert(!is_native_ptr, "native pointer op cannot use a java address");
2525 if (alias_type->field()->type()->is_klass()) {
2526 sharpened_klass = alias_type->field()->type()->as_klass();
2527 }
2528 }
2529
2530 // See if it is a narrow oop array.
2531 if (adr_type->isa_aryptr()) {
2532 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2533 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2534 if (elem_type != NULL) {
2535 sharpened_klass = elem_type->klass();
2536 }
2537 }
2538 }
2539
2540 // The sharpened class might be unloaded if there is no class loader
2541 // contraint in place.
2542 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2543 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2544
2545 #ifndef PRODUCT
2546 if (C->print_intrinsics() || C->print_inlining()) {
2547 tty->print(" from base type: "); adr_type->dump();
2548 tty->print(" sharpened value: "); tjp->dump();
2549 }
2550 #endif
2551 // Sharpen the value type.
2552 return tjp;
2553 }
2554 return NULL;
2555 }
2556
2557 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2558 if (callee()->is_static()) return false; // caller must have the capability!
2559 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2560
2561 #ifndef PRODUCT
2562 {
2563 ResourceMark rm;
2564 // Check the signatures.
2565 ciSignature* sig = callee()->signature();
2566 #ifdef ASSERT
2567 if (!is_store) {
2568 // Object getObject(Object base, int/long offset), etc.
2569 BasicType rtype = sig->return_type()->basic_type();
2570 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2571 rtype = T_ADDRESS; // it is really a C void*
2572 assert(rtype == type, "getter must return the expected value");
2573 if (!is_native_ptr) {
2574 assert(sig->count() == 2, "oop getter has 2 arguments");
2575 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2576 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2577 } else {
2578 assert(sig->count() == 1, "native getter has 1 argument");
2579 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2580 }
2581 } else {
2582 // void putObject(Object base, int/long offset, Object x), etc.
2583 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2584 if (!is_native_ptr) {
2585 assert(sig->count() == 3, "oop putter has 3 arguments");
2586 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2587 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2588 } else {
2589 assert(sig->count() == 2, "native putter has 2 arguments");
2590 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2591 }
2592 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2593 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2594 vtype = T_ADDRESS; // it is really a C void*
2595 assert(vtype == type, "putter must accept the expected value");
2596 }
2597 #endif // ASSERT
2598 }
2599 #endif //PRODUCT
2600
2601 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2602
2603 Node* receiver = argument(0); // type: oop
2604
2605 // Build address expression. See the code in inline_unsafe_prefetch.
2606 Node* adr;
2607 Node* heap_base_oop = top();
2608 Node* offset = top();
2609 Node* val;
2610
2611 if (!is_native_ptr) {
2612 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2613 Node* base = argument(1); // type: oop
2614 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2615 offset = argument(2); // type: long
2616 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2617 // to be plain byte offsets, which are also the same as those accepted
2618 // by oopDesc::field_base.
2619 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2620 "fieldOffset must be byte-scaled");
2621 // 32-bit machines ignore the high half!
2622 offset = ConvL2X(offset);
2623 adr = make_unsafe_address(base, offset);
2624 heap_base_oop = base;
2625 val = is_store ? argument(4) : NULL;
2626 } else {
2627 Node* ptr = argument(1); // type: long
2628 ptr = ConvL2X(ptr); // adjust Java long to machine word
2629 adr = make_unsafe_address(NULL, ptr);
2630 val = is_store ? argument(3) : NULL;
2631 }
2632
2633 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2634
2635 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2636 // there was not enough information to nail it down.
2637 Compile::AliasType* alias_type = C->alias_type(adr_type);
2638 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2639
2640 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM ||
2641 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown");
2642 bool mismatched = false;
2643 BasicType bt = alias_type->basic_type();
2644 if (bt != T_ILLEGAL) {
2645 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2646 // Alias type doesn't differentiate between byte[] and boolean[]).
2647 // Use address type to get the element type.
2648 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2649 }
2650 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2651 // accessing an array field with getObject is not a mismatch
2652 bt = T_OBJECT;
2653 }
2654 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2655 // Don't intrinsify mismatched object accesses
2656 return false;
2657 }
2658 mismatched = (bt != type);
2659 }
2660
2661 // First guess at the value type.
2662 const Type *value_type = Type::get_const_basic_type(type);
2663
2664 // We will need memory barriers unless we can determine a unique
2665 // alias category for this reference. (Note: If for some reason
2666 // the barriers get omitted and the unsafe reference begins to "pollute"
2667 // the alias analysis of the rest of the graph, either Compile::can_alias
2668 // or Compile::must_alias will throw a diagnostic assert.)
2669 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2670
2671 // If we are reading the value of the referent field of a Reference
2672 // object (either by using Unsafe directly or through reflection)
2673 // then, if G1 is enabled, we need to record the referent in an
2674 // SATB log buffer using the pre-barrier mechanism.
2675 // Also we need to add memory barrier to prevent commoning reads
2676 // from this field across safepoint since GC can change its value.
2677 bool need_read_barrier = !is_native_ptr && !is_store &&
2678 offset != top() && heap_base_oop != top();
2679
2680 if (!is_store && type == T_OBJECT) {
2681 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2682 if (tjp != NULL) {
2683 value_type = tjp;
2684 }
2685 }
2686
2687 receiver = null_check(receiver);
2688 if (stopped()) {
2689 return true;
2690 }
2691 // Heap pointers get a null-check from the interpreter,
2692 // as a courtesy. However, this is not guaranteed by Unsafe,
2693 // and it is not possible to fully distinguish unintended nulls
2694 // from intended ones in this API.
2695
2696 if (is_volatile) {
2697 // We need to emit leading and trailing CPU membars (see below) in
2698 // addition to memory membars when is_volatile. This is a little
2699 // too strong, but avoids the need to insert per-alias-type
2700 // volatile membars (for stores; compare Parse::do_put_xxx), which
2701 // we cannot do effectively here because we probably only have a
2702 // rough approximation of type.
2703 need_mem_bar = true;
2704 // For Stores, place a memory ordering barrier now.
2705 if (is_store) {
2706 insert_mem_bar(Op_MemBarRelease);
2707 } else {
2708 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2709 insert_mem_bar(Op_MemBarVolatile);
2710 }
2711 }
2712 }
2713
2714 // Memory barrier to prevent normal and 'unsafe' accesses from
2715 // bypassing each other. Happens after null checks, so the
2716 // exception paths do not take memory state from the memory barrier,
2717 // so there's no problems making a strong assert about mixing users
2718 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2719 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2720 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2721
2722 if (!is_store) {
2723 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2724 // To be valid, unsafe loads may depend on other conditions than
2725 // the one that guards them: pin the Load node
2726 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2727 // load value
2728 switch (type) {
2729 case T_BOOLEAN:
2730 case T_CHAR:
2731 case T_BYTE:
2732 case T_SHORT:
2733 case T_INT:
2734 case T_LONG:
2735 case T_FLOAT:
2736 case T_DOUBLE:
2737 break;
2738 case T_OBJECT:
2739 if (need_read_barrier) {
2740 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2741 }
2742 break;
2743 case T_ADDRESS:
2744 // Cast to an int type.
2745 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2746 p = ConvX2UL(p);
2747 break;
2748 default:
2749 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2750 break;
2751 }
2752 // The load node has the control of the preceding MemBarCPUOrder. All
2753 // following nodes will have the control of the MemBarCPUOrder inserted at
2754 // the end of this method. So, pushing the load onto the stack at a later
2755 // point is fine.
2756 set_result(p);
2757 } else {
2758 // place effect of store into memory
2759 switch (type) {
2760 case T_DOUBLE:
2761 val = dstore_rounding(val);
2762 break;
2763 case T_ADDRESS:
2764 // Repackage the long as a pointer.
2765 val = ConvL2X(val);
2766 val = _gvn.transform(new (C) CastX2PNode(val));
2767 break;
2768 }
2769
2770 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2771 if (type != T_OBJECT ) {
2772 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2773 } else {
2774 // Possibly an oop being stored to Java heap or native memory
2775 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2776 // oop to Java heap.
2777 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2778 } else {
2779 // We can't tell at compile time if we are storing in the Java heap or outside
2780 // of it. So we need to emit code to conditionally do the proper type of
2781 // store.
2782
2783 IdealKit ideal(this);
2784 #define __ ideal.
2785 // QQQ who knows what probability is here??
2786 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2787 // Sync IdealKit and graphKit.
2788 sync_kit(ideal);
2789 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2790 // Update IdealKit memory.
2791 __ sync_kit(this);
2792 } __ else_(); {
2793 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched);
2794 } __ end_if();
2795 // Final sync IdealKit and GraphKit.
2796 final_sync(ideal);
2797 #undef __
2798 }
2799 }
2800 }
2801
2802 if (is_volatile) {
2803 if (!is_store) {
2804 insert_mem_bar(Op_MemBarAcquire);
2805 } else {
2806 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2807 insert_mem_bar(Op_MemBarVolatile);
2808 }
2809 }
2810 }
2811
2812 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2813
2814 return true;
2815 }
2816
2817 //----------------------------inline_unsafe_prefetch----------------------------
2818
2819 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2820 #ifndef PRODUCT
2821 {
2822 ResourceMark rm;
2823 // Check the signatures.
2824 ciSignature* sig = callee()->signature();
2825 #ifdef ASSERT
2826 // Object getObject(Object base, int/long offset), etc.
2827 BasicType rtype = sig->return_type()->basic_type();
2828 if (!is_native_ptr) {
2829 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2830 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2831 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2832 } else {
2833 assert(sig->count() == 1, "native prefetch has 1 argument");
2834 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2835 }
2836 #endif // ASSERT
2837 }
2838 #endif // !PRODUCT
2839
2840 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2841
2842 const int idx = is_static ? 0 : 1;
2843 if (!is_static) {
2844 null_check_receiver();
2845 if (stopped()) {
2846 return true;
2847 }
2848 }
2849
2850 // Build address expression. See the code in inline_unsafe_access.
2851 Node *adr;
2852 if (!is_native_ptr) {
2853 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2854 Node* base = argument(idx + 0); // type: oop
2855 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2856 Node* offset = argument(idx + 1); // type: long
2857 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2858 // to be plain byte offsets, which are also the same as those accepted
2859 // by oopDesc::field_base.
2860 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2861 "fieldOffset must be byte-scaled");
2862 // 32-bit machines ignore the high half!
2863 offset = ConvL2X(offset);
2864 adr = make_unsafe_address(base, offset);
2865 } else {
2866 Node* ptr = argument(idx + 0); // type: long
2867 ptr = ConvL2X(ptr); // adjust Java long to machine word
2868 adr = make_unsafe_address(NULL, ptr);
2869 }
2870
2871 // Generate the read or write prefetch
2872 Node *prefetch;
2873 if (is_store) {
2874 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2875 } else {
2876 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2877 }
2878 prefetch->init_req(0, control());
2879 set_i_o(_gvn.transform(prefetch));
2880
2881 return true;
2882 }
2883
2884 //----------------------------inline_unsafe_load_store----------------------------
2885 // This method serves a couple of different customers (depending on LoadStoreKind):
2886 //
2887 // LS_cmpxchg:
2888 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2889 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2890 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2891 //
2892 // LS_xadd:
2893 // public int getAndAddInt( Object o, long offset, int delta)
2894 // public long getAndAddLong(Object o, long offset, long delta)
2895 //
2896 // LS_xchg:
2897 // int getAndSet(Object o, long offset, int newValue)
2898 // long getAndSet(Object o, long offset, long newValue)
2899 // Object getAndSet(Object o, long offset, Object newValue)
2900 //
2901 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2902 // This basic scheme here is the same as inline_unsafe_access, but
2903 // differs in enough details that combining them would make the code
2904 // overly confusing. (This is a true fact! I originally combined
2905 // them, but even I was confused by it!) As much code/comments as
2906 // possible are retained from inline_unsafe_access though to make
2907 // the correspondences clearer. - dl
2908
2909 if (callee()->is_static()) return false; // caller must have the capability!
2910
2911 #ifndef PRODUCT
2912 BasicType rtype;
2913 {
2914 ResourceMark rm;
2915 // Check the signatures.
2916 ciSignature* sig = callee()->signature();
2917 rtype = sig->return_type()->basic_type();
2918 if (kind == LS_xadd || kind == LS_xchg) {
2919 // Check the signatures.
2920 #ifdef ASSERT
2921 assert(rtype == type, "get and set must return the expected type");
2922 assert(sig->count() == 3, "get and set has 3 arguments");
2923 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2924 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2925 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2926 #endif // ASSERT
2927 } else if (kind == LS_cmpxchg) {
2928 // Check the signatures.
2929 #ifdef ASSERT
2930 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2931 assert(sig->count() == 4, "CAS has 4 arguments");
2932 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2933 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2934 #endif // ASSERT
2935 } else {
2936 ShouldNotReachHere();
2937 }
2938 }
2939 #endif //PRODUCT
2940
2941 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2942
2943 // Get arguments:
2944 Node* receiver = NULL;
2945 Node* base = NULL;
2946 Node* offset = NULL;
2947 Node* oldval = NULL;
2948 Node* newval = NULL;
2949 if (kind == LS_cmpxchg) {
2950 const bool two_slot_type = type2size[type] == 2;
2951 receiver = argument(0); // type: oop
2952 base = argument(1); // type: oop
2953 offset = argument(2); // type: long
2954 oldval = argument(4); // type: oop, int, or long
2955 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2956 } else if (kind == LS_xadd || kind == LS_xchg){
2957 receiver = argument(0); // type: oop
2958 base = argument(1); // type: oop
2959 offset = argument(2); // type: long
2960 oldval = NULL;
2961 newval = argument(4); // type: oop, int, or long
2962 }
2963
2964 // Null check receiver.
2965 receiver = null_check(receiver);
2966 if (stopped()) {
2967 return true;
2968 }
2969
2970 // Build field offset expression.
2971 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2972 // to be plain byte offsets, which are also the same as those accepted
2973 // by oopDesc::field_base.
2974 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2975 // 32-bit machines ignore the high half of long offsets
2976 offset = ConvL2X(offset);
2977 Node* adr = make_unsafe_address(base, offset);
2978 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2979
2980 Compile::AliasType* alias_type = C->alias_type(adr_type);
2981 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM ||
2982 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown");
2983 BasicType bt = alias_type->basic_type();
2984 if (bt != T_ILLEGAL &&
2985 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2986 // Don't intrinsify mismatched object accesses.
2987 return false;
2988 }
2989
2990 // For CAS, unlike inline_unsafe_access, there seems no point in
2991 // trying to refine types. Just use the coarse types here.
2992 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2993 const Type *value_type = Type::get_const_basic_type(type);
2994
2995 if (kind == LS_xchg && type == T_OBJECT) {
2996 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2997 if (tjp != NULL) {
2998 value_type = tjp;
2999 }
3000 }
3001
3002 int alias_idx = C->get_alias_index(adr_type);
3003
3004 // Memory-model-wise, a LoadStore acts like a little synchronized
3005 // block, so needs barriers on each side. These don't translate
3006 // into actual barriers on most machines, but we still need rest of
3007 // compiler to respect ordering.
3008
3009 insert_mem_bar(Op_MemBarRelease);
3010 insert_mem_bar(Op_MemBarCPUOrder);
3011
3012 // 4984716: MemBars must be inserted before this
3013 // memory node in order to avoid a false
3014 // dependency which will confuse the scheduler.
3015 Node *mem = memory(alias_idx);
3016
3017 // For now, we handle only those cases that actually exist: ints,
3018 // longs, and Object. Adding others should be straightforward.
3019 Node* load_store = NULL;
3020 switch(type) {
3021 case T_INT:
3022 if (kind == LS_xadd) {
3023 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3024 } else if (kind == LS_xchg) {
3025 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3026 } else if (kind == LS_cmpxchg) {
3027 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3028 } else {
3029 ShouldNotReachHere();
3030 }
3031 break;
3032 case T_LONG:
3033 if (kind == LS_xadd) {
3034 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3035 } else if (kind == LS_xchg) {
3036 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3037 } else if (kind == LS_cmpxchg) {
3038 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3039 } else {
3040 ShouldNotReachHere();
3041 }
3042 break;
3043 case T_OBJECT:
3044 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3045 // could be delayed during Parse (for example, in adjust_map_after_if()).
3046 // Execute transformation here to avoid barrier generation in such case.
3047 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3048 newval = _gvn.makecon(TypePtr::NULL_PTR);
3049
3050 // Reference stores need a store barrier.
3051 if (kind == LS_xchg) {
3052 // If pre-barrier must execute before the oop store, old value will require do_load here.
3053 if (!can_move_pre_barrier()) {
3054 pre_barrier(true /* do_load*/,
3055 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3056 NULL /* pre_val*/,
3057 T_OBJECT);
3058 } // Else move pre_barrier to use load_store value, see below.
3059 } else if (kind == LS_cmpxchg) {
3060 // Same as for newval above:
3061 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3062 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3063 }
3064 // The only known value which might get overwritten is oldval.
3065 pre_barrier(false /* do_load */,
3066 control(), NULL, NULL, max_juint, NULL, NULL,
3067 oldval /* pre_val */,
3068 T_OBJECT);
3069 } else {
3070 ShouldNotReachHere();
3071 }
3072
3073 #ifdef _LP64
3074 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3075 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3076 if (kind == LS_xchg) {
3077 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3078 newval_enc, adr_type, value_type->make_narrowoop()));
3079 } else {
3080 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3081 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3082 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3083 newval_enc, oldval_enc));
3084 }
3085 } else
3086 #endif
3087 {
3088 if (kind == LS_xchg) {
3089 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3090 } else {
3091 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3092 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3093 }
3094 }
3095 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3096 break;
3097 default:
3098 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3099 break;
3100 }
3101
3102 // SCMemProjNodes represent the memory state of a LoadStore. Their
3103 // main role is to prevent LoadStore nodes from being optimized away
3104 // when their results aren't used.
3105 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3106 set_memory(proj, alias_idx);
3107
3108 if (type == T_OBJECT && kind == LS_xchg) {
3109 #ifdef _LP64
3110 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3111 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3112 }
3113 #endif
3114 if (can_move_pre_barrier()) {
3115 // Don't need to load pre_val. The old value is returned by load_store.
3116 // The pre_barrier can execute after the xchg as long as no safepoint
3117 // gets inserted between them.
3118 pre_barrier(false /* do_load */,
3119 control(), NULL, NULL, max_juint, NULL, NULL,
3120 load_store /* pre_val */,
3121 T_OBJECT);
3122 }
3123 }
3124
3125 // Add the trailing membar surrounding the access
3126 insert_mem_bar(Op_MemBarCPUOrder);
3127 insert_mem_bar(Op_MemBarAcquire);
3128
3129 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3130 set_result(load_store);
3131 return true;
3132 }
3133
3134 //----------------------------inline_unsafe_ordered_store----------------------
3135 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3136 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3137 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3138 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3139 // This is another variant of inline_unsafe_access, differing in
3140 // that it always issues store-store ("release") barrier and ensures
3141 // store-atomicity (which only matters for "long").
3142
3143 if (callee()->is_static()) return false; // caller must have the capability!
3144
3145 #ifndef PRODUCT
3146 {
3147 ResourceMark rm;
3148 // Check the signatures.
3149 ciSignature* sig = callee()->signature();
3150 #ifdef ASSERT
3151 BasicType rtype = sig->return_type()->basic_type();
3152 assert(rtype == T_VOID, "must return void");
3153 assert(sig->count() == 3, "has 3 arguments");
3154 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3155 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3156 #endif // ASSERT
3157 }
3158 #endif //PRODUCT
3159
3160 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3161
3162 // Get arguments:
3163 Node* receiver = argument(0); // type: oop
3164 Node* base = argument(1); // type: oop
3165 Node* offset = argument(2); // type: long
3166 Node* val = argument(4); // type: oop, int, or long
3167
3168 // Null check receiver.
3169 receiver = null_check(receiver);
3170 if (stopped()) {
3171 return true;
3172 }
3173
3174 // Build field offset expression.
3175 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3176 // 32-bit machines ignore the high half of long offsets
3177 offset = ConvL2X(offset);
3178 Node* adr = make_unsafe_address(base, offset);
3179 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3180 const Type *value_type = Type::get_const_basic_type(type);
3181 Compile::AliasType* alias_type = C->alias_type(adr_type);
3182
3183 insert_mem_bar(Op_MemBarRelease);
3184 insert_mem_bar(Op_MemBarCPUOrder);
3185 // Ensure that the store is atomic for longs:
3186 const bool require_atomic_access = true;
3187 Node* store;
3188 if (type == T_OBJECT) // reference stores need a store barrier.
3189 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3190 else {
3191 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3192 }
3193 insert_mem_bar(Op_MemBarCPUOrder);
3194 return true;
3195 }
3196
3197 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3198 // Regardless of form, don't allow previous ld/st to move down,
3199 // then issue acquire, release, or volatile mem_bar.
3200 insert_mem_bar(Op_MemBarCPUOrder);
3201 switch(id) {
3202 case vmIntrinsics::_loadFence:
3203 insert_mem_bar(Op_LoadFence);
3204 return true;
3205 case vmIntrinsics::_storeFence:
3206 insert_mem_bar(Op_StoreFence);
3207 return true;
3208 case vmIntrinsics::_fullFence:
3209 insert_mem_bar(Op_MemBarVolatile);
3210 return true;
3211 default:
3212 fatal_unexpected_iid(id);
3213 return false;
3214 }
3215 }
3216
3217 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3218 if (!kls->is_Con()) {
3219 return true;
3220 }
3221 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3222 if (klsptr == NULL) {
3223 return true;
3224 }
3225 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3226 // don't need a guard for a klass that is already initialized
3227 return !ik->is_initialized();
3228 }
3229
3230 //----------------------------inline_unsafe_allocate---------------------------
3231 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3232 bool LibraryCallKit::inline_unsafe_allocate() {
3233 if (callee()->is_static()) return false; // caller must have the capability!
3234
3235 null_check_receiver(); // null-check, then ignore
3236 Node* cls = null_check(argument(1));
3237 if (stopped()) return true;
3238
3239 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3240 kls = null_check(kls);
3241 if (stopped()) return true; // argument was like int.class
3242
3243 Node* test = NULL;
3244 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3245 // Note: The argument might still be an illegal value like
3246 // Serializable.class or Object[].class. The runtime will handle it.
3247 // But we must make an explicit check for initialization.
3248 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3249 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3250 // can generate code to load it as unsigned byte.
3251 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3252 Node* bits = intcon(InstanceKlass::fully_initialized);
3253 test = _gvn.transform(new (C) SubINode(inst, bits));
3254 // The 'test' is non-zero if we need to take a slow path.
3255 }
3256
3257 Node* obj = new_instance(kls, test);
3258 set_result(obj);
3259 return true;
3260 }
3261
3262 #ifdef TRACE_HAVE_INTRINSICS
3263 /*
3264 * oop -> myklass
3265 * myklass->trace_id |= USED
3266 * return myklass->trace_id & ~0x3
3267 */
3268 bool LibraryCallKit::inline_native_classID() {
3269 null_check_receiver(); // null-check, then ignore
3270 Node* cls = null_check(argument(1), T_OBJECT);
3271 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3272 kls = null_check(kls, T_OBJECT);
3273 ByteSize offset = TRACE_ID_OFFSET;
3274 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3275 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3276 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3277 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3278 Node* clsused = longcon(0x01l); // set the class bit
3279 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3280
3281 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3282 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3283 set_result(andl);
3284 return true;
3285 }
3286
3287 bool LibraryCallKit::inline_native_threadID() {
3288 Node* tls_ptr = NULL;
3289 Node* cur_thr = generate_current_thread(tls_ptr);
3290 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3291 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3292 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3293
3294 Node* threadid = NULL;
3295 size_t thread_id_size = OSThread::thread_id_size();
3296 if (thread_id_size == (size_t) BytesPerLong) {
3297 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3298 } else if (thread_id_size == (size_t) BytesPerInt) {
3299 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3300 } else {
3301 ShouldNotReachHere();
3302 }
3303 set_result(threadid);
3304 return true;
3305 }
3306 #endif
3307
3308 //------------------------inline_native_time_funcs--------------
3309 // inline code for System.currentTimeMillis() and System.nanoTime()
3310 // these have the same type and signature
3311 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3312 const TypeFunc* tf = OptoRuntime::void_long_Type();
3313 const TypePtr* no_memory_effects = NULL;
3314 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3315 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3316 #ifdef ASSERT
3317 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3318 assert(value_top == top(), "second value must be top");
3319 #endif
3320 set_result(value);
3321 return true;
3322 }
3323
3324 //------------------------inline_native_currentThread------------------
3325 bool LibraryCallKit::inline_native_currentThread() {
3326 Node* junk = NULL;
3327 set_result(generate_current_thread(junk));
3328 return true;
3329 }
3330
3331 //------------------------inline_native_isInterrupted------------------
3332 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3333 bool LibraryCallKit::inline_native_isInterrupted() {
3334 // Add a fast path to t.isInterrupted(clear_int):
3335 // (t == Thread.current() &&
3336 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3337 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3338 // So, in the common case that the interrupt bit is false,
3339 // we avoid making a call into the VM. Even if the interrupt bit
3340 // is true, if the clear_int argument is false, we avoid the VM call.
3341 // However, if the receiver is not currentThread, we must call the VM,
3342 // because there must be some locking done around the operation.
3343
3344 // We only go to the fast case code if we pass two guards.
3345 // Paths which do not pass are accumulated in the slow_region.
3346
3347 enum {
3348 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3349 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3350 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3351 PATH_LIMIT
3352 };
3353
3354 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3355 // out of the function.
3356 insert_mem_bar(Op_MemBarCPUOrder);
3357
3358 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3359 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3360
3361 RegionNode* slow_region = new (C) RegionNode(1);
3362 record_for_igvn(slow_region);
3363
3364 // (a) Receiving thread must be the current thread.
3365 Node* rec_thr = argument(0);
3366 Node* tls_ptr = NULL;
3367 Node* cur_thr = generate_current_thread(tls_ptr);
3368 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3369 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3370
3371 generate_slow_guard(bol_thr, slow_region);
3372
3373 // (b) Interrupt bit on TLS must be false.
3374 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3375 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3376 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3377
3378 // Set the control input on the field _interrupted read to prevent it floating up.
3379 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3380 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3381 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3382
3383 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3384
3385 // First fast path: if (!TLS._interrupted) return false;
3386 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3387 result_rgn->init_req(no_int_result_path, false_bit);
3388 result_val->init_req(no_int_result_path, intcon(0));
3389
3390 // drop through to next case
3391 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3392
3393 #ifndef TARGET_OS_FAMILY_windows
3394 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3395 Node* clr_arg = argument(1);
3396 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3397 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3398 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3399
3400 // Second fast path: ... else if (!clear_int) return true;
3401 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3402 result_rgn->init_req(no_clear_result_path, false_arg);
3403 result_val->init_req(no_clear_result_path, intcon(1));
3404
3405 // drop through to next case
3406 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3407 #else
3408 // To return true on Windows you must read the _interrupted field
3409 // and check the the event state i.e. take the slow path.
3410 #endif // TARGET_OS_FAMILY_windows
3411
3412 // (d) Otherwise, go to the slow path.
3413 slow_region->add_req(control());
3414 set_control( _gvn.transform(slow_region));
3415
3416 if (stopped()) {
3417 // There is no slow path.
3418 result_rgn->init_req(slow_result_path, top());
3419 result_val->init_req(slow_result_path, top());
3420 } else {
3421 // non-virtual because it is a private non-static
3422 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3423
3424 Node* slow_val = set_results_for_java_call(slow_call);
3425 // this->control() comes from set_results_for_java_call
3426
3427 Node* fast_io = slow_call->in(TypeFunc::I_O);
3428 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3429
3430 // These two phis are pre-filled with copies of of the fast IO and Memory
3431 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3432 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3433
3434 result_rgn->init_req(slow_result_path, control());
3435 result_io ->init_req(slow_result_path, i_o());
3436 result_mem->init_req(slow_result_path, reset_memory());
3437 result_val->init_req(slow_result_path, slow_val);
3438
3439 set_all_memory(_gvn.transform(result_mem));
3440 set_i_o( _gvn.transform(result_io));
3441 }
3442
3443 C->set_has_split_ifs(true); // Has chance for split-if optimization
3444 set_result(result_rgn, result_val);
3445 return true;
3446 }
3447
3448 //---------------------------load_mirror_from_klass----------------------------
3449 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3450 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3451 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3452 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3453 }
3454
3455 //-----------------------load_klass_from_mirror_common-------------------------
3456 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3457 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3458 // and branch to the given path on the region.
3459 // If never_see_null, take an uncommon trap on null, so we can optimistically
3460 // compile for the non-null case.
3461 // If the region is NULL, force never_see_null = true.
3462 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3463 bool never_see_null,
3464 RegionNode* region,
3465 int null_path,
3466 int offset) {
3467 if (region == NULL) never_see_null = true;
3468 Node* p = basic_plus_adr(mirror, offset);
3469 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3470 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3471 Node* null_ctl = top();
3472 kls = null_check_oop(kls, &null_ctl, never_see_null);
3473 if (region != NULL) {
3474 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3475 region->init_req(null_path, null_ctl);
3476 } else {
3477 assert(null_ctl == top(), "no loose ends");
3478 }
3479 return kls;
3480 }
3481
3482 //--------------------(inline_native_Class_query helpers)---------------------
3483 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3484 // Fall through if (mods & mask) == bits, take the guard otherwise.
3485 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3486 // Branch around if the given klass has the given modifier bit set.
3487 // Like generate_guard, adds a new path onto the region.
3488 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3489 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3490 Node* mask = intcon(modifier_mask);
3491 Node* bits = intcon(modifier_bits);
3492 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3493 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3494 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3495 return generate_fair_guard(bol, region);
3496 }
3497 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3498 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3499 }
3500
3501 //-------------------------inline_native_Class_query-------------------
3502 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3503 const Type* return_type = TypeInt::BOOL;
3504 Node* prim_return_value = top(); // what happens if it's a primitive class?
3505 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3506 bool expect_prim = false; // most of these guys expect to work on refs
3507
3508 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3509
3510 Node* mirror = argument(0);
3511 Node* obj = top();
3512
3513 switch (id) {
3514 case vmIntrinsics::_isInstance:
3515 // nothing is an instance of a primitive type
3516 prim_return_value = intcon(0);
3517 obj = argument(1);
3518 break;
3519 case vmIntrinsics::_getModifiers:
3520 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3521 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3522 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3523 break;
3524 case vmIntrinsics::_isInterface:
3525 prim_return_value = intcon(0);
3526 break;
3527 case vmIntrinsics::_isArray:
3528 prim_return_value = intcon(0);
3529 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3530 break;
3531 case vmIntrinsics::_isPrimitive:
3532 prim_return_value = intcon(1);
3533 expect_prim = true; // obviously
3534 break;
3535 case vmIntrinsics::_getSuperclass:
3536 prim_return_value = null();
3537 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3538 break;
3539 case vmIntrinsics::_getComponentType:
3540 prim_return_value = null();
3541 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3542 break;
3543 case vmIntrinsics::_getClassAccessFlags:
3544 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3545 return_type = TypeInt::INT; // not bool! 6297094
3546 break;
3547 default:
3548 fatal_unexpected_iid(id);
3549 break;
3550 }
3551
3552 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3553 if (mirror_con == NULL) return false; // cannot happen?
3554
3555 #ifndef PRODUCT
3556 if (C->print_intrinsics() || C->print_inlining()) {
3557 ciType* k = mirror_con->java_mirror_type();
3558 if (k) {
3559 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3560 k->print_name();
3561 tty->cr();
3562 }
3563 }
3564 #endif
3565
3566 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3567 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3568 record_for_igvn(region);
3569 PhiNode* phi = new (C) PhiNode(region, return_type);
3570
3571 // The mirror will never be null of Reflection.getClassAccessFlags, however
3572 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3573 // if it is. See bug 4774291.
3574
3575 // For Reflection.getClassAccessFlags(), the null check occurs in
3576 // the wrong place; see inline_unsafe_access(), above, for a similar
3577 // situation.
3578 mirror = null_check(mirror);
3579 // If mirror or obj is dead, only null-path is taken.
3580 if (stopped()) return true;
3581
3582 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3583
3584 // Now load the mirror's klass metaobject, and null-check it.
3585 // Side-effects region with the control path if the klass is null.
3586 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3587 // If kls is null, we have a primitive mirror.
3588 phi->init_req(_prim_path, prim_return_value);
3589 if (stopped()) { set_result(region, phi); return true; }
3590 bool safe_for_replace = (region->in(_prim_path) == top());
3591
3592 Node* p; // handy temp
3593 Node* null_ctl;
3594
3595 // Now that we have the non-null klass, we can perform the real query.
3596 // For constant classes, the query will constant-fold in LoadNode::Value.
3597 Node* query_value = top();
3598 switch (id) {
3599 case vmIntrinsics::_isInstance:
3600 // nothing is an instance of a primitive type
3601 query_value = gen_instanceof(obj, kls, safe_for_replace);
3602 break;
3603
3604 case vmIntrinsics::_getModifiers:
3605 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3606 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3607 break;
3608
3609 case vmIntrinsics::_isInterface:
3610 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3611 if (generate_interface_guard(kls, region) != NULL)
3612 // A guard was added. If the guard is taken, it was an interface.
3613 phi->add_req(intcon(1));
3614 // If we fall through, it's a plain class.
3615 query_value = intcon(0);
3616 break;
3617
3618 case vmIntrinsics::_isArray:
3619 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3620 if (generate_array_guard(kls, region) != NULL)
3621 // A guard was added. If the guard is taken, it was an array.
3622 phi->add_req(intcon(1));
3623 // If we fall through, it's a plain class.
3624 query_value = intcon(0);
3625 break;
3626
3627 case vmIntrinsics::_isPrimitive:
3628 query_value = intcon(0); // "normal" path produces false
3629 break;
3630
3631 case vmIntrinsics::_getSuperclass:
3632 // The rules here are somewhat unfortunate, but we can still do better
3633 // with random logic than with a JNI call.
3634 // Interfaces store null or Object as _super, but must report null.
3635 // Arrays store an intermediate super as _super, but must report Object.
3636 // Other types can report the actual _super.
3637 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3638 if (generate_interface_guard(kls, region) != NULL)
3639 // A guard was added. If the guard is taken, it was an interface.
3640 phi->add_req(null());
3641 if (generate_array_guard(kls, region) != NULL)
3642 // A guard was added. If the guard is taken, it was an array.
3643 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3644 // If we fall through, it's a plain class. Get its _super.
3645 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3646 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3647 null_ctl = top();
3648 kls = null_check_oop(kls, &null_ctl);
3649 if (null_ctl != top()) {
3650 // If the guard is taken, Object.superClass is null (both klass and mirror).
3651 region->add_req(null_ctl);
3652 phi ->add_req(null());
3653 }
3654 if (!stopped()) {
3655 query_value = load_mirror_from_klass(kls);
3656 }
3657 break;
3658
3659 case vmIntrinsics::_getComponentType:
3660 if (generate_array_guard(kls, region) != NULL) {
3661 // Be sure to pin the oop load to the guard edge just created:
3662 Node* is_array_ctrl = region->in(region->req()-1);
3663 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3664 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3665 phi->add_req(cmo);
3666 }
3667 query_value = null(); // non-array case is null
3668 break;
3669
3670 case vmIntrinsics::_getClassAccessFlags:
3671 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3672 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3673 break;
3674
3675 default:
3676 fatal_unexpected_iid(id);
3677 break;
3678 }
3679
3680 // Fall-through is the normal case of a query to a real class.
3681 phi->init_req(1, query_value);
3682 region->init_req(1, control());
3683
3684 C->set_has_split_ifs(true); // Has chance for split-if optimization
3685 set_result(region, phi);
3686 return true;
3687 }
3688
3689 //--------------------------inline_native_subtype_check------------------------
3690 // This intrinsic takes the JNI calls out of the heart of
3691 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3692 bool LibraryCallKit::inline_native_subtype_check() {
3693 // Pull both arguments off the stack.
3694 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3695 args[0] = argument(0);
3696 args[1] = argument(1);
3697 Node* klasses[2]; // corresponding Klasses: superk, subk
3698 klasses[0] = klasses[1] = top();
3699
3700 enum {
3701 // A full decision tree on {superc is prim, subc is prim}:
3702 _prim_0_path = 1, // {P,N} => false
3703 // {P,P} & superc!=subc => false
3704 _prim_same_path, // {P,P} & superc==subc => true
3705 _prim_1_path, // {N,P} => false
3706 _ref_subtype_path, // {N,N} & subtype check wins => true
3707 _both_ref_path, // {N,N} & subtype check loses => false
3708 PATH_LIMIT
3709 };
3710
3711 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3712 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3713 record_for_igvn(region);
3714
3715 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3716 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3717 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3718
3719 // First null-check both mirrors and load each mirror's klass metaobject.
3720 int which_arg;
3721 for (which_arg = 0; which_arg <= 1; which_arg++) {
3722 Node* arg = args[which_arg];
3723 arg = null_check(arg);
3724 if (stopped()) break;
3725 args[which_arg] = arg;
3726
3727 Node* p = basic_plus_adr(arg, class_klass_offset);
3728 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3729 klasses[which_arg] = _gvn.transform(kls);
3730 }
3731
3732 // Having loaded both klasses, test each for null.
3733 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3734 for (which_arg = 0; which_arg <= 1; which_arg++) {
3735 Node* kls = klasses[which_arg];
3736 Node* null_ctl = top();
3737 kls = null_check_oop(kls, &null_ctl, never_see_null);
3738 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3739 region->init_req(prim_path, null_ctl);
3740 if (stopped()) break;
3741 klasses[which_arg] = kls;
3742 }
3743
3744 if (!stopped()) {
3745 // now we have two reference types, in klasses[0..1]
3746 Node* subk = klasses[1]; // the argument to isAssignableFrom
3747 Node* superk = klasses[0]; // the receiver
3748 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3749 // now we have a successful reference subtype check
3750 region->set_req(_ref_subtype_path, control());
3751 }
3752
3753 // If both operands are primitive (both klasses null), then
3754 // we must return true when they are identical primitives.
3755 // It is convenient to test this after the first null klass check.
3756 set_control(region->in(_prim_0_path)); // go back to first null check
3757 if (!stopped()) {
3758 // Since superc is primitive, make a guard for the superc==subc case.
3759 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3760 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3761 generate_guard(bol_eq, region, PROB_FAIR);
3762 if (region->req() == PATH_LIMIT+1) {
3763 // A guard was added. If the added guard is taken, superc==subc.
3764 region->swap_edges(PATH_LIMIT, _prim_same_path);
3765 region->del_req(PATH_LIMIT);
3766 }
3767 region->set_req(_prim_0_path, control()); // Not equal after all.
3768 }
3769
3770 // these are the only paths that produce 'true':
3771 phi->set_req(_prim_same_path, intcon(1));
3772 phi->set_req(_ref_subtype_path, intcon(1));
3773
3774 // pull together the cases:
3775 assert(region->req() == PATH_LIMIT, "sane region");
3776 for (uint i = 1; i < region->req(); i++) {
3777 Node* ctl = region->in(i);
3778 if (ctl == NULL || ctl == top()) {
3779 region->set_req(i, top());
3780 phi ->set_req(i, top());
3781 } else if (phi->in(i) == NULL) {
3782 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3783 }
3784 }
3785
3786 set_control(_gvn.transform(region));
3787 set_result(_gvn.transform(phi));
3788 return true;
3789 }
3790
3791 //---------------------generate_array_guard_common------------------------
3792 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3793 bool obj_array, bool not_array) {
3794 // If obj_array/non_array==false/false:
3795 // Branch around if the given klass is in fact an array (either obj or prim).
3796 // If obj_array/non_array==false/true:
3797 // Branch around if the given klass is not an array klass of any kind.
3798 // If obj_array/non_array==true/true:
3799 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3800 // If obj_array/non_array==true/false:
3801 // Branch around if the kls is an oop array (Object[] or subtype)
3802 //
3803 // Like generate_guard, adds a new path onto the region.
3804 jint layout_con = 0;
3805 Node* layout_val = get_layout_helper(kls, layout_con);
3806 if (layout_val == NULL) {
3807 bool query = (obj_array
3808 ? Klass::layout_helper_is_objArray(layout_con)
3809 : Klass::layout_helper_is_array(layout_con));
3810 if (query == not_array) {
3811 return NULL; // never a branch
3812 } else { // always a branch
3813 Node* always_branch = control();
3814 if (region != NULL)
3815 region->add_req(always_branch);
3816 set_control(top());
3817 return always_branch;
3818 }
3819 }
3820 // Now test the correct condition.
3821 jint nval = (obj_array
3822 ? ((jint)Klass::_lh_array_tag_type_value
3823 << Klass::_lh_array_tag_shift)
3824 : Klass::_lh_neutral_value);
3825 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3826 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3827 // invert the test if we are looking for a non-array
3828 if (not_array) btest = BoolTest(btest).negate();
3829 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3830 return generate_fair_guard(bol, region);
3831 }
3832
3833
3834 //-----------------------inline_native_newArray--------------------------
3835 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3836 bool LibraryCallKit::inline_native_newArray() {
3837 Node* mirror = argument(0);
3838 Node* count_val = argument(1);
3839
3840 mirror = null_check(mirror);
3841 // If mirror or obj is dead, only null-path is taken.
3842 if (stopped()) return true;
3843
3844 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3845 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3846 PhiNode* result_val = new(C) PhiNode(result_reg,
3847 TypeInstPtr::NOTNULL);
3848 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3849 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3850 TypePtr::BOTTOM);
3851
3852 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3853 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3854 result_reg, _slow_path);
3855 Node* normal_ctl = control();
3856 Node* no_array_ctl = result_reg->in(_slow_path);
3857
3858 // Generate code for the slow case. We make a call to newArray().
3859 set_control(no_array_ctl);
3860 if (!stopped()) {
3861 // Either the input type is void.class, or else the
3862 // array klass has not yet been cached. Either the
3863 // ensuing call will throw an exception, or else it
3864 // will cache the array klass for next time.
3865 PreserveJVMState pjvms(this);
3866 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3867 Node* slow_result = set_results_for_java_call(slow_call);
3868 // this->control() comes from set_results_for_java_call
3869 result_reg->set_req(_slow_path, control());
3870 result_val->set_req(_slow_path, slow_result);
3871 result_io ->set_req(_slow_path, i_o());
3872 result_mem->set_req(_slow_path, reset_memory());
3873 }
3874
3875 set_control(normal_ctl);
3876 if (!stopped()) {
3877 // Normal case: The array type has been cached in the java.lang.Class.
3878 // The following call works fine even if the array type is polymorphic.
3879 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3880 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3881 result_reg->init_req(_normal_path, control());
3882 result_val->init_req(_normal_path, obj);
3883 result_io ->init_req(_normal_path, i_o());
3884 result_mem->init_req(_normal_path, reset_memory());
3885 }
3886
3887 // Return the combined state.
3888 set_i_o( _gvn.transform(result_io) );
3889 set_all_memory( _gvn.transform(result_mem));
3890
3891 C->set_has_split_ifs(true); // Has chance for split-if optimization
3892 set_result(result_reg, result_val);
3893 return true;
3894 }
3895
3896 //----------------------inline_native_getLength--------------------------
3897 // public static native int java.lang.reflect.Array.getLength(Object array);
3898 bool LibraryCallKit::inline_native_getLength() {
3899 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3900
3901 Node* array = null_check(argument(0));
3902 // If array is dead, only null-path is taken.
3903 if (stopped()) return true;
3904
3905 // Deoptimize if it is a non-array.
3906 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3907
3908 if (non_array != NULL) {
3909 PreserveJVMState pjvms(this);
3910 set_control(non_array);
3911 uncommon_trap(Deoptimization::Reason_intrinsic,
3912 Deoptimization::Action_maybe_recompile);
3913 }
3914
3915 // If control is dead, only non-array-path is taken.
3916 if (stopped()) return true;
3917
3918 // The works fine even if the array type is polymorphic.
3919 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3920 Node* result = load_array_length(array);
3921
3922 C->set_has_split_ifs(true); // Has chance for split-if optimization
3923 set_result(result);
3924 return true;
3925 }
3926
3927 //------------------------inline_array_copyOf----------------------------
3928 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3929 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3930 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3931 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3932
3933 // Get the arguments.
3934 Node* original = argument(0);
3935 Node* start = is_copyOfRange? argument(1): intcon(0);
3936 Node* end = is_copyOfRange? argument(2): argument(1);
3937 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3938
3939 Node* newcopy = NULL;
3940
3941 // Set the original stack and the reexecute bit for the interpreter to reexecute
3942 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3943 { PreserveReexecuteState preexecs(this);
3944 jvms()->set_should_reexecute(true);
3945
3946 array_type_mirror = null_check(array_type_mirror);
3947 original = null_check(original);
3948
3949 // Check if a null path was taken unconditionally.
3950 if (stopped()) return true;
3951
3952 Node* orig_length = load_array_length(original);
3953
3954 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3955 klass_node = null_check(klass_node);
3956
3957 RegionNode* bailout = new (C) RegionNode(1);
3958 record_for_igvn(bailout);
3959
3960 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3961 // Bail out if that is so.
3962 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3963 if (not_objArray != NULL) {
3964 // Improve the klass node's type from the new optimistic assumption:
3965 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3966 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3967 Node* cast = new (C) CastPPNode(klass_node, akls);
3968 cast->init_req(0, control());
3969 klass_node = _gvn.transform(cast);
3970 }
3971
3972 // Bail out if either start or end is negative.
3973 generate_negative_guard(start, bailout, &start);
3974 generate_negative_guard(end, bailout, &end);
3975
3976 Node* length = end;
3977 if (_gvn.type(start) != TypeInt::ZERO) {
3978 length = _gvn.transform(new (C) SubINode(end, start));
3979 }
3980
3981 // Bail out if length is negative.
3982 // Without this the new_array would throw
3983 // NegativeArraySizeException but IllegalArgumentException is what
3984 // should be thrown
3985 generate_negative_guard(length, bailout, &length);
3986
3987 if (bailout->req() > 1) {
3988 PreserveJVMState pjvms(this);
3989 set_control(_gvn.transform(bailout));
3990 uncommon_trap(Deoptimization::Reason_intrinsic,
3991 Deoptimization::Action_maybe_recompile);
3992 }
3993
3994 if (!stopped()) {
3995 // How many elements will we copy from the original?
3996 // The answer is MinI(orig_length - start, length).
3997 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3998 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3999
4000 newcopy = new_array(klass_node, length, 0); // no argments to push
4001
4002 // Generate a direct call to the right arraycopy function(s).
4003 // We know the copy is disjoint but we might not know if the
4004 // oop stores need checking.
4005 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4006 // This will fail a store-check if x contains any non-nulls.
4007 bool disjoint_bases = true;
4008 // if start > orig_length then the length of the copy may be
4009 // negative.
4010 bool length_never_negative = !is_copyOfRange;
4011 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4012 original, start, newcopy, intcon(0), moved,
4013 disjoint_bases, length_never_negative);
4014 }
4015 } // original reexecute is set back here
4016
4017 C->set_has_split_ifs(true); // Has chance for split-if optimization
4018 if (!stopped()) {
4019 set_result(newcopy);
4020 }
4021 return true;
4022 }
4023
4024
4025 //----------------------generate_virtual_guard---------------------------
4026 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4027 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4028 RegionNode* slow_region) {
4029 ciMethod* method = callee();
4030 int vtable_index = method->vtable_index();
4031 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4032 err_msg_res("bad index %d", vtable_index));
4033 // Get the Method* out of the appropriate vtable entry.
4034 int entry_offset = (InstanceKlass::vtable_start_offset() +
4035 vtable_index*vtableEntry::size()) * wordSize +
4036 vtableEntry::method_offset_in_bytes();
4037 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4038 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4039
4040 // Compare the target method with the expected method (e.g., Object.hashCode).
4041 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4042
4043 Node* native_call = makecon(native_call_addr);
4044 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4045 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4046
4047 return generate_slow_guard(test_native, slow_region);
4048 }
4049
4050 //-----------------------generate_method_call----------------------------
4051 // Use generate_method_call to make a slow-call to the real
4052 // method if the fast path fails. An alternative would be to
4053 // use a stub like OptoRuntime::slow_arraycopy_Java.
4054 // This only works for expanding the current library call,
4055 // not another intrinsic. (E.g., don't use this for making an
4056 // arraycopy call inside of the copyOf intrinsic.)
4057 CallJavaNode*
4058 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4059 // When compiling the intrinsic method itself, do not use this technique.
4060 guarantee(callee() != C->method(), "cannot make slow-call to self");
4061
4062 ciMethod* method = callee();
4063 // ensure the JVMS we have will be correct for this call
4064 guarantee(method_id == method->intrinsic_id(), "must match");
4065
4066 const TypeFunc* tf = TypeFunc::make(method);
4067 CallJavaNode* slow_call;
4068 if (is_static) {
4069 assert(!is_virtual, "");
4070 slow_call = new(C) CallStaticJavaNode(C, tf,
4071 SharedRuntime::get_resolve_static_call_stub(),
4072 method, bci());
4073 } else if (is_virtual) {
4074 null_check_receiver();
4075 int vtable_index = Method::invalid_vtable_index;
4076 if (UseInlineCaches) {
4077 // Suppress the vtable call
4078 } else {
4079 // hashCode and clone are not a miranda methods,
4080 // so the vtable index is fixed.
4081 // No need to use the linkResolver to get it.
4082 vtable_index = method->vtable_index();
4083 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4084 err_msg_res("bad index %d", vtable_index));
4085 }
4086 slow_call = new(C) CallDynamicJavaNode(tf,
4087 SharedRuntime::get_resolve_virtual_call_stub(),
4088 method, vtable_index, bci());
4089 } else { // neither virtual nor static: opt_virtual
4090 null_check_receiver();
4091 slow_call = new(C) CallStaticJavaNode(C, tf,
4092 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4093 method, bci());
4094 slow_call->set_optimized_virtual(true);
4095 }
4096 set_arguments_for_java_call(slow_call);
4097 set_edges_for_java_call(slow_call);
4098 return slow_call;
4099 }
4100
4101
4102 /**
4103 * Build special case code for calls to hashCode on an object. This call may
4104 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4105 * slightly different code.
4106 */
4107 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4108 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4109 assert(!(is_virtual && is_static), "either virtual, special, or static");
4110
4111 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4112
4113 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4114 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4115 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4116 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4117 Node* obj = NULL;
4118 if (!is_static) {
4119 // Check for hashing null object
4120 obj = null_check_receiver();
4121 if (stopped()) return true; // unconditionally null
4122 result_reg->init_req(_null_path, top());
4123 result_val->init_req(_null_path, top());
4124 } else {
4125 // Do a null check, and return zero if null.
4126 // System.identityHashCode(null) == 0
4127 obj = argument(0);
4128 Node* null_ctl = top();
4129 obj = null_check_oop(obj, &null_ctl);
4130 result_reg->init_req(_null_path, null_ctl);
4131 result_val->init_req(_null_path, _gvn.intcon(0));
4132 }
4133
4134 // Unconditionally null? Then return right away.
4135 if (stopped()) {
4136 set_control( result_reg->in(_null_path));
4137 if (!stopped())
4138 set_result(result_val->in(_null_path));
4139 return true;
4140 }
4141
4142 // We only go to the fast case code if we pass a number of guards. The
4143 // paths which do not pass are accumulated in the slow_region.
4144 RegionNode* slow_region = new (C) RegionNode(1);
4145 record_for_igvn(slow_region);
4146
4147 // If this is a virtual call, we generate a funny guard. We pull out
4148 // the vtable entry corresponding to hashCode() from the target object.
4149 // If the target method which we are calling happens to be the native
4150 // Object hashCode() method, we pass the guard. We do not need this
4151 // guard for non-virtual calls -- the caller is known to be the native
4152 // Object hashCode().
4153 if (is_virtual) {
4154 // After null check, get the object's klass.
4155 Node* obj_klass = load_object_klass(obj);
4156 generate_virtual_guard(obj_klass, slow_region);
4157 }
4158
4159 // Get the header out of the object, use LoadMarkNode when available
4160 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4161 // The control of the load must be NULL. Otherwise, the load can move before
4162 // the null check after castPP removal.
4163 Node* no_ctrl = NULL;
4164 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4165
4166 // Test the header to see if it is unlocked.
4167 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4168 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4169 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4170 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4171 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4172
4173 generate_slow_guard(test_unlocked, slow_region);
4174
4175 // Get the hash value and check to see that it has been properly assigned.
4176 // We depend on hash_mask being at most 32 bits and avoid the use of
4177 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4178 // vm: see markOop.hpp.
4179 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4180 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4181 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4182 // This hack lets the hash bits live anywhere in the mark object now, as long
4183 // as the shift drops the relevant bits into the low 32 bits. Note that
4184 // Java spec says that HashCode is an int so there's no point in capturing
4185 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4186 hshifted_header = ConvX2I(hshifted_header);
4187 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4188
4189 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4190 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4191 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4192
4193 generate_slow_guard(test_assigned, slow_region);
4194
4195 Node* init_mem = reset_memory();
4196 // fill in the rest of the null path:
4197 result_io ->init_req(_null_path, i_o());
4198 result_mem->init_req(_null_path, init_mem);
4199
4200 result_val->init_req(_fast_path, hash_val);
4201 result_reg->init_req(_fast_path, control());
4202 result_io ->init_req(_fast_path, i_o());
4203 result_mem->init_req(_fast_path, init_mem);
4204
4205 // Generate code for the slow case. We make a call to hashCode().
4206 set_control(_gvn.transform(slow_region));
4207 if (!stopped()) {
4208 // No need for PreserveJVMState, because we're using up the present state.
4209 set_all_memory(init_mem);
4210 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4211 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4212 Node* slow_result = set_results_for_java_call(slow_call);
4213 // this->control() comes from set_results_for_java_call
4214 result_reg->init_req(_slow_path, control());
4215 result_val->init_req(_slow_path, slow_result);
4216 result_io ->set_req(_slow_path, i_o());
4217 result_mem ->set_req(_slow_path, reset_memory());
4218 }
4219
4220 // Return the combined state.
4221 set_i_o( _gvn.transform(result_io) );
4222 set_all_memory( _gvn.transform(result_mem));
4223
4224 set_result(result_reg, result_val);
4225 return true;
4226 }
4227
4228 //---------------------------inline_native_getClass----------------------------
4229 // public final native Class<?> java.lang.Object.getClass();
4230 //
4231 // Build special case code for calls to getClass on an object.
4232 bool LibraryCallKit::inline_native_getClass() {
4233 Node* obj = null_check_receiver();
4234 if (stopped()) return true;
4235 set_result(load_mirror_from_klass(load_object_klass(obj)));
4236 return true;
4237 }
4238
4239 //-----------------inline_native_Reflection_getCallerClass---------------------
4240 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4241 //
4242 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4243 //
4244 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4245 // in that it must skip particular security frames and checks for
4246 // caller sensitive methods.
4247 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4248 #ifndef PRODUCT
4249 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4250 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4251 }
4252 #endif
4253
4254 if (!jvms()->has_method()) {
4255 #ifndef PRODUCT
4256 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4257 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4258 }
4259 #endif
4260 return false;
4261 }
4262
4263 // Walk back up the JVM state to find the caller at the required
4264 // depth.
4265 JVMState* caller_jvms = jvms();
4266
4267 // Cf. JVM_GetCallerClass
4268 // NOTE: Start the loop at depth 1 because the current JVM state does
4269 // not include the Reflection.getCallerClass() frame.
4270 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4271 ciMethod* m = caller_jvms->method();
4272 switch (n) {
4273 case 0:
4274 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4275 break;
4276 case 1:
4277 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4278 if (!m->caller_sensitive()) {
4279 #ifndef PRODUCT
4280 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4281 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4282 }
4283 #endif
4284 return false; // bail-out; let JVM_GetCallerClass do the work
4285 }
4286 break;
4287 default:
4288 if (!m->is_ignored_by_security_stack_walk()) {
4289 // We have reached the desired frame; return the holder class.
4290 // Acquire method holder as java.lang.Class and push as constant.
4291 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4292 ciInstance* caller_mirror = caller_klass->java_mirror();
4293 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4294
4295 #ifndef PRODUCT
4296 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4297 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());
4298 tty->print_cr(" JVM state at this point:");
4299 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4300 ciMethod* m = jvms()->of_depth(i)->method();
4301 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4302 }
4303 }
4304 #endif
4305 return true;
4306 }
4307 break;
4308 }
4309 }
4310
4311 #ifndef PRODUCT
4312 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4313 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4314 tty->print_cr(" JVM state at this point:");
4315 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4316 ciMethod* m = jvms()->of_depth(i)->method();
4317 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4318 }
4319 }
4320 #endif
4321
4322 return false; // bail-out; let JVM_GetCallerClass do the work
4323 }
4324
4325 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4326 Node* arg = argument(0);
4327 Node* result = NULL;
4328
4329 switch (id) {
4330 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4331 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4332 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4333 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4334
4335 case vmIntrinsics::_doubleToLongBits: {
4336 // two paths (plus control) merge in a wood
4337 RegionNode *r = new (C) RegionNode(3);
4338 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4339
4340 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4341 // Build the boolean node
4342 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4343
4344 // Branch either way.
4345 // NaN case is less traveled, which makes all the difference.
4346 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4347 Node *opt_isnan = _gvn.transform(ifisnan);
4348 assert( opt_isnan->is_If(), "Expect an IfNode");
4349 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4350 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4351
4352 set_control(iftrue);
4353
4354 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4355 Node *slow_result = longcon(nan_bits); // return NaN
4356 phi->init_req(1, _gvn.transform( slow_result ));
4357 r->init_req(1, iftrue);
4358
4359 // Else fall through
4360 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4361 set_control(iffalse);
4362
4363 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4364 r->init_req(2, iffalse);
4365
4366 // Post merge
4367 set_control(_gvn.transform(r));
4368 record_for_igvn(r);
4369
4370 C->set_has_split_ifs(true); // Has chance for split-if optimization
4371 result = phi;
4372 assert(result->bottom_type()->isa_long(), "must be");
4373 break;
4374 }
4375
4376 case vmIntrinsics::_floatToIntBits: {
4377 // two paths (plus control) merge in a wood
4378 RegionNode *r = new (C) RegionNode(3);
4379 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4380
4381 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4382 // Build the boolean node
4383 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4384
4385 // Branch either way.
4386 // NaN case is less traveled, which makes all the difference.
4387 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4388 Node *opt_isnan = _gvn.transform(ifisnan);
4389 assert( opt_isnan->is_If(), "Expect an IfNode");
4390 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4391 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4392
4393 set_control(iftrue);
4394
4395 static const jint nan_bits = 0x7fc00000;
4396 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4397 phi->init_req(1, _gvn.transform( slow_result ));
4398 r->init_req(1, iftrue);
4399
4400 // Else fall through
4401 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4402 set_control(iffalse);
4403
4404 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4405 r->init_req(2, iffalse);
4406
4407 // Post merge
4408 set_control(_gvn.transform(r));
4409 record_for_igvn(r);
4410
4411 C->set_has_split_ifs(true); // Has chance for split-if optimization
4412 result = phi;
4413 assert(result->bottom_type()->isa_int(), "must be");
4414 break;
4415 }
4416
4417 default:
4418 fatal_unexpected_iid(id);
4419 break;
4420 }
4421 set_result(_gvn.transform(result));
4422 return true;
4423 }
4424
4425 #ifdef _LP64
4426 #define XTOP ,top() /*additional argument*/
4427 #else //_LP64
4428 #define XTOP /*no additional argument*/
4429 #endif //_LP64
4430
4431 //----------------------inline_unsafe_copyMemory-------------------------
4432 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4433 bool LibraryCallKit::inline_unsafe_copyMemory() {
4434 if (callee()->is_static()) return false; // caller must have the capability!
4435 null_check_receiver(); // null-check receiver
4436 if (stopped()) return true;
4437
4438 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4439
4440 Node* src_ptr = argument(1); // type: oop
4441 Node* src_off = ConvL2X(argument(2)); // type: long
4442 Node* dst_ptr = argument(4); // type: oop
4443 Node* dst_off = ConvL2X(argument(5)); // type: long
4444 Node* size = ConvL2X(argument(7)); // type: long
4445
4446 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4447 "fieldOffset must be byte-scaled");
4448
4449 Node* src = make_unsafe_address(src_ptr, src_off);
4450 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4451
4452 // Conservatively insert a memory barrier on all memory slices.
4453 // Do not let writes of the copy source or destination float below the copy.
4454 insert_mem_bar(Op_MemBarCPUOrder);
4455
4456 // Call it. Note that the length argument is not scaled.
4457 make_runtime_call(RC_LEAF|RC_NO_FP,
4458 OptoRuntime::fast_arraycopy_Type(),
4459 StubRoutines::unsafe_arraycopy(),
4460 "unsafe_arraycopy",
4461 TypeRawPtr::BOTTOM,
4462 src, dst, size XTOP);
4463
4464 // Do not let reads of the copy destination float above the copy.
4465 insert_mem_bar(Op_MemBarCPUOrder);
4466
4467 return true;
4468 }
4469
4470 //------------------------clone_coping-----------------------------------
4471 // Helper function for inline_native_clone.
4472 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4473 assert(obj_size != NULL, "");
4474 Node* raw_obj = alloc_obj->in(1);
4475 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4476
4477 AllocateNode* alloc = NULL;
4478 if (ReduceBulkZeroing) {
4479 // We will be completely responsible for initializing this object -
4480 // mark Initialize node as complete.
4481 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4482 // The object was just allocated - there should be no any stores!
4483 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4484 // Mark as complete_with_arraycopy so that on AllocateNode
4485 // expansion, we know this AllocateNode is initialized by an array
4486 // copy and a StoreStore barrier exists after the array copy.
4487 alloc->initialization()->set_complete_with_arraycopy();
4488 }
4489
4490 // Copy the fastest available way.
4491 // TODO: generate fields copies for small objects instead.
4492 Node* src = obj;
4493 Node* dest = alloc_obj;
4494 Node* size = _gvn.transform(obj_size);
4495
4496 // Exclude the header but include array length to copy by 8 bytes words.
4497 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4498 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4499 instanceOopDesc::base_offset_in_bytes();
4500 // base_off:
4501 // 8 - 32-bit VM
4502 // 12 - 64-bit VM, compressed klass
4503 // 16 - 64-bit VM, normal klass
4504 if (base_off % BytesPerLong != 0) {
4505 assert(UseCompressedClassPointers, "");
4506 if (is_array) {
4507 // Exclude length to copy by 8 bytes words.
4508 base_off += sizeof(int);
4509 } else {
4510 // Include klass to copy by 8 bytes words.
4511 base_off = instanceOopDesc::klass_offset_in_bytes();
4512 }
4513 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4514 }
4515 src = basic_plus_adr(src, base_off);
4516 dest = basic_plus_adr(dest, base_off);
4517
4518 // Compute the length also, if needed:
4519 Node* countx = size;
4520 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4521 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4522
4523 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4524 bool disjoint_bases = true;
4525 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4526 src, NULL, dest, NULL, countx,
4527 /*dest_uninitialized*/true);
4528
4529 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4530 if (card_mark) {
4531 assert(!is_array, "");
4532 // Put in store barrier for any and all oops we are sticking
4533 // into this object. (We could avoid this if we could prove
4534 // that the object type contains no oop fields at all.)
4535 Node* no_particular_value = NULL;
4536 Node* no_particular_field = NULL;
4537 int raw_adr_idx = Compile::AliasIdxRaw;
4538 post_barrier(control(),
4539 memory(raw_adr_type),
4540 alloc_obj,
4541 no_particular_field,
4542 raw_adr_idx,
4543 no_particular_value,
4544 T_OBJECT,
4545 false);
4546 }
4547
4548 // Do not let reads from the cloned object float above the arraycopy.
4549 if (alloc != NULL) {
4550 // Do not let stores that initialize this object be reordered with
4551 // a subsequent store that would make this object accessible by
4552 // other threads.
4553 // Record what AllocateNode this StoreStore protects so that
4554 // escape analysis can go from the MemBarStoreStoreNode to the
4555 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4556 // based on the escape status of the AllocateNode.
4557 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4558 } else {
4559 insert_mem_bar(Op_MemBarCPUOrder);
4560 }
4561 }
4562
4563 //------------------------inline_native_clone----------------------------
4564 // protected native Object java.lang.Object.clone();
4565 //
4566 // Here are the simple edge cases:
4567 // null receiver => normal trap
4568 // virtual and clone was overridden => slow path to out-of-line clone
4569 // not cloneable or finalizer => slow path to out-of-line Object.clone
4570 //
4571 // The general case has two steps, allocation and copying.
4572 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4573 //
4574 // Copying also has two cases, oop arrays and everything else.
4575 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4576 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4577 //
4578 // These steps fold up nicely if and when the cloned object's klass
4579 // can be sharply typed as an object array, a type array, or an instance.
4580 //
4581 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4582 PhiNode* result_val;
4583
4584 // Set the reexecute bit for the interpreter to reexecute
4585 // the bytecode that invokes Object.clone if deoptimization happens.
4586 { PreserveReexecuteState preexecs(this);
4587 jvms()->set_should_reexecute(true);
4588
4589 Node* obj = null_check_receiver();
4590 if (stopped()) return true;
4591
4592 Node* obj_klass = load_object_klass(obj);
4593 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4594 const TypeOopPtr* toop = ((tklass != NULL)
4595 ? tklass->as_instance_type()
4596 : TypeInstPtr::NOTNULL);
4597
4598 // Conservatively insert a memory barrier on all memory slices.
4599 // Do not let writes into the original float below the clone.
4600 insert_mem_bar(Op_MemBarCPUOrder);
4601
4602 // paths into result_reg:
4603 enum {
4604 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4605 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4606 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4607 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4608 PATH_LIMIT
4609 };
4610 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4611 result_val = new(C) PhiNode(result_reg,
4612 TypeInstPtr::NOTNULL);
4613 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4614 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4615 TypePtr::BOTTOM);
4616 record_for_igvn(result_reg);
4617
4618 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4619 int raw_adr_idx = Compile::AliasIdxRaw;
4620
4621 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4622 if (array_ctl != NULL) {
4623 // It's an array.
4624 PreserveJVMState pjvms(this);
4625 set_control(array_ctl);
4626 Node* obj_length = load_array_length(obj);
4627 Node* obj_size = NULL;
4628 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4629
4630 if (!use_ReduceInitialCardMarks()) {
4631 // If it is an oop array, it requires very special treatment,
4632 // because card marking is required on each card of the array.
4633 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4634 if (is_obja != NULL) {
4635 PreserveJVMState pjvms2(this);
4636 set_control(is_obja);
4637 // Generate a direct call to the right arraycopy function(s).
4638 bool disjoint_bases = true;
4639 bool length_never_negative = true;
4640 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4641 obj, intcon(0), alloc_obj, intcon(0),
4642 obj_length,
4643 disjoint_bases, length_never_negative);
4644 result_reg->init_req(_objArray_path, control());
4645 result_val->init_req(_objArray_path, alloc_obj);
4646 result_i_o ->set_req(_objArray_path, i_o());
4647 result_mem ->set_req(_objArray_path, reset_memory());
4648 }
4649 }
4650 // Otherwise, there are no card marks to worry about.
4651 // (We can dispense with card marks if we know the allocation
4652 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4653 // causes the non-eden paths to take compensating steps to
4654 // simulate a fresh allocation, so that no further
4655 // card marks are required in compiled code to initialize
4656 // the object.)
4657
4658 if (!stopped()) {
4659 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4660
4661 // Present the results of the copy.
4662 result_reg->init_req(_array_path, control());
4663 result_val->init_req(_array_path, alloc_obj);
4664 result_i_o ->set_req(_array_path, i_o());
4665 result_mem ->set_req(_array_path, reset_memory());
4666 }
4667 }
4668
4669 // We only go to the instance fast case code if we pass a number of guards.
4670 // The paths which do not pass are accumulated in the slow_region.
4671 RegionNode* slow_region = new (C) RegionNode(1);
4672 record_for_igvn(slow_region);
4673 if (!stopped()) {
4674 // It's an instance (we did array above). Make the slow-path tests.
4675 // If this is a virtual call, we generate a funny guard. We grab
4676 // the vtable entry corresponding to clone() from the target object.
4677 // If the target method which we are calling happens to be the
4678 // Object clone() method, we pass the guard. We do not need this
4679 // guard for non-virtual calls; the caller is known to be the native
4680 // Object clone().
4681 if (is_virtual) {
4682 generate_virtual_guard(obj_klass, slow_region);
4683 }
4684
4685 // The object must be cloneable and must not have a finalizer.
4686 // Both of these conditions may be checked in a single test.
4687 // We could optimize the cloneable test further, but we don't care.
4688 generate_access_flags_guard(obj_klass,
4689 // Test both conditions:
4690 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4691 // Must be cloneable but not finalizer:
4692 JVM_ACC_IS_CLONEABLE,
4693 slow_region);
4694 }
4695
4696 if (!stopped()) {
4697 // It's an instance, and it passed the slow-path tests.
4698 PreserveJVMState pjvms(this);
4699 Node* obj_size = NULL;
4700 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4701 // is reexecuted if deoptimization occurs and there could be problems when merging
4702 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4703 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4704
4705 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4706
4707 // Present the results of the slow call.
4708 result_reg->init_req(_instance_path, control());
4709 result_val->init_req(_instance_path, alloc_obj);
4710 result_i_o ->set_req(_instance_path, i_o());
4711 result_mem ->set_req(_instance_path, reset_memory());
4712 }
4713
4714 // Generate code for the slow case. We make a call to clone().
4715 set_control(_gvn.transform(slow_region));
4716 if (!stopped()) {
4717 PreserveJVMState pjvms(this);
4718 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4719 Node* slow_result = set_results_for_java_call(slow_call);
4720 // this->control() comes from set_results_for_java_call
4721 result_reg->init_req(_slow_path, control());
4722 result_val->init_req(_slow_path, slow_result);
4723 result_i_o ->set_req(_slow_path, i_o());
4724 result_mem ->set_req(_slow_path, reset_memory());
4725 }
4726
4727 // Return the combined state.
4728 set_control( _gvn.transform(result_reg));
4729 set_i_o( _gvn.transform(result_i_o));
4730 set_all_memory( _gvn.transform(result_mem));
4731 } // original reexecute is set back here
4732
4733 set_result(_gvn.transform(result_val));
4734 return true;
4735 }
4736
4737 //------------------------------basictype2arraycopy----------------------------
4738 address LibraryCallKit::basictype2arraycopy(BasicType t,
4739 Node* src_offset,
4740 Node* dest_offset,
4741 bool disjoint_bases,
4742 const char* &name,
4743 bool dest_uninitialized) {
4744 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4745 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4746
4747 bool aligned = false;
4748 bool disjoint = disjoint_bases;
4749
4750 // if the offsets are the same, we can treat the memory regions as
4751 // disjoint, because either the memory regions are in different arrays,
4752 // or they are identical (which we can treat as disjoint.) We can also
4753 // treat a copy with a destination index less that the source index
4754 // as disjoint since a low->high copy will work correctly in this case.
4755 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4756 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4757 // both indices are constants
4758 int s_offs = src_offset_inttype->get_con();
4759 int d_offs = dest_offset_inttype->get_con();
4760 int element_size = type2aelembytes(t);
4761 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4762 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4763 if (s_offs >= d_offs) disjoint = true;
4764 } else if (src_offset == dest_offset && src_offset != NULL) {
4765 // This can occur if the offsets are identical non-constants.
4766 disjoint = true;
4767 }
4768
4769 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4770 }
4771
4772
4773 //------------------------------inline_arraycopy-----------------------
4774 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4775 // Object dest, int destPos,
4776 // int length);
4777 bool LibraryCallKit::inline_arraycopy() {
4778 // Get the arguments.
4779 Node* src = argument(0); // type: oop
4780 Node* src_offset = argument(1); // type: int
4781 Node* dest = argument(2); // type: oop
4782 Node* dest_offset = argument(3); // type: int
4783 Node* length = argument(4); // type: int
4784
4785 // Compile time checks. If any of these checks cannot be verified at compile time,
4786 // we do not make a fast path for this call. Instead, we let the call remain as it
4787 // is. The checks we choose to mandate at compile time are:
4788 //
4789 // (1) src and dest are arrays.
4790 const Type* src_type = src->Value(&_gvn);
4791 const Type* dest_type = dest->Value(&_gvn);
4792 const TypeAryPtr* top_src = src_type->isa_aryptr();
4793 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4794
4795 // Do we have the type of src?
4796 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4797 // Do we have the type of dest?
4798 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4799 // Is the type for src from speculation?
4800 bool src_spec = false;
4801 // Is the type for dest from speculation?
4802 bool dest_spec = false;
4803
4804 if (!has_src || !has_dest) {
4805 // We don't have sufficient type information, let's see if
4806 // speculative types can help. We need to have types for both src
4807 // and dest so that it pays off.
4808
4809 // Do we already have or could we have type information for src
4810 bool could_have_src = has_src;
4811 // Do we already have or could we have type information for dest
4812 bool could_have_dest = has_dest;
4813
4814 ciKlass* src_k = NULL;
4815 if (!has_src) {
4816 src_k = src_type->speculative_type();
4817 if (src_k != NULL && src_k->is_array_klass()) {
4818 could_have_src = true;
4819 }
4820 }
4821
4822 ciKlass* dest_k = NULL;
4823 if (!has_dest) {
4824 dest_k = dest_type->speculative_type();
4825 if (dest_k != NULL && dest_k->is_array_klass()) {
4826 could_have_dest = true;
4827 }
4828 }
4829
4830 if (could_have_src && could_have_dest) {
4831 // This is going to pay off so emit the required guards
4832 if (!has_src) {
4833 src = maybe_cast_profiled_obj(src, src_k);
4834 src_type = _gvn.type(src);
4835 top_src = src_type->isa_aryptr();
4836 has_src = (top_src != NULL && top_src->klass() != NULL);
4837 src_spec = true;
4838 }
4839 if (!has_dest) {
4840 dest = maybe_cast_profiled_obj(dest, dest_k);
4841 dest_type = _gvn.type(dest);
4842 top_dest = dest_type->isa_aryptr();
4843 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4844 dest_spec = true;
4845 }
4846 }
4847 }
4848
4849 if (!has_src || !has_dest) {
4850 // Conservatively insert a memory barrier on all memory slices.
4851 // Do not let writes into the source float below the arraycopy.
4852 insert_mem_bar(Op_MemBarCPUOrder);
4853
4854 // Call StubRoutines::generic_arraycopy stub.
4855 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4856 src, src_offset, dest, dest_offset, length);
4857
4858 // Do not let reads from the destination float above the arraycopy.
4859 // Since we cannot type the arrays, we don't know which slices
4860 // might be affected. We could restrict this barrier only to those
4861 // memory slices which pertain to array elements--but don't bother.
4862 if (!InsertMemBarAfterArraycopy)
4863 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4864 insert_mem_bar(Op_MemBarCPUOrder);
4865 return true;
4866 }
4867
4868 // (2) src and dest arrays must have elements of the same BasicType
4869 // Figure out the size and type of the elements we will be copying.
4870 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4871 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4872 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4873 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4874
4875 if (src_elem != dest_elem || dest_elem == T_VOID) {
4876 // The component types are not the same or are not recognized. Punt.
4877 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4878 generate_slow_arraycopy(TypePtr::BOTTOM,
4879 src, src_offset, dest, dest_offset, length,
4880 /*dest_uninitialized*/false);
4881 return true;
4882 }
4883
4884 if (src_elem == T_OBJECT) {
4885 // If both arrays are object arrays then having the exact types
4886 // for both will remove the need for a subtype check at runtime
4887 // before the call and may make it possible to pick a faster copy
4888 // routine (without a subtype check on every element)
4889 // Do we have the exact type of src?
4890 bool could_have_src = src_spec;
4891 // Do we have the exact type of dest?
4892 bool could_have_dest = dest_spec;
4893 ciKlass* src_k = top_src->klass();
4894 ciKlass* dest_k = top_dest->klass();
4895 if (!src_spec) {
4896 src_k = src_type->speculative_type();
4897 if (src_k != NULL && src_k->is_array_klass()) {
4898 could_have_src = true;
4899 }
4900 }
4901 if (!dest_spec) {
4902 dest_k = dest_type->speculative_type();
4903 if (dest_k != NULL && dest_k->is_array_klass()) {
4904 could_have_dest = true;
4905 }
4906 }
4907 if (could_have_src && could_have_dest) {
4908 // If we can have both exact types, emit the missing guards
4909 if (could_have_src && !src_spec) {
4910 src = maybe_cast_profiled_obj(src, src_k);
4911 }
4912 if (could_have_dest && !dest_spec) {
4913 dest = maybe_cast_profiled_obj(dest, dest_k);
4914 }
4915 }
4916 }
4917
4918 //---------------------------------------------------------------------------
4919 // We will make a fast path for this call to arraycopy.
4920
4921 // We have the following tests left to perform:
4922 //
4923 // (3) src and dest must not be null.
4924 // (4) src_offset must not be negative.
4925 // (5) dest_offset must not be negative.
4926 // (6) length must not be negative.
4927 // (7) src_offset + length must not exceed length of src.
4928 // (8) dest_offset + length must not exceed length of dest.
4929 // (9) each element of an oop array must be assignable
4930
4931 RegionNode* slow_region = new (C) RegionNode(1);
4932 record_for_igvn(slow_region);
4933
4934 // (3) operands must not be null
4935 // We currently perform our null checks with the null_check routine.
4936 // This means that the null exceptions will be reported in the caller
4937 // rather than (correctly) reported inside of the native arraycopy call.
4938 // This should be corrected, given time. We do our null check with the
4939 // stack pointer restored.
4940 src = null_check(src, T_ARRAY);
4941 dest = null_check(dest, T_ARRAY);
4942
4943 // (4) src_offset must not be negative.
4944 generate_negative_guard(src_offset, slow_region);
4945
4946 // (5) dest_offset must not be negative.
4947 generate_negative_guard(dest_offset, slow_region);
4948
4949 // (6) length must not be negative (moved to generate_arraycopy()).
4950 // generate_negative_guard(length, slow_region);
4951
4952 // (7) src_offset + length must not exceed length of src.
4953 generate_limit_guard(src_offset, length,
4954 load_array_length(src),
4955 slow_region);
4956
4957 // (8) dest_offset + length must not exceed length of dest.
4958 generate_limit_guard(dest_offset, length,
4959 load_array_length(dest),
4960 slow_region);
4961
4962 // (9) each element of an oop array must be assignable
4963 // The generate_arraycopy subroutine checks this.
4964
4965 // This is where the memory effects are placed:
4966 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4967 generate_arraycopy(adr_type, dest_elem,
4968 src, src_offset, dest, dest_offset, length,
4969 false, false, slow_region);
4970
4971 return true;
4972 }
4973
4974 //-----------------------------generate_arraycopy----------------------
4975 // Generate an optimized call to arraycopy.
4976 // Caller must guard against non-arrays.
4977 // Caller must determine a common array basic-type for both arrays.
4978 // Caller must validate offsets against array bounds.
4979 // The slow_region has already collected guard failure paths
4980 // (such as out of bounds length or non-conformable array types).
4981 // The generated code has this shape, in general:
4982 //
4983 // if (length == 0) return // via zero_path
4984 // slowval = -1
4985 // if (types unknown) {
4986 // slowval = call generic copy loop
4987 // if (slowval == 0) return // via checked_path
4988 // } else if (indexes in bounds) {
4989 // if ((is object array) && !(array type check)) {
4990 // slowval = call checked copy loop
4991 // if (slowval == 0) return // via checked_path
4992 // } else {
4993 // call bulk copy loop
4994 // return // via fast_path
4995 // }
4996 // }
4997 // // adjust params for remaining work:
4998 // if (slowval != -1) {
4999 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
5000 // }
5001 // slow_region:
5002 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
5003 // return // via slow_call_path
5004 //
5005 // This routine is used from several intrinsics: System.arraycopy,
5006 // Object.clone (the array subcase), and Arrays.copyOf[Range].
5007 //
5008 void
5009 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
5010 BasicType basic_elem_type,
5011 Node* src, Node* src_offset,
5012 Node* dest, Node* dest_offset,
5013 Node* copy_length,
5014 bool disjoint_bases,
5015 bool length_never_negative,
5016 RegionNode* slow_region) {
5017
5018 if (slow_region == NULL) {
5019 slow_region = new(C) RegionNode(1);
5020 record_for_igvn(slow_region);
5021 }
5022
5023 Node* original_dest = dest;
5024 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5025 bool dest_uninitialized = false;
5026
5027 // See if this is the initialization of a newly-allocated array.
5028 // If so, we will take responsibility here for initializing it to zero.
5029 // (Note: Because tightly_coupled_allocation performs checks on the
5030 // out-edges of the dest, we need to avoid making derived pointers
5031 // from it until we have checked its uses.)
5032 if (ReduceBulkZeroing
5033 && !ZeroTLAB // pointless if already zeroed
5034 && basic_elem_type != T_CONFLICT // avoid corner case
5035 && !src->eqv_uncast(dest)
5036 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5037 != NULL)
5038 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5039 && alloc->maybe_set_complete(&_gvn)) {
5040 // "You break it, you buy it."
5041 InitializeNode* init = alloc->initialization();
5042 assert(init->is_complete(), "we just did this");
5043 init->set_complete_with_arraycopy();
5044 assert(dest->is_CheckCastPP(), "sanity");
5045 assert(dest->in(0)->in(0) == init, "dest pinned");
5046 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5047 // From this point on, every exit path is responsible for
5048 // initializing any non-copied parts of the object to zero.
5049 // Also, if this flag is set we make sure that arraycopy interacts properly
5050 // with G1, eliding pre-barriers. See CR 6627983.
5051 dest_uninitialized = true;
5052 } else {
5053 // No zeroing elimination here.
5054 alloc = NULL;
5055 //original_dest = dest;
5056 //dest_uninitialized = false;
5057 }
5058
5059 // Results are placed here:
5060 enum { fast_path = 1, // normal void-returning assembly stub
5061 checked_path = 2, // special assembly stub with cleanup
5062 slow_call_path = 3, // something went wrong; call the VM
5063 zero_path = 4, // bypass when length of copy is zero
5064 bcopy_path = 5, // copy primitive array by 64-bit blocks
5065 PATH_LIMIT = 6
5066 };
5067 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5068 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5069 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5070 record_for_igvn(result_region);
5071 _gvn.set_type_bottom(result_i_o);
5072 _gvn.set_type_bottom(result_memory);
5073 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5074
5075 // The slow_control path:
5076 Node* slow_control;
5077 Node* slow_i_o = i_o();
5078 Node* slow_mem = memory(adr_type);
5079 debug_only(slow_control = (Node*) badAddress);
5080
5081 // Checked control path:
5082 Node* checked_control = top();
5083 Node* checked_mem = NULL;
5084 Node* checked_i_o = NULL;
5085 Node* checked_value = NULL;
5086
5087 if (basic_elem_type == T_CONFLICT) {
5088 assert(!dest_uninitialized, "");
5089 Node* cv = generate_generic_arraycopy(adr_type,
5090 src, src_offset, dest, dest_offset,
5091 copy_length, dest_uninitialized);
5092 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5093 checked_control = control();
5094 checked_i_o = i_o();
5095 checked_mem = memory(adr_type);
5096 checked_value = cv;
5097 set_control(top()); // no fast path
5098 }
5099
5100 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5101 if (not_pos != NULL) {
5102 PreserveJVMState pjvms(this);
5103 set_control(not_pos);
5104
5105 // (6) length must not be negative.
5106 if (!length_never_negative) {
5107 generate_negative_guard(copy_length, slow_region);
5108 }
5109
5110 // copy_length is 0.
5111 if (!stopped() && dest_uninitialized) {
5112 Node* dest_length = alloc->in(AllocateNode::ALength);
5113 if (copy_length->eqv_uncast(dest_length)
5114 || _gvn.find_int_con(dest_length, 1) <= 0) {
5115 // There is no zeroing to do. No need for a secondary raw memory barrier.
5116 } else {
5117 // Clear the whole thing since there are no source elements to copy.
5118 generate_clear_array(adr_type, dest, basic_elem_type,
5119 intcon(0), NULL,
5120 alloc->in(AllocateNode::AllocSize));
5121 // Use a secondary InitializeNode as raw memory barrier.
5122 // Currently it is needed only on this path since other
5123 // paths have stub or runtime calls as raw memory barriers.
5124 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5125 Compile::AliasIdxRaw,
5126 top())->as_Initialize();
5127 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5128 }
5129 }
5130
5131 // Present the results of the fast call.
5132 result_region->init_req(zero_path, control());
5133 result_i_o ->init_req(zero_path, i_o());
5134 result_memory->init_req(zero_path, memory(adr_type));
5135 }
5136
5137 if (!stopped() && dest_uninitialized) {
5138 // We have to initialize the *uncopied* part of the array to zero.
5139 // The copy destination is the slice dest[off..off+len]. The other slices
5140 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5141 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5142 Node* dest_length = alloc->in(AllocateNode::ALength);
5143 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5144 copy_length));
5145
5146 // If there is a head section that needs zeroing, do it now.
5147 if (find_int_con(dest_offset, -1) != 0) {
5148 generate_clear_array(adr_type, dest, basic_elem_type,
5149 intcon(0), dest_offset,
5150 NULL);
5151 }
5152
5153 // Next, perform a dynamic check on the tail length.
5154 // It is often zero, and we can win big if we prove this.
5155 // There are two wins: Avoid generating the ClearArray
5156 // with its attendant messy index arithmetic, and upgrade
5157 // the copy to a more hardware-friendly word size of 64 bits.
5158 Node* tail_ctl = NULL;
5159 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5160 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5161 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5162 tail_ctl = generate_slow_guard(bol_lt, NULL);
5163 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5164 }
5165
5166 // At this point, let's assume there is no tail.
5167 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5168 // There is no tail. Try an upgrade to a 64-bit copy.
5169 bool didit = false;
5170 { PreserveJVMState pjvms(this);
5171 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5172 src, src_offset, dest, dest_offset,
5173 dest_size, dest_uninitialized);
5174 if (didit) {
5175 // Present the results of the block-copying fast call.
5176 result_region->init_req(bcopy_path, control());
5177 result_i_o ->init_req(bcopy_path, i_o());
5178 result_memory->init_req(bcopy_path, memory(adr_type));
5179 }
5180 }
5181 if (didit)
5182 set_control(top()); // no regular fast path
5183 }
5184
5185 // Clear the tail, if any.
5186 if (tail_ctl != NULL) {
5187 Node* notail_ctl = stopped() ? NULL : control();
5188 set_control(tail_ctl);
5189 if (notail_ctl == NULL) {
5190 generate_clear_array(adr_type, dest, basic_elem_type,
5191 dest_tail, NULL,
5192 dest_size);
5193 } else {
5194 // Make a local merge.
5195 Node* done_ctl = new(C) RegionNode(3);
5196 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5197 done_ctl->init_req(1, notail_ctl);
5198 done_mem->init_req(1, memory(adr_type));
5199 generate_clear_array(adr_type, dest, basic_elem_type,
5200 dest_tail, NULL,
5201 dest_size);
5202 done_ctl->init_req(2, control());
5203 done_mem->init_req(2, memory(adr_type));
5204 set_control( _gvn.transform(done_ctl));
5205 set_memory( _gvn.transform(done_mem), adr_type );
5206 }
5207 }
5208 }
5209
5210 BasicType copy_type = basic_elem_type;
5211 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5212 if (!stopped() && copy_type == T_OBJECT) {
5213 // If src and dest have compatible element types, we can copy bits.
5214 // Types S[] and D[] are compatible if D is a supertype of S.
5215 //
5216 // If they are not, we will use checked_oop_disjoint_arraycopy,
5217 // which performs a fast optimistic per-oop check, and backs off
5218 // further to JVM_ArrayCopy on the first per-oop check that fails.
5219 // (Actually, we don't move raw bits only; the GC requires card marks.)
5220
5221 // Get the Klass* for both src and dest
5222 Node* src_klass = load_object_klass(src);
5223 Node* dest_klass = load_object_klass(dest);
5224
5225 // Generate the subtype check.
5226 // This might fold up statically, or then again it might not.
5227 //
5228 // Non-static example: Copying List<String>.elements to a new String[].
5229 // The backing store for a List<String> is always an Object[],
5230 // but its elements are always type String, if the generic types
5231 // are correct at the source level.
5232 //
5233 // Test S[] against D[], not S against D, because (probably)
5234 // the secondary supertype cache is less busy for S[] than S.
5235 // This usually only matters when D is an interface.
5236 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5237 // Plug failing path into checked_oop_disjoint_arraycopy
5238 if (not_subtype_ctrl != top()) {
5239 PreserveJVMState pjvms(this);
5240 set_control(not_subtype_ctrl);
5241 // (At this point we can assume disjoint_bases, since types differ.)
5242 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5243 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5244 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5245 Node* dest_elem_klass = _gvn.transform(n1);
5246 Node* cv = generate_checkcast_arraycopy(adr_type,
5247 dest_elem_klass,
5248 src, src_offset, dest, dest_offset,
5249 ConvI2X(copy_length), dest_uninitialized);
5250 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5251 checked_control = control();
5252 checked_i_o = i_o();
5253 checked_mem = memory(adr_type);
5254 checked_value = cv;
5255 }
5256 // At this point we know we do not need type checks on oop stores.
5257
5258 // Let's see if we need card marks:
5259 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5260 // If we do not need card marks, copy using the jint or jlong stub.
5261 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5262 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5263 "sizes agree");
5264 }
5265 }
5266
5267 if (!stopped()) {
5268 // Generate the fast path, if possible.
5269 PreserveJVMState pjvms(this);
5270 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5271 src, src_offset, dest, dest_offset,
5272 ConvI2X(copy_length), dest_uninitialized);
5273
5274 // Present the results of the fast call.
5275 result_region->init_req(fast_path, control());
5276 result_i_o ->init_req(fast_path, i_o());
5277 result_memory->init_req(fast_path, memory(adr_type));
5278 }
5279
5280 // Here are all the slow paths up to this point, in one bundle:
5281 slow_control = top();
5282 if (slow_region != NULL)
5283 slow_control = _gvn.transform(slow_region);
5284 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5285
5286 set_control(checked_control);
5287 if (!stopped()) {
5288 // Clean up after the checked call.
5289 // The returned value is either 0 or -1^K,
5290 // where K = number of partially transferred array elements.
5291 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5292 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5293 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5294
5295 // If it is 0, we are done, so transfer to the end.
5296 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5297 result_region->init_req(checked_path, checks_done);
5298 result_i_o ->init_req(checked_path, checked_i_o);
5299 result_memory->init_req(checked_path, checked_mem);
5300
5301 // If it is not zero, merge into the slow call.
5302 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5303 RegionNode* slow_reg2 = new(C) RegionNode(3);
5304 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5305 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5306 record_for_igvn(slow_reg2);
5307 slow_reg2 ->init_req(1, slow_control);
5308 slow_i_o2 ->init_req(1, slow_i_o);
5309 slow_mem2 ->init_req(1, slow_mem);
5310 slow_reg2 ->init_req(2, control());
5311 slow_i_o2 ->init_req(2, checked_i_o);
5312 slow_mem2 ->init_req(2, checked_mem);
5313
5314 slow_control = _gvn.transform(slow_reg2);
5315 slow_i_o = _gvn.transform(slow_i_o2);
5316 slow_mem = _gvn.transform(slow_mem2);
5317
5318 if (alloc != NULL) {
5319 // We'll restart from the very beginning, after zeroing the whole thing.
5320 // This can cause double writes, but that's OK since dest is brand new.
5321 // So we ignore the low 31 bits of the value returned from the stub.
5322 } else {
5323 // We must continue the copy exactly where it failed, or else
5324 // another thread might see the wrong number of writes to dest.
5325 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5326 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5327 slow_offset->init_req(1, intcon(0));
5328 slow_offset->init_req(2, checked_offset);
5329 slow_offset = _gvn.transform(slow_offset);
5330
5331 // Adjust the arguments by the conditionally incoming offset.
5332 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5333 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5334 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5335
5336 // Tweak the node variables to adjust the code produced below:
5337 src_offset = src_off_plus;
5338 dest_offset = dest_off_plus;
5339 copy_length = length_minus;
5340 }
5341 }
5342
5343 set_control(slow_control);
5344 if (!stopped()) {
5345 // Generate the slow path, if needed.
5346 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5347
5348 set_memory(slow_mem, adr_type);
5349 set_i_o(slow_i_o);
5350
5351 if (dest_uninitialized) {
5352 generate_clear_array(adr_type, dest, basic_elem_type,
5353 intcon(0), NULL,
5354 alloc->in(AllocateNode::AllocSize));
5355 }
5356
5357 generate_slow_arraycopy(adr_type,
5358 src, src_offset, dest, dest_offset,
5359 copy_length, /*dest_uninitialized*/false);
5360
5361 result_region->init_req(slow_call_path, control());
5362 result_i_o ->init_req(slow_call_path, i_o());
5363 result_memory->init_req(slow_call_path, memory(adr_type));
5364 }
5365
5366 // Remove unused edges.
5367 for (uint i = 1; i < result_region->req(); i++) {
5368 if (result_region->in(i) == NULL)
5369 result_region->init_req(i, top());
5370 }
5371
5372 // Finished; return the combined state.
5373 set_control( _gvn.transform(result_region));
5374 set_i_o( _gvn.transform(result_i_o) );
5375 set_memory( _gvn.transform(result_memory), adr_type );
5376
5377 // The memory edges above are precise in order to model effects around
5378 // array copies accurately to allow value numbering of field loads around
5379 // arraycopy. Such field loads, both before and after, are common in Java
5380 // collections and similar classes involving header/array data structures.
5381 //
5382 // But with low number of register or when some registers are used or killed
5383 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5384 // The next memory barrier is added to avoid it. If the arraycopy can be
5385 // optimized away (which it can, sometimes) then we can manually remove
5386 // the membar also.
5387 //
5388 // Do not let reads from the cloned object float above the arraycopy.
5389 if (alloc != NULL) {
5390 // Do not let stores that initialize this object be reordered with
5391 // a subsequent store that would make this object accessible by
5392 // other threads.
5393 // Record what AllocateNode this StoreStore protects so that
5394 // escape analysis can go from the MemBarStoreStoreNode to the
5395 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5396 // based on the escape status of the AllocateNode.
5397 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5398 } else if (InsertMemBarAfterArraycopy)
5399 insert_mem_bar(Op_MemBarCPUOrder);
5400 }
5401
5402
5403 // Helper function which determines if an arraycopy immediately follows
5404 // an allocation, with no intervening tests or other escapes for the object.
5405 AllocateArrayNode*
5406 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5407 RegionNode* slow_region) {
5408 if (stopped()) return NULL; // no fast path
5409 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5410
5411 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5412 if (alloc == NULL) return NULL;
5413
5414 Node* rawmem = memory(Compile::AliasIdxRaw);
5415 // Is the allocation's memory state untouched?
5416 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5417 // Bail out if there have been raw-memory effects since the allocation.
5418 // (Example: There might have been a call or safepoint.)
5419 return NULL;
5420 }
5421 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5422 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5423 return NULL;
5424 }
5425
5426 // There must be no unexpected observers of this allocation.
5427 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5428 Node* obs = ptr->fast_out(i);
5429 if (obs != this->map()) {
5430 return NULL;
5431 }
5432 }
5433
5434 // This arraycopy must unconditionally follow the allocation of the ptr.
5435 Node* alloc_ctl = ptr->in(0);
5436 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5437
5438 Node* ctl = control();
5439 while (ctl != alloc_ctl) {
5440 // There may be guards which feed into the slow_region.
5441 // Any other control flow means that we might not get a chance
5442 // to finish initializing the allocated object.
5443 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5444 IfNode* iff = ctl->in(0)->as_If();
5445 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5446 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5447 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5448 ctl = iff->in(0); // This test feeds the known slow_region.
5449 continue;
5450 }
5451 // One more try: Various low-level checks bottom out in
5452 // uncommon traps. If the debug-info of the trap omits
5453 // any reference to the allocation, as we've already
5454 // observed, then there can be no objection to the trap.
5455 bool found_trap = false;
5456 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5457 Node* obs = not_ctl->fast_out(j);
5458 if (obs->in(0) == not_ctl && obs->is_Call() &&
5459 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5460 found_trap = true; break;
5461 }
5462 }
5463 if (found_trap) {
5464 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5465 continue;
5466 }
5467 }
5468 return NULL;
5469 }
5470
5471 // If we get this far, we have an allocation which immediately
5472 // precedes the arraycopy, and we can take over zeroing the new object.
5473 // The arraycopy will finish the initialization, and provide
5474 // a new control state to which we will anchor the destination pointer.
5475
5476 return alloc;
5477 }
5478
5479 // Helper for initialization of arrays, creating a ClearArray.
5480 // It writes zero bits in [start..end), within the body of an array object.
5481 // The memory effects are all chained onto the 'adr_type' alias category.
5482 //
5483 // Since the object is otherwise uninitialized, we are free
5484 // to put a little "slop" around the edges of the cleared area,
5485 // as long as it does not go back into the array's header,
5486 // or beyond the array end within the heap.
5487 //
5488 // The lower edge can be rounded down to the nearest jint and the
5489 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5490 //
5491 // Arguments:
5492 // adr_type memory slice where writes are generated
5493 // dest oop of the destination array
5494 // basic_elem_type element type of the destination
5495 // slice_idx array index of first element to store
5496 // slice_len number of elements to store (or NULL)
5497 // dest_size total size in bytes of the array object
5498 //
5499 // Exactly one of slice_len or dest_size must be non-NULL.
5500 // If dest_size is non-NULL, zeroing extends to the end of the object.
5501 // If slice_len is non-NULL, the slice_idx value must be a constant.
5502 void
5503 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5504 Node* dest,
5505 BasicType basic_elem_type,
5506 Node* slice_idx,
5507 Node* slice_len,
5508 Node* dest_size) {
5509 // one or the other but not both of slice_len and dest_size:
5510 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5511 if (slice_len == NULL) slice_len = top();
5512 if (dest_size == NULL) dest_size = top();
5513
5514 // operate on this memory slice:
5515 Node* mem = memory(adr_type); // memory slice to operate on
5516
5517 // scaling and rounding of indexes:
5518 int scale = exact_log2(type2aelembytes(basic_elem_type));
5519 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5520 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5521 int bump_bit = (-1 << scale) & BytesPerInt;
5522
5523 // determine constant starts and ends
5524 const intptr_t BIG_NEG = -128;
5525 assert(BIG_NEG + 2*abase < 0, "neg enough");
5526 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5527 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5528 if (slice_len_con == 0) {
5529 return; // nothing to do here
5530 }
5531 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5532 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5533 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5534 assert(end_con < 0, "not two cons");
5535 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5536 BytesPerLong);
5537 }
5538
5539 if (start_con >= 0 && end_con >= 0) {
5540 // Constant start and end. Simple.
5541 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5542 start_con, end_con, &_gvn);
5543 } else if (start_con >= 0 && dest_size != top()) {
5544 // Constant start, pre-rounded end after the tail of the array.
5545 Node* end = dest_size;
5546 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5547 start_con, end, &_gvn);
5548 } else if (start_con >= 0 && slice_len != top()) {
5549 // Constant start, non-constant end. End needs rounding up.
5550 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5551 intptr_t end_base = abase + (slice_idx_con << scale);
5552 int end_round = (-1 << scale) & (BytesPerLong - 1);
5553 Node* end = ConvI2X(slice_len);
5554 if (scale != 0)
5555 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5556 end_base += end_round;
5557 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5558 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5559 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5560 start_con, end, &_gvn);
5561 } else if (start_con < 0 && dest_size != top()) {
5562 // Non-constant start, pre-rounded end after the tail of the array.
5563 // This is almost certainly a "round-to-end" operation.
5564 Node* start = slice_idx;
5565 start = ConvI2X(start);
5566 if (scale != 0)
5567 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5568 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5569 if ((bump_bit | clear_low) != 0) {
5570 int to_clear = (bump_bit | clear_low);
5571 // Align up mod 8, then store a jint zero unconditionally
5572 // just before the mod-8 boundary.
5573 if (((abase + bump_bit) & ~to_clear) - bump_bit
5574 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5575 bump_bit = 0;
5576 assert((abase & to_clear) == 0, "array base must be long-aligned");
5577 } else {
5578 // Bump 'start' up to (or past) the next jint boundary:
5579 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5580 assert((abase & clear_low) == 0, "array base must be int-aligned");
5581 }
5582 // Round bumped 'start' down to jlong boundary in body of array.
5583 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5584 if (bump_bit != 0) {
5585 // Store a zero to the immediately preceding jint:
5586 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5587 Node* p1 = basic_plus_adr(dest, x1);
5588 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5589 mem = _gvn.transform(mem);
5590 }
5591 }
5592 Node* end = dest_size; // pre-rounded
5593 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5594 start, end, &_gvn);
5595 } else {
5596 // Non-constant start, unrounded non-constant end.
5597 // (Nobody zeroes a random midsection of an array using this routine.)
5598 ShouldNotReachHere(); // fix caller
5599 }
5600
5601 // Done.
5602 set_memory(mem, adr_type);
5603 }
5604
5605
5606 bool
5607 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5608 BasicType basic_elem_type,
5609 AllocateNode* alloc,
5610 Node* src, Node* src_offset,
5611 Node* dest, Node* dest_offset,
5612 Node* dest_size, bool dest_uninitialized) {
5613 // See if there is an advantage from block transfer.
5614 int scale = exact_log2(type2aelembytes(basic_elem_type));
5615 if (scale >= LogBytesPerLong)
5616 return false; // it is already a block transfer
5617
5618 // Look at the alignment of the starting offsets.
5619 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5620
5621 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5622 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5623 if (src_off_con < 0 || dest_off_con < 0)
5624 // At present, we can only understand constants.
5625 return false;
5626
5627 intptr_t src_off = abase + (src_off_con << scale);
5628 intptr_t dest_off = abase + (dest_off_con << scale);
5629
5630 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5631 // Non-aligned; too bad.
5632 // One more chance: Pick off an initial 32-bit word.
5633 // This is a common case, since abase can be odd mod 8.
5634 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5635 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5636 Node* sptr = basic_plus_adr(src, src_off);
5637 Node* dptr = basic_plus_adr(dest, dest_off);
5638 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5639 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5640 src_off += BytesPerInt;
5641 dest_off += BytesPerInt;
5642 } else {
5643 return false;
5644 }
5645 }
5646 assert(src_off % BytesPerLong == 0, "");
5647 assert(dest_off % BytesPerLong == 0, "");
5648
5649 // Do this copy by giant steps.
5650 Node* sptr = basic_plus_adr(src, src_off);
5651 Node* dptr = basic_plus_adr(dest, dest_off);
5652 Node* countx = dest_size;
5653 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5654 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5655
5656 bool disjoint_bases = true; // since alloc != NULL
5657 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5658 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5659
5660 return true;
5661 }
5662
5663
5664 // Helper function; generates code for the slow case.
5665 // We make a call to a runtime method which emulates the native method,
5666 // but without the native wrapper overhead.
5667 void
5668 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5669 Node* src, Node* src_offset,
5670 Node* dest, Node* dest_offset,
5671 Node* copy_length, bool dest_uninitialized) {
5672 assert(!dest_uninitialized, "Invariant");
5673 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5674 OptoRuntime::slow_arraycopy_Type(),
5675 OptoRuntime::slow_arraycopy_Java(),
5676 "slow_arraycopy", adr_type,
5677 src, src_offset, dest, dest_offset,
5678 copy_length);
5679
5680 // Handle exceptions thrown by this fellow:
5681 make_slow_call_ex(call, env()->Throwable_klass(), false);
5682 }
5683
5684 // Helper function; generates code for cases requiring runtime checks.
5685 Node*
5686 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5687 Node* dest_elem_klass,
5688 Node* src, Node* src_offset,
5689 Node* dest, Node* dest_offset,
5690 Node* copy_length, bool dest_uninitialized) {
5691 if (stopped()) return NULL;
5692
5693 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5694 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5695 return NULL;
5696 }
5697
5698 // Pick out the parameters required to perform a store-check
5699 // for the target array. This is an optimistic check. It will
5700 // look in each non-null element's class, at the desired klass's
5701 // super_check_offset, for the desired klass.
5702 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5703 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5704 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5705 Node* check_offset = ConvI2X(_gvn.transform(n3));
5706 Node* check_value = dest_elem_klass;
5707
5708 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5709 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5710
5711 // (We know the arrays are never conjoint, because their types differ.)
5712 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5713 OptoRuntime::checkcast_arraycopy_Type(),
5714 copyfunc_addr, "checkcast_arraycopy", adr_type,
5715 // five arguments, of which two are
5716 // intptr_t (jlong in LP64)
5717 src_start, dest_start,
5718 copy_length XTOP,
5719 check_offset XTOP,
5720 check_value);
5721
5722 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5723 }
5724
5725
5726 // Helper function; generates code for cases requiring runtime checks.
5727 Node*
5728 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5729 Node* src, Node* src_offset,
5730 Node* dest, Node* dest_offset,
5731 Node* copy_length, bool dest_uninitialized) {
5732 assert(!dest_uninitialized, "Invariant");
5733 if (stopped()) return NULL;
5734 address copyfunc_addr = StubRoutines::generic_arraycopy();
5735 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5736 return NULL;
5737 }
5738
5739 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5740 OptoRuntime::generic_arraycopy_Type(),
5741 copyfunc_addr, "generic_arraycopy", adr_type,
5742 src, src_offset, dest, dest_offset, copy_length);
5743
5744 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5745 }
5746
5747 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5748 void
5749 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5750 BasicType basic_elem_type,
5751 bool disjoint_bases,
5752 Node* src, Node* src_offset,
5753 Node* dest, Node* dest_offset,
5754 Node* copy_length, bool dest_uninitialized) {
5755 if (stopped()) return; // nothing to do
5756
5757 Node* src_start = src;
5758 Node* dest_start = dest;
5759 if (src_offset != NULL || dest_offset != NULL) {
5760 assert(src_offset != NULL && dest_offset != NULL, "");
5761 src_start = array_element_address(src, src_offset, basic_elem_type);
5762 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5763 }
5764
5765 // Figure out which arraycopy runtime method to call.
5766 const char* copyfunc_name = "arraycopy";
5767 address copyfunc_addr =
5768 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5769 disjoint_bases, copyfunc_name, dest_uninitialized);
5770
5771 // Call it. Note that the count_ix value is not scaled to a byte-size.
5772 make_runtime_call(RC_LEAF|RC_NO_FP,
5773 OptoRuntime::fast_arraycopy_Type(),
5774 copyfunc_addr, copyfunc_name, adr_type,
5775 src_start, dest_start, copy_length XTOP);
5776 }
5777
5778 //-------------inline_encodeISOArray-----------------------------------
5779 // encode char[] to byte[] in ISO_8859_1
5780 bool LibraryCallKit::inline_encodeISOArray() {
5781 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5782 // no receiver since it is static method
5783 Node *src = argument(0);
5784 Node *src_offset = argument(1);
5785 Node *dst = argument(2);
5786 Node *dst_offset = argument(3);
5787 Node *length = argument(4);
5788
5789 const Type* src_type = src->Value(&_gvn);
5790 const Type* dst_type = dst->Value(&_gvn);
5791 const TypeAryPtr* top_src = src_type->isa_aryptr();
5792 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5793 if (top_src == NULL || top_src->klass() == NULL ||
5794 top_dest == NULL || top_dest->klass() == NULL) {
5795 // failed array check
5796 return false;
5797 }
5798
5799 // Figure out the size and type of the elements we will be copying.
5800 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5801 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5802 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5803 return false;
5804 }
5805 Node* src_start = array_element_address(src, src_offset, src_elem);
5806 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5807 // 'src_start' points to src array + scaled offset
5808 // 'dst_start' points to dst array + scaled offset
5809
5810 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5811 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5812 enc = _gvn.transform(enc);
5813 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5814 set_memory(res_mem, mtype);
5815 set_result(enc);
5816 return true;
5817 }
5818
5819 //-------------inline_multiplyToLen-----------------------------------
5820 bool LibraryCallKit::inline_multiplyToLen() {
5821 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5822
5823 address stubAddr = StubRoutines::multiplyToLen();
5824 if (stubAddr == NULL) {
5825 return false; // Intrinsic's stub is not implemented on this platform
5826 }
5827 const char* stubName = "multiplyToLen";
5828
5829 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5830
5831 // no receiver because it is a static method
5832 Node* x = argument(0);
5833 Node* xlen = argument(1);
5834 Node* y = argument(2);
5835 Node* ylen = argument(3);
5836 Node* z = argument(4);
5837
5838 const Type* x_type = x->Value(&_gvn);
5839 const Type* y_type = y->Value(&_gvn);
5840 const TypeAryPtr* top_x = x_type->isa_aryptr();
5841 const TypeAryPtr* top_y = y_type->isa_aryptr();
5842 if (top_x == NULL || top_x->klass() == NULL ||
5843 top_y == NULL || top_y->klass() == NULL) {
5844 // failed array check
5845 return false;
5846 }
5847
5848 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5849 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5850 if (x_elem != T_INT || y_elem != T_INT) {
5851 return false;
5852 }
5853
5854 // Set the original stack and the reexecute bit for the interpreter to reexecute
5855 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5856 // on the return from z array allocation in runtime.
5857 { PreserveReexecuteState preexecs(this);
5858 jvms()->set_should_reexecute(true);
5859
5860 Node* x_start = array_element_address(x, intcon(0), x_elem);
5861 Node* y_start = array_element_address(y, intcon(0), y_elem);
5862 // 'x_start' points to x array + scaled xlen
5863 // 'y_start' points to y array + scaled ylen
5864
5865 // Allocate the result array
5866 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5867 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5868 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5869
5870 IdealKit ideal(this);
5871
5872 #define __ ideal.
5873 Node* one = __ ConI(1);
5874 Node* zero = __ ConI(0);
5875 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5876 __ set(need_alloc, zero);
5877 __ set(z_alloc, z);
5878 __ if_then(z, BoolTest::eq, null()); {
5879 __ increment (need_alloc, one);
5880 } __ else_(); {
5881 // Update graphKit memory and control from IdealKit.
5882 sync_kit(ideal);
5883 Node* zlen_arg = load_array_length(z);
5884 // Update IdealKit memory and control from graphKit.
5885 __ sync_kit(this);
5886 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5887 __ increment (need_alloc, one);
5888 } __ end_if();
5889 } __ end_if();
5890
5891 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5892 // Update graphKit memory and control from IdealKit.
5893 sync_kit(ideal);
5894 Node * narr = new_array(klass_node, zlen, 1);
5895 // Update IdealKit memory and control from graphKit.
5896 __ sync_kit(this);
5897 __ set(z_alloc, narr);
5898 } __ end_if();
5899
5900 sync_kit(ideal);
5901 z = __ value(z_alloc);
5902 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5903 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5904 // Final sync IdealKit and GraphKit.
5905 final_sync(ideal);
5906 #undef __
5907
5908 Node* z_start = array_element_address(z, intcon(0), T_INT);
5909
5910 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5911 OptoRuntime::multiplyToLen_Type(),
5912 stubAddr, stubName, TypePtr::BOTTOM,
5913 x_start, xlen, y_start, ylen, z_start, zlen);
5914 } // original reexecute is set back here
5915
5916 C->set_has_split_ifs(true); // Has chance for split-if optimization
5917 set_result(z);
5918 return true;
5919 }
5920
5921 //-------------inline_squareToLen------------------------------------
5922 bool LibraryCallKit::inline_squareToLen() {
5923 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5924
5925 address stubAddr = StubRoutines::squareToLen();
5926 if (stubAddr == NULL) {
5927 return false; // Intrinsic's stub is not implemented on this platform
5928 }
5929 const char* stubName = "squareToLen";
5930
5931 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5932
5933 Node* x = argument(0);
5934 Node* len = argument(1);
5935 Node* z = argument(2);
5936 Node* zlen = argument(3);
5937
5938 const Type* x_type = x->Value(&_gvn);
5939 const Type* z_type = z->Value(&_gvn);
5940 const TypeAryPtr* top_x = x_type->isa_aryptr();
5941 const TypeAryPtr* top_z = z_type->isa_aryptr();
5942 if (top_x == NULL || top_x->klass() == NULL ||
5943 top_z == NULL || top_z->klass() == NULL) {
5944 // failed array check
5945 return false;
5946 }
5947
5948 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5949 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5950 if (x_elem != T_INT || z_elem != T_INT) {
5951 return false;
5952 }
5953
5954
5955 Node* x_start = array_element_address(x, intcon(0), x_elem);
5956 Node* z_start = array_element_address(z, intcon(0), z_elem);
5957
5958 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5959 OptoRuntime::squareToLen_Type(),
5960 stubAddr, stubName, TypePtr::BOTTOM,
5961 x_start, len, z_start, zlen);
5962
5963 set_result(z);
5964 return true;
5965 }
5966
5967 //-------------inline_mulAdd------------------------------------------
5968 bool LibraryCallKit::inline_mulAdd() {
5969 assert(UseMulAddIntrinsic, "not implementated on this platform");
5970
5971 address stubAddr = StubRoutines::mulAdd();
5972 if (stubAddr == NULL) {
5973 return false; // Intrinsic's stub is not implemented on this platform
5974 }
5975 const char* stubName = "mulAdd";
5976
5977 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5978
5979 Node* out = argument(0);
5980 Node* in = argument(1);
5981 Node* offset = argument(2);
5982 Node* len = argument(3);
5983 Node* k = argument(4);
5984
5985 const Type* out_type = out->Value(&_gvn);
5986 const Type* in_type = in->Value(&_gvn);
5987 const TypeAryPtr* top_out = out_type->isa_aryptr();
5988 const TypeAryPtr* top_in = in_type->isa_aryptr();
5989 if (top_out == NULL || top_out->klass() == NULL ||
5990 top_in == NULL || top_in->klass() == NULL) {
5991 // failed array check
5992 return false;
5993 }
5994
5995 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5996 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5997 if (out_elem != T_INT || in_elem != T_INT) {
5998 return false;
5999 }
6000
6001 Node* outlen = load_array_length(out);
6002 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
6003 Node* out_start = array_element_address(out, intcon(0), out_elem);
6004 Node* in_start = array_element_address(in, intcon(0), in_elem);
6005
6006 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6007 OptoRuntime::mulAdd_Type(),
6008 stubAddr, stubName, TypePtr::BOTTOM,
6009 out_start,in_start, new_offset, len, k);
6010 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6011 set_result(result);
6012 return true;
6013 }
6014
6015 //-------------inline_montgomeryMultiply-----------------------------------
6016 bool LibraryCallKit::inline_montgomeryMultiply() {
6017 address stubAddr = StubRoutines::montgomeryMultiply();
6018 if (stubAddr == NULL) {
6019 return false; // Intrinsic's stub is not implemented on this platform
6020 }
6021
6022 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6023 const char* stubName = "montgomery_square";
6024
6025 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6026
6027 Node* a = argument(0);
6028 Node* b = argument(1);
6029 Node* n = argument(2);
6030 Node* len = argument(3);
6031 Node* inv = argument(4);
6032 Node* m = argument(6);
6033
6034 const Type* a_type = a->Value(&_gvn);
6035 const TypeAryPtr* top_a = a_type->isa_aryptr();
6036 const Type* b_type = b->Value(&_gvn);
6037 const TypeAryPtr* top_b = b_type->isa_aryptr();
6038 const Type* n_type = a->Value(&_gvn);
6039 const TypeAryPtr* top_n = n_type->isa_aryptr();
6040 const Type* m_type = a->Value(&_gvn);
6041 const TypeAryPtr* top_m = m_type->isa_aryptr();
6042 if (top_a == NULL || top_a->klass() == NULL ||
6043 top_b == NULL || top_b->klass() == NULL ||
6044 top_n == NULL || top_n->klass() == NULL ||
6045 top_m == NULL || top_m->klass() == NULL) {
6046 // failed array check
6047 return false;
6048 }
6049
6050 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6051 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6052 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6053 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6054 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6055 return false;
6056 }
6057
6058 // Make the call
6059 {
6060 Node* a_start = array_element_address(a, intcon(0), a_elem);
6061 Node* b_start = array_element_address(b, intcon(0), b_elem);
6062 Node* n_start = array_element_address(n, intcon(0), n_elem);
6063 Node* m_start = array_element_address(m, intcon(0), m_elem);
6064
6065 Node* call = make_runtime_call(RC_LEAF,
6066 OptoRuntime::montgomeryMultiply_Type(),
6067 stubAddr, stubName, TypePtr::BOTTOM,
6068 a_start, b_start, n_start, len, inv, top(),
6069 m_start);
6070 set_result(m);
6071 }
6072
6073 return true;
6074 }
6075
6076 bool LibraryCallKit::inline_montgomerySquare() {
6077 address stubAddr = StubRoutines::montgomerySquare();
6078 if (stubAddr == NULL) {
6079 return false; // Intrinsic's stub is not implemented on this platform
6080 }
6081
6082 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6083 const char* stubName = "montgomery_square";
6084
6085 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6086
6087 Node* a = argument(0);
6088 Node* n = argument(1);
6089 Node* len = argument(2);
6090 Node* inv = argument(3);
6091 Node* m = argument(5);
6092
6093 const Type* a_type = a->Value(&_gvn);
6094 const TypeAryPtr* top_a = a_type->isa_aryptr();
6095 const Type* n_type = a->Value(&_gvn);
6096 const TypeAryPtr* top_n = n_type->isa_aryptr();
6097 const Type* m_type = a->Value(&_gvn);
6098 const TypeAryPtr* top_m = m_type->isa_aryptr();
6099 if (top_a == NULL || top_a->klass() == NULL ||
6100 top_n == NULL || top_n->klass() == NULL ||
6101 top_m == NULL || top_m->klass() == NULL) {
6102 // failed array check
6103 return false;
6104 }
6105
6106 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6107 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6108 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6109 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6110 return false;
6111 }
6112
6113 // Make the call
6114 {
6115 Node* a_start = array_element_address(a, intcon(0), a_elem);
6116 Node* n_start = array_element_address(n, intcon(0), n_elem);
6117 Node* m_start = array_element_address(m, intcon(0), m_elem);
6118
6119 Node* call = make_runtime_call(RC_LEAF,
6120 OptoRuntime::montgomerySquare_Type(),
6121 stubAddr, stubName, TypePtr::BOTTOM,
6122 a_start, n_start, len, inv, top(),
6123 m_start);
6124 set_result(m);
6125 }
6126
6127 return true;
6128 }
6129
6130
6131 /**
6132 * Calculate CRC32 for byte.
6133 * int java.util.zip.CRC32.update(int crc, int b)
6134 */
6135 bool LibraryCallKit::inline_updateCRC32() {
6136 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6137 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6138 // no receiver since it is static method
6139 Node* crc = argument(0); // type: int
6140 Node* b = argument(1); // type: int
6141
6142 /*
6143 * int c = ~ crc;
6144 * b = timesXtoThe32[(b ^ c) & 0xFF];
6145 * b = b ^ (c >>> 8);
6146 * crc = ~b;
6147 */
6148
6149 Node* M1 = intcon(-1);
6150 crc = _gvn.transform(new (C) XorINode(crc, M1));
6151 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6152 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6153
6154 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6155 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6156 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6157 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6158
6159 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6160 result = _gvn.transform(new (C) XorINode(crc, result));
6161 result = _gvn.transform(new (C) XorINode(result, M1));
6162 set_result(result);
6163 return true;
6164 }
6165
6166 /**
6167 * Calculate CRC32 for byte[] array.
6168 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6169 */
6170 bool LibraryCallKit::inline_updateBytesCRC32() {
6171 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6172 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6173 // no receiver since it is static method
6174 Node* crc = argument(0); // type: int
6175 Node* src = argument(1); // type: oop
6176 Node* offset = argument(2); // type: int
6177 Node* length = argument(3); // type: int
6178
6179 const Type* src_type = src->Value(&_gvn);
6180 const TypeAryPtr* top_src = src_type->isa_aryptr();
6181 if (top_src == NULL || top_src->klass() == NULL) {
6182 // failed array check
6183 return false;
6184 }
6185
6186 // Figure out the size and type of the elements we will be copying.
6187 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6188 if (src_elem != T_BYTE) {
6189 return false;
6190 }
6191
6192 // 'src_start' points to src array + scaled offset
6193 Node* src_start = array_element_address(src, offset, src_elem);
6194
6195 // We assume that range check is done by caller.
6196 // TODO: generate range check (offset+length < src.length) in debug VM.
6197
6198 // Call the stub.
6199 address stubAddr = StubRoutines::updateBytesCRC32();
6200 const char *stubName = "updateBytesCRC32";
6201
6202 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6203 stubAddr, stubName, TypePtr::BOTTOM,
6204 crc, src_start, length);
6205 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6206 set_result(result);
6207 return true;
6208 }
6209
6210 /**
6211 * Calculate CRC32 for ByteBuffer.
6212 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6213 */
6214 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6215 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6216 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6217 // no receiver since it is static method
6218 Node* crc = argument(0); // type: int
6219 Node* src = argument(1); // type: long
6220 Node* offset = argument(3); // type: int
6221 Node* length = argument(4); // type: int
6222
6223 src = ConvL2X(src); // adjust Java long to machine word
6224 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6225 offset = ConvI2X(offset);
6226
6227 // 'src_start' points to src array + scaled offset
6228 Node* src_start = basic_plus_adr(top(), base, offset);
6229
6230 // Call the stub.
6231 address stubAddr = StubRoutines::updateBytesCRC32();
6232 const char *stubName = "updateBytesCRC32";
6233
6234 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6235 stubAddr, stubName, TypePtr::BOTTOM,
6236 crc, src_start, length);
6237 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6238 set_result(result);
6239 return true;
6240 }
6241
6242 //----------------------------inline_reference_get----------------------------
6243 // public T java.lang.ref.Reference.get();
6244 bool LibraryCallKit::inline_reference_get() {
6245 const int referent_offset = java_lang_ref_Reference::referent_offset;
6246 guarantee(referent_offset > 0, "should have already been set");
6247
6248 // Get the argument:
6249 Node* reference_obj = null_check_receiver();
6250 if (stopped()) return true;
6251
6252 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6253
6254 ciInstanceKlass* klass = env()->Object_klass();
6255 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6256
6257 Node* no_ctrl = NULL;
6258 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6259
6260 // Use the pre-barrier to record the value in the referent field
6261 pre_barrier(false /* do_load */,
6262 control(),
6263 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6264 result /* pre_val */,
6265 T_OBJECT);
6266
6267 // Add memory barrier to prevent commoning reads from this field
6268 // across safepoint since GC can change its value.
6269 insert_mem_bar(Op_MemBarCPUOrder);
6270
6271 set_result(result);
6272 return true;
6273 }
6274
6275
6276 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6277 bool is_exact=true, bool is_static=false) {
6278
6279 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6280 assert(tinst != NULL, "obj is null");
6281 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6282 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6283
6284 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6285 ciSymbol::make(fieldTypeString),
6286 is_static);
6287 if (field == NULL) return (Node *) NULL;
6288 assert (field != NULL, "undefined field");
6289
6290 // Next code copied from Parse::do_get_xxx():
6291
6292 // Compute address and memory type.
6293 int offset = field->offset_in_bytes();
6294 bool is_vol = field->is_volatile();
6295 ciType* field_klass = field->type();
6296 assert(field_klass->is_loaded(), "should be loaded");
6297 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6298 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6299 BasicType bt = field->layout_type();
6300
6301 // Build the resultant type of the load
6302 const Type *type;
6303 if (bt == T_OBJECT) {
6304 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6305 } else {
6306 type = Type::get_const_basic_type(bt);
6307 }
6308
6309 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6310 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6311 }
6312 // Build the load.
6313 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6314 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6315 // If reference is volatile, prevent following memory ops from
6316 // floating up past the volatile read. Also prevents commoning
6317 // another volatile read.
6318 if (is_vol) {
6319 // Memory barrier includes bogus read of value to force load BEFORE membar
6320 insert_mem_bar(Op_MemBarAcquire, loadedField);
6321 }
6322 return loadedField;
6323 }
6324
6325
6326 //------------------------------inline_aescrypt_Block-----------------------
6327 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6328 address stubAddr = NULL;
6329 const char *stubName;
6330 assert(UseAES, "need AES instruction support");
6331
6332 switch(id) {
6333 case vmIntrinsics::_aescrypt_encryptBlock:
6334 stubAddr = StubRoutines::aescrypt_encryptBlock();
6335 stubName = "aescrypt_encryptBlock";
6336 break;
6337 case vmIntrinsics::_aescrypt_decryptBlock:
6338 stubAddr = StubRoutines::aescrypt_decryptBlock();
6339 stubName = "aescrypt_decryptBlock";
6340 break;
6341 }
6342 if (stubAddr == NULL) return false;
6343
6344 Node* aescrypt_object = argument(0);
6345 Node* src = argument(1);
6346 Node* src_offset = argument(2);
6347 Node* dest = argument(3);
6348 Node* dest_offset = argument(4);
6349
6350 // (1) src and dest are arrays.
6351 const Type* src_type = src->Value(&_gvn);
6352 const Type* dest_type = dest->Value(&_gvn);
6353 const TypeAryPtr* top_src = src_type->isa_aryptr();
6354 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6355 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6356
6357 // for the quick and dirty code we will skip all the checks.
6358 // we are just trying to get the call to be generated.
6359 Node* src_start = src;
6360 Node* dest_start = dest;
6361 if (src_offset != NULL || dest_offset != NULL) {
6362 assert(src_offset != NULL && dest_offset != NULL, "");
6363 src_start = array_element_address(src, src_offset, T_BYTE);
6364 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6365 }
6366
6367 // now need to get the start of its expanded key array
6368 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6369 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6370 if (k_start == NULL) return false;
6371
6372 if (Matcher::pass_original_key_for_aes()) {
6373 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6374 // compatibility issues between Java key expansion and SPARC crypto instructions
6375 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6376 if (original_k_start == NULL) return false;
6377
6378 // Call the stub.
6379 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6380 stubAddr, stubName, TypePtr::BOTTOM,
6381 src_start, dest_start, k_start, original_k_start);
6382 } else {
6383 // Call the stub.
6384 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6385 stubAddr, stubName, TypePtr::BOTTOM,
6386 src_start, dest_start, k_start);
6387 }
6388
6389 return true;
6390 }
6391
6392 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6393 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6394 address stubAddr = NULL;
6395 const char *stubName = NULL;
6396
6397 assert(UseAES, "need AES instruction support");
6398
6399 switch(id) {
6400 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6401 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6402 stubName = "cipherBlockChaining_encryptAESCrypt";
6403 break;
6404 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6405 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6406 stubName = "cipherBlockChaining_decryptAESCrypt";
6407 break;
6408 }
6409 if (stubAddr == NULL) return false;
6410
6411 Node* cipherBlockChaining_object = argument(0);
6412 Node* src = argument(1);
6413 Node* src_offset = argument(2);
6414 Node* len = argument(3);
6415 Node* dest = argument(4);
6416 Node* dest_offset = argument(5);
6417
6418 // (1) src and dest are arrays.
6419 const Type* src_type = src->Value(&_gvn);
6420 const Type* dest_type = dest->Value(&_gvn);
6421 const TypeAryPtr* top_src = src_type->isa_aryptr();
6422 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6423 assert (top_src != NULL && top_src->klass() != NULL
6424 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6425
6426 // checks are the responsibility of the caller
6427 Node* src_start = src;
6428 Node* dest_start = dest;
6429 if (src_offset != NULL || dest_offset != NULL) {
6430 assert(src_offset != NULL && dest_offset != NULL, "");
6431 src_start = array_element_address(src, src_offset, T_BYTE);
6432 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6433 }
6434
6435 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6436 // (because of the predicated logic executed earlier).
6437 // so we cast it here safely.
6438 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6439
6440 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6441 if (embeddedCipherObj == NULL) return false;
6442
6443 // cast it to what we know it will be at runtime
6444 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6445 assert(tinst != NULL, "CBC obj is null");
6446 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6447 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6448 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6449
6450 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6451 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6452 const TypeOopPtr* xtype = aklass->as_instance_type();
6453 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6454 aescrypt_object = _gvn.transform(aescrypt_object);
6455
6456 // we need to get the start of the aescrypt_object's expanded key array
6457 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6458 if (k_start == NULL) return false;
6459
6460 // similarly, get the start address of the r vector
6461 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6462 if (objRvec == NULL) return false;
6463 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6464
6465 Node* cbcCrypt;
6466 if (Matcher::pass_original_key_for_aes()) {
6467 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6468 // compatibility issues between Java key expansion and SPARC crypto instructions
6469 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6470 if (original_k_start == NULL) return false;
6471
6472 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6473 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6474 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6475 stubAddr, stubName, TypePtr::BOTTOM,
6476 src_start, dest_start, k_start, r_start, len, original_k_start);
6477 } else {
6478 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6479 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6480 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6481 stubAddr, stubName, TypePtr::BOTTOM,
6482 src_start, dest_start, k_start, r_start, len);
6483 }
6484
6485 // return cipher length (int)
6486 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6487 set_result(retvalue);
6488 return true;
6489 }
6490
6491 //------------------------------get_key_start_from_aescrypt_object-----------------------
6492 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6493 #ifdef PPC64
6494 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6495 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6496 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6497 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6498 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6499 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6500 if (objSessionK == NULL) {
6501 return (Node *) NULL;
6502 }
6503 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6504 #else
6505 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6506 #endif // PPC64
6507 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6508 if (objAESCryptKey == NULL) return (Node *) NULL;
6509
6510 // now have the array, need to get the start address of the K array
6511 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6512 return k_start;
6513 }
6514
6515 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6516 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6517 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6518 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6519 if (objAESCryptKey == NULL) return (Node *) NULL;
6520
6521 // now have the array, need to get the start address of the lastKey array
6522 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6523 return original_k_start;
6524 }
6525
6526 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6527 // Return node representing slow path of predicate check.
6528 // the pseudo code we want to emulate with this predicate is:
6529 // for encryption:
6530 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6531 // for decryption:
6532 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6533 // note cipher==plain is more conservative than the original java code but that's OK
6534 //
6535 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6536 // The receiver was checked for NULL already.
6537 Node* objCBC = argument(0);
6538
6539 // Load embeddedCipher field of CipherBlockChaining object.
6540 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6541
6542 // get AESCrypt klass for instanceOf check
6543 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6544 // will have same classloader as CipherBlockChaining object
6545 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6546 assert(tinst != NULL, "CBCobj is null");
6547 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6548
6549 // we want to do an instanceof comparison against the AESCrypt class
6550 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6551 if (!klass_AESCrypt->is_loaded()) {
6552 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6553 Node* ctrl = control();
6554 set_control(top()); // no regular fast path
6555 return ctrl;
6556 }
6557 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6558
6559 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6560 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6561 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6562
6563 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6564
6565 // for encryption, we are done
6566 if (!decrypting)
6567 return instof_false; // even if it is NULL
6568
6569 // for decryption, we need to add a further check to avoid
6570 // taking the intrinsic path when cipher and plain are the same
6571 // see the original java code for why.
6572 RegionNode* region = new(C) RegionNode(3);
6573 region->init_req(1, instof_false);
6574 Node* src = argument(1);
6575 Node* dest = argument(4);
6576 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6577 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6578 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6579 region->init_req(2, src_dest_conjoint);
6580
6581 record_for_igvn(region);
6582 return _gvn.transform(region);
6583 }
6584
6585 //------------------------------inline_sha_implCompress-----------------------
6586 //
6587 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6588 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6589 //
6590 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6591 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6592 //
6593 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6594 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6595 //
6596 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6597 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6598
6599 Node* sha_obj = argument(0);
6600 Node* src = argument(1); // type oop
6601 Node* ofs = argument(2); // type int
6602
6603 const Type* src_type = src->Value(&_gvn);
6604 const TypeAryPtr* top_src = src_type->isa_aryptr();
6605 if (top_src == NULL || top_src->klass() == NULL) {
6606 // failed array check
6607 return false;
6608 }
6609 // Figure out the size and type of the elements we will be copying.
6610 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6611 if (src_elem != T_BYTE) {
6612 return false;
6613 }
6614 // 'src_start' points to src array + offset
6615 Node* src_start = array_element_address(src, ofs, src_elem);
6616 Node* state = NULL;
6617 address stubAddr;
6618 const char *stubName;
6619
6620 switch(id) {
6621 case vmIntrinsics::_sha_implCompress:
6622 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6623 state = get_state_from_sha_object(sha_obj);
6624 stubAddr = StubRoutines::sha1_implCompress();
6625 stubName = "sha1_implCompress";
6626 break;
6627 case vmIntrinsics::_sha2_implCompress:
6628 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6629 state = get_state_from_sha_object(sha_obj);
6630 stubAddr = StubRoutines::sha256_implCompress();
6631 stubName = "sha256_implCompress";
6632 break;
6633 case vmIntrinsics::_sha5_implCompress:
6634 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6635 state = get_state_from_sha5_object(sha_obj);
6636 stubAddr = StubRoutines::sha512_implCompress();
6637 stubName = "sha512_implCompress";
6638 break;
6639 default:
6640 fatal_unexpected_iid(id);
6641 return false;
6642 }
6643 if (state == NULL) return false;
6644
6645 // Call the stub.
6646 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6647 stubAddr, stubName, TypePtr::BOTTOM,
6648 src_start, state);
6649
6650 return true;
6651 }
6652
6653 //------------------------------inline_digestBase_implCompressMB-----------------------
6654 //
6655 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6656 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6657 //
6658 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6659 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6660 "need SHA1/SHA256/SHA512 instruction support");
6661 assert((uint)predicate < 3, "sanity");
6662 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6663
6664 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6665 Node* src = argument(1); // byte[] array
6666 Node* ofs = argument(2); // type int
6667 Node* limit = argument(3); // type int
6668
6669 const Type* src_type = src->Value(&_gvn);
6670 const TypeAryPtr* top_src = src_type->isa_aryptr();
6671 if (top_src == NULL || top_src->klass() == NULL) {
6672 // failed array check
6673 return false;
6674 }
6675 // Figure out the size and type of the elements we will be copying.
6676 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6677 if (src_elem != T_BYTE) {
6678 return false;
6679 }
6680 // 'src_start' points to src array + offset
6681 Node* src_start = array_element_address(src, ofs, src_elem);
6682
6683 const char* klass_SHA_name = NULL;
6684 const char* stub_name = NULL;
6685 address stub_addr = NULL;
6686 bool long_state = false;
6687
6688 switch (predicate) {
6689 case 0:
6690 if (UseSHA1Intrinsics) {
6691 klass_SHA_name = "sun/security/provider/SHA";
6692 stub_name = "sha1_implCompressMB";
6693 stub_addr = StubRoutines::sha1_implCompressMB();
6694 }
6695 break;
6696 case 1:
6697 if (UseSHA256Intrinsics) {
6698 klass_SHA_name = "sun/security/provider/SHA2";
6699 stub_name = "sha256_implCompressMB";
6700 stub_addr = StubRoutines::sha256_implCompressMB();
6701 }
6702 break;
6703 case 2:
6704 if (UseSHA512Intrinsics) {
6705 klass_SHA_name = "sun/security/provider/SHA5";
6706 stub_name = "sha512_implCompressMB";
6707 stub_addr = StubRoutines::sha512_implCompressMB();
6708 long_state = true;
6709 }
6710 break;
6711 default:
6712 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6713 }
6714 if (klass_SHA_name != NULL) {
6715 // get DigestBase klass to lookup for SHA klass
6716 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6717 assert(tinst != NULL, "digestBase_obj is not instance???");
6718 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6719
6720 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6721 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6722 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6723 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6724 }
6725 return false;
6726 }
6727 //------------------------------inline_sha_implCompressMB-----------------------
6728 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6729 bool long_state, address stubAddr, const char *stubName,
6730 Node* src_start, Node* ofs, Node* limit) {
6731 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6732 const TypeOopPtr* xtype = aklass->as_instance_type();
6733 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6734 sha_obj = _gvn.transform(sha_obj);
6735
6736 Node* state;
6737 if (long_state) {
6738 state = get_state_from_sha5_object(sha_obj);
6739 } else {
6740 state = get_state_from_sha_object(sha_obj);
6741 }
6742 if (state == NULL) return false;
6743
6744 // Call the stub.
6745 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6746 OptoRuntime::digestBase_implCompressMB_Type(),
6747 stubAddr, stubName, TypePtr::BOTTOM,
6748 src_start, state, ofs, limit);
6749 // return ofs (int)
6750 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6751 set_result(result);
6752
6753 return true;
6754 }
6755
6756 //------------------------------get_state_from_sha_object-----------------------
6757 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6758 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6759 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6760 if (sha_state == NULL) return (Node *) NULL;
6761
6762 // now have the array, need to get the start address of the state array
6763 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6764 return state;
6765 }
6766
6767 //------------------------------get_state_from_sha5_object-----------------------
6768 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6769 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6770 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6771 if (sha_state == NULL) return (Node *) NULL;
6772
6773 // now have the array, need to get the start address of the state array
6774 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6775 return state;
6776 }
6777
6778 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6779 // Return node representing slow path of predicate check.
6780 // the pseudo code we want to emulate with this predicate is:
6781 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6782 //
6783 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6784 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6785 "need SHA1/SHA256/SHA512 instruction support");
6786 assert((uint)predicate < 3, "sanity");
6787
6788 // The receiver was checked for NULL already.
6789 Node* digestBaseObj = argument(0);
6790
6791 // get DigestBase klass for instanceOf check
6792 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6793 assert(tinst != NULL, "digestBaseObj is null");
6794 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6795
6796 const char* klass_SHA_name = NULL;
6797 switch (predicate) {
6798 case 0:
6799 if (UseSHA1Intrinsics) {
6800 // we want to do an instanceof comparison against the SHA class
6801 klass_SHA_name = "sun/security/provider/SHA";
6802 }
6803 break;
6804 case 1:
6805 if (UseSHA256Intrinsics) {
6806 // we want to do an instanceof comparison against the SHA2 class
6807 klass_SHA_name = "sun/security/provider/SHA2";
6808 }
6809 break;
6810 case 2:
6811 if (UseSHA512Intrinsics) {
6812 // we want to do an instanceof comparison against the SHA5 class
6813 klass_SHA_name = "sun/security/provider/SHA5";
6814 }
6815 break;
6816 default:
6817 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6818 }
6819
6820 ciKlass* klass_SHA = NULL;
6821 if (klass_SHA_name != NULL) {
6822 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6823 }
6824 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6825 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6826 Node* ctrl = control();
6827 set_control(top()); // no intrinsic path
6828 return ctrl;
6829 }
6830 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6831
6832 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6833 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6834 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6835 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6836
6837 return instof_false; // even if it is NULL
6838 }
6839
6840 bool LibraryCallKit::inline_profileBoolean() {
6841 Node* counts = argument(1);
6842 const TypeAryPtr* ary = NULL;
6843 ciArray* aobj = NULL;
6844 if (counts->is_Con()
6845 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6846 && (aobj = ary->const_oop()->as_array()) != NULL
6847 && (aobj->length() == 2)) {
6848 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6849 jint false_cnt = aobj->element_value(0).as_int();
6850 jint true_cnt = aobj->element_value(1).as_int();
6851
6852 if (C->log() != NULL) {
6853 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6854 false_cnt, true_cnt);
6855 }
6856
6857 if (false_cnt + true_cnt == 0) {
6858 // According to profile, never executed.
6859 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6860 Deoptimization::Action_reinterpret);
6861 return true;
6862 }
6863
6864 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6865 // is a number of each value occurrences.
6866 Node* result = argument(0);
6867 if (false_cnt == 0 || true_cnt == 0) {
6868 // According to profile, one value has been never seen.
6869 int expected_val = (false_cnt == 0) ? 1 : 0;
6870
6871 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6872 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6873
6874 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6875 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6876 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6877
6878 { // Slow path: uncommon trap for never seen value and then reexecute
6879 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6880 // the value has been seen at least once.
6881 PreserveJVMState pjvms(this);
6882 PreserveReexecuteState preexecs(this);
6883 jvms()->set_should_reexecute(true);
6884
6885 set_control(slow_path);
6886 set_i_o(i_o());
6887
6888 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6889 Deoptimization::Action_reinterpret);
6890 }
6891 // The guard for never seen value enables sharpening of the result and
6892 // returning a constant. It allows to eliminate branches on the same value
6893 // later on.
6894 set_control(fast_path);
6895 result = intcon(expected_val);
6896 }
6897 // Stop profiling.
6898 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6899 // By replacing method body with profile data (represented as ProfileBooleanNode
6900 // on IR level) we effectively disable profiling.
6901 // It enables full speed execution once optimized code is generated.
6902 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6903 C->record_for_igvn(profile);
6904 set_result(profile);
6905 return true;
6906 } else {
6907 // Continue profiling.
6908 // Profile data isn't available at the moment. So, execute method's bytecode version.
6909 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6910 // is compiled and counters aren't available since corresponding MethodHandle
6911 // isn't a compile-time constant.
6912 return false;
6913 }
6914 }