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