1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mathexactnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "prims/nativeLookup.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "trace/traceMacros.hpp"
43
44 class LibraryIntrinsic : public InlineCallGenerator {
45 // Extend the set of intrinsics known to the runtime:
46 public:
47 private:
48 bool _is_virtual;
49 bool _is_predicted;
50 bool _does_virtual_dispatch;
51 vmIntrinsics::ID _intrinsic_id;
52
53 public:
54 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
55 : InlineCallGenerator(m),
56 _is_virtual(is_virtual),
57 _is_predicted(is_predicted),
58 _does_virtual_dispatch(does_virtual_dispatch),
59 _intrinsic_id(id)
60 {
61 }
62 virtual bool is_intrinsic() const { return true; }
63 virtual bool is_virtual() const { return _is_virtual; }
64 virtual bool is_predicted() const { return _is_predicted; }
65 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
66 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
67 virtual Node* generate_predicate(JVMState* jvms);
68 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
69 };
70
71
72 // Local helper class for LibraryIntrinsic:
73 class LibraryCallKit : public GraphKit {
74 private:
75 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
76 Node* _result; // the result node, if any
77 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
78
79 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
80
81 public:
82 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
83 : GraphKit(jvms),
84 _intrinsic(intrinsic),
85 _result(NULL)
86 {
87 // Check if this is a root compile. In that case we don't have a caller.
88 if (!jvms->has_method()) {
89 _reexecute_sp = sp();
90 } else {
91 // Find out how many arguments the interpreter needs when deoptimizing
92 // and save the stack pointer value so it can used by uncommon_trap.
93 // We find the argument count by looking at the declared signature.
94 bool ignored_will_link;
95 ciSignature* declared_signature = NULL;
96 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
97 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
98 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
99 }
100 }
101
102 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
103
104 ciMethod* caller() const { return jvms()->method(); }
105 int bci() const { return jvms()->bci(); }
106 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
107 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
108 ciMethod* callee() const { return _intrinsic->method(); }
109
110 bool try_to_inline();
111 Node* try_to_predicate();
112
113 void push_result() {
114 // Push the result onto the stack.
115 if (!stopped() && result() != NULL) {
116 BasicType bt = result()->bottom_type()->basic_type();
117 push_node(bt, result());
118 }
119 }
120
121 private:
122 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
123 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
124 }
125
126 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
127 void set_result(RegionNode* region, PhiNode* value);
128 Node* result() { return _result; }
129
130 virtual int reexecute_sp() { return _reexecute_sp; }
131
132 // Helper functions to inline natives
133 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
134 Node* generate_slow_guard(Node* test, RegionNode* region);
135 Node* generate_fair_guard(Node* test, RegionNode* region);
136 Node* generate_negative_guard(Node* index, RegionNode* region,
137 // resulting CastII of index:
138 Node* *pos_index = NULL);
139 Node* generate_nonpositive_guard(Node* index, bool never_negative,
140 // resulting CastII of index:
141 Node* *pos_index = NULL);
142 Node* generate_limit_guard(Node* offset, Node* subseq_length,
143 Node* array_length,
144 RegionNode* region);
145 Node* generate_current_thread(Node* &tls_output);
146 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
147 bool disjoint_bases, const char* &name, bool dest_uninitialized);
148 Node* load_mirror_from_klass(Node* klass);
149 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
150 RegionNode* region, int null_path,
151 int offset);
152 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
153 RegionNode* region, int null_path) {
154 int offset = java_lang_Class::klass_offset_in_bytes();
155 return load_klass_from_mirror_common(mirror, never_see_null,
156 region, null_path,
157 offset);
158 }
159 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::array_klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* generate_access_flags_guard(Node* kls,
167 int modifier_mask, int modifier_bits,
168 RegionNode* region);
169 Node* generate_interface_guard(Node* kls, RegionNode* region);
170 Node* generate_array_guard(Node* kls, RegionNode* region) {
171 return generate_array_guard_common(kls, region, false, false);
172 }
173 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
174 return generate_array_guard_common(kls, region, false, true);
175 }
176 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, true, false);
178 }
179 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, true, true);
181 }
182 Node* generate_array_guard_common(Node* kls, RegionNode* region,
183 bool obj_array, bool not_array);
184 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
185 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
186 bool is_virtual = false, bool is_static = false);
187 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
188 return generate_method_call(method_id, false, true);
189 }
190 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
191 return generate_method_call(method_id, true, false);
192 }
193 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
194
195 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
196 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
197 bool inline_string_compareTo();
198 bool inline_string_indexOf();
199 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
200 bool inline_string_equals();
201 Node* round_double_node(Node* n);
202 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
203 bool inline_math_native(vmIntrinsics::ID id);
204 bool inline_trig(vmIntrinsics::ID id);
205 bool inline_math(vmIntrinsics::ID id);
206 void inline_math_mathExact(Node* math);
207 bool inline_math_addExact();
208 bool inline_math_addExactL();
209 bool inline_math_decrementExact();
210 bool inline_math_decrementExactL();
211 bool inline_math_incrementExact();
212 bool inline_math_incrementExactL();
213 bool inline_math_multiplyExact();
214 bool inline_math_multiplyExactL();
215 bool inline_math_negateExact();
216 bool inline_math_negateExactL();
217 bool inline_math_subtractExact();
218 bool inline_math_subtractExactL();
219 bool inline_exp();
220 bool inline_pow();
221 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
222 bool inline_min_max(vmIntrinsics::ID id);
223 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
224 // This returns Type::AnyPtr, RawPtr, or OopPtr.
225 int classify_unsafe_addr(Node* &base, Node* &offset);
226 Node* make_unsafe_address(Node* base, Node* offset);
227 // Helper for inline_unsafe_access.
228 // Generates the guards that check whether the result of
229 // Unsafe.getObject should be recorded in an SATB log buffer.
230 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
231 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
232 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
233 static bool klass_needs_init_guard(Node* kls);
234 bool inline_unsafe_allocate();
235 bool inline_unsafe_copyMemory();
236 bool inline_native_currentThread();
237 #ifdef TRACE_HAVE_INTRINSICS
238 bool inline_native_classID();
239 bool inline_native_threadID();
240 #endif
241 bool inline_native_time_funcs(address method, const char* funcName);
242 bool inline_native_isInterrupted();
243 bool inline_native_Class_query(vmIntrinsics::ID id);
244 bool inline_native_subtype_check();
245
246 bool inline_native_newArray();
247 bool inline_native_getLength();
248 bool inline_array_copyOf(bool is_copyOfRange);
249 bool inline_array_equals();
250 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
251 bool inline_native_clone(bool is_virtual);
252 bool inline_native_Reflection_getCallerClass();
253 // Helper function for inlining native object hash method
254 bool inline_native_hashcode(bool is_virtual, bool is_static);
255 bool inline_native_getClass();
256
257 // Helper functions for inlining arraycopy
258 bool inline_arraycopy();
259 void generate_arraycopy(const TypePtr* adr_type,
260 BasicType basic_elem_type,
261 Node* src, Node* src_offset,
262 Node* dest, Node* dest_offset,
263 Node* copy_length,
264 bool disjoint_bases = false,
265 bool length_never_negative = false,
266 RegionNode* slow_region = NULL);
267 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
268 RegionNode* slow_region);
269 void generate_clear_array(const TypePtr* adr_type,
270 Node* dest,
271 BasicType basic_elem_type,
272 Node* slice_off,
273 Node* slice_len,
274 Node* slice_end);
275 bool generate_block_arraycopy(const TypePtr* adr_type,
276 BasicType basic_elem_type,
277 AllocateNode* alloc,
278 Node* src, Node* src_offset,
279 Node* dest, Node* dest_offset,
280 Node* dest_size, bool dest_uninitialized);
281 void generate_slow_arraycopy(const TypePtr* adr_type,
282 Node* src, Node* src_offset,
283 Node* dest, Node* dest_offset,
284 Node* copy_length, bool dest_uninitialized);
285 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
286 Node* dest_elem_klass,
287 Node* src, Node* src_offset,
288 Node* dest, Node* dest_offset,
289 Node* copy_length, bool dest_uninitialized);
290 Node* generate_generic_arraycopy(const TypePtr* adr_type,
291 Node* src, Node* src_offset,
292 Node* dest, Node* dest_offset,
293 Node* copy_length, bool dest_uninitialized);
294 void generate_unchecked_arraycopy(const TypePtr* adr_type,
295 BasicType basic_elem_type,
296 bool disjoint_bases,
297 Node* src, Node* src_offset,
298 Node* dest, Node* dest_offset,
299 Node* copy_length, bool dest_uninitialized);
300 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
301 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
302 bool inline_unsafe_ordered_store(BasicType type);
303 bool inline_unsafe_fence(vmIntrinsics::ID id);
304 bool inline_fp_conversions(vmIntrinsics::ID id);
305 bool inline_number_methods(vmIntrinsics::ID id);
306 bool inline_reference_get();
307 bool inline_aescrypt_Block(vmIntrinsics::ID id);
308 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
309 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
310 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
311 bool inline_encodeISOArray();
312 bool inline_updateCRC32();
313 bool inline_updateBytesCRC32();
314 bool inline_updateByteBufferCRC32();
315 };
316
317
318 //---------------------------make_vm_intrinsic----------------------------
319 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
320 vmIntrinsics::ID id = m->intrinsic_id();
321 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
322
323 if (DisableIntrinsic[0] != '\0'
324 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
325 // disabled by a user request on the command line:
326 // example: -XX:DisableIntrinsic=_hashCode,_getClass
327 return NULL;
328 }
329
330 if (!m->is_loaded()) {
331 // do not attempt to inline unloaded methods
332 return NULL;
333 }
334
335 // Only a few intrinsics implement a virtual dispatch.
336 // They are expensive calls which are also frequently overridden.
337 if (is_virtual) {
338 switch (id) {
339 case vmIntrinsics::_hashCode:
340 case vmIntrinsics::_clone:
341 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
342 break;
343 default:
344 return NULL;
345 }
346 }
347
348 // -XX:-InlineNatives disables nearly all intrinsics:
349 if (!InlineNatives) {
350 switch (id) {
351 case vmIntrinsics::_indexOf:
352 case vmIntrinsics::_compareTo:
353 case vmIntrinsics::_equals:
354 case vmIntrinsics::_equalsC:
355 case vmIntrinsics::_getAndAddInt:
356 case vmIntrinsics::_getAndAddLong:
357 case vmIntrinsics::_getAndSetInt:
358 case vmIntrinsics::_getAndSetLong:
359 case vmIntrinsics::_getAndSetObject:
360 case vmIntrinsics::_loadFence:
361 case vmIntrinsics::_storeFence:
362 case vmIntrinsics::_fullFence:
363 break; // InlineNatives does not control String.compareTo
364 case vmIntrinsics::_Reference_get:
365 break; // InlineNatives does not control Reference.get
366 default:
367 return NULL;
368 }
369 }
370
371 bool is_predicted = false;
372 bool does_virtual_dispatch = false;
373
374 switch (id) {
375 case vmIntrinsics::_compareTo:
376 if (!SpecialStringCompareTo) return NULL;
377 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
378 break;
379 case vmIntrinsics::_indexOf:
380 if (!SpecialStringIndexOf) return NULL;
381 break;
382 case vmIntrinsics::_equals:
383 if (!SpecialStringEquals) return NULL;
384 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
385 break;
386 case vmIntrinsics::_equalsC:
387 if (!SpecialArraysEquals) return NULL;
388 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
389 break;
390 case vmIntrinsics::_arraycopy:
391 if (!InlineArrayCopy) return NULL;
392 break;
393 case vmIntrinsics::_copyMemory:
394 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
395 if (!InlineArrayCopy) return NULL;
396 break;
397 case vmIntrinsics::_hashCode:
398 if (!InlineObjectHash) return NULL;
399 does_virtual_dispatch = true;
400 break;
401 case vmIntrinsics::_clone:
402 does_virtual_dispatch = true;
403 case vmIntrinsics::_copyOf:
404 case vmIntrinsics::_copyOfRange:
405 if (!InlineObjectCopy) return NULL;
406 // These also use the arraycopy intrinsic mechanism:
407 if (!InlineArrayCopy) return NULL;
408 break;
409 case vmIntrinsics::_encodeISOArray:
410 if (!SpecialEncodeISOArray) return NULL;
411 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
412 break;
413 case vmIntrinsics::_checkIndex:
414 // We do not intrinsify this. The optimizer does fine with it.
415 return NULL;
416
417 case vmIntrinsics::_getCallerClass:
418 if (!UseNewReflection) return NULL;
419 if (!InlineReflectionGetCallerClass) return NULL;
420 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
421 break;
422
423 case vmIntrinsics::_bitCount_i:
424 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
425 break;
426
427 case vmIntrinsics::_bitCount_l:
428 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
429 break;
430
431 case vmIntrinsics::_numberOfLeadingZeros_i:
432 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
433 break;
434
435 case vmIntrinsics::_numberOfLeadingZeros_l:
436 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
437 break;
438
439 case vmIntrinsics::_numberOfTrailingZeros_i:
440 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
441 break;
442
443 case vmIntrinsics::_numberOfTrailingZeros_l:
444 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
445 break;
446
447 case vmIntrinsics::_reverseBytes_c:
448 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
449 break;
450 case vmIntrinsics::_reverseBytes_s:
451 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
452 break;
453 case vmIntrinsics::_reverseBytes_i:
454 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
455 break;
456 case vmIntrinsics::_reverseBytes_l:
457 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
458 break;
459
460 case vmIntrinsics::_Reference_get:
461 // Use the intrinsic version of Reference.get() so that the value in
462 // the referent field can be registered by the G1 pre-barrier code.
463 // Also add memory barrier to prevent commoning reads from this field
464 // across safepoint since GC can change it value.
465 break;
466
467 case vmIntrinsics::_compareAndSwapObject:
468 #ifdef _LP64
469 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
470 #endif
471 break;
472
473 case vmIntrinsics::_compareAndSwapLong:
474 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
475 break;
476
477 case vmIntrinsics::_getAndAddInt:
478 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
479 break;
480
481 case vmIntrinsics::_getAndAddLong:
482 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
483 break;
484
485 case vmIntrinsics::_getAndSetInt:
486 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
487 break;
488
489 case vmIntrinsics::_getAndSetLong:
490 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
491 break;
492
493 case vmIntrinsics::_getAndSetObject:
494 #ifdef _LP64
495 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
496 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
497 break;
498 #else
499 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
500 break;
501 #endif
502
503 case vmIntrinsics::_aescrypt_encryptBlock:
504 case vmIntrinsics::_aescrypt_decryptBlock:
505 if (!UseAESIntrinsics) return NULL;
506 break;
507
508 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
509 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
510 if (!UseAESIntrinsics) return NULL;
511 // these two require the predicated logic
512 is_predicted = true;
513 break;
514
515 case vmIntrinsics::_updateCRC32:
516 case vmIntrinsics::_updateBytesCRC32:
517 case vmIntrinsics::_updateByteBufferCRC32:
518 if (!UseCRC32Intrinsics) return NULL;
519 break;
520
521 case vmIntrinsics::_incrementExact:
522 case vmIntrinsics::_addExact:
523 if (!Matcher::match_rule_supported(Op_AddExactI) || !UseMathExactIntrinsics) return NULL;
524 break;
525 case vmIntrinsics::_incrementExactL:
526 case vmIntrinsics::_addExactL:
527 if (!Matcher::match_rule_supported(Op_AddExactL) || !UseMathExactIntrinsics) return NULL;
528 break;
529 case vmIntrinsics::_decrementExact:
530 case vmIntrinsics::_subtractExact:
531 if (!Matcher::match_rule_supported(Op_SubExactI) || !UseMathExactIntrinsics) return NULL;
532 break;
533 case vmIntrinsics::_decrementExactL:
534 case vmIntrinsics::_subtractExactL:
535 if (!Matcher::match_rule_supported(Op_SubExactL) || !UseMathExactIntrinsics) return NULL;
536 break;
537 case vmIntrinsics::_negateExact:
538 if (!Matcher::match_rule_supported(Op_NegExactI) || !UseMathExactIntrinsics) return NULL;
539 break;
540 case vmIntrinsics::_negateExactL:
541 if (!Matcher::match_rule_supported(Op_NegExactL) || !UseMathExactIntrinsics) return NULL;
542 break;
543 case vmIntrinsics::_multiplyExact:
544 if (!Matcher::match_rule_supported(Op_MulExactI) || !UseMathExactIntrinsics) return NULL;
545 break;
546 case vmIntrinsics::_multiplyExactL:
547 if (!Matcher::match_rule_supported(Op_MulExactL) || !UseMathExactIntrinsics) return NULL;
548 break;
549
550 default:
551 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
552 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
553 break;
554 }
555
556 // -XX:-InlineClassNatives disables natives from the Class class.
557 // The flag applies to all reflective calls, notably Array.newArray
558 // (visible to Java programmers as Array.newInstance).
559 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
560 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
561 if (!InlineClassNatives) return NULL;
562 }
563
564 // -XX:-InlineThreadNatives disables natives from the Thread class.
565 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
566 if (!InlineThreadNatives) return NULL;
567 }
568
569 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
570 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
571 m->holder()->name() == ciSymbol::java_lang_Float() ||
572 m->holder()->name() == ciSymbol::java_lang_Double()) {
573 if (!InlineMathNatives) return NULL;
574 }
575
576 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
577 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
578 if (!InlineUnsafeOps) return NULL;
579 }
580
581 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
582 }
583
584 //----------------------register_library_intrinsics-----------------------
585 // Initialize this file's data structures, for each Compile instance.
586 void Compile::register_library_intrinsics() {
587 // Nothing to do here.
588 }
589
590 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
591 LibraryCallKit kit(jvms, this);
592 Compile* C = kit.C;
593 int nodes = C->unique();
594 #ifndef PRODUCT
595 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
596 char buf[1000];
597 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
598 tty->print_cr("Intrinsic %s", str);
599 }
600 #endif
601 ciMethod* callee = kit.callee();
602 const int bci = kit.bci();
603
604 // Try to inline the intrinsic.
605 if (kit.try_to_inline()) {
606 if (C->print_intrinsics() || C->print_inlining()) {
607 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
608 }
609 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
610 if (C->log()) {
611 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
612 vmIntrinsics::name_at(intrinsic_id()),
613 (is_virtual() ? " virtual='1'" : ""),
614 C->unique() - nodes);
615 }
616 // Push the result from the inlined method onto the stack.
617 kit.push_result();
618 return kit.transfer_exceptions_into_jvms();
619 }
620
621 // The intrinsic bailed out
622 if (C->print_intrinsics() || C->print_inlining()) {
623 if (jvms->has_method()) {
624 // Not a root compile.
625 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
626 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
627 } else {
628 // Root compile
629 tty->print("Did not generate intrinsic %s%s at bci:%d in",
630 vmIntrinsics::name_at(intrinsic_id()),
631 (is_virtual() ? " (virtual)" : ""), bci);
632 }
633 }
634 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
635 return NULL;
636 }
637
638 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
639 LibraryCallKit kit(jvms, this);
640 Compile* C = kit.C;
641 int nodes = C->unique();
642 #ifndef PRODUCT
643 assert(is_predicted(), "sanity");
644 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
645 char buf[1000];
646 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
647 tty->print_cr("Predicate for intrinsic %s", str);
648 }
649 #endif
650 ciMethod* callee = kit.callee();
651 const int bci = kit.bci();
652
653 Node* slow_ctl = kit.try_to_predicate();
654 if (!kit.failing()) {
655 if (C->print_intrinsics() || C->print_inlining()) {
656 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
657 }
658 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
659 if (C->log()) {
660 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
661 vmIntrinsics::name_at(intrinsic_id()),
662 (is_virtual() ? " virtual='1'" : ""),
663 C->unique() - nodes);
664 }
665 return slow_ctl; // Could be NULL if the check folds.
666 }
667
668 // The intrinsic bailed out
669 if (C->print_intrinsics() || C->print_inlining()) {
670 if (jvms->has_method()) {
671 // Not a root compile.
672 const char* msg = "failed to generate predicate for intrinsic";
673 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
674 } else {
675 // Root compile
676 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
677 vmIntrinsics::name_at(intrinsic_id()),
678 (is_virtual() ? " (virtual)" : ""), bci);
679 }
680 }
681 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
682 return NULL;
683 }
684
685 bool LibraryCallKit::try_to_inline() {
686 // Handle symbolic names for otherwise undistinguished boolean switches:
687 const bool is_store = true;
688 const bool is_native_ptr = true;
689 const bool is_static = true;
690 const bool is_volatile = true;
691
692 if (!jvms()->has_method()) {
693 // Root JVMState has a null method.
694 assert(map()->memory()->Opcode() == Op_Parm, "");
695 // Insert the memory aliasing node
696 set_all_memory(reset_memory());
697 }
698 assert(merged_memory(), "");
699
700
701 switch (intrinsic_id()) {
702 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
703 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
704 case vmIntrinsics::_getClass: return inline_native_getClass();
705
706 case vmIntrinsics::_dsin:
707 case vmIntrinsics::_dcos:
708 case vmIntrinsics::_dtan:
709 case vmIntrinsics::_dabs:
710 case vmIntrinsics::_datan2:
711 case vmIntrinsics::_dsqrt:
712 case vmIntrinsics::_dexp:
713 case vmIntrinsics::_dlog:
714 case vmIntrinsics::_dlog10:
715 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
716
717 case vmIntrinsics::_min:
718 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
719
720 case vmIntrinsics::_addExact: return inline_math_addExact();
721 case vmIntrinsics::_addExactL: return inline_math_addExactL();
722 case vmIntrinsics::_decrementExact: return inline_math_decrementExact();
723 case vmIntrinsics::_decrementExactL: return inline_math_decrementExactL();
724 case vmIntrinsics::_incrementExact: return inline_math_incrementExact();
725 case vmIntrinsics::_incrementExactL: return inline_math_incrementExactL();
726 case vmIntrinsics::_multiplyExact: return inline_math_multiplyExact();
727 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
728 case vmIntrinsics::_negateExact: return inline_math_negateExact();
729 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
730 case vmIntrinsics::_subtractExact: return inline_math_subtractExact();
731 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL();
732
733 case vmIntrinsics::_arraycopy: return inline_arraycopy();
734
735 case vmIntrinsics::_compareTo: return inline_string_compareTo();
736 case vmIntrinsics::_indexOf: return inline_string_indexOf();
737 case vmIntrinsics::_equals: return inline_string_equals();
738
739 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
740 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
741 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
742 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
743 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
744 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
745 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
746 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
747 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
748
749 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
750 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
751 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
752 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
753 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
754 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
755 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
756 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
757 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
758
759 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
760 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
761 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
762 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
763 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
764 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
765 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
766 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
767
768 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
769 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
770 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
771 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
772 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
773 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
774 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
775 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
776
777 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
778 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
779 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
780 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
781 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
782 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
783 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
784 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
785 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
786
787 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
788 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
789 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
790 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
791 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
792 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
793 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
794 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
795 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
796
797 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
798 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
799 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
800 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
801
802 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
803 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
804 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
805
806 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
807 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
808 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
809
810 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
811 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
812 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
813 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
814 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
815
816 case vmIntrinsics::_loadFence:
817 case vmIntrinsics::_storeFence:
818 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
819
820 case vmIntrinsics::_currentThread: return inline_native_currentThread();
821 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
822
823 #ifdef TRACE_HAVE_INTRINSICS
824 case vmIntrinsics::_classID: return inline_native_classID();
825 case vmIntrinsics::_threadID: return inline_native_threadID();
826 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
827 #endif
828 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
829 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
830 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
831 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
832 case vmIntrinsics::_newArray: return inline_native_newArray();
833 case vmIntrinsics::_getLength: return inline_native_getLength();
834 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
835 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
836 case vmIntrinsics::_equalsC: return inline_array_equals();
837 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
838
839 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
840
841 case vmIntrinsics::_isInstance:
842 case vmIntrinsics::_getModifiers:
843 case vmIntrinsics::_isInterface:
844 case vmIntrinsics::_isArray:
845 case vmIntrinsics::_isPrimitive:
846 case vmIntrinsics::_getSuperclass:
847 case vmIntrinsics::_getComponentType:
848 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
849
850 case vmIntrinsics::_floatToRawIntBits:
851 case vmIntrinsics::_floatToIntBits:
852 case vmIntrinsics::_intBitsToFloat:
853 case vmIntrinsics::_doubleToRawLongBits:
854 case vmIntrinsics::_doubleToLongBits:
855 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
856
857 case vmIntrinsics::_numberOfLeadingZeros_i:
858 case vmIntrinsics::_numberOfLeadingZeros_l:
859 case vmIntrinsics::_numberOfTrailingZeros_i:
860 case vmIntrinsics::_numberOfTrailingZeros_l:
861 case vmIntrinsics::_bitCount_i:
862 case vmIntrinsics::_bitCount_l:
863 case vmIntrinsics::_reverseBytes_i:
864 case vmIntrinsics::_reverseBytes_l:
865 case vmIntrinsics::_reverseBytes_s:
866 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
867
868 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
869
870 case vmIntrinsics::_Reference_get: return inline_reference_get();
871
872 case vmIntrinsics::_aescrypt_encryptBlock:
873 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
874
875 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
876 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
877 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
878
879 case vmIntrinsics::_encodeISOArray:
880 return inline_encodeISOArray();
881
882 case vmIntrinsics::_updateCRC32:
883 return inline_updateCRC32();
884 case vmIntrinsics::_updateBytesCRC32:
885 return inline_updateBytesCRC32();
886 case vmIntrinsics::_updateByteBufferCRC32:
887 return inline_updateByteBufferCRC32();
888
889 default:
890 // If you get here, it may be that someone has added a new intrinsic
891 // to the list in vmSymbols.hpp without implementing it here.
892 #ifndef PRODUCT
893 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
894 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
895 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
896 }
897 #endif
898 return false;
899 }
900 }
901
902 Node* LibraryCallKit::try_to_predicate() {
903 if (!jvms()->has_method()) {
904 // Root JVMState has a null method.
905 assert(map()->memory()->Opcode() == Op_Parm, "");
906 // Insert the memory aliasing node
907 set_all_memory(reset_memory());
908 }
909 assert(merged_memory(), "");
910
911 switch (intrinsic_id()) {
912 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
913 return inline_cipherBlockChaining_AESCrypt_predicate(false);
914 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
915 return inline_cipherBlockChaining_AESCrypt_predicate(true);
916
917 default:
918 // If you get here, it may be that someone has added a new intrinsic
919 // to the list in vmSymbols.hpp without implementing it here.
920 #ifndef PRODUCT
921 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
922 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
923 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
924 }
925 #endif
926 Node* slow_ctl = control();
927 set_control(top()); // No fast path instrinsic
928 return slow_ctl;
929 }
930 }
931
932 //------------------------------set_result-------------------------------
933 // Helper function for finishing intrinsics.
934 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
935 record_for_igvn(region);
936 set_control(_gvn.transform(region));
937 set_result( _gvn.transform(value));
938 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
939 }
940
941 //------------------------------generate_guard---------------------------
942 // Helper function for generating guarded fast-slow graph structures.
943 // The given 'test', if true, guards a slow path. If the test fails
944 // then a fast path can be taken. (We generally hope it fails.)
945 // In all cases, GraphKit::control() is updated to the fast path.
946 // The returned value represents the control for the slow path.
947 // The return value is never 'top'; it is either a valid control
948 // or NULL if it is obvious that the slow path can never be taken.
949 // Also, if region and the slow control are not NULL, the slow edge
950 // is appended to the region.
951 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
952 if (stopped()) {
953 // Already short circuited.
954 return NULL;
955 }
956
957 // Build an if node and its projections.
958 // If test is true we take the slow path, which we assume is uncommon.
959 if (_gvn.type(test) == TypeInt::ZERO) {
960 // The slow branch is never taken. No need to build this guard.
961 return NULL;
962 }
963
964 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
965
966 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
967 if (if_slow == top()) {
968 // The slow branch is never taken. No need to build this guard.
969 return NULL;
970 }
971
972 if (region != NULL)
973 region->add_req(if_slow);
974
975 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
976 set_control(if_fast);
977
978 return if_slow;
979 }
980
981 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
982 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
983 }
984 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
985 return generate_guard(test, region, PROB_FAIR);
986 }
987
988 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
989 Node* *pos_index) {
990 if (stopped())
991 return NULL; // already stopped
992 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
993 return NULL; // index is already adequately typed
994 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
995 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
996 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
997 if (is_neg != NULL && pos_index != NULL) {
998 // Emulate effect of Parse::adjust_map_after_if.
999 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1000 ccast->set_req(0, control());
1001 (*pos_index) = _gvn.transform(ccast);
1002 }
1003 return is_neg;
1004 }
1005
1006 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1007 Node* *pos_index) {
1008 if (stopped())
1009 return NULL; // already stopped
1010 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1011 return NULL; // index is already adequately typed
1012 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1013 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1014 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1015 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1016 if (is_notp != NULL && pos_index != NULL) {
1017 // Emulate effect of Parse::adjust_map_after_if.
1018 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1019 ccast->set_req(0, control());
1020 (*pos_index) = _gvn.transform(ccast);
1021 }
1022 return is_notp;
1023 }
1024
1025 // Make sure that 'position' is a valid limit index, in [0..length].
1026 // There are two equivalent plans for checking this:
1027 // A. (offset + copyLength) unsigned<= arrayLength
1028 // B. offset <= (arrayLength - copyLength)
1029 // We require that all of the values above, except for the sum and
1030 // difference, are already known to be non-negative.
1031 // Plan A is robust in the face of overflow, if offset and copyLength
1032 // are both hugely positive.
1033 //
1034 // Plan B is less direct and intuitive, but it does not overflow at
1035 // all, since the difference of two non-negatives is always
1036 // representable. Whenever Java methods must perform the equivalent
1037 // check they generally use Plan B instead of Plan A.
1038 // For the moment we use Plan A.
1039 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1040 Node* subseq_length,
1041 Node* array_length,
1042 RegionNode* region) {
1043 if (stopped())
1044 return NULL; // already stopped
1045 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1046 if (zero_offset && subseq_length->eqv_uncast(array_length))
1047 return NULL; // common case of whole-array copy
1048 Node* last = subseq_length;
1049 if (!zero_offset) // last += offset
1050 last = _gvn.transform(new (C) AddINode(last, offset));
1051 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1052 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1053 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1054 return is_over;
1055 }
1056
1057
1058 //--------------------------generate_current_thread--------------------
1059 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1060 ciKlass* thread_klass = env()->Thread_klass();
1061 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1062 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1063 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1064 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1065 tls_output = thread;
1066 return threadObj;
1067 }
1068
1069
1070 //------------------------------make_string_method_node------------------------
1071 // Helper method for String intrinsic functions. This version is called
1072 // with str1 and str2 pointing to String object nodes.
1073 //
1074 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1075 Node* no_ctrl = NULL;
1076
1077 // Get start addr of string
1078 Node* str1_value = load_String_value(no_ctrl, str1);
1079 Node* str1_offset = load_String_offset(no_ctrl, str1);
1080 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1081
1082 // Get length of string 1
1083 Node* str1_len = load_String_length(no_ctrl, str1);
1084
1085 Node* str2_value = load_String_value(no_ctrl, str2);
1086 Node* str2_offset = load_String_offset(no_ctrl, str2);
1087 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1088
1089 Node* str2_len = NULL;
1090 Node* result = NULL;
1091
1092 switch (opcode) {
1093 case Op_StrIndexOf:
1094 // Get length of string 2
1095 str2_len = load_String_length(no_ctrl, str2);
1096
1097 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1098 str1_start, str1_len, str2_start, str2_len);
1099 break;
1100 case Op_StrComp:
1101 // Get length of string 2
1102 str2_len = load_String_length(no_ctrl, str2);
1103
1104 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1105 str1_start, str1_len, str2_start, str2_len);
1106 break;
1107 case Op_StrEquals:
1108 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1109 str1_start, str2_start, str1_len);
1110 break;
1111 default:
1112 ShouldNotReachHere();
1113 return NULL;
1114 }
1115
1116 // All these intrinsics have checks.
1117 C->set_has_split_ifs(true); // Has chance for split-if optimization
1118
1119 return _gvn.transform(result);
1120 }
1121
1122 // Helper method for String intrinsic functions. This version is called
1123 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1124 // to Int nodes containing the lenghts of str1 and str2.
1125 //
1126 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1127 Node* result = NULL;
1128 switch (opcode) {
1129 case Op_StrIndexOf:
1130 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1131 str1_start, cnt1, str2_start, cnt2);
1132 break;
1133 case Op_StrComp:
1134 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1135 str1_start, cnt1, str2_start, cnt2);
1136 break;
1137 case Op_StrEquals:
1138 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1139 str1_start, str2_start, cnt1);
1140 break;
1141 default:
1142 ShouldNotReachHere();
1143 return NULL;
1144 }
1145
1146 // All these intrinsics have checks.
1147 C->set_has_split_ifs(true); // Has chance for split-if optimization
1148
1149 return _gvn.transform(result);
1150 }
1151
1152 //------------------------------inline_string_compareTo------------------------
1153 // public int java.lang.String.compareTo(String anotherString);
1154 bool LibraryCallKit::inline_string_compareTo() {
1155 Node* receiver = null_check(argument(0));
1156 Node* arg = null_check(argument(1));
1157 if (stopped()) {
1158 return true;
1159 }
1160 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1161 return true;
1162 }
1163
1164 //------------------------------inline_string_equals------------------------
1165 bool LibraryCallKit::inline_string_equals() {
1166 Node* receiver = null_check_receiver();
1167 // NOTE: Do not null check argument for String.equals() because spec
1168 // allows to specify NULL as argument.
1169 Node* argument = this->argument(1);
1170 if (stopped()) {
1171 return true;
1172 }
1173
1174 // paths (plus control) merge
1175 RegionNode* region = new (C) RegionNode(5);
1176 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1177
1178 // does source == target string?
1179 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1180 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1181
1182 Node* if_eq = generate_slow_guard(bol, NULL);
1183 if (if_eq != NULL) {
1184 // receiver == argument
1185 phi->init_req(2, intcon(1));
1186 region->init_req(2, if_eq);
1187 }
1188
1189 // get String klass for instanceOf
1190 ciInstanceKlass* klass = env()->String_klass();
1191
1192 if (!stopped()) {
1193 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1194 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1195 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1196
1197 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1198 //instanceOf == true, fallthrough
1199
1200 if (inst_false != NULL) {
1201 phi->init_req(3, intcon(0));
1202 region->init_req(3, inst_false);
1203 }
1204 }
1205
1206 if (!stopped()) {
1207 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1208
1209 // Properly cast the argument to String
1210 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1211 // This path is taken only when argument's type is String:NotNull.
1212 argument = cast_not_null(argument, false);
1213
1214 Node* no_ctrl = NULL;
1215
1216 // Get start addr of receiver
1217 Node* receiver_val = load_String_value(no_ctrl, receiver);
1218 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1219 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1220
1221 // Get length of receiver
1222 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1223
1224 // Get start addr of argument
1225 Node* argument_val = load_String_value(no_ctrl, argument);
1226 Node* argument_offset = load_String_offset(no_ctrl, argument);
1227 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1228
1229 // Get length of argument
1230 Node* argument_cnt = load_String_length(no_ctrl, argument);
1231
1232 // Check for receiver count != argument count
1233 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1234 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1235 Node* if_ne = generate_slow_guard(bol, NULL);
1236 if (if_ne != NULL) {
1237 phi->init_req(4, intcon(0));
1238 region->init_req(4, if_ne);
1239 }
1240
1241 // Check for count == 0 is done by assembler code for StrEquals.
1242
1243 if (!stopped()) {
1244 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1245 phi->init_req(1, equals);
1246 region->init_req(1, control());
1247 }
1248 }
1249
1250 // post merge
1251 set_control(_gvn.transform(region));
1252 record_for_igvn(region);
1253
1254 set_result(_gvn.transform(phi));
1255 return true;
1256 }
1257
1258 //------------------------------inline_array_equals----------------------------
1259 bool LibraryCallKit::inline_array_equals() {
1260 Node* arg1 = argument(0);
1261 Node* arg2 = argument(1);
1262 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1263 return true;
1264 }
1265
1266 // Java version of String.indexOf(constant string)
1267 // class StringDecl {
1268 // StringDecl(char[] ca) {
1269 // offset = 0;
1270 // count = ca.length;
1271 // value = ca;
1272 // }
1273 // int offset;
1274 // int count;
1275 // char[] value;
1276 // }
1277 //
1278 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1279 // int targetOffset, int cache_i, int md2) {
1280 // int cache = cache_i;
1281 // int sourceOffset = string_object.offset;
1282 // int sourceCount = string_object.count;
1283 // int targetCount = target_object.length;
1284 //
1285 // int targetCountLess1 = targetCount - 1;
1286 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1287 //
1288 // char[] source = string_object.value;
1289 // char[] target = target_object;
1290 // int lastChar = target[targetCountLess1];
1291 //
1292 // outer_loop:
1293 // for (int i = sourceOffset; i < sourceEnd; ) {
1294 // int src = source[i + targetCountLess1];
1295 // if (src == lastChar) {
1296 // // With random strings and a 4-character alphabet,
1297 // // reverse matching at this point sets up 0.8% fewer
1298 // // frames, but (paradoxically) makes 0.3% more probes.
1299 // // Since those probes are nearer the lastChar probe,
1300 // // there is may be a net D$ win with reverse matching.
1301 // // But, reversing loop inhibits unroll of inner loop
1302 // // for unknown reason. So, does running outer loop from
1303 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1304 // for (int j = 0; j < targetCountLess1; j++) {
1305 // if (target[targetOffset + j] != source[i+j]) {
1306 // if ((cache & (1 << source[i+j])) == 0) {
1307 // if (md2 < j+1) {
1308 // i += j+1;
1309 // continue outer_loop;
1310 // }
1311 // }
1312 // i += md2;
1313 // continue outer_loop;
1314 // }
1315 // }
1316 // return i - sourceOffset;
1317 // }
1318 // if ((cache & (1 << src)) == 0) {
1319 // i += targetCountLess1;
1320 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1321 // i++;
1322 // }
1323 // return -1;
1324 // }
1325
1326 //------------------------------string_indexOf------------------------
1327 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1328 jint cache_i, jint md2_i) {
1329
1330 Node* no_ctrl = NULL;
1331 float likely = PROB_LIKELY(0.9);
1332 float unlikely = PROB_UNLIKELY(0.9);
1333
1334 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1335
1336 Node* source = load_String_value(no_ctrl, string_object);
1337 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1338 Node* sourceCount = load_String_length(no_ctrl, string_object);
1339
1340 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1341 jint target_length = target_array->length();
1342 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1343 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1344
1345 // String.value field is known to be @Stable.
1346 if (UseImplicitStableValues) {
1347 target = cast_array_to_stable(target, target_type);
1348 }
1349
1350 IdealKit kit(this, false, true);
1351 #define __ kit.
1352 Node* zero = __ ConI(0);
1353 Node* one = __ ConI(1);
1354 Node* cache = __ ConI(cache_i);
1355 Node* md2 = __ ConI(md2_i);
1356 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1357 Node* targetCount = __ ConI(target_length);
1358 Node* targetCountLess1 = __ ConI(target_length - 1);
1359 Node* targetOffset = __ ConI(targetOffset_i);
1360 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1361
1362 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1363 Node* outer_loop = __ make_label(2 /* goto */);
1364 Node* return_ = __ make_label(1);
1365
1366 __ set(rtn,__ ConI(-1));
1367 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1368 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1369 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1370 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1371 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1372 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1373 Node* tpj = __ AddI(targetOffset, __ value(j));
1374 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1375 Node* ipj = __ AddI(__ value(i), __ value(j));
1376 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1377 __ if_then(targ, BoolTest::ne, src2); {
1378 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1379 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1380 __ increment(i, __ AddI(__ value(j), one));
1381 __ goto_(outer_loop);
1382 } __ end_if(); __ dead(j);
1383 }__ end_if(); __ dead(j);
1384 __ increment(i, md2);
1385 __ goto_(outer_loop);
1386 }__ end_if();
1387 __ increment(j, one);
1388 }__ end_loop(); __ dead(j);
1389 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1390 __ goto_(return_);
1391 }__ end_if();
1392 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1393 __ increment(i, targetCountLess1);
1394 }__ end_if();
1395 __ increment(i, one);
1396 __ bind(outer_loop);
1397 }__ end_loop(); __ dead(i);
1398 __ bind(return_);
1399
1400 // Final sync IdealKit and GraphKit.
1401 final_sync(kit);
1402 Node* result = __ value(rtn);
1403 #undef __
1404 C->set_has_loops(true);
1405 return result;
1406 }
1407
1408 //------------------------------inline_string_indexOf------------------------
1409 bool LibraryCallKit::inline_string_indexOf() {
1410 Node* receiver = argument(0);
1411 Node* arg = argument(1);
1412
1413 Node* result;
1414 // Disable the use of pcmpestri until it can be guaranteed that
1415 // the load doesn't cross into the uncommited space.
1416 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1417 UseSSE42Intrinsics) {
1418 // Generate SSE4.2 version of indexOf
1419 // We currently only have match rules that use SSE4.2
1420
1421 receiver = null_check(receiver);
1422 arg = null_check(arg);
1423 if (stopped()) {
1424 return true;
1425 }
1426
1427 ciInstanceKlass* str_klass = env()->String_klass();
1428 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1429
1430 // Make the merge point
1431 RegionNode* result_rgn = new (C) RegionNode(4);
1432 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1433 Node* no_ctrl = NULL;
1434
1435 // Get start addr of source string
1436 Node* source = load_String_value(no_ctrl, receiver);
1437 Node* source_offset = load_String_offset(no_ctrl, receiver);
1438 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1439
1440 // Get length of source string
1441 Node* source_cnt = load_String_length(no_ctrl, receiver);
1442
1443 // Get start addr of substring
1444 Node* substr = load_String_value(no_ctrl, arg);
1445 Node* substr_offset = load_String_offset(no_ctrl, arg);
1446 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1447
1448 // Get length of source string
1449 Node* substr_cnt = load_String_length(no_ctrl, arg);
1450
1451 // Check for substr count > string count
1452 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1453 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1454 Node* if_gt = generate_slow_guard(bol, NULL);
1455 if (if_gt != NULL) {
1456 result_phi->init_req(2, intcon(-1));
1457 result_rgn->init_req(2, if_gt);
1458 }
1459
1460 if (!stopped()) {
1461 // Check for substr count == 0
1462 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1463 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1464 Node* if_zero = generate_slow_guard(bol, NULL);
1465 if (if_zero != NULL) {
1466 result_phi->init_req(3, intcon(0));
1467 result_rgn->init_req(3, if_zero);
1468 }
1469 }
1470
1471 if (!stopped()) {
1472 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1473 result_phi->init_req(1, result);
1474 result_rgn->init_req(1, control());
1475 }
1476 set_control(_gvn.transform(result_rgn));
1477 record_for_igvn(result_rgn);
1478 result = _gvn.transform(result_phi);
1479
1480 } else { // Use LibraryCallKit::string_indexOf
1481 // don't intrinsify if argument isn't a constant string.
1482 if (!arg->is_Con()) {
1483 return false;
1484 }
1485 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1486 if (str_type == NULL) {
1487 return false;
1488 }
1489 ciInstanceKlass* klass = env()->String_klass();
1490 ciObject* str_const = str_type->const_oop();
1491 if (str_const == NULL || str_const->klass() != klass) {
1492 return false;
1493 }
1494 ciInstance* str = str_const->as_instance();
1495 assert(str != NULL, "must be instance");
1496
1497 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1498 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1499
1500 int o;
1501 int c;
1502 if (java_lang_String::has_offset_field()) {
1503 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1504 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1505 } else {
1506 o = 0;
1507 c = pat->length();
1508 }
1509
1510 // constant strings have no offset and count == length which
1511 // simplifies the resulting code somewhat so lets optimize for that.
1512 if (o != 0 || c != pat->length()) {
1513 return false;
1514 }
1515
1516 receiver = null_check(receiver, T_OBJECT);
1517 // NOTE: No null check on the argument is needed since it's a constant String oop.
1518 if (stopped()) {
1519 return true;
1520 }
1521
1522 // The null string as a pattern always returns 0 (match at beginning of string)
1523 if (c == 0) {
1524 set_result(intcon(0));
1525 return true;
1526 }
1527
1528 // Generate default indexOf
1529 jchar lastChar = pat->char_at(o + (c - 1));
1530 int cache = 0;
1531 int i;
1532 for (i = 0; i < c - 1; i++) {
1533 assert(i < pat->length(), "out of range");
1534 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1535 }
1536
1537 int md2 = c;
1538 for (i = 0; i < c - 1; i++) {
1539 assert(i < pat->length(), "out of range");
1540 if (pat->char_at(o + i) == lastChar) {
1541 md2 = (c - 1) - i;
1542 }
1543 }
1544
1545 result = string_indexOf(receiver, pat, o, cache, md2);
1546 }
1547 set_result(result);
1548 return true;
1549 }
1550
1551 //--------------------------round_double_node--------------------------------
1552 // Round a double node if necessary.
1553 Node* LibraryCallKit::round_double_node(Node* n) {
1554 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1555 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1556 return n;
1557 }
1558
1559 //------------------------------inline_math-----------------------------------
1560 // public static double Math.abs(double)
1561 // public static double Math.sqrt(double)
1562 // public static double Math.log(double)
1563 // public static double Math.log10(double)
1564 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1565 Node* arg = round_double_node(argument(0));
1566 Node* n;
1567 switch (id) {
1568 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1569 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1570 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1571 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1572 default: fatal_unexpected_iid(id); break;
1573 }
1574 set_result(_gvn.transform(n));
1575 return true;
1576 }
1577
1578 //------------------------------inline_trig----------------------------------
1579 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1580 // argument reduction which will turn into a fast/slow diamond.
1581 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1582 Node* arg = round_double_node(argument(0));
1583 Node* n = NULL;
1584
1585 switch (id) {
1586 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1587 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1588 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1589 default: fatal_unexpected_iid(id); break;
1590 }
1591 n = _gvn.transform(n);
1592
1593 // Rounding required? Check for argument reduction!
1594 if (Matcher::strict_fp_requires_explicit_rounding) {
1595 static const double pi_4 = 0.7853981633974483;
1596 static const double neg_pi_4 = -0.7853981633974483;
1597 // pi/2 in 80-bit extended precision
1598 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1599 // -pi/2 in 80-bit extended precision
1600 // 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};
1601 // Cutoff value for using this argument reduction technique
1602 //static const double pi_2_minus_epsilon = 1.564660403643354;
1603 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1604
1605 // Pseudocode for sin:
1606 // if (x <= Math.PI / 4.0) {
1607 // if (x >= -Math.PI / 4.0) return fsin(x);
1608 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1609 // } else {
1610 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1611 // }
1612 // return StrictMath.sin(x);
1613
1614 // Pseudocode for cos:
1615 // if (x <= Math.PI / 4.0) {
1616 // if (x >= -Math.PI / 4.0) return fcos(x);
1617 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1618 // } else {
1619 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1620 // }
1621 // return StrictMath.cos(x);
1622
1623 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1624 // requires a special machine instruction to load it. Instead we'll try
1625 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1626 // probably do the math inside the SIN encoding.
1627
1628 // Make the merge point
1629 RegionNode* r = new (C) RegionNode(3);
1630 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1631
1632 // Flatten arg so we need only 1 test
1633 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1634 // Node for PI/4 constant
1635 Node *pi4 = makecon(TypeD::make(pi_4));
1636 // Check PI/4 : abs(arg)
1637 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1638 // Check: If PI/4 < abs(arg) then go slow
1639 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1640 // Branch either way
1641 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1642 set_control(opt_iff(r,iff));
1643
1644 // Set fast path result
1645 phi->init_req(2, n);
1646
1647 // Slow path - non-blocking leaf call
1648 Node* call = NULL;
1649 switch (id) {
1650 case vmIntrinsics::_dsin:
1651 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1652 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1653 "Sin", NULL, arg, top());
1654 break;
1655 case vmIntrinsics::_dcos:
1656 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1657 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1658 "Cos", NULL, arg, top());
1659 break;
1660 case vmIntrinsics::_dtan:
1661 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1662 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1663 "Tan", NULL, arg, top());
1664 break;
1665 }
1666 assert(control()->in(0) == call, "");
1667 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1668 r->init_req(1, control());
1669 phi->init_req(1, slow_result);
1670
1671 // Post-merge
1672 set_control(_gvn.transform(r));
1673 record_for_igvn(r);
1674 n = _gvn.transform(phi);
1675
1676 C->set_has_split_ifs(true); // Has chance for split-if optimization
1677 }
1678 set_result(n);
1679 return true;
1680 }
1681
1682 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1683 //-------------------
1684 //result=(result.isNaN())? funcAddr():result;
1685 // Check: If isNaN() by checking result!=result? then either trap
1686 // or go to runtime
1687 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1688 // Build the boolean node
1689 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1690
1691 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1692 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1693 // The pow or exp intrinsic returned a NaN, which requires a call
1694 // to the runtime. Recompile with the runtime call.
1695 uncommon_trap(Deoptimization::Reason_intrinsic,
1696 Deoptimization::Action_make_not_entrant);
1697 }
1698 set_result(result);
1699 } else {
1700 // If this inlining ever returned NaN in the past, we compile a call
1701 // to the runtime to properly handle corner cases
1702
1703 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1704 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1705 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1706
1707 if (!if_slow->is_top()) {
1708 RegionNode* result_region = new (C) RegionNode(3);
1709 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1710
1711 result_region->init_req(1, if_fast);
1712 result_val->init_req(1, result);
1713
1714 set_control(if_slow);
1715
1716 const TypePtr* no_memory_effects = NULL;
1717 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1718 no_memory_effects,
1719 x, top(), y, y ? top() : NULL);
1720 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1721 #ifdef ASSERT
1722 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1723 assert(value_top == top(), "second value must be top");
1724 #endif
1725
1726 result_region->init_req(2, control());
1727 result_val->init_req(2, value);
1728 set_result(result_region, result_val);
1729 } else {
1730 set_result(result);
1731 }
1732 }
1733 }
1734
1735 //------------------------------inline_exp-------------------------------------
1736 // Inline exp instructions, if possible. The Intel hardware only misses
1737 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1738 bool LibraryCallKit::inline_exp() {
1739 Node* arg = round_double_node(argument(0));
1740 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1741
1742 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1743
1744 C->set_has_split_ifs(true); // Has chance for split-if optimization
1745 return true;
1746 }
1747
1748 //------------------------------inline_pow-------------------------------------
1749 // Inline power instructions, if possible.
1750 bool LibraryCallKit::inline_pow() {
1751 // Pseudocode for pow
1752 // if (x <= 0.0) {
1753 // long longy = (long)y;
1754 // if ((double)longy == y) { // if y is long
1755 // if (y + 1 == y) longy = 0; // huge number: even
1756 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1757 // } else {
1758 // result = NaN;
1759 // }
1760 // } else {
1761 // result = DPow(x,y);
1762 // }
1763 // if (result != result)? {
1764 // result = uncommon_trap() or runtime_call();
1765 // }
1766 // return result;
1767
1768 Node* x = round_double_node(argument(0));
1769 Node* y = round_double_node(argument(2));
1770
1771 Node* result = NULL;
1772
1773 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1774 // Short form: skip the fancy tests and just check for NaN result.
1775 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1776 } else {
1777 // If this inlining ever returned NaN in the past, include all
1778 // checks + call to the runtime.
1779
1780 // Set the merge point for If node with condition of (x <= 0.0)
1781 // There are four possible paths to region node and phi node
1782 RegionNode *r = new (C) RegionNode(4);
1783 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1784
1785 // Build the first if node: if (x <= 0.0)
1786 // Node for 0 constant
1787 Node *zeronode = makecon(TypeD::ZERO);
1788 // Check x:0
1789 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1790 // Check: If (x<=0) then go complex path
1791 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1792 // Branch either way
1793 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1794 // Fast path taken; set region slot 3
1795 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1796 r->init_req(3,fast_taken); // Capture fast-control
1797
1798 // Fast path not-taken, i.e. slow path
1799 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1800
1801 // Set fast path result
1802 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1803 phi->init_req(3, fast_result);
1804
1805 // Complex path
1806 // Build the second if node (if y is long)
1807 // Node for (long)y
1808 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1809 // Node for (double)((long) y)
1810 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1811 // Check (double)((long) y) : y
1812 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1813 // Check if (y isn't long) then go to slow path
1814
1815 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1816 // Branch either way
1817 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1818 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1819
1820 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1821
1822 // Calculate DPow(abs(x), y)*(1 & (long)y)
1823 // Node for constant 1
1824 Node *conone = longcon(1);
1825 // 1& (long)y
1826 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1827
1828 // A huge number is always even. Detect a huge number by checking
1829 // if y + 1 == y and set integer to be tested for parity to 0.
1830 // Required for corner case:
1831 // (long)9.223372036854776E18 = max_jlong
1832 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1833 // max_jlong is odd but 9.223372036854776E18 is even
1834 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1835 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1836 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1837 Node* correctedsign = NULL;
1838 if (ConditionalMoveLimit != 0) {
1839 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1840 } else {
1841 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1842 RegionNode *r = new (C) RegionNode(3);
1843 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1844 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1845 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1846 phi->init_req(1, signnode);
1847 phi->init_req(2, longcon(0));
1848 correctedsign = _gvn.transform(phi);
1849 ylong_path = _gvn.transform(r);
1850 record_for_igvn(r);
1851 }
1852
1853 // zero node
1854 Node *conzero = longcon(0);
1855 // Check (1&(long)y)==0?
1856 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1857 // Check if (1&(long)y)!=0?, if so the result is negative
1858 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1859 // abs(x)
1860 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1861 // abs(x)^y
1862 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1863 // -abs(x)^y
1864 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1865 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1866 Node *signresult = NULL;
1867 if (ConditionalMoveLimit != 0) {
1868 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1869 } else {
1870 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1871 RegionNode *r = new (C) RegionNode(3);
1872 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1873 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1874 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1875 phi->init_req(1, absxpowy);
1876 phi->init_req(2, negabsxpowy);
1877 signresult = _gvn.transform(phi);
1878 ylong_path = _gvn.transform(r);
1879 record_for_igvn(r);
1880 }
1881 // Set complex path fast result
1882 r->init_req(2, ylong_path);
1883 phi->init_req(2, signresult);
1884
1885 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1886 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1887 r->init_req(1,slow_path);
1888 phi->init_req(1,slow_result);
1889
1890 // Post merge
1891 set_control(_gvn.transform(r));
1892 record_for_igvn(r);
1893 result = _gvn.transform(phi);
1894 }
1895
1896 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1897
1898 C->set_has_split_ifs(true); // Has chance for split-if optimization
1899 return true;
1900 }
1901
1902 //------------------------------runtime_math-----------------------------
1903 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1904 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1905 "must be (DD)D or (D)D type");
1906
1907 // Inputs
1908 Node* a = round_double_node(argument(0));
1909 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1910
1911 const TypePtr* no_memory_effects = NULL;
1912 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1913 no_memory_effects,
1914 a, top(), b, b ? top() : NULL);
1915 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1916 #ifdef ASSERT
1917 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1918 assert(value_top == top(), "second value must be top");
1919 #endif
1920
1921 set_result(value);
1922 return true;
1923 }
1924
1925 //------------------------------inline_math_native-----------------------------
1926 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1927 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1928 switch (id) {
1929 // These intrinsics are not properly supported on all hardware
1930 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1931 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1932 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1933 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1934 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1935 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1936
1937 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1938 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1939 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1940 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1941
1942 // These intrinsics are supported on all hardware
1943 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1944 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1945
1946 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1947 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1948 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1949 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1950 #undef FN_PTR
1951
1952 // These intrinsics are not yet correctly implemented
1953 case vmIntrinsics::_datan2:
1954 return false;
1955
1956 default:
1957 fatal_unexpected_iid(id);
1958 return false;
1959 }
1960 }
1961
1962 static bool is_simple_name(Node* n) {
1963 return (n->req() == 1 // constant
1964 || (n->is_Type() && n->as_Type()->type()->singleton())
1965 || n->is_Proj() // parameter or return value
1966 || n->is_Phi() // local of some sort
1967 );
1968 }
1969
1970 //----------------------------inline_min_max-----------------------------------
1971 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1972 set_result(generate_min_max(id, argument(0), argument(1)));
1973 return true;
1974 }
1975
1976 void LibraryCallKit::inline_math_mathExact(Node* math) {
1977 // If we didn't get the expected opcode it means we have optimized
1978 // the node to something else and don't need the exception edge.
1979 if (MathExactNode::is_MathExactOpcode(math->Opcode()) == false) {
1980 set_result(math);
1981 return;
1982 }
1983
1984 Node* result = _gvn.transform( new(C) ProjNode(math, MathExactNode::result_proj_node));
1985 Node* flags = _gvn.transform( new(C) FlagsProjNode(math, MathExactNode::flags_proj_node));
1986
1987 Node* bol = _gvn.transform( new (C) BoolNode(flags, BoolTest::overflow) );
1988 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1989 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1990 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1991
1992 {
1993 PreserveJVMState pjvms(this);
1994 PreserveReexecuteState preexecs(this);
1995 jvms()->set_should_reexecute(true);
1996
1997 set_control(slow_path);
1998 set_i_o(i_o());
1999
2000 uncommon_trap(Deoptimization::Reason_intrinsic,
2001 Deoptimization::Action_none);
2002 }
2003
2004 set_control(fast_path);
2005 set_result(result);
2006 }
2007
2008 bool LibraryCallKit::inline_math_addExact() {
2009 Node* arg1 = argument(0);
2010 Node* arg2 = argument(1);
2011
2012 Node* add = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2) );
2013 inline_math_mathExact(add);
2014 return true;
2015 }
2016
2017 bool LibraryCallKit::inline_math_addExactL() {
2018 Node* arg1 = argument(0); // type long
2019 // argument(1) == TOP
2020 Node* arg2 = argument(2); // type long
2021 // argument(4) == TOP
2022
2023 Node* add = _gvn.transform( new(C) AddExactLNode(NULL, arg1, arg2) );
2024 inline_math_mathExact(add);
2025 return true;
2026 }
2027
2028 bool LibraryCallKit::inline_math_subtractExactL() {
2029 Node* arg1 = argument(0); // type long
2030 // argument(1) == TOP
2031 Node* arg2 = argument(2); // type long
2032 // argument(4) == TOP
2033
2034 Node* sub = _gvn.transform( new(C) SubExactLNode(NULL, arg1, arg2) );
2035 inline_math_mathExact(sub);
2036 return true;
2037 }
2038
2039 bool LibraryCallKit::inline_math_subtractExact() {
2040 Node* arg1 = argument(0);
2041 Node* arg2 = argument(1);
2042
2043 Node* sub = _gvn.transform( new(C) SubExactINode(NULL, arg1, arg2) );
2044 inline_math_mathExact(sub);
2045 return true;
2046 }
2047
2048 bool LibraryCallKit::inline_math_negateExact() {
2049 Node* arg1 = argument(0);
2050
2051 Node* neg = _gvn.transform( new(C) NegExactINode(NULL, arg1));
2052 inline_math_mathExact(neg);
2053 return true;
2054 }
2055
2056 bool LibraryCallKit::inline_math_negateExactL() {
2057 Node* arg1 = argument(0);
2058 // argument(1) == TOP
2059
2060 Node* neg = _gvn.transform( new(C) NegExactLNode(NULL, arg1));
2061 inline_math_mathExact(neg);
2062 return true;
2063 }
2064
2065 bool LibraryCallKit::inline_math_multiplyExact() {
2066 Node* arg1 = argument(0);
2067 Node* arg2 = argument(1);
2068
2069 Node* mul = _gvn.transform( new(C) MulExactINode(NULL, arg1, arg2) );
2070 inline_math_mathExact(mul);
2071 return true;
2072 }
2073
2074 bool LibraryCallKit::inline_math_multiplyExactL() {
2075 Node* arg1 = argument(0);
2076 // argument(1) == TOP
2077 Node* arg2 = argument(2);
2078 // argument(3) == TOP
2079
2080 Node* mul = _gvn.transform( new(C) MulExactLNode(NULL, arg1, arg2) );
2081 inline_math_mathExact(mul);
2082 return true;
2083 }
2084
2085 bool LibraryCallKit::inline_math_incrementExact() {
2086 Node* arg1 = argument(0);
2087 Node* arg2 = intcon(1);
2088
2089 Node* inc = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2));
2090 inline_math_mathExact(inc);
2091 return true;
2092 }
2093
2094 bool LibraryCallKit::inline_math_incrementExactL() {
2095 Node* arg1 = argument(0);
2096 // argument(1) == TOP
2097 Node* arg2 = longcon(1);
2098
2099 Node* inc = _gvn.transform( new(C) AddExactLNode(NULL, arg1, arg2));
2100 inline_math_mathExact(inc);
2101 return true;
2102 }
2103
2104 bool LibraryCallKit::inline_math_decrementExact() {
2105 Node* arg1 = argument(0);
2106 Node* arg2 = intcon(1);
2107
2108 Node* dec = _gvn.transform( new(C) SubExactINode(NULL, arg1, arg2));
2109 inline_math_mathExact(dec);
2110 return true;
2111 }
2112
2113 bool LibraryCallKit::inline_math_decrementExactL() {
2114 Node* arg1 = argument(0);
2115 // argument(1) == TOP
2116 Node* arg2 = longcon(1);
2117
2118 Node* dec = _gvn.transform( new(C) SubExactLNode(NULL, arg1, arg2));
2119 inline_math_mathExact(dec);
2120 return true;
2121 }
2122
2123 Node*
2124 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2125 // These are the candidate return value:
2126 Node* xvalue = x0;
2127 Node* yvalue = y0;
2128
2129 if (xvalue == yvalue) {
2130 return xvalue;
2131 }
2132
2133 bool want_max = (id == vmIntrinsics::_max);
2134
2135 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2136 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2137 if (txvalue == NULL || tyvalue == NULL) return top();
2138 // This is not really necessary, but it is consistent with a
2139 // hypothetical MaxINode::Value method:
2140 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2141
2142 // %%% This folding logic should (ideally) be in a different place.
2143 // Some should be inside IfNode, and there to be a more reliable
2144 // transformation of ?: style patterns into cmoves. We also want
2145 // more powerful optimizations around cmove and min/max.
2146
2147 // Try to find a dominating comparison of these guys.
2148 // It can simplify the index computation for Arrays.copyOf
2149 // and similar uses of System.arraycopy.
2150 // First, compute the normalized version of CmpI(x, y).
2151 int cmp_op = Op_CmpI;
2152 Node* xkey = xvalue;
2153 Node* ykey = yvalue;
2154 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2155 if (ideal_cmpxy->is_Cmp()) {
2156 // E.g., if we have CmpI(length - offset, count),
2157 // it might idealize to CmpI(length, count + offset)
2158 cmp_op = ideal_cmpxy->Opcode();
2159 xkey = ideal_cmpxy->in(1);
2160 ykey = ideal_cmpxy->in(2);
2161 }
2162
2163 // Start by locating any relevant comparisons.
2164 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2165 Node* cmpxy = NULL;
2166 Node* cmpyx = NULL;
2167 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2168 Node* cmp = start_from->fast_out(k);
2169 if (cmp->outcnt() > 0 && // must have prior uses
2170 cmp->in(0) == NULL && // must be context-independent
2171 cmp->Opcode() == cmp_op) { // right kind of compare
2172 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2173 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2174 }
2175 }
2176
2177 const int NCMPS = 2;
2178 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2179 int cmpn;
2180 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2181 if (cmps[cmpn] != NULL) break; // find a result
2182 }
2183 if (cmpn < NCMPS) {
2184 // Look for a dominating test that tells us the min and max.
2185 int depth = 0; // Limit search depth for speed
2186 Node* dom = control();
2187 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2188 if (++depth >= 100) break;
2189 Node* ifproj = dom;
2190 if (!ifproj->is_Proj()) continue;
2191 Node* iff = ifproj->in(0);
2192 if (!iff->is_If()) continue;
2193 Node* bol = iff->in(1);
2194 if (!bol->is_Bool()) continue;
2195 Node* cmp = bol->in(1);
2196 if (cmp == NULL) continue;
2197 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2198 if (cmps[cmpn] == cmp) break;
2199 if (cmpn == NCMPS) continue;
2200 BoolTest::mask btest = bol->as_Bool()->_test._test;
2201 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2202 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2203 // At this point, we know that 'x btest y' is true.
2204 switch (btest) {
2205 case BoolTest::eq:
2206 // They are proven equal, so we can collapse the min/max.
2207 // Either value is the answer. Choose the simpler.
2208 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2209 return yvalue;
2210 return xvalue;
2211 case BoolTest::lt: // x < y
2212 case BoolTest::le: // x <= y
2213 return (want_max ? yvalue : xvalue);
2214 case BoolTest::gt: // x > y
2215 case BoolTest::ge: // x >= y
2216 return (want_max ? xvalue : yvalue);
2217 }
2218 }
2219 }
2220
2221 // We failed to find a dominating test.
2222 // Let's pick a test that might GVN with prior tests.
2223 Node* best_bol = NULL;
2224 BoolTest::mask best_btest = BoolTest::illegal;
2225 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2226 Node* cmp = cmps[cmpn];
2227 if (cmp == NULL) continue;
2228 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2229 Node* bol = cmp->fast_out(j);
2230 if (!bol->is_Bool()) continue;
2231 BoolTest::mask btest = bol->as_Bool()->_test._test;
2232 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2233 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2234 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2235 best_bol = bol->as_Bool();
2236 best_btest = btest;
2237 }
2238 }
2239 }
2240
2241 Node* answer_if_true = NULL;
2242 Node* answer_if_false = NULL;
2243 switch (best_btest) {
2244 default:
2245 if (cmpxy == NULL)
2246 cmpxy = ideal_cmpxy;
2247 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2248 // and fall through:
2249 case BoolTest::lt: // x < y
2250 case BoolTest::le: // x <= y
2251 answer_if_true = (want_max ? yvalue : xvalue);
2252 answer_if_false = (want_max ? xvalue : yvalue);
2253 break;
2254 case BoolTest::gt: // x > y
2255 case BoolTest::ge: // x >= y
2256 answer_if_true = (want_max ? xvalue : yvalue);
2257 answer_if_false = (want_max ? yvalue : xvalue);
2258 break;
2259 }
2260
2261 jint hi, lo;
2262 if (want_max) {
2263 // We can sharpen the minimum.
2264 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2265 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2266 } else {
2267 // We can sharpen the maximum.
2268 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2269 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2270 }
2271
2272 // Use a flow-free graph structure, to avoid creating excess control edges
2273 // which could hinder other optimizations.
2274 // Since Math.min/max is often used with arraycopy, we want
2275 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2276 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2277 answer_if_false, answer_if_true,
2278 TypeInt::make(lo, hi, widen));
2279
2280 return _gvn.transform(cmov);
2281
2282 /*
2283 // This is not as desirable as it may seem, since Min and Max
2284 // nodes do not have a full set of optimizations.
2285 // And they would interfere, anyway, with 'if' optimizations
2286 // and with CMoveI canonical forms.
2287 switch (id) {
2288 case vmIntrinsics::_min:
2289 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2290 case vmIntrinsics::_max:
2291 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2292 default:
2293 ShouldNotReachHere();
2294 }
2295 */
2296 }
2297
2298 inline int
2299 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2300 const TypePtr* base_type = TypePtr::NULL_PTR;
2301 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2302 if (base_type == NULL) {
2303 // Unknown type.
2304 return Type::AnyPtr;
2305 } else if (base_type == TypePtr::NULL_PTR) {
2306 // Since this is a NULL+long form, we have to switch to a rawptr.
2307 base = _gvn.transform(new (C) CastX2PNode(offset));
2308 offset = MakeConX(0);
2309 return Type::RawPtr;
2310 } else if (base_type->base() == Type::RawPtr) {
2311 return Type::RawPtr;
2312 } else if (base_type->isa_oopptr()) {
2313 // Base is never null => always a heap address.
2314 if (base_type->ptr() == TypePtr::NotNull) {
2315 return Type::OopPtr;
2316 }
2317 // Offset is small => always a heap address.
2318 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2319 if (offset_type != NULL &&
2320 base_type->offset() == 0 && // (should always be?)
2321 offset_type->_lo >= 0 &&
2322 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2323 return Type::OopPtr;
2324 }
2325 // Otherwise, it might either be oop+off or NULL+addr.
2326 return Type::AnyPtr;
2327 } else {
2328 // No information:
2329 return Type::AnyPtr;
2330 }
2331 }
2332
2333 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2334 int kind = classify_unsafe_addr(base, offset);
2335 if (kind == Type::RawPtr) {
2336 return basic_plus_adr(top(), base, offset);
2337 } else {
2338 return basic_plus_adr(base, offset);
2339 }
2340 }
2341
2342 //--------------------------inline_number_methods-----------------------------
2343 // inline int Integer.numberOfLeadingZeros(int)
2344 // inline int Long.numberOfLeadingZeros(long)
2345 //
2346 // inline int Integer.numberOfTrailingZeros(int)
2347 // inline int Long.numberOfTrailingZeros(long)
2348 //
2349 // inline int Integer.bitCount(int)
2350 // inline int Long.bitCount(long)
2351 //
2352 // inline char Character.reverseBytes(char)
2353 // inline short Short.reverseBytes(short)
2354 // inline int Integer.reverseBytes(int)
2355 // inline long Long.reverseBytes(long)
2356 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2357 Node* arg = argument(0);
2358 Node* n;
2359 switch (id) {
2360 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2361 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2362 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2363 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2364 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2365 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2366 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2367 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2368 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2369 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2370 default: fatal_unexpected_iid(id); break;
2371 }
2372 set_result(_gvn.transform(n));
2373 return true;
2374 }
2375
2376 //----------------------------inline_unsafe_access----------------------------
2377
2378 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2379
2380 // Helper that guards and inserts a pre-barrier.
2381 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2382 Node* pre_val, bool need_mem_bar) {
2383 // We could be accessing the referent field of a reference object. If so, when G1
2384 // is enabled, we need to log the value in the referent field in an SATB buffer.
2385 // This routine performs some compile time filters and generates suitable
2386 // runtime filters that guard the pre-barrier code.
2387 // Also add memory barrier for non volatile load from the referent field
2388 // to prevent commoning of loads across safepoint.
2389 if (!UseG1GC && !need_mem_bar)
2390 return;
2391
2392 // Some compile time checks.
2393
2394 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2395 const TypeX* otype = offset->find_intptr_t_type();
2396 if (otype != NULL && otype->is_con() &&
2397 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2398 // Constant offset but not the reference_offset so just return
2399 return;
2400 }
2401
2402 // We only need to generate the runtime guards for instances.
2403 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2404 if (btype != NULL) {
2405 if (btype->isa_aryptr()) {
2406 // Array type so nothing to do
2407 return;
2408 }
2409
2410 const TypeInstPtr* itype = btype->isa_instptr();
2411 if (itype != NULL) {
2412 // Can the klass of base_oop be statically determined to be
2413 // _not_ a sub-class of Reference and _not_ Object?
2414 ciKlass* klass = itype->klass();
2415 if ( klass->is_loaded() &&
2416 !klass->is_subtype_of(env()->Reference_klass()) &&
2417 !env()->Object_klass()->is_subtype_of(klass)) {
2418 return;
2419 }
2420 }
2421 }
2422
2423 // The compile time filters did not reject base_oop/offset so
2424 // we need to generate the following runtime filters
2425 //
2426 // if (offset == java_lang_ref_Reference::_reference_offset) {
2427 // if (instance_of(base, java.lang.ref.Reference)) {
2428 // pre_barrier(_, pre_val, ...);
2429 // }
2430 // }
2431
2432 float likely = PROB_LIKELY( 0.999);
2433 float unlikely = PROB_UNLIKELY(0.999);
2434
2435 IdealKit ideal(this);
2436 #define __ ideal.
2437
2438 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2439
2440 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2441 // Update graphKit memory and control from IdealKit.
2442 sync_kit(ideal);
2443
2444 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2445 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2446
2447 // Update IdealKit memory and control from graphKit.
2448 __ sync_kit(this);
2449
2450 Node* one = __ ConI(1);
2451 // is_instof == 0 if base_oop == NULL
2452 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2453
2454 // Update graphKit from IdeakKit.
2455 sync_kit(ideal);
2456
2457 // Use the pre-barrier to record the value in the referent field
2458 pre_barrier(false /* do_load */,
2459 __ ctrl(),
2460 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2461 pre_val /* pre_val */,
2462 T_OBJECT);
2463 if (need_mem_bar) {
2464 // Add memory barrier to prevent commoning reads from this field
2465 // across safepoint since GC can change its value.
2466 insert_mem_bar(Op_MemBarCPUOrder);
2467 }
2468 // Update IdealKit from graphKit.
2469 __ sync_kit(this);
2470
2471 } __ end_if(); // _ref_type != ref_none
2472 } __ end_if(); // offset == referent_offset
2473
2474 // Final sync IdealKit and GraphKit.
2475 final_sync(ideal);
2476 #undef __
2477 }
2478
2479
2480 // Interpret Unsafe.fieldOffset cookies correctly:
2481 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2482
2483 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2484 // Attempt to infer a sharper value type from the offset and base type.
2485 ciKlass* sharpened_klass = NULL;
2486
2487 // See if it is an instance field, with an object type.
2488 if (alias_type->field() != NULL) {
2489 assert(!is_native_ptr, "native pointer op cannot use a java address");
2490 if (alias_type->field()->type()->is_klass()) {
2491 sharpened_klass = alias_type->field()->type()->as_klass();
2492 }
2493 }
2494
2495 // See if it is a narrow oop array.
2496 if (adr_type->isa_aryptr()) {
2497 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2498 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2499 if (elem_type != NULL) {
2500 sharpened_klass = elem_type->klass();
2501 }
2502 }
2503 }
2504
2505 // The sharpened class might be unloaded if there is no class loader
2506 // contraint in place.
2507 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2508 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2509
2510 #ifndef PRODUCT
2511 if (C->print_intrinsics() || C->print_inlining()) {
2512 tty->print(" from base type: "); adr_type->dump();
2513 tty->print(" sharpened value: "); tjp->dump();
2514 }
2515 #endif
2516 // Sharpen the value type.
2517 return tjp;
2518 }
2519 return NULL;
2520 }
2521
2522 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2523 if (callee()->is_static()) return false; // caller must have the capability!
2524
2525 #ifndef PRODUCT
2526 {
2527 ResourceMark rm;
2528 // Check the signatures.
2529 ciSignature* sig = callee()->signature();
2530 #ifdef ASSERT
2531 if (!is_store) {
2532 // Object getObject(Object base, int/long offset), etc.
2533 BasicType rtype = sig->return_type()->basic_type();
2534 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2535 rtype = T_ADDRESS; // it is really a C void*
2536 assert(rtype == type, "getter must return the expected value");
2537 if (!is_native_ptr) {
2538 assert(sig->count() == 2, "oop getter has 2 arguments");
2539 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2540 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2541 } else {
2542 assert(sig->count() == 1, "native getter has 1 argument");
2543 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2544 }
2545 } else {
2546 // void putObject(Object base, int/long offset, Object x), etc.
2547 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2548 if (!is_native_ptr) {
2549 assert(sig->count() == 3, "oop putter has 3 arguments");
2550 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2551 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2552 } else {
2553 assert(sig->count() == 2, "native putter has 2 arguments");
2554 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2555 }
2556 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2557 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2558 vtype = T_ADDRESS; // it is really a C void*
2559 assert(vtype == type, "putter must accept the expected value");
2560 }
2561 #endif // ASSERT
2562 }
2563 #endif //PRODUCT
2564
2565 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2566
2567 Node* receiver = argument(0); // type: oop
2568
2569 // Build address expression. See the code in inline_unsafe_prefetch.
2570 Node* adr;
2571 Node* heap_base_oop = top();
2572 Node* offset = top();
2573 Node* val;
2574
2575 if (!is_native_ptr) {
2576 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2577 Node* base = argument(1); // type: oop
2578 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2579 offset = argument(2); // type: long
2580 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2581 // to be plain byte offsets, which are also the same as those accepted
2582 // by oopDesc::field_base.
2583 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2584 "fieldOffset must be byte-scaled");
2585 // 32-bit machines ignore the high half!
2586 offset = ConvL2X(offset);
2587 adr = make_unsafe_address(base, offset);
2588 heap_base_oop = base;
2589 val = is_store ? argument(4) : NULL;
2590 } else {
2591 Node* ptr = argument(1); // type: long
2592 ptr = ConvL2X(ptr); // adjust Java long to machine word
2593 adr = make_unsafe_address(NULL, ptr);
2594 val = is_store ? argument(3) : NULL;
2595 }
2596
2597 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2598
2599 // First guess at the value type.
2600 const Type *value_type = Type::get_const_basic_type(type);
2601
2602 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2603 // there was not enough information to nail it down.
2604 Compile::AliasType* alias_type = C->alias_type(adr_type);
2605 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2606
2607 // We will need memory barriers unless we can determine a unique
2608 // alias category for this reference. (Note: If for some reason
2609 // the barriers get omitted and the unsafe reference begins to "pollute"
2610 // the alias analysis of the rest of the graph, either Compile::can_alias
2611 // or Compile::must_alias will throw a diagnostic assert.)
2612 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2613
2614 // If we are reading the value of the referent field of a Reference
2615 // object (either by using Unsafe directly or through reflection)
2616 // then, if G1 is enabled, we need to record the referent in an
2617 // SATB log buffer using the pre-barrier mechanism.
2618 // Also we need to add memory barrier to prevent commoning reads
2619 // from this field across safepoint since GC can change its value.
2620 bool need_read_barrier = !is_native_ptr && !is_store &&
2621 offset != top() && heap_base_oop != top();
2622
2623 if (!is_store && type == T_OBJECT) {
2624 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2625 if (tjp != NULL) {
2626 value_type = tjp;
2627 }
2628 }
2629
2630 receiver = null_check(receiver);
2631 if (stopped()) {
2632 return true;
2633 }
2634 // Heap pointers get a null-check from the interpreter,
2635 // as a courtesy. However, this is not guaranteed by Unsafe,
2636 // and it is not possible to fully distinguish unintended nulls
2637 // from intended ones in this API.
2638
2639 if (is_volatile) {
2640 // We need to emit leading and trailing CPU membars (see below) in
2641 // addition to memory membars when is_volatile. This is a little
2642 // too strong, but avoids the need to insert per-alias-type
2643 // volatile membars (for stores; compare Parse::do_put_xxx), which
2644 // we cannot do effectively here because we probably only have a
2645 // rough approximation of type.
2646 need_mem_bar = true;
2647 // For Stores, place a memory ordering barrier now.
2648 if (is_store)
2649 insert_mem_bar(Op_MemBarRelease);
2650 }
2651
2652 // Memory barrier to prevent normal and 'unsafe' accesses from
2653 // bypassing each other. Happens after null checks, so the
2654 // exception paths do not take memory state from the memory barrier,
2655 // so there's no problems making a strong assert about mixing users
2656 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2657 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2658 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2659
2660 if (!is_store) {
2661 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2662 // load value
2663 switch (type) {
2664 case T_BOOLEAN:
2665 case T_CHAR:
2666 case T_BYTE:
2667 case T_SHORT:
2668 case T_INT:
2669 case T_LONG:
2670 case T_FLOAT:
2671 case T_DOUBLE:
2672 break;
2673 case T_OBJECT:
2674 if (need_read_barrier) {
2675 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2676 }
2677 break;
2678 case T_ADDRESS:
2679 // Cast to an int type.
2680 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2681 p = ConvX2L(p);
2682 break;
2683 default:
2684 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2685 break;
2686 }
2687 // The load node has the control of the preceding MemBarCPUOrder. All
2688 // following nodes will have the control of the MemBarCPUOrder inserted at
2689 // the end of this method. So, pushing the load onto the stack at a later
2690 // point is fine.
2691 set_result(p);
2692 } else {
2693 // place effect of store into memory
2694 switch (type) {
2695 case T_DOUBLE:
2696 val = dstore_rounding(val);
2697 break;
2698 case T_ADDRESS:
2699 // Repackage the long as a pointer.
2700 val = ConvL2X(val);
2701 val = _gvn.transform(new (C) CastX2PNode(val));
2702 break;
2703 }
2704
2705 if (type != T_OBJECT ) {
2706 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2707 } else {
2708 // Possibly an oop being stored to Java heap or native memory
2709 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2710 // oop to Java heap.
2711 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2712 } else {
2713 // We can't tell at compile time if we are storing in the Java heap or outside
2714 // of it. So we need to emit code to conditionally do the proper type of
2715 // store.
2716
2717 IdealKit ideal(this);
2718 #define __ ideal.
2719 // QQQ who knows what probability is here??
2720 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2721 // Sync IdealKit and graphKit.
2722 sync_kit(ideal);
2723 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2724 // Update IdealKit memory.
2725 __ sync_kit(this);
2726 } __ else_(); {
2727 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2728 } __ end_if();
2729 // Final sync IdealKit and GraphKit.
2730 final_sync(ideal);
2731 #undef __
2732 }
2733 }
2734 }
2735
2736 if (is_volatile) {
2737 if (!is_store)
2738 insert_mem_bar(Op_MemBarAcquire);
2739 else
2740 insert_mem_bar(Op_MemBarVolatile);
2741 }
2742
2743 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2744
2745 return true;
2746 }
2747
2748 //----------------------------inline_unsafe_prefetch----------------------------
2749
2750 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2751 #ifndef PRODUCT
2752 {
2753 ResourceMark rm;
2754 // Check the signatures.
2755 ciSignature* sig = callee()->signature();
2756 #ifdef ASSERT
2757 // Object getObject(Object base, int/long offset), etc.
2758 BasicType rtype = sig->return_type()->basic_type();
2759 if (!is_native_ptr) {
2760 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2761 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2762 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2763 } else {
2764 assert(sig->count() == 1, "native prefetch has 1 argument");
2765 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2766 }
2767 #endif // ASSERT
2768 }
2769 #endif // !PRODUCT
2770
2771 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2772
2773 const int idx = is_static ? 0 : 1;
2774 if (!is_static) {
2775 null_check_receiver();
2776 if (stopped()) {
2777 return true;
2778 }
2779 }
2780
2781 // Build address expression. See the code in inline_unsafe_access.
2782 Node *adr;
2783 if (!is_native_ptr) {
2784 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2785 Node* base = argument(idx + 0); // type: oop
2786 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2787 Node* offset = argument(idx + 1); // type: long
2788 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2789 // to be plain byte offsets, which are also the same as those accepted
2790 // by oopDesc::field_base.
2791 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2792 "fieldOffset must be byte-scaled");
2793 // 32-bit machines ignore the high half!
2794 offset = ConvL2X(offset);
2795 adr = make_unsafe_address(base, offset);
2796 } else {
2797 Node* ptr = argument(idx + 0); // type: long
2798 ptr = ConvL2X(ptr); // adjust Java long to machine word
2799 adr = make_unsafe_address(NULL, ptr);
2800 }
2801
2802 // Generate the read or write prefetch
2803 Node *prefetch;
2804 if (is_store) {
2805 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2806 } else {
2807 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2808 }
2809 prefetch->init_req(0, control());
2810 set_i_o(_gvn.transform(prefetch));
2811
2812 return true;
2813 }
2814
2815 //----------------------------inline_unsafe_load_store----------------------------
2816 // This method serves a couple of different customers (depending on LoadStoreKind):
2817 //
2818 // LS_cmpxchg:
2819 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2820 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2821 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2822 //
2823 // LS_xadd:
2824 // public int getAndAddInt( Object o, long offset, int delta)
2825 // public long getAndAddLong(Object o, long offset, long delta)
2826 //
2827 // LS_xchg:
2828 // int getAndSet(Object o, long offset, int newValue)
2829 // long getAndSet(Object o, long offset, long newValue)
2830 // Object getAndSet(Object o, long offset, Object newValue)
2831 //
2832 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2833 // This basic scheme here is the same as inline_unsafe_access, but
2834 // differs in enough details that combining them would make the code
2835 // overly confusing. (This is a true fact! I originally combined
2836 // them, but even I was confused by it!) As much code/comments as
2837 // possible are retained from inline_unsafe_access though to make
2838 // the correspondences clearer. - dl
2839
2840 if (callee()->is_static()) return false; // caller must have the capability!
2841
2842 #ifndef PRODUCT
2843 BasicType rtype;
2844 {
2845 ResourceMark rm;
2846 // Check the signatures.
2847 ciSignature* sig = callee()->signature();
2848 rtype = sig->return_type()->basic_type();
2849 if (kind == LS_xadd || kind == LS_xchg) {
2850 // Check the signatures.
2851 #ifdef ASSERT
2852 assert(rtype == type, "get and set must return the expected type");
2853 assert(sig->count() == 3, "get and set has 3 arguments");
2854 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2855 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2856 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2857 #endif // ASSERT
2858 } else if (kind == LS_cmpxchg) {
2859 // Check the signatures.
2860 #ifdef ASSERT
2861 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2862 assert(sig->count() == 4, "CAS has 4 arguments");
2863 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2864 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2865 #endif // ASSERT
2866 } else {
2867 ShouldNotReachHere();
2868 }
2869 }
2870 #endif //PRODUCT
2871
2872 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2873
2874 // Get arguments:
2875 Node* receiver = NULL;
2876 Node* base = NULL;
2877 Node* offset = NULL;
2878 Node* oldval = NULL;
2879 Node* newval = NULL;
2880 if (kind == LS_cmpxchg) {
2881 const bool two_slot_type = type2size[type] == 2;
2882 receiver = argument(0); // type: oop
2883 base = argument(1); // type: oop
2884 offset = argument(2); // type: long
2885 oldval = argument(4); // type: oop, int, or long
2886 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2887 } else if (kind == LS_xadd || kind == LS_xchg){
2888 receiver = argument(0); // type: oop
2889 base = argument(1); // type: oop
2890 offset = argument(2); // type: long
2891 oldval = NULL;
2892 newval = argument(4); // type: oop, int, or long
2893 }
2894
2895 // Null check receiver.
2896 receiver = null_check(receiver);
2897 if (stopped()) {
2898 return true;
2899 }
2900
2901 // Build field offset expression.
2902 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2903 // to be plain byte offsets, which are also the same as those accepted
2904 // by oopDesc::field_base.
2905 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2906 // 32-bit machines ignore the high half of long offsets
2907 offset = ConvL2X(offset);
2908 Node* adr = make_unsafe_address(base, offset);
2909 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2910
2911 // For CAS, unlike inline_unsafe_access, there seems no point in
2912 // trying to refine types. Just use the coarse types here.
2913 const Type *value_type = Type::get_const_basic_type(type);
2914 Compile::AliasType* alias_type = C->alias_type(adr_type);
2915 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2916
2917 if (kind == LS_xchg && type == T_OBJECT) {
2918 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2919 if (tjp != NULL) {
2920 value_type = tjp;
2921 }
2922 }
2923
2924 int alias_idx = C->get_alias_index(adr_type);
2925
2926 // Memory-model-wise, a LoadStore acts like a little synchronized
2927 // block, so needs barriers on each side. These don't translate
2928 // into actual barriers on most machines, but we still need rest of
2929 // compiler to respect ordering.
2930
2931 insert_mem_bar(Op_MemBarRelease);
2932 insert_mem_bar(Op_MemBarCPUOrder);
2933
2934 // 4984716: MemBars must be inserted before this
2935 // memory node in order to avoid a false
2936 // dependency which will confuse the scheduler.
2937 Node *mem = memory(alias_idx);
2938
2939 // For now, we handle only those cases that actually exist: ints,
2940 // longs, and Object. Adding others should be straightforward.
2941 Node* load_store;
2942 switch(type) {
2943 case T_INT:
2944 if (kind == LS_xadd) {
2945 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2946 } else if (kind == LS_xchg) {
2947 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2948 } else if (kind == LS_cmpxchg) {
2949 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2950 } else {
2951 ShouldNotReachHere();
2952 }
2953 break;
2954 case T_LONG:
2955 if (kind == LS_xadd) {
2956 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2957 } else if (kind == LS_xchg) {
2958 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2959 } else if (kind == LS_cmpxchg) {
2960 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2961 } else {
2962 ShouldNotReachHere();
2963 }
2964 break;
2965 case T_OBJECT:
2966 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2967 // could be delayed during Parse (for example, in adjust_map_after_if()).
2968 // Execute transformation here to avoid barrier generation in such case.
2969 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2970 newval = _gvn.makecon(TypePtr::NULL_PTR);
2971
2972 // Reference stores need a store barrier.
2973 if (kind == LS_xchg) {
2974 // If pre-barrier must execute before the oop store, old value will require do_load here.
2975 if (!can_move_pre_barrier()) {
2976 pre_barrier(true /* do_load*/,
2977 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2978 NULL /* pre_val*/,
2979 T_OBJECT);
2980 } // Else move pre_barrier to use load_store value, see below.
2981 } else if (kind == LS_cmpxchg) {
2982 // Same as for newval above:
2983 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2984 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2985 }
2986 // The only known value which might get overwritten is oldval.
2987 pre_barrier(false /* do_load */,
2988 control(), NULL, NULL, max_juint, NULL, NULL,
2989 oldval /* pre_val */,
2990 T_OBJECT);
2991 } else {
2992 ShouldNotReachHere();
2993 }
2994
2995 #ifdef _LP64
2996 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2997 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2998 if (kind == LS_xchg) {
2999 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3000 newval_enc, adr_type, value_type->make_narrowoop()));
3001 } else {
3002 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3003 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3004 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3005 newval_enc, oldval_enc));
3006 }
3007 } else
3008 #endif
3009 {
3010 if (kind == LS_xchg) {
3011 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3012 } else {
3013 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3014 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3015 }
3016 }
3017 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3018 break;
3019 default:
3020 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3021 break;
3022 }
3023
3024 // SCMemProjNodes represent the memory state of a LoadStore. Their
3025 // main role is to prevent LoadStore nodes from being optimized away
3026 // when their results aren't used.
3027 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3028 set_memory(proj, alias_idx);
3029
3030 if (type == T_OBJECT && kind == LS_xchg) {
3031 #ifdef _LP64
3032 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3033 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3034 }
3035 #endif
3036 if (can_move_pre_barrier()) {
3037 // Don't need to load pre_val. The old value is returned by load_store.
3038 // The pre_barrier can execute after the xchg as long as no safepoint
3039 // gets inserted between them.
3040 pre_barrier(false /* do_load */,
3041 control(), NULL, NULL, max_juint, NULL, NULL,
3042 load_store /* pre_val */,
3043 T_OBJECT);
3044 }
3045 }
3046
3047 // Add the trailing membar surrounding the access
3048 insert_mem_bar(Op_MemBarCPUOrder);
3049 insert_mem_bar(Op_MemBarAcquire);
3050
3051 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3052 set_result(load_store);
3053 return true;
3054 }
3055
3056 //----------------------------inline_unsafe_ordered_store----------------------
3057 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3058 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3059 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3060 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3061 // This is another variant of inline_unsafe_access, differing in
3062 // that it always issues store-store ("release") barrier and ensures
3063 // store-atomicity (which only matters for "long").
3064
3065 if (callee()->is_static()) return false; // caller must have the capability!
3066
3067 #ifndef PRODUCT
3068 {
3069 ResourceMark rm;
3070 // Check the signatures.
3071 ciSignature* sig = callee()->signature();
3072 #ifdef ASSERT
3073 BasicType rtype = sig->return_type()->basic_type();
3074 assert(rtype == T_VOID, "must return void");
3075 assert(sig->count() == 3, "has 3 arguments");
3076 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3077 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3078 #endif // ASSERT
3079 }
3080 #endif //PRODUCT
3081
3082 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3083
3084 // Get arguments:
3085 Node* receiver = argument(0); // type: oop
3086 Node* base = argument(1); // type: oop
3087 Node* offset = argument(2); // type: long
3088 Node* val = argument(4); // type: oop, int, or long
3089
3090 // Null check receiver.
3091 receiver = null_check(receiver);
3092 if (stopped()) {
3093 return true;
3094 }
3095
3096 // Build field offset expression.
3097 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3098 // 32-bit machines ignore the high half of long offsets
3099 offset = ConvL2X(offset);
3100 Node* adr = make_unsafe_address(base, offset);
3101 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3102 const Type *value_type = Type::get_const_basic_type(type);
3103 Compile::AliasType* alias_type = C->alias_type(adr_type);
3104
3105 insert_mem_bar(Op_MemBarRelease);
3106 insert_mem_bar(Op_MemBarCPUOrder);
3107 // Ensure that the store is atomic for longs:
3108 const bool require_atomic_access = true;
3109 Node* store;
3110 if (type == T_OBJECT) // reference stores need a store barrier.
3111 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
3112 else {
3113 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
3114 }
3115 insert_mem_bar(Op_MemBarCPUOrder);
3116 return true;
3117 }
3118
3119 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3120 // Regardless of form, don't allow previous ld/st to move down,
3121 // then issue acquire, release, or volatile mem_bar.
3122 insert_mem_bar(Op_MemBarCPUOrder);
3123 switch(id) {
3124 case vmIntrinsics::_loadFence:
3125 insert_mem_bar(Op_MemBarAcquire);
3126 return true;
3127 case vmIntrinsics::_storeFence:
3128 insert_mem_bar(Op_MemBarRelease);
3129 return true;
3130 case vmIntrinsics::_fullFence:
3131 insert_mem_bar(Op_MemBarVolatile);
3132 return true;
3133 default:
3134 fatal_unexpected_iid(id);
3135 return false;
3136 }
3137 }
3138
3139 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3140 if (!kls->is_Con()) {
3141 return true;
3142 }
3143 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3144 if (klsptr == NULL) {
3145 return true;
3146 }
3147 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3148 // don't need a guard for a klass that is already initialized
3149 return !ik->is_initialized();
3150 }
3151
3152 //----------------------------inline_unsafe_allocate---------------------------
3153 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3154 bool LibraryCallKit::inline_unsafe_allocate() {
3155 if (callee()->is_static()) return false; // caller must have the capability!
3156
3157 null_check_receiver(); // null-check, then ignore
3158 Node* cls = null_check(argument(1));
3159 if (stopped()) return true;
3160
3161 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3162 kls = null_check(kls);
3163 if (stopped()) return true; // argument was like int.class
3164
3165 Node* test = NULL;
3166 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3167 // Note: The argument might still be an illegal value like
3168 // Serializable.class or Object[].class. The runtime will handle it.
3169 // But we must make an explicit check for initialization.
3170 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3171 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3172 // can generate code to load it as unsigned byte.
3173 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3174 Node* bits = intcon(InstanceKlass::fully_initialized);
3175 test = _gvn.transform(new (C) SubINode(inst, bits));
3176 // The 'test' is non-zero if we need to take a slow path.
3177 }
3178
3179 Node* obj = new_instance(kls, test);
3180 set_result(obj);
3181 return true;
3182 }
3183
3184 #ifdef TRACE_HAVE_INTRINSICS
3185 /*
3186 * oop -> myklass
3187 * myklass->trace_id |= USED
3188 * return myklass->trace_id & ~0x3
3189 */
3190 bool LibraryCallKit::inline_native_classID() {
3191 null_check_receiver(); // null-check, then ignore
3192 Node* cls = null_check(argument(1), T_OBJECT);
3193 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3194 kls = null_check(kls, T_OBJECT);
3195 ByteSize offset = TRACE_ID_OFFSET;
3196 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3197 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3198 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3199 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3200 Node* clsused = longcon(0x01l); // set the class bit
3201 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3202
3203 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3204 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3205 set_result(andl);
3206 return true;
3207 }
3208
3209 bool LibraryCallKit::inline_native_threadID() {
3210 Node* tls_ptr = NULL;
3211 Node* cur_thr = generate_current_thread(tls_ptr);
3212 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3213 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3214 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3215
3216 Node* threadid = NULL;
3217 size_t thread_id_size = OSThread::thread_id_size();
3218 if (thread_id_size == (size_t) BytesPerLong) {
3219 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3220 } else if (thread_id_size == (size_t) BytesPerInt) {
3221 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3222 } else {
3223 ShouldNotReachHere();
3224 }
3225 set_result(threadid);
3226 return true;
3227 }
3228 #endif
3229
3230 //------------------------inline_native_time_funcs--------------
3231 // inline code for System.currentTimeMillis() and System.nanoTime()
3232 // these have the same type and signature
3233 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3234 const TypeFunc* tf = OptoRuntime::void_long_Type();
3235 const TypePtr* no_memory_effects = NULL;
3236 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3237 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3238 #ifdef ASSERT
3239 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3240 assert(value_top == top(), "second value must be top");
3241 #endif
3242 set_result(value);
3243 return true;
3244 }
3245
3246 //------------------------inline_native_currentThread------------------
3247 bool LibraryCallKit::inline_native_currentThread() {
3248 Node* junk = NULL;
3249 set_result(generate_current_thread(junk));
3250 return true;
3251 }
3252
3253 //------------------------inline_native_isInterrupted------------------
3254 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3255 bool LibraryCallKit::inline_native_isInterrupted() {
3256 // Add a fast path to t.isInterrupted(clear_int):
3257 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3258 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3259 // So, in the common case that the interrupt bit is false,
3260 // we avoid making a call into the VM. Even if the interrupt bit
3261 // is true, if the clear_int argument is false, we avoid the VM call.
3262 // However, if the receiver is not currentThread, we must call the VM,
3263 // because there must be some locking done around the operation.
3264
3265 // We only go to the fast case code if we pass two guards.
3266 // Paths which do not pass are accumulated in the slow_region.
3267
3268 enum {
3269 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3270 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3271 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3272 PATH_LIMIT
3273 };
3274
3275 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3276 // out of the function.
3277 insert_mem_bar(Op_MemBarCPUOrder);
3278
3279 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3280 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3281
3282 RegionNode* slow_region = new (C) RegionNode(1);
3283 record_for_igvn(slow_region);
3284
3285 // (a) Receiving thread must be the current thread.
3286 Node* rec_thr = argument(0);
3287 Node* tls_ptr = NULL;
3288 Node* cur_thr = generate_current_thread(tls_ptr);
3289 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3290 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3291
3292 generate_slow_guard(bol_thr, slow_region);
3293
3294 // (b) Interrupt bit on TLS must be false.
3295 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3296 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3297 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3298
3299 // Set the control input on the field _interrupted read to prevent it floating up.
3300 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3301 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3302 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3303
3304 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3305
3306 // First fast path: if (!TLS._interrupted) return false;
3307 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3308 result_rgn->init_req(no_int_result_path, false_bit);
3309 result_val->init_req(no_int_result_path, intcon(0));
3310
3311 // drop through to next case
3312 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3313
3314 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3315 Node* clr_arg = argument(1);
3316 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3317 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3318 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3319
3320 // Second fast path: ... else if (!clear_int) return true;
3321 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3322 result_rgn->init_req(no_clear_result_path, false_arg);
3323 result_val->init_req(no_clear_result_path, intcon(1));
3324
3325 // drop through to next case
3326 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3327
3328 // (d) Otherwise, go to the slow path.
3329 slow_region->add_req(control());
3330 set_control( _gvn.transform(slow_region));
3331
3332 if (stopped()) {
3333 // There is no slow path.
3334 result_rgn->init_req(slow_result_path, top());
3335 result_val->init_req(slow_result_path, top());
3336 } else {
3337 // non-virtual because it is a private non-static
3338 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3339
3340 Node* slow_val = set_results_for_java_call(slow_call);
3341 // this->control() comes from set_results_for_java_call
3342
3343 Node* fast_io = slow_call->in(TypeFunc::I_O);
3344 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3345
3346 // These two phis are pre-filled with copies of of the fast IO and Memory
3347 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3348 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3349
3350 result_rgn->init_req(slow_result_path, control());
3351 result_io ->init_req(slow_result_path, i_o());
3352 result_mem->init_req(slow_result_path, reset_memory());
3353 result_val->init_req(slow_result_path, slow_val);
3354
3355 set_all_memory(_gvn.transform(result_mem));
3356 set_i_o( _gvn.transform(result_io));
3357 }
3358
3359 C->set_has_split_ifs(true); // Has chance for split-if optimization
3360 set_result(result_rgn, result_val);
3361 return true;
3362 }
3363
3364 //---------------------------load_mirror_from_klass----------------------------
3365 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3366 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3367 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3368 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3369 }
3370
3371 //-----------------------load_klass_from_mirror_common-------------------------
3372 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3373 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3374 // and branch to the given path on the region.
3375 // If never_see_null, take an uncommon trap on null, so we can optimistically
3376 // compile for the non-null case.
3377 // If the region is NULL, force never_see_null = true.
3378 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3379 bool never_see_null,
3380 RegionNode* region,
3381 int null_path,
3382 int offset) {
3383 if (region == NULL) never_see_null = true;
3384 Node* p = basic_plus_adr(mirror, offset);
3385 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3386 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3387 Node* null_ctl = top();
3388 kls = null_check_oop(kls, &null_ctl, never_see_null);
3389 if (region != NULL) {
3390 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3391 region->init_req(null_path, null_ctl);
3392 } else {
3393 assert(null_ctl == top(), "no loose ends");
3394 }
3395 return kls;
3396 }
3397
3398 //--------------------(inline_native_Class_query helpers)---------------------
3399 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3400 // Fall through if (mods & mask) == bits, take the guard otherwise.
3401 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3402 // Branch around if the given klass has the given modifier bit set.
3403 // Like generate_guard, adds a new path onto the region.
3404 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3405 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3406 Node* mask = intcon(modifier_mask);
3407 Node* bits = intcon(modifier_bits);
3408 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3409 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3410 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3411 return generate_fair_guard(bol, region);
3412 }
3413 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3414 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3415 }
3416
3417 //-------------------------inline_native_Class_query-------------------
3418 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3419 const Type* return_type = TypeInt::BOOL;
3420 Node* prim_return_value = top(); // what happens if it's a primitive class?
3421 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3422 bool expect_prim = false; // most of these guys expect to work on refs
3423
3424 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3425
3426 Node* mirror = argument(0);
3427 Node* obj = top();
3428
3429 switch (id) {
3430 case vmIntrinsics::_isInstance:
3431 // nothing is an instance of a primitive type
3432 prim_return_value = intcon(0);
3433 obj = argument(1);
3434 break;
3435 case vmIntrinsics::_getModifiers:
3436 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3437 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3438 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3439 break;
3440 case vmIntrinsics::_isInterface:
3441 prim_return_value = intcon(0);
3442 break;
3443 case vmIntrinsics::_isArray:
3444 prim_return_value = intcon(0);
3445 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3446 break;
3447 case vmIntrinsics::_isPrimitive:
3448 prim_return_value = intcon(1);
3449 expect_prim = true; // obviously
3450 break;
3451 case vmIntrinsics::_getSuperclass:
3452 prim_return_value = null();
3453 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3454 break;
3455 case vmIntrinsics::_getComponentType:
3456 prim_return_value = null();
3457 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3458 break;
3459 case vmIntrinsics::_getClassAccessFlags:
3460 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3461 return_type = TypeInt::INT; // not bool! 6297094
3462 break;
3463 default:
3464 fatal_unexpected_iid(id);
3465 break;
3466 }
3467
3468 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3469 if (mirror_con == NULL) return false; // cannot happen?
3470
3471 #ifndef PRODUCT
3472 if (C->print_intrinsics() || C->print_inlining()) {
3473 ciType* k = mirror_con->java_mirror_type();
3474 if (k) {
3475 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3476 k->print_name();
3477 tty->cr();
3478 }
3479 }
3480 #endif
3481
3482 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3483 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3484 record_for_igvn(region);
3485 PhiNode* phi = new (C) PhiNode(region, return_type);
3486
3487 // The mirror will never be null of Reflection.getClassAccessFlags, however
3488 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3489 // if it is. See bug 4774291.
3490
3491 // For Reflection.getClassAccessFlags(), the null check occurs in
3492 // the wrong place; see inline_unsafe_access(), above, for a similar
3493 // situation.
3494 mirror = null_check(mirror);
3495 // If mirror or obj is dead, only null-path is taken.
3496 if (stopped()) return true;
3497
3498 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3499
3500 // Now load the mirror's klass metaobject, and null-check it.
3501 // Side-effects region with the control path if the klass is null.
3502 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3503 // If kls is null, we have a primitive mirror.
3504 phi->init_req(_prim_path, prim_return_value);
3505 if (stopped()) { set_result(region, phi); return true; }
3506
3507 Node* p; // handy temp
3508 Node* null_ctl;
3509
3510 // Now that we have the non-null klass, we can perform the real query.
3511 // For constant classes, the query will constant-fold in LoadNode::Value.
3512 Node* query_value = top();
3513 switch (id) {
3514 case vmIntrinsics::_isInstance:
3515 // nothing is an instance of a primitive type
3516 query_value = gen_instanceof(obj, kls);
3517 break;
3518
3519 case vmIntrinsics::_getModifiers:
3520 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3521 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3522 break;
3523
3524 case vmIntrinsics::_isInterface:
3525 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3526 if (generate_interface_guard(kls, region) != NULL)
3527 // A guard was added. If the guard is taken, it was an interface.
3528 phi->add_req(intcon(1));
3529 // If we fall through, it's a plain class.
3530 query_value = intcon(0);
3531 break;
3532
3533 case vmIntrinsics::_isArray:
3534 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3535 if (generate_array_guard(kls, region) != NULL)
3536 // A guard was added. If the guard is taken, it was an array.
3537 phi->add_req(intcon(1));
3538 // If we fall through, it's a plain class.
3539 query_value = intcon(0);
3540 break;
3541
3542 case vmIntrinsics::_isPrimitive:
3543 query_value = intcon(0); // "normal" path produces false
3544 break;
3545
3546 case vmIntrinsics::_getSuperclass:
3547 // The rules here are somewhat unfortunate, but we can still do better
3548 // with random logic than with a JNI call.
3549 // Interfaces store null or Object as _super, but must report null.
3550 // Arrays store an intermediate super as _super, but must report Object.
3551 // Other types can report the actual _super.
3552 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3553 if (generate_interface_guard(kls, region) != NULL)
3554 // A guard was added. If the guard is taken, it was an interface.
3555 phi->add_req(null());
3556 if (generate_array_guard(kls, region) != NULL)
3557 // A guard was added. If the guard is taken, it was an array.
3558 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3559 // If we fall through, it's a plain class. Get its _super.
3560 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3561 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3562 null_ctl = top();
3563 kls = null_check_oop(kls, &null_ctl);
3564 if (null_ctl != top()) {
3565 // If the guard is taken, Object.superClass is null (both klass and mirror).
3566 region->add_req(null_ctl);
3567 phi ->add_req(null());
3568 }
3569 if (!stopped()) {
3570 query_value = load_mirror_from_klass(kls);
3571 }
3572 break;
3573
3574 case vmIntrinsics::_getComponentType:
3575 if (generate_array_guard(kls, region) != NULL) {
3576 // Be sure to pin the oop load to the guard edge just created:
3577 Node* is_array_ctrl = region->in(region->req()-1);
3578 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3579 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3580 phi->add_req(cmo);
3581 }
3582 query_value = null(); // non-array case is null
3583 break;
3584
3585 case vmIntrinsics::_getClassAccessFlags:
3586 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3587 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3588 break;
3589
3590 default:
3591 fatal_unexpected_iid(id);
3592 break;
3593 }
3594
3595 // Fall-through is the normal case of a query to a real class.
3596 phi->init_req(1, query_value);
3597 region->init_req(1, control());
3598
3599 C->set_has_split_ifs(true); // Has chance for split-if optimization
3600 set_result(region, phi);
3601 return true;
3602 }
3603
3604 //--------------------------inline_native_subtype_check------------------------
3605 // This intrinsic takes the JNI calls out of the heart of
3606 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3607 bool LibraryCallKit::inline_native_subtype_check() {
3608 // Pull both arguments off the stack.
3609 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3610 args[0] = argument(0);
3611 args[1] = argument(1);
3612 Node* klasses[2]; // corresponding Klasses: superk, subk
3613 klasses[0] = klasses[1] = top();
3614
3615 enum {
3616 // A full decision tree on {superc is prim, subc is prim}:
3617 _prim_0_path = 1, // {P,N} => false
3618 // {P,P} & superc!=subc => false
3619 _prim_same_path, // {P,P} & superc==subc => true
3620 _prim_1_path, // {N,P} => false
3621 _ref_subtype_path, // {N,N} & subtype check wins => true
3622 _both_ref_path, // {N,N} & subtype check loses => false
3623 PATH_LIMIT
3624 };
3625
3626 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3627 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3628 record_for_igvn(region);
3629
3630 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3631 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3632 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3633
3634 // First null-check both mirrors and load each mirror's klass metaobject.
3635 int which_arg;
3636 for (which_arg = 0; which_arg <= 1; which_arg++) {
3637 Node* arg = args[which_arg];
3638 arg = null_check(arg);
3639 if (stopped()) break;
3640 args[which_arg] = arg;
3641
3642 Node* p = basic_plus_adr(arg, class_klass_offset);
3643 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3644 klasses[which_arg] = _gvn.transform(kls);
3645 }
3646
3647 // Having loaded both klasses, test each for null.
3648 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3649 for (which_arg = 0; which_arg <= 1; which_arg++) {
3650 Node* kls = klasses[which_arg];
3651 Node* null_ctl = top();
3652 kls = null_check_oop(kls, &null_ctl, never_see_null);
3653 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3654 region->init_req(prim_path, null_ctl);
3655 if (stopped()) break;
3656 klasses[which_arg] = kls;
3657 }
3658
3659 if (!stopped()) {
3660 // now we have two reference types, in klasses[0..1]
3661 Node* subk = klasses[1]; // the argument to isAssignableFrom
3662 Node* superk = klasses[0]; // the receiver
3663 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3664 // now we have a successful reference subtype check
3665 region->set_req(_ref_subtype_path, control());
3666 }
3667
3668 // If both operands are primitive (both klasses null), then
3669 // we must return true when they are identical primitives.
3670 // It is convenient to test this after the first null klass check.
3671 set_control(region->in(_prim_0_path)); // go back to first null check
3672 if (!stopped()) {
3673 // Since superc is primitive, make a guard for the superc==subc case.
3674 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3675 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3676 generate_guard(bol_eq, region, PROB_FAIR);
3677 if (region->req() == PATH_LIMIT+1) {
3678 // A guard was added. If the added guard is taken, superc==subc.
3679 region->swap_edges(PATH_LIMIT, _prim_same_path);
3680 region->del_req(PATH_LIMIT);
3681 }
3682 region->set_req(_prim_0_path, control()); // Not equal after all.
3683 }
3684
3685 // these are the only paths that produce 'true':
3686 phi->set_req(_prim_same_path, intcon(1));
3687 phi->set_req(_ref_subtype_path, intcon(1));
3688
3689 // pull together the cases:
3690 assert(region->req() == PATH_LIMIT, "sane region");
3691 for (uint i = 1; i < region->req(); i++) {
3692 Node* ctl = region->in(i);
3693 if (ctl == NULL || ctl == top()) {
3694 region->set_req(i, top());
3695 phi ->set_req(i, top());
3696 } else if (phi->in(i) == NULL) {
3697 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3698 }
3699 }
3700
3701 set_control(_gvn.transform(region));
3702 set_result(_gvn.transform(phi));
3703 return true;
3704 }
3705
3706 //---------------------generate_array_guard_common------------------------
3707 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3708 bool obj_array, bool not_array) {
3709 // If obj_array/non_array==false/false:
3710 // Branch around if the given klass is in fact an array (either obj or prim).
3711 // If obj_array/non_array==false/true:
3712 // Branch around if the given klass is not an array klass of any kind.
3713 // If obj_array/non_array==true/true:
3714 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3715 // If obj_array/non_array==true/false:
3716 // Branch around if the kls is an oop array (Object[] or subtype)
3717 //
3718 // Like generate_guard, adds a new path onto the region.
3719 jint layout_con = 0;
3720 Node* layout_val = get_layout_helper(kls, layout_con);
3721 if (layout_val == NULL) {
3722 bool query = (obj_array
3723 ? Klass::layout_helper_is_objArray(layout_con)
3724 : Klass::layout_helper_is_array(layout_con));
3725 if (query == not_array) {
3726 return NULL; // never a branch
3727 } else { // always a branch
3728 Node* always_branch = control();
3729 if (region != NULL)
3730 region->add_req(always_branch);
3731 set_control(top());
3732 return always_branch;
3733 }
3734 }
3735 // Now test the correct condition.
3736 jint nval = (obj_array
3737 ? ((jint)Klass::_lh_array_tag_type_value
3738 << Klass::_lh_array_tag_shift)
3739 : Klass::_lh_neutral_value);
3740 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3741 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3742 // invert the test if we are looking for a non-array
3743 if (not_array) btest = BoolTest(btest).negate();
3744 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3745 return generate_fair_guard(bol, region);
3746 }
3747
3748
3749 //-----------------------inline_native_newArray--------------------------
3750 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3751 bool LibraryCallKit::inline_native_newArray() {
3752 Node* mirror = argument(0);
3753 Node* count_val = argument(1);
3754
3755 mirror = null_check(mirror);
3756 // If mirror or obj is dead, only null-path is taken.
3757 if (stopped()) return true;
3758
3759 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3760 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3761 PhiNode* result_val = new(C) PhiNode(result_reg,
3762 TypeInstPtr::NOTNULL);
3763 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3764 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3765 TypePtr::BOTTOM);
3766
3767 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3768 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3769 result_reg, _slow_path);
3770 Node* normal_ctl = control();
3771 Node* no_array_ctl = result_reg->in(_slow_path);
3772
3773 // Generate code for the slow case. We make a call to newArray().
3774 set_control(no_array_ctl);
3775 if (!stopped()) {
3776 // Either the input type is void.class, or else the
3777 // array klass has not yet been cached. Either the
3778 // ensuing call will throw an exception, or else it
3779 // will cache the array klass for next time.
3780 PreserveJVMState pjvms(this);
3781 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3782 Node* slow_result = set_results_for_java_call(slow_call);
3783 // this->control() comes from set_results_for_java_call
3784 result_reg->set_req(_slow_path, control());
3785 result_val->set_req(_slow_path, slow_result);
3786 result_io ->set_req(_slow_path, i_o());
3787 result_mem->set_req(_slow_path, reset_memory());
3788 }
3789
3790 set_control(normal_ctl);
3791 if (!stopped()) {
3792 // Normal case: The array type has been cached in the java.lang.Class.
3793 // The following call works fine even if the array type is polymorphic.
3794 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3795 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3796 result_reg->init_req(_normal_path, control());
3797 result_val->init_req(_normal_path, obj);
3798 result_io ->init_req(_normal_path, i_o());
3799 result_mem->init_req(_normal_path, reset_memory());
3800 }
3801
3802 // Return the combined state.
3803 set_i_o( _gvn.transform(result_io) );
3804 set_all_memory( _gvn.transform(result_mem));
3805
3806 C->set_has_split_ifs(true); // Has chance for split-if optimization
3807 set_result(result_reg, result_val);
3808 return true;
3809 }
3810
3811 //----------------------inline_native_getLength--------------------------
3812 // public static native int java.lang.reflect.Array.getLength(Object array);
3813 bool LibraryCallKit::inline_native_getLength() {
3814 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3815
3816 Node* array = null_check(argument(0));
3817 // If array is dead, only null-path is taken.
3818 if (stopped()) return true;
3819
3820 // Deoptimize if it is a non-array.
3821 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3822
3823 if (non_array != NULL) {
3824 PreserveJVMState pjvms(this);
3825 set_control(non_array);
3826 uncommon_trap(Deoptimization::Reason_intrinsic,
3827 Deoptimization::Action_maybe_recompile);
3828 }
3829
3830 // If control is dead, only non-array-path is taken.
3831 if (stopped()) return true;
3832
3833 // The works fine even if the array type is polymorphic.
3834 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3835 Node* result = load_array_length(array);
3836
3837 C->set_has_split_ifs(true); // Has chance for split-if optimization
3838 set_result(result);
3839 return true;
3840 }
3841
3842 //------------------------inline_array_copyOf----------------------------
3843 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3844 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3845 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3846 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3847
3848 // Get the arguments.
3849 Node* original = argument(0);
3850 Node* start = is_copyOfRange? argument(1): intcon(0);
3851 Node* end = is_copyOfRange? argument(2): argument(1);
3852 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3853
3854 Node* newcopy;
3855
3856 // Set the original stack and the reexecute bit for the interpreter to reexecute
3857 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3858 { PreserveReexecuteState preexecs(this);
3859 jvms()->set_should_reexecute(true);
3860
3861 array_type_mirror = null_check(array_type_mirror);
3862 original = null_check(original);
3863
3864 // Check if a null path was taken unconditionally.
3865 if (stopped()) return true;
3866
3867 Node* orig_length = load_array_length(original);
3868
3869 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3870 klass_node = null_check(klass_node);
3871
3872 RegionNode* bailout = new (C) RegionNode(1);
3873 record_for_igvn(bailout);
3874
3875 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3876 // Bail out if that is so.
3877 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3878 if (not_objArray != NULL) {
3879 // Improve the klass node's type from the new optimistic assumption:
3880 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3881 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3882 Node* cast = new (C) CastPPNode(klass_node, akls);
3883 cast->init_req(0, control());
3884 klass_node = _gvn.transform(cast);
3885 }
3886
3887 // Bail out if either start or end is negative.
3888 generate_negative_guard(start, bailout, &start);
3889 generate_negative_guard(end, bailout, &end);
3890
3891 Node* length = end;
3892 if (_gvn.type(start) != TypeInt::ZERO) {
3893 length = _gvn.transform(new (C) SubINode(end, start));
3894 }
3895
3896 // Bail out if length is negative.
3897 // Without this the new_array would throw
3898 // NegativeArraySizeException but IllegalArgumentException is what
3899 // should be thrown
3900 generate_negative_guard(length, bailout, &length);
3901
3902 if (bailout->req() > 1) {
3903 PreserveJVMState pjvms(this);
3904 set_control(_gvn.transform(bailout));
3905 uncommon_trap(Deoptimization::Reason_intrinsic,
3906 Deoptimization::Action_maybe_recompile);
3907 }
3908
3909 if (!stopped()) {
3910 // How many elements will we copy from the original?
3911 // The answer is MinI(orig_length - start, length).
3912 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3913 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3914
3915 newcopy = new_array(klass_node, length, 0); // no argments to push
3916
3917 // Generate a direct call to the right arraycopy function(s).
3918 // We know the copy is disjoint but we might not know if the
3919 // oop stores need checking.
3920 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3921 // This will fail a store-check if x contains any non-nulls.
3922 bool disjoint_bases = true;
3923 // if start > orig_length then the length of the copy may be
3924 // negative.
3925 bool length_never_negative = !is_copyOfRange;
3926 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3927 original, start, newcopy, intcon(0), moved,
3928 disjoint_bases, length_never_negative);
3929 }
3930 } // original reexecute is set back here
3931
3932 C->set_has_split_ifs(true); // Has chance for split-if optimization
3933 if (!stopped()) {
3934 set_result(newcopy);
3935 }
3936 return true;
3937 }
3938
3939
3940 //----------------------generate_virtual_guard---------------------------
3941 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3942 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3943 RegionNode* slow_region) {
3944 ciMethod* method = callee();
3945 int vtable_index = method->vtable_index();
3946 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3947 err_msg_res("bad index %d", vtable_index));
3948 // Get the Method* out of the appropriate vtable entry.
3949 int entry_offset = (InstanceKlass::vtable_start_offset() +
3950 vtable_index*vtableEntry::size()) * wordSize +
3951 vtableEntry::method_offset_in_bytes();
3952 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3953 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3954
3955 // Compare the target method with the expected method (e.g., Object.hashCode).
3956 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3957
3958 Node* native_call = makecon(native_call_addr);
3959 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3960 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3961
3962 return generate_slow_guard(test_native, slow_region);
3963 }
3964
3965 //-----------------------generate_method_call----------------------------
3966 // Use generate_method_call to make a slow-call to the real
3967 // method if the fast path fails. An alternative would be to
3968 // use a stub like OptoRuntime::slow_arraycopy_Java.
3969 // This only works for expanding the current library call,
3970 // not another intrinsic. (E.g., don't use this for making an
3971 // arraycopy call inside of the copyOf intrinsic.)
3972 CallJavaNode*
3973 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3974 // When compiling the intrinsic method itself, do not use this technique.
3975 guarantee(callee() != C->method(), "cannot make slow-call to self");
3976
3977 ciMethod* method = callee();
3978 // ensure the JVMS we have will be correct for this call
3979 guarantee(method_id == method->intrinsic_id(), "must match");
3980
3981 const TypeFunc* tf = TypeFunc::make(method);
3982 CallJavaNode* slow_call;
3983 if (is_static) {
3984 assert(!is_virtual, "");
3985 slow_call = new(C) CallStaticJavaNode(C, tf,
3986 SharedRuntime::get_resolve_static_call_stub(),
3987 method, bci());
3988 } else if (is_virtual) {
3989 null_check_receiver();
3990 int vtable_index = Method::invalid_vtable_index;
3991 if (UseInlineCaches) {
3992 // Suppress the vtable call
3993 } else {
3994 // hashCode and clone are not a miranda methods,
3995 // so the vtable index is fixed.
3996 // No need to use the linkResolver to get it.
3997 vtable_index = method->vtable_index();
3998 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3999 err_msg_res("bad index %d", vtable_index));
4000 }
4001 slow_call = new(C) CallDynamicJavaNode(tf,
4002 SharedRuntime::get_resolve_virtual_call_stub(),
4003 method, vtable_index, bci());
4004 } else { // neither virtual nor static: opt_virtual
4005 null_check_receiver();
4006 slow_call = new(C) CallStaticJavaNode(C, tf,
4007 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4008 method, bci());
4009 slow_call->set_optimized_virtual(true);
4010 }
4011 set_arguments_for_java_call(slow_call);
4012 set_edges_for_java_call(slow_call);
4013 return slow_call;
4014 }
4015
4016
4017 //------------------------------inline_native_hashcode--------------------
4018 // Build special case code for calls to hashCode on an object.
4019 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4020 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4021 assert(!(is_virtual && is_static), "either virtual, special, or static");
4022
4023 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4024
4025 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4026 PhiNode* result_val = new(C) PhiNode(result_reg,
4027 TypeInt::INT);
4028 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4029 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4030 TypePtr::BOTTOM);
4031 Node* obj = NULL;
4032 if (!is_static) {
4033 // Check for hashing null object
4034 obj = null_check_receiver();
4035 if (stopped()) return true; // unconditionally null
4036 result_reg->init_req(_null_path, top());
4037 result_val->init_req(_null_path, top());
4038 } else {
4039 // Do a null check, and return zero if null.
4040 // System.identityHashCode(null) == 0
4041 obj = argument(0);
4042 Node* null_ctl = top();
4043 obj = null_check_oop(obj, &null_ctl);
4044 result_reg->init_req(_null_path, null_ctl);
4045 result_val->init_req(_null_path, _gvn.intcon(0));
4046 }
4047
4048 // Unconditionally null? Then return right away.
4049 if (stopped()) {
4050 set_control( result_reg->in(_null_path));
4051 if (!stopped())
4052 set_result(result_val->in(_null_path));
4053 return true;
4054 }
4055
4056 // After null check, get the object's klass.
4057 Node* obj_klass = load_object_klass(obj);
4058
4059 // This call may be virtual (invokevirtual) or bound (invokespecial).
4060 // For each case we generate slightly different code.
4061
4062 // We only go to the fast case code if we pass a number of guards. The
4063 // paths which do not pass are accumulated in the slow_region.
4064 RegionNode* slow_region = new (C) RegionNode(1);
4065 record_for_igvn(slow_region);
4066
4067 // If this is a virtual call, we generate a funny guard. We pull out
4068 // the vtable entry corresponding to hashCode() from the target object.
4069 // If the target method which we are calling happens to be the native
4070 // Object hashCode() method, we pass the guard. We do not need this
4071 // guard for non-virtual calls -- the caller is known to be the native
4072 // Object hashCode().
4073 if (is_virtual) {
4074 generate_virtual_guard(obj_klass, slow_region);
4075 }
4076
4077 // Get the header out of the object, use LoadMarkNode when available
4078 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4079 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
4080
4081 // Test the header to see if it is unlocked.
4082 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4083 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4084 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4085 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4086 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4087
4088 generate_slow_guard(test_unlocked, slow_region);
4089
4090 // Get the hash value and check to see that it has been properly assigned.
4091 // We depend on hash_mask being at most 32 bits and avoid the use of
4092 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4093 // vm: see markOop.hpp.
4094 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4095 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4096 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4097 // This hack lets the hash bits live anywhere in the mark object now, as long
4098 // as the shift drops the relevant bits into the low 32 bits. Note that
4099 // Java spec says that HashCode is an int so there's no point in capturing
4100 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4101 hshifted_header = ConvX2I(hshifted_header);
4102 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4103
4104 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4105 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4106 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4107
4108 generate_slow_guard(test_assigned, slow_region);
4109
4110 Node* init_mem = reset_memory();
4111 // fill in the rest of the null path:
4112 result_io ->init_req(_null_path, i_o());
4113 result_mem->init_req(_null_path, init_mem);
4114
4115 result_val->init_req(_fast_path, hash_val);
4116 result_reg->init_req(_fast_path, control());
4117 result_io ->init_req(_fast_path, i_o());
4118 result_mem->init_req(_fast_path, init_mem);
4119
4120 // Generate code for the slow case. We make a call to hashCode().
4121 set_control(_gvn.transform(slow_region));
4122 if (!stopped()) {
4123 // No need for PreserveJVMState, because we're using up the present state.
4124 set_all_memory(init_mem);
4125 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4126 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4127 Node* slow_result = set_results_for_java_call(slow_call);
4128 // this->control() comes from set_results_for_java_call
4129 result_reg->init_req(_slow_path, control());
4130 result_val->init_req(_slow_path, slow_result);
4131 result_io ->set_req(_slow_path, i_o());
4132 result_mem ->set_req(_slow_path, reset_memory());
4133 }
4134
4135 // Return the combined state.
4136 set_i_o( _gvn.transform(result_io) );
4137 set_all_memory( _gvn.transform(result_mem));
4138
4139 set_result(result_reg, result_val);
4140 return true;
4141 }
4142
4143 //---------------------------inline_native_getClass----------------------------
4144 // public final native Class<?> java.lang.Object.getClass();
4145 //
4146 // Build special case code for calls to getClass on an object.
4147 bool LibraryCallKit::inline_native_getClass() {
4148 Node* obj = null_check_receiver();
4149 if (stopped()) return true;
4150 set_result(load_mirror_from_klass(load_object_klass(obj)));
4151 return true;
4152 }
4153
4154 //-----------------inline_native_Reflection_getCallerClass---------------------
4155 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4156 //
4157 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4158 //
4159 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4160 // in that it must skip particular security frames and checks for
4161 // caller sensitive methods.
4162 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4163 #ifndef PRODUCT
4164 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4165 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4166 }
4167 #endif
4168
4169 if (!jvms()->has_method()) {
4170 #ifndef PRODUCT
4171 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4172 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4173 }
4174 #endif
4175 return false;
4176 }
4177
4178 // Walk back up the JVM state to find the caller at the required
4179 // depth.
4180 JVMState* caller_jvms = jvms();
4181
4182 // Cf. JVM_GetCallerClass
4183 // NOTE: Start the loop at depth 1 because the current JVM state does
4184 // not include the Reflection.getCallerClass() frame.
4185 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4186 ciMethod* m = caller_jvms->method();
4187 switch (n) {
4188 case 0:
4189 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4190 break;
4191 case 1:
4192 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4193 if (!m->caller_sensitive()) {
4194 #ifndef PRODUCT
4195 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4196 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4197 }
4198 #endif
4199 return false; // bail-out; let JVM_GetCallerClass do the work
4200 }
4201 break;
4202 default:
4203 if (!m->is_ignored_by_security_stack_walk()) {
4204 // We have reached the desired frame; return the holder class.
4205 // Acquire method holder as java.lang.Class and push as constant.
4206 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4207 ciInstance* caller_mirror = caller_klass->java_mirror();
4208 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4209
4210 #ifndef PRODUCT
4211 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4212 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());
4213 tty->print_cr(" JVM state at this point:");
4214 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4215 ciMethod* m = jvms()->of_depth(i)->method();
4216 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4217 }
4218 }
4219 #endif
4220 return true;
4221 }
4222 break;
4223 }
4224 }
4225
4226 #ifndef PRODUCT
4227 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4228 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4229 tty->print_cr(" JVM state at this point:");
4230 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4231 ciMethod* m = jvms()->of_depth(i)->method();
4232 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4233 }
4234 }
4235 #endif
4236
4237 return false; // bail-out; let JVM_GetCallerClass do the work
4238 }
4239
4240 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4241 Node* arg = argument(0);
4242 Node* result;
4243
4244 switch (id) {
4245 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4246 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4247 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4248 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4249
4250 case vmIntrinsics::_doubleToLongBits: {
4251 // two paths (plus control) merge in a wood
4252 RegionNode *r = new (C) RegionNode(3);
4253 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4254
4255 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4256 // Build the boolean node
4257 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4258
4259 // Branch either way.
4260 // NaN case is less traveled, which makes all the difference.
4261 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4262 Node *opt_isnan = _gvn.transform(ifisnan);
4263 assert( opt_isnan->is_If(), "Expect an IfNode");
4264 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4265 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4266
4267 set_control(iftrue);
4268
4269 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4270 Node *slow_result = longcon(nan_bits); // return NaN
4271 phi->init_req(1, _gvn.transform( slow_result ));
4272 r->init_req(1, iftrue);
4273
4274 // Else fall through
4275 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4276 set_control(iffalse);
4277
4278 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4279 r->init_req(2, iffalse);
4280
4281 // Post merge
4282 set_control(_gvn.transform(r));
4283 record_for_igvn(r);
4284
4285 C->set_has_split_ifs(true); // Has chance for split-if optimization
4286 result = phi;
4287 assert(result->bottom_type()->isa_long(), "must be");
4288 break;
4289 }
4290
4291 case vmIntrinsics::_floatToIntBits: {
4292 // two paths (plus control) merge in a wood
4293 RegionNode *r = new (C) RegionNode(3);
4294 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4295
4296 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4297 // Build the boolean node
4298 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4299
4300 // Branch either way.
4301 // NaN case is less traveled, which makes all the difference.
4302 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4303 Node *opt_isnan = _gvn.transform(ifisnan);
4304 assert( opt_isnan->is_If(), "Expect an IfNode");
4305 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4306 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4307
4308 set_control(iftrue);
4309
4310 static const jint nan_bits = 0x7fc00000;
4311 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4312 phi->init_req(1, _gvn.transform( slow_result ));
4313 r->init_req(1, iftrue);
4314
4315 // Else fall through
4316 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4317 set_control(iffalse);
4318
4319 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4320 r->init_req(2, iffalse);
4321
4322 // Post merge
4323 set_control(_gvn.transform(r));
4324 record_for_igvn(r);
4325
4326 C->set_has_split_ifs(true); // Has chance for split-if optimization
4327 result = phi;
4328 assert(result->bottom_type()->isa_int(), "must be");
4329 break;
4330 }
4331
4332 default:
4333 fatal_unexpected_iid(id);
4334 break;
4335 }
4336 set_result(_gvn.transform(result));
4337 return true;
4338 }
4339
4340 #ifdef _LP64
4341 #define XTOP ,top() /*additional argument*/
4342 #else //_LP64
4343 #define XTOP /*no additional argument*/
4344 #endif //_LP64
4345
4346 //----------------------inline_unsafe_copyMemory-------------------------
4347 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4348 bool LibraryCallKit::inline_unsafe_copyMemory() {
4349 if (callee()->is_static()) return false; // caller must have the capability!
4350 null_check_receiver(); // null-check receiver
4351 if (stopped()) return true;
4352
4353 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4354
4355 Node* src_ptr = argument(1); // type: oop
4356 Node* src_off = ConvL2X(argument(2)); // type: long
4357 Node* dst_ptr = argument(4); // type: oop
4358 Node* dst_off = ConvL2X(argument(5)); // type: long
4359 Node* size = ConvL2X(argument(7)); // type: long
4360
4361 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4362 "fieldOffset must be byte-scaled");
4363
4364 Node* src = make_unsafe_address(src_ptr, src_off);
4365 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4366
4367 // Conservatively insert a memory barrier on all memory slices.
4368 // Do not let writes of the copy source or destination float below the copy.
4369 insert_mem_bar(Op_MemBarCPUOrder);
4370
4371 // Call it. Note that the length argument is not scaled.
4372 make_runtime_call(RC_LEAF|RC_NO_FP,
4373 OptoRuntime::fast_arraycopy_Type(),
4374 StubRoutines::unsafe_arraycopy(),
4375 "unsafe_arraycopy",
4376 TypeRawPtr::BOTTOM,
4377 src, dst, size XTOP);
4378
4379 // Do not let reads of the copy destination float above the copy.
4380 insert_mem_bar(Op_MemBarCPUOrder);
4381
4382 return true;
4383 }
4384
4385 //------------------------clone_coping-----------------------------------
4386 // Helper function for inline_native_clone.
4387 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4388 assert(obj_size != NULL, "");
4389 Node* raw_obj = alloc_obj->in(1);
4390 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4391
4392 AllocateNode* alloc = NULL;
4393 if (ReduceBulkZeroing) {
4394 // We will be completely responsible for initializing this object -
4395 // mark Initialize node as complete.
4396 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4397 // The object was just allocated - there should be no any stores!
4398 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4399 // Mark as complete_with_arraycopy so that on AllocateNode
4400 // expansion, we know this AllocateNode is initialized by an array
4401 // copy and a StoreStore barrier exists after the array copy.
4402 alloc->initialization()->set_complete_with_arraycopy();
4403 }
4404
4405 // Copy the fastest available way.
4406 // TODO: generate fields copies for small objects instead.
4407 Node* src = obj;
4408 Node* dest = alloc_obj;
4409 Node* size = _gvn.transform(obj_size);
4410
4411 // Exclude the header but include array length to copy by 8 bytes words.
4412 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4413 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4414 instanceOopDesc::base_offset_in_bytes();
4415 // base_off:
4416 // 8 - 32-bit VM
4417 // 12 - 64-bit VM, compressed klass
4418 // 16 - 64-bit VM, normal klass
4419 if (base_off % BytesPerLong != 0) {
4420 assert(UseCompressedClassPointers, "");
4421 if (is_array) {
4422 // Exclude length to copy by 8 bytes words.
4423 base_off += sizeof(int);
4424 } else {
4425 // Include klass to copy by 8 bytes words.
4426 base_off = instanceOopDesc::klass_offset_in_bytes();
4427 }
4428 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4429 }
4430 src = basic_plus_adr(src, base_off);
4431 dest = basic_plus_adr(dest, base_off);
4432
4433 // Compute the length also, if needed:
4434 Node* countx = size;
4435 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4436 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4437
4438 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4439 bool disjoint_bases = true;
4440 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4441 src, NULL, dest, NULL, countx,
4442 /*dest_uninitialized*/true);
4443
4444 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4445 if (card_mark) {
4446 assert(!is_array, "");
4447 // Put in store barrier for any and all oops we are sticking
4448 // into this object. (We could avoid this if we could prove
4449 // that the object type contains no oop fields at all.)
4450 Node* no_particular_value = NULL;
4451 Node* no_particular_field = NULL;
4452 int raw_adr_idx = Compile::AliasIdxRaw;
4453 post_barrier(control(),
4454 memory(raw_adr_type),
4455 alloc_obj,
4456 no_particular_field,
4457 raw_adr_idx,
4458 no_particular_value,
4459 T_OBJECT,
4460 false);
4461 }
4462
4463 // Do not let reads from the cloned object float above the arraycopy.
4464 if (alloc != NULL) {
4465 // Do not let stores that initialize this object be reordered with
4466 // a subsequent store that would make this object accessible by
4467 // other threads.
4468 // Record what AllocateNode this StoreStore protects so that
4469 // escape analysis can go from the MemBarStoreStoreNode to the
4470 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4471 // based on the escape status of the AllocateNode.
4472 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4473 } else {
4474 insert_mem_bar(Op_MemBarCPUOrder);
4475 }
4476 }
4477
4478 //------------------------inline_native_clone----------------------------
4479 // protected native Object java.lang.Object.clone();
4480 //
4481 // Here are the simple edge cases:
4482 // null receiver => normal trap
4483 // virtual and clone was overridden => slow path to out-of-line clone
4484 // not cloneable or finalizer => slow path to out-of-line Object.clone
4485 //
4486 // The general case has two steps, allocation and copying.
4487 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4488 //
4489 // Copying also has two cases, oop arrays and everything else.
4490 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4491 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4492 //
4493 // These steps fold up nicely if and when the cloned object's klass
4494 // can be sharply typed as an object array, a type array, or an instance.
4495 //
4496 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4497 PhiNode* result_val;
4498
4499 // Set the reexecute bit for the interpreter to reexecute
4500 // the bytecode that invokes Object.clone if deoptimization happens.
4501 { PreserveReexecuteState preexecs(this);
4502 jvms()->set_should_reexecute(true);
4503
4504 Node* obj = null_check_receiver();
4505 if (stopped()) return true;
4506
4507 Node* obj_klass = load_object_klass(obj);
4508 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4509 const TypeOopPtr* toop = ((tklass != NULL)
4510 ? tklass->as_instance_type()
4511 : TypeInstPtr::NOTNULL);
4512
4513 // Conservatively insert a memory barrier on all memory slices.
4514 // Do not let writes into the original float below the clone.
4515 insert_mem_bar(Op_MemBarCPUOrder);
4516
4517 // paths into result_reg:
4518 enum {
4519 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4520 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4521 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4522 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4523 PATH_LIMIT
4524 };
4525 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4526 result_val = new(C) PhiNode(result_reg,
4527 TypeInstPtr::NOTNULL);
4528 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4529 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4530 TypePtr::BOTTOM);
4531 record_for_igvn(result_reg);
4532
4533 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4534 int raw_adr_idx = Compile::AliasIdxRaw;
4535
4536 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4537 if (array_ctl != NULL) {
4538 // It's an array.
4539 PreserveJVMState pjvms(this);
4540 set_control(array_ctl);
4541 Node* obj_length = load_array_length(obj);
4542 Node* obj_size = NULL;
4543 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4544
4545 if (!use_ReduceInitialCardMarks()) {
4546 // If it is an oop array, it requires very special treatment,
4547 // because card marking is required on each card of the array.
4548 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4549 if (is_obja != NULL) {
4550 PreserveJVMState pjvms2(this);
4551 set_control(is_obja);
4552 // Generate a direct call to the right arraycopy function(s).
4553 bool disjoint_bases = true;
4554 bool length_never_negative = true;
4555 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4556 obj, intcon(0), alloc_obj, intcon(0),
4557 obj_length,
4558 disjoint_bases, length_never_negative);
4559 result_reg->init_req(_objArray_path, control());
4560 result_val->init_req(_objArray_path, alloc_obj);
4561 result_i_o ->set_req(_objArray_path, i_o());
4562 result_mem ->set_req(_objArray_path, reset_memory());
4563 }
4564 }
4565 // Otherwise, there are no card marks to worry about.
4566 // (We can dispense with card marks if we know the allocation
4567 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4568 // causes the non-eden paths to take compensating steps to
4569 // simulate a fresh allocation, so that no further
4570 // card marks are required in compiled code to initialize
4571 // the object.)
4572
4573 if (!stopped()) {
4574 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4575
4576 // Present the results of the copy.
4577 result_reg->init_req(_array_path, control());
4578 result_val->init_req(_array_path, alloc_obj);
4579 result_i_o ->set_req(_array_path, i_o());
4580 result_mem ->set_req(_array_path, reset_memory());
4581 }
4582 }
4583
4584 // We only go to the instance fast case code if we pass a number of guards.
4585 // The paths which do not pass are accumulated in the slow_region.
4586 RegionNode* slow_region = new (C) RegionNode(1);
4587 record_for_igvn(slow_region);
4588 if (!stopped()) {
4589 // It's an instance (we did array above). Make the slow-path tests.
4590 // If this is a virtual call, we generate a funny guard. We grab
4591 // the vtable entry corresponding to clone() from the target object.
4592 // If the target method which we are calling happens to be the
4593 // Object clone() method, we pass the guard. We do not need this
4594 // guard for non-virtual calls; the caller is known to be the native
4595 // Object clone().
4596 if (is_virtual) {
4597 generate_virtual_guard(obj_klass, slow_region);
4598 }
4599
4600 // The object must be cloneable and must not have a finalizer.
4601 // Both of these conditions may be checked in a single test.
4602 // We could optimize the cloneable test further, but we don't care.
4603 generate_access_flags_guard(obj_klass,
4604 // Test both conditions:
4605 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4606 // Must be cloneable but not finalizer:
4607 JVM_ACC_IS_CLONEABLE,
4608 slow_region);
4609 }
4610
4611 if (!stopped()) {
4612 // It's an instance, and it passed the slow-path tests.
4613 PreserveJVMState pjvms(this);
4614 Node* obj_size = NULL;
4615 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4616
4617 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4618
4619 // Present the results of the slow call.
4620 result_reg->init_req(_instance_path, control());
4621 result_val->init_req(_instance_path, alloc_obj);
4622 result_i_o ->set_req(_instance_path, i_o());
4623 result_mem ->set_req(_instance_path, reset_memory());
4624 }
4625
4626 // Generate code for the slow case. We make a call to clone().
4627 set_control(_gvn.transform(slow_region));
4628 if (!stopped()) {
4629 PreserveJVMState pjvms(this);
4630 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4631 Node* slow_result = set_results_for_java_call(slow_call);
4632 // this->control() comes from set_results_for_java_call
4633 result_reg->init_req(_slow_path, control());
4634 result_val->init_req(_slow_path, slow_result);
4635 result_i_o ->set_req(_slow_path, i_o());
4636 result_mem ->set_req(_slow_path, reset_memory());
4637 }
4638
4639 // Return the combined state.
4640 set_control( _gvn.transform(result_reg));
4641 set_i_o( _gvn.transform(result_i_o));
4642 set_all_memory( _gvn.transform(result_mem));
4643 } // original reexecute is set back here
4644
4645 set_result(_gvn.transform(result_val));
4646 return true;
4647 }
4648
4649 //------------------------------basictype2arraycopy----------------------------
4650 address LibraryCallKit::basictype2arraycopy(BasicType t,
4651 Node* src_offset,
4652 Node* dest_offset,
4653 bool disjoint_bases,
4654 const char* &name,
4655 bool dest_uninitialized) {
4656 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4657 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4658
4659 bool aligned = false;
4660 bool disjoint = disjoint_bases;
4661
4662 // if the offsets are the same, we can treat the memory regions as
4663 // disjoint, because either the memory regions are in different arrays,
4664 // or they are identical (which we can treat as disjoint.) We can also
4665 // treat a copy with a destination index less that the source index
4666 // as disjoint since a low->high copy will work correctly in this case.
4667 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4668 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4669 // both indices are constants
4670 int s_offs = src_offset_inttype->get_con();
4671 int d_offs = dest_offset_inttype->get_con();
4672 int element_size = type2aelembytes(t);
4673 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4674 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4675 if (s_offs >= d_offs) disjoint = true;
4676 } else if (src_offset == dest_offset && src_offset != NULL) {
4677 // This can occur if the offsets are identical non-constants.
4678 disjoint = true;
4679 }
4680
4681 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4682 }
4683
4684
4685 //------------------------------inline_arraycopy-----------------------
4686 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4687 // Object dest, int destPos,
4688 // int length);
4689 bool LibraryCallKit::inline_arraycopy() {
4690 // Get the arguments.
4691 Node* src = argument(0); // type: oop
4692 Node* src_offset = argument(1); // type: int
4693 Node* dest = argument(2); // type: oop
4694 Node* dest_offset = argument(3); // type: int
4695 Node* length = argument(4); // type: int
4696
4697 // Compile time checks. If any of these checks cannot be verified at compile time,
4698 // we do not make a fast path for this call. Instead, we let the call remain as it
4699 // is. The checks we choose to mandate at compile time are:
4700 //
4701 // (1) src and dest are arrays.
4702 const Type* src_type = src->Value(&_gvn);
4703 const Type* dest_type = dest->Value(&_gvn);
4704 const TypeAryPtr* top_src = src_type->isa_aryptr();
4705 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4706 if (top_src == NULL || top_src->klass() == NULL ||
4707 top_dest == NULL || top_dest->klass() == NULL) {
4708 // Conservatively insert a memory barrier on all memory slices.
4709 // Do not let writes into the source float below the arraycopy.
4710 insert_mem_bar(Op_MemBarCPUOrder);
4711
4712 // Call StubRoutines::generic_arraycopy stub.
4713 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4714 src, src_offset, dest, dest_offset, length);
4715
4716 // Do not let reads from the destination float above the arraycopy.
4717 // Since we cannot type the arrays, we don't know which slices
4718 // might be affected. We could restrict this barrier only to those
4719 // memory slices which pertain to array elements--but don't bother.
4720 if (!InsertMemBarAfterArraycopy)
4721 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4722 insert_mem_bar(Op_MemBarCPUOrder);
4723 return true;
4724 }
4725
4726 // (2) src and dest arrays must have elements of the same BasicType
4727 // Figure out the size and type of the elements we will be copying.
4728 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4729 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4730 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4731 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4732
4733 if (src_elem != dest_elem || dest_elem == T_VOID) {
4734 // The component types are not the same or are not recognized. Punt.
4735 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4736 generate_slow_arraycopy(TypePtr::BOTTOM,
4737 src, src_offset, dest, dest_offset, length,
4738 /*dest_uninitialized*/false);
4739 return true;
4740 }
4741
4742 //---------------------------------------------------------------------------
4743 // We will make a fast path for this call to arraycopy.
4744
4745 // We have the following tests left to perform:
4746 //
4747 // (3) src and dest must not be null.
4748 // (4) src_offset must not be negative.
4749 // (5) dest_offset must not be negative.
4750 // (6) length must not be negative.
4751 // (7) src_offset + length must not exceed length of src.
4752 // (8) dest_offset + length must not exceed length of dest.
4753 // (9) each element of an oop array must be assignable
4754
4755 RegionNode* slow_region = new (C) RegionNode(1);
4756 record_for_igvn(slow_region);
4757
4758 // (3) operands must not be null
4759 // We currently perform our null checks with the null_check routine.
4760 // This means that the null exceptions will be reported in the caller
4761 // rather than (correctly) reported inside of the native arraycopy call.
4762 // This should be corrected, given time. We do our null check with the
4763 // stack pointer restored.
4764 src = null_check(src, T_ARRAY);
4765 dest = null_check(dest, T_ARRAY);
4766
4767 // (4) src_offset must not be negative.
4768 generate_negative_guard(src_offset, slow_region);
4769
4770 // (5) dest_offset must not be negative.
4771 generate_negative_guard(dest_offset, slow_region);
4772
4773 // (6) length must not be negative (moved to generate_arraycopy()).
4774 // generate_negative_guard(length, slow_region);
4775
4776 // (7) src_offset + length must not exceed length of src.
4777 generate_limit_guard(src_offset, length,
4778 load_array_length(src),
4779 slow_region);
4780
4781 // (8) dest_offset + length must not exceed length of dest.
4782 generate_limit_guard(dest_offset, length,
4783 load_array_length(dest),
4784 slow_region);
4785
4786 // (9) each element of an oop array must be assignable
4787 // The generate_arraycopy subroutine checks this.
4788
4789 // This is where the memory effects are placed:
4790 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4791 generate_arraycopy(adr_type, dest_elem,
4792 src, src_offset, dest, dest_offset, length,
4793 false, false, slow_region);
4794
4795 return true;
4796 }
4797
4798 //-----------------------------generate_arraycopy----------------------
4799 // Generate an optimized call to arraycopy.
4800 // Caller must guard against non-arrays.
4801 // Caller must determine a common array basic-type for both arrays.
4802 // Caller must validate offsets against array bounds.
4803 // The slow_region has already collected guard failure paths
4804 // (such as out of bounds length or non-conformable array types).
4805 // The generated code has this shape, in general:
4806 //
4807 // if (length == 0) return // via zero_path
4808 // slowval = -1
4809 // if (types unknown) {
4810 // slowval = call generic copy loop
4811 // if (slowval == 0) return // via checked_path
4812 // } else if (indexes in bounds) {
4813 // if ((is object array) && !(array type check)) {
4814 // slowval = call checked copy loop
4815 // if (slowval == 0) return // via checked_path
4816 // } else {
4817 // call bulk copy loop
4818 // return // via fast_path
4819 // }
4820 // }
4821 // // adjust params for remaining work:
4822 // if (slowval != -1) {
4823 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4824 // }
4825 // slow_region:
4826 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4827 // return // via slow_call_path
4828 //
4829 // This routine is used from several intrinsics: System.arraycopy,
4830 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4831 //
4832 void
4833 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4834 BasicType basic_elem_type,
4835 Node* src, Node* src_offset,
4836 Node* dest, Node* dest_offset,
4837 Node* copy_length,
4838 bool disjoint_bases,
4839 bool length_never_negative,
4840 RegionNode* slow_region) {
4841
4842 if (slow_region == NULL) {
4843 slow_region = new(C) RegionNode(1);
4844 record_for_igvn(slow_region);
4845 }
4846
4847 Node* original_dest = dest;
4848 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4849 bool dest_uninitialized = false;
4850
4851 // See if this is the initialization of a newly-allocated array.
4852 // If so, we will take responsibility here for initializing it to zero.
4853 // (Note: Because tightly_coupled_allocation performs checks on the
4854 // out-edges of the dest, we need to avoid making derived pointers
4855 // from it until we have checked its uses.)
4856 if (ReduceBulkZeroing
4857 && !ZeroTLAB // pointless if already zeroed
4858 && basic_elem_type != T_CONFLICT // avoid corner case
4859 && !src->eqv_uncast(dest)
4860 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4861 != NULL)
4862 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4863 && alloc->maybe_set_complete(&_gvn)) {
4864 // "You break it, you buy it."
4865 InitializeNode* init = alloc->initialization();
4866 assert(init->is_complete(), "we just did this");
4867 init->set_complete_with_arraycopy();
4868 assert(dest->is_CheckCastPP(), "sanity");
4869 assert(dest->in(0)->in(0) == init, "dest pinned");
4870 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4871 // From this point on, every exit path is responsible for
4872 // initializing any non-copied parts of the object to zero.
4873 // Also, if this flag is set we make sure that arraycopy interacts properly
4874 // with G1, eliding pre-barriers. See CR 6627983.
4875 dest_uninitialized = true;
4876 } else {
4877 // No zeroing elimination here.
4878 alloc = NULL;
4879 //original_dest = dest;
4880 //dest_uninitialized = false;
4881 }
4882
4883 // Results are placed here:
4884 enum { fast_path = 1, // normal void-returning assembly stub
4885 checked_path = 2, // special assembly stub with cleanup
4886 slow_call_path = 3, // something went wrong; call the VM
4887 zero_path = 4, // bypass when length of copy is zero
4888 bcopy_path = 5, // copy primitive array by 64-bit blocks
4889 PATH_LIMIT = 6
4890 };
4891 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4892 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4893 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4894 record_for_igvn(result_region);
4895 _gvn.set_type_bottom(result_i_o);
4896 _gvn.set_type_bottom(result_memory);
4897 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4898
4899 // The slow_control path:
4900 Node* slow_control;
4901 Node* slow_i_o = i_o();
4902 Node* slow_mem = memory(adr_type);
4903 debug_only(slow_control = (Node*) badAddress);
4904
4905 // Checked control path:
4906 Node* checked_control = top();
4907 Node* checked_mem = NULL;
4908 Node* checked_i_o = NULL;
4909 Node* checked_value = NULL;
4910
4911 if (basic_elem_type == T_CONFLICT) {
4912 assert(!dest_uninitialized, "");
4913 Node* cv = generate_generic_arraycopy(adr_type,
4914 src, src_offset, dest, dest_offset,
4915 copy_length, dest_uninitialized);
4916 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4917 checked_control = control();
4918 checked_i_o = i_o();
4919 checked_mem = memory(adr_type);
4920 checked_value = cv;
4921 set_control(top()); // no fast path
4922 }
4923
4924 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4925 if (not_pos != NULL) {
4926 PreserveJVMState pjvms(this);
4927 set_control(not_pos);
4928
4929 // (6) length must not be negative.
4930 if (!length_never_negative) {
4931 generate_negative_guard(copy_length, slow_region);
4932 }
4933
4934 // copy_length is 0.
4935 if (!stopped() && dest_uninitialized) {
4936 Node* dest_length = alloc->in(AllocateNode::ALength);
4937 if (copy_length->eqv_uncast(dest_length)
4938 || _gvn.find_int_con(dest_length, 1) <= 0) {
4939 // There is no zeroing to do. No need for a secondary raw memory barrier.
4940 } else {
4941 // Clear the whole thing since there are no source elements to copy.
4942 generate_clear_array(adr_type, dest, basic_elem_type,
4943 intcon(0), NULL,
4944 alloc->in(AllocateNode::AllocSize));
4945 // Use a secondary InitializeNode as raw memory barrier.
4946 // Currently it is needed only on this path since other
4947 // paths have stub or runtime calls as raw memory barriers.
4948 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4949 Compile::AliasIdxRaw,
4950 top())->as_Initialize();
4951 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4952 }
4953 }
4954
4955 // Present the results of the fast call.
4956 result_region->init_req(zero_path, control());
4957 result_i_o ->init_req(zero_path, i_o());
4958 result_memory->init_req(zero_path, memory(adr_type));
4959 }
4960
4961 if (!stopped() && dest_uninitialized) {
4962 // We have to initialize the *uncopied* part of the array to zero.
4963 // The copy destination is the slice dest[off..off+len]. The other slices
4964 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4965 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4966 Node* dest_length = alloc->in(AllocateNode::ALength);
4967 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4968 copy_length));
4969
4970 // If there is a head section that needs zeroing, do it now.
4971 if (find_int_con(dest_offset, -1) != 0) {
4972 generate_clear_array(adr_type, dest, basic_elem_type,
4973 intcon(0), dest_offset,
4974 NULL);
4975 }
4976
4977 // Next, perform a dynamic check on the tail length.
4978 // It is often zero, and we can win big if we prove this.
4979 // There are two wins: Avoid generating the ClearArray
4980 // with its attendant messy index arithmetic, and upgrade
4981 // the copy to a more hardware-friendly word size of 64 bits.
4982 Node* tail_ctl = NULL;
4983 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4984 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
4985 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
4986 tail_ctl = generate_slow_guard(bol_lt, NULL);
4987 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4988 }
4989
4990 // At this point, let's assume there is no tail.
4991 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4992 // There is no tail. Try an upgrade to a 64-bit copy.
4993 bool didit = false;
4994 { PreserveJVMState pjvms(this);
4995 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4996 src, src_offset, dest, dest_offset,
4997 dest_size, dest_uninitialized);
4998 if (didit) {
4999 // Present the results of the block-copying fast call.
5000 result_region->init_req(bcopy_path, control());
5001 result_i_o ->init_req(bcopy_path, i_o());
5002 result_memory->init_req(bcopy_path, memory(adr_type));
5003 }
5004 }
5005 if (didit)
5006 set_control(top()); // no regular fast path
5007 }
5008
5009 // Clear the tail, if any.
5010 if (tail_ctl != NULL) {
5011 Node* notail_ctl = stopped() ? NULL : control();
5012 set_control(tail_ctl);
5013 if (notail_ctl == NULL) {
5014 generate_clear_array(adr_type, dest, basic_elem_type,
5015 dest_tail, NULL,
5016 dest_size);
5017 } else {
5018 // Make a local merge.
5019 Node* done_ctl = new(C) RegionNode(3);
5020 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5021 done_ctl->init_req(1, notail_ctl);
5022 done_mem->init_req(1, memory(adr_type));
5023 generate_clear_array(adr_type, dest, basic_elem_type,
5024 dest_tail, NULL,
5025 dest_size);
5026 done_ctl->init_req(2, control());
5027 done_mem->init_req(2, memory(adr_type));
5028 set_control( _gvn.transform(done_ctl));
5029 set_memory( _gvn.transform(done_mem), adr_type );
5030 }
5031 }
5032 }
5033
5034 BasicType copy_type = basic_elem_type;
5035 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5036 if (!stopped() && copy_type == T_OBJECT) {
5037 // If src and dest have compatible element types, we can copy bits.
5038 // Types S[] and D[] are compatible if D is a supertype of S.
5039 //
5040 // If they are not, we will use checked_oop_disjoint_arraycopy,
5041 // which performs a fast optimistic per-oop check, and backs off
5042 // further to JVM_ArrayCopy on the first per-oop check that fails.
5043 // (Actually, we don't move raw bits only; the GC requires card marks.)
5044
5045 // Get the Klass* for both src and dest
5046 Node* src_klass = load_object_klass(src);
5047 Node* dest_klass = load_object_klass(dest);
5048
5049 // Generate the subtype check.
5050 // This might fold up statically, or then again it might not.
5051 //
5052 // Non-static example: Copying List<String>.elements to a new String[].
5053 // The backing store for a List<String> is always an Object[],
5054 // but its elements are always type String, if the generic types
5055 // are correct at the source level.
5056 //
5057 // Test S[] against D[], not S against D, because (probably)
5058 // the secondary supertype cache is less busy for S[] than S.
5059 // This usually only matters when D is an interface.
5060 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5061 // Plug failing path into checked_oop_disjoint_arraycopy
5062 if (not_subtype_ctrl != top()) {
5063 PreserveJVMState pjvms(this);
5064 set_control(not_subtype_ctrl);
5065 // (At this point we can assume disjoint_bases, since types differ.)
5066 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5067 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5068 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5069 Node* dest_elem_klass = _gvn.transform(n1);
5070 Node* cv = generate_checkcast_arraycopy(adr_type,
5071 dest_elem_klass,
5072 src, src_offset, dest, dest_offset,
5073 ConvI2X(copy_length), dest_uninitialized);
5074 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5075 checked_control = control();
5076 checked_i_o = i_o();
5077 checked_mem = memory(adr_type);
5078 checked_value = cv;
5079 }
5080 // At this point we know we do not need type checks on oop stores.
5081
5082 // Let's see if we need card marks:
5083 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5084 // If we do not need card marks, copy using the jint or jlong stub.
5085 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5086 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5087 "sizes agree");
5088 }
5089 }
5090
5091 if (!stopped()) {
5092 // Generate the fast path, if possible.
5093 PreserveJVMState pjvms(this);
5094 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5095 src, src_offset, dest, dest_offset,
5096 ConvI2X(copy_length), dest_uninitialized);
5097
5098 // Present the results of the fast call.
5099 result_region->init_req(fast_path, control());
5100 result_i_o ->init_req(fast_path, i_o());
5101 result_memory->init_req(fast_path, memory(adr_type));
5102 }
5103
5104 // Here are all the slow paths up to this point, in one bundle:
5105 slow_control = top();
5106 if (slow_region != NULL)
5107 slow_control = _gvn.transform(slow_region);
5108 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5109
5110 set_control(checked_control);
5111 if (!stopped()) {
5112 // Clean up after the checked call.
5113 // The returned value is either 0 or -1^K,
5114 // where K = number of partially transferred array elements.
5115 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5116 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5117 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5118
5119 // If it is 0, we are done, so transfer to the end.
5120 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5121 result_region->init_req(checked_path, checks_done);
5122 result_i_o ->init_req(checked_path, checked_i_o);
5123 result_memory->init_req(checked_path, checked_mem);
5124
5125 // If it is not zero, merge into the slow call.
5126 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5127 RegionNode* slow_reg2 = new(C) RegionNode(3);
5128 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5129 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5130 record_for_igvn(slow_reg2);
5131 slow_reg2 ->init_req(1, slow_control);
5132 slow_i_o2 ->init_req(1, slow_i_o);
5133 slow_mem2 ->init_req(1, slow_mem);
5134 slow_reg2 ->init_req(2, control());
5135 slow_i_o2 ->init_req(2, checked_i_o);
5136 slow_mem2 ->init_req(2, checked_mem);
5137
5138 slow_control = _gvn.transform(slow_reg2);
5139 slow_i_o = _gvn.transform(slow_i_o2);
5140 slow_mem = _gvn.transform(slow_mem2);
5141
5142 if (alloc != NULL) {
5143 // We'll restart from the very beginning, after zeroing the whole thing.
5144 // This can cause double writes, but that's OK since dest is brand new.
5145 // So we ignore the low 31 bits of the value returned from the stub.
5146 } else {
5147 // We must continue the copy exactly where it failed, or else
5148 // another thread might see the wrong number of writes to dest.
5149 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5150 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5151 slow_offset->init_req(1, intcon(0));
5152 slow_offset->init_req(2, checked_offset);
5153 slow_offset = _gvn.transform(slow_offset);
5154
5155 // Adjust the arguments by the conditionally incoming offset.
5156 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5157 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5158 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5159
5160 // Tweak the node variables to adjust the code produced below:
5161 src_offset = src_off_plus;
5162 dest_offset = dest_off_plus;
5163 copy_length = length_minus;
5164 }
5165 }
5166
5167 set_control(slow_control);
5168 if (!stopped()) {
5169 // Generate the slow path, if needed.
5170 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5171
5172 set_memory(slow_mem, adr_type);
5173 set_i_o(slow_i_o);
5174
5175 if (dest_uninitialized) {
5176 generate_clear_array(adr_type, dest, basic_elem_type,
5177 intcon(0), NULL,
5178 alloc->in(AllocateNode::AllocSize));
5179 }
5180
5181 generate_slow_arraycopy(adr_type,
5182 src, src_offset, dest, dest_offset,
5183 copy_length, /*dest_uninitialized*/false);
5184
5185 result_region->init_req(slow_call_path, control());
5186 result_i_o ->init_req(slow_call_path, i_o());
5187 result_memory->init_req(slow_call_path, memory(adr_type));
5188 }
5189
5190 // Remove unused edges.
5191 for (uint i = 1; i < result_region->req(); i++) {
5192 if (result_region->in(i) == NULL)
5193 result_region->init_req(i, top());
5194 }
5195
5196 // Finished; return the combined state.
5197 set_control( _gvn.transform(result_region));
5198 set_i_o( _gvn.transform(result_i_o) );
5199 set_memory( _gvn.transform(result_memory), adr_type );
5200
5201 // The memory edges above are precise in order to model effects around
5202 // array copies accurately to allow value numbering of field loads around
5203 // arraycopy. Such field loads, both before and after, are common in Java
5204 // collections and similar classes involving header/array data structures.
5205 //
5206 // But with low number of register or when some registers are used or killed
5207 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5208 // The next memory barrier is added to avoid it. If the arraycopy can be
5209 // optimized away (which it can, sometimes) then we can manually remove
5210 // the membar also.
5211 //
5212 // Do not let reads from the cloned object float above the arraycopy.
5213 if (alloc != NULL) {
5214 // Do not let stores that initialize this object be reordered with
5215 // a subsequent store that would make this object accessible by
5216 // other threads.
5217 // Record what AllocateNode this StoreStore protects so that
5218 // escape analysis can go from the MemBarStoreStoreNode to the
5219 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5220 // based on the escape status of the AllocateNode.
5221 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5222 } else if (InsertMemBarAfterArraycopy)
5223 insert_mem_bar(Op_MemBarCPUOrder);
5224 }
5225
5226
5227 // Helper function which determines if an arraycopy immediately follows
5228 // an allocation, with no intervening tests or other escapes for the object.
5229 AllocateArrayNode*
5230 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5231 RegionNode* slow_region) {
5232 if (stopped()) return NULL; // no fast path
5233 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5234
5235 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5236 if (alloc == NULL) return NULL;
5237
5238 Node* rawmem = memory(Compile::AliasIdxRaw);
5239 // Is the allocation's memory state untouched?
5240 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5241 // Bail out if there have been raw-memory effects since the allocation.
5242 // (Example: There might have been a call or safepoint.)
5243 return NULL;
5244 }
5245 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5246 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5247 return NULL;
5248 }
5249
5250 // There must be no unexpected observers of this allocation.
5251 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5252 Node* obs = ptr->fast_out(i);
5253 if (obs != this->map()) {
5254 return NULL;
5255 }
5256 }
5257
5258 // This arraycopy must unconditionally follow the allocation of the ptr.
5259 Node* alloc_ctl = ptr->in(0);
5260 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5261
5262 Node* ctl = control();
5263 while (ctl != alloc_ctl) {
5264 // There may be guards which feed into the slow_region.
5265 // Any other control flow means that we might not get a chance
5266 // to finish initializing the allocated object.
5267 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5268 IfNode* iff = ctl->in(0)->as_If();
5269 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5270 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5271 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5272 ctl = iff->in(0); // This test feeds the known slow_region.
5273 continue;
5274 }
5275 // One more try: Various low-level checks bottom out in
5276 // uncommon traps. If the debug-info of the trap omits
5277 // any reference to the allocation, as we've already
5278 // observed, then there can be no objection to the trap.
5279 bool found_trap = false;
5280 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5281 Node* obs = not_ctl->fast_out(j);
5282 if (obs->in(0) == not_ctl && obs->is_Call() &&
5283 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5284 found_trap = true; break;
5285 }
5286 }
5287 if (found_trap) {
5288 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5289 continue;
5290 }
5291 }
5292 return NULL;
5293 }
5294
5295 // If we get this far, we have an allocation which immediately
5296 // precedes the arraycopy, and we can take over zeroing the new object.
5297 // The arraycopy will finish the initialization, and provide
5298 // a new control state to which we will anchor the destination pointer.
5299
5300 return alloc;
5301 }
5302
5303 // Helper for initialization of arrays, creating a ClearArray.
5304 // It writes zero bits in [start..end), within the body of an array object.
5305 // The memory effects are all chained onto the 'adr_type' alias category.
5306 //
5307 // Since the object is otherwise uninitialized, we are free
5308 // to put a little "slop" around the edges of the cleared area,
5309 // as long as it does not go back into the array's header,
5310 // or beyond the array end within the heap.
5311 //
5312 // The lower edge can be rounded down to the nearest jint and the
5313 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5314 //
5315 // Arguments:
5316 // adr_type memory slice where writes are generated
5317 // dest oop of the destination array
5318 // basic_elem_type element type of the destination
5319 // slice_idx array index of first element to store
5320 // slice_len number of elements to store (or NULL)
5321 // dest_size total size in bytes of the array object
5322 //
5323 // Exactly one of slice_len or dest_size must be non-NULL.
5324 // If dest_size is non-NULL, zeroing extends to the end of the object.
5325 // If slice_len is non-NULL, the slice_idx value must be a constant.
5326 void
5327 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5328 Node* dest,
5329 BasicType basic_elem_type,
5330 Node* slice_idx,
5331 Node* slice_len,
5332 Node* dest_size) {
5333 // one or the other but not both of slice_len and dest_size:
5334 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5335 if (slice_len == NULL) slice_len = top();
5336 if (dest_size == NULL) dest_size = top();
5337
5338 // operate on this memory slice:
5339 Node* mem = memory(adr_type); // memory slice to operate on
5340
5341 // scaling and rounding of indexes:
5342 int scale = exact_log2(type2aelembytes(basic_elem_type));
5343 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5344 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5345 int bump_bit = (-1 << scale) & BytesPerInt;
5346
5347 // determine constant starts and ends
5348 const intptr_t BIG_NEG = -128;
5349 assert(BIG_NEG + 2*abase < 0, "neg enough");
5350 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5351 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5352 if (slice_len_con == 0) {
5353 return; // nothing to do here
5354 }
5355 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5356 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5357 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5358 assert(end_con < 0, "not two cons");
5359 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5360 BytesPerLong);
5361 }
5362
5363 if (start_con >= 0 && end_con >= 0) {
5364 // Constant start and end. Simple.
5365 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5366 start_con, end_con, &_gvn);
5367 } else if (start_con >= 0 && dest_size != top()) {
5368 // Constant start, pre-rounded end after the tail of the array.
5369 Node* end = dest_size;
5370 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5371 start_con, end, &_gvn);
5372 } else if (start_con >= 0 && slice_len != top()) {
5373 // Constant start, non-constant end. End needs rounding up.
5374 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5375 intptr_t end_base = abase + (slice_idx_con << scale);
5376 int end_round = (-1 << scale) & (BytesPerLong - 1);
5377 Node* end = ConvI2X(slice_len);
5378 if (scale != 0)
5379 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5380 end_base += end_round;
5381 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5382 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5383 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5384 start_con, end, &_gvn);
5385 } else if (start_con < 0 && dest_size != top()) {
5386 // Non-constant start, pre-rounded end after the tail of the array.
5387 // This is almost certainly a "round-to-end" operation.
5388 Node* start = slice_idx;
5389 start = ConvI2X(start);
5390 if (scale != 0)
5391 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5392 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5393 if ((bump_bit | clear_low) != 0) {
5394 int to_clear = (bump_bit | clear_low);
5395 // Align up mod 8, then store a jint zero unconditionally
5396 // just before the mod-8 boundary.
5397 if (((abase + bump_bit) & ~to_clear) - bump_bit
5398 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5399 bump_bit = 0;
5400 assert((abase & to_clear) == 0, "array base must be long-aligned");
5401 } else {
5402 // Bump 'start' up to (or past) the next jint boundary:
5403 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5404 assert((abase & clear_low) == 0, "array base must be int-aligned");
5405 }
5406 // Round bumped 'start' down to jlong boundary in body of array.
5407 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5408 if (bump_bit != 0) {
5409 // Store a zero to the immediately preceding jint:
5410 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5411 Node* p1 = basic_plus_adr(dest, x1);
5412 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5413 mem = _gvn.transform(mem);
5414 }
5415 }
5416 Node* end = dest_size; // pre-rounded
5417 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5418 start, end, &_gvn);
5419 } else {
5420 // Non-constant start, unrounded non-constant end.
5421 // (Nobody zeroes a random midsection of an array using this routine.)
5422 ShouldNotReachHere(); // fix caller
5423 }
5424
5425 // Done.
5426 set_memory(mem, adr_type);
5427 }
5428
5429
5430 bool
5431 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5432 BasicType basic_elem_type,
5433 AllocateNode* alloc,
5434 Node* src, Node* src_offset,
5435 Node* dest, Node* dest_offset,
5436 Node* dest_size, bool dest_uninitialized) {
5437 // See if there is an advantage from block transfer.
5438 int scale = exact_log2(type2aelembytes(basic_elem_type));
5439 if (scale >= LogBytesPerLong)
5440 return false; // it is already a block transfer
5441
5442 // Look at the alignment of the starting offsets.
5443 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5444
5445 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5446 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5447 if (src_off_con < 0 || dest_off_con < 0)
5448 // At present, we can only understand constants.
5449 return false;
5450
5451 intptr_t src_off = abase + (src_off_con << scale);
5452 intptr_t dest_off = abase + (dest_off_con << scale);
5453
5454 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5455 // Non-aligned; too bad.
5456 // One more chance: Pick off an initial 32-bit word.
5457 // This is a common case, since abase can be odd mod 8.
5458 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5459 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5460 Node* sptr = basic_plus_adr(src, src_off);
5461 Node* dptr = basic_plus_adr(dest, dest_off);
5462 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5463 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5464 src_off += BytesPerInt;
5465 dest_off += BytesPerInt;
5466 } else {
5467 return false;
5468 }
5469 }
5470 assert(src_off % BytesPerLong == 0, "");
5471 assert(dest_off % BytesPerLong == 0, "");
5472
5473 // Do this copy by giant steps.
5474 Node* sptr = basic_plus_adr(src, src_off);
5475 Node* dptr = basic_plus_adr(dest, dest_off);
5476 Node* countx = dest_size;
5477 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5478 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5479
5480 bool disjoint_bases = true; // since alloc != NULL
5481 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5482 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5483
5484 return true;
5485 }
5486
5487
5488 // Helper function; generates code for the slow case.
5489 // We make a call to a runtime method which emulates the native method,
5490 // but without the native wrapper overhead.
5491 void
5492 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5493 Node* src, Node* src_offset,
5494 Node* dest, Node* dest_offset,
5495 Node* copy_length, bool dest_uninitialized) {
5496 assert(!dest_uninitialized, "Invariant");
5497 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5498 OptoRuntime::slow_arraycopy_Type(),
5499 OptoRuntime::slow_arraycopy_Java(),
5500 "slow_arraycopy", adr_type,
5501 src, src_offset, dest, dest_offset,
5502 copy_length);
5503
5504 // Handle exceptions thrown by this fellow:
5505 make_slow_call_ex(call, env()->Throwable_klass(), false);
5506 }
5507
5508 // Helper function; generates code for cases requiring runtime checks.
5509 Node*
5510 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5511 Node* dest_elem_klass,
5512 Node* src, Node* src_offset,
5513 Node* dest, Node* dest_offset,
5514 Node* copy_length, bool dest_uninitialized) {
5515 if (stopped()) return NULL;
5516
5517 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5518 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5519 return NULL;
5520 }
5521
5522 // Pick out the parameters required to perform a store-check
5523 // for the target array. This is an optimistic check. It will
5524 // look in each non-null element's class, at the desired klass's
5525 // super_check_offset, for the desired klass.
5526 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5527 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5528 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5529 Node* check_offset = ConvI2X(_gvn.transform(n3));
5530 Node* check_value = dest_elem_klass;
5531
5532 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5533 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5534
5535 // (We know the arrays are never conjoint, because their types differ.)
5536 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5537 OptoRuntime::checkcast_arraycopy_Type(),
5538 copyfunc_addr, "checkcast_arraycopy", adr_type,
5539 // five arguments, of which two are
5540 // intptr_t (jlong in LP64)
5541 src_start, dest_start,
5542 copy_length XTOP,
5543 check_offset XTOP,
5544 check_value);
5545
5546 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5547 }
5548
5549
5550 // Helper function; generates code for cases requiring runtime checks.
5551 Node*
5552 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5553 Node* src, Node* src_offset,
5554 Node* dest, Node* dest_offset,
5555 Node* copy_length, bool dest_uninitialized) {
5556 assert(!dest_uninitialized, "Invariant");
5557 if (stopped()) return NULL;
5558 address copyfunc_addr = StubRoutines::generic_arraycopy();
5559 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5560 return NULL;
5561 }
5562
5563 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5564 OptoRuntime::generic_arraycopy_Type(),
5565 copyfunc_addr, "generic_arraycopy", adr_type,
5566 src, src_offset, dest, dest_offset, copy_length);
5567
5568 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5569 }
5570
5571 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5572 void
5573 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5574 BasicType basic_elem_type,
5575 bool disjoint_bases,
5576 Node* src, Node* src_offset,
5577 Node* dest, Node* dest_offset,
5578 Node* copy_length, bool dest_uninitialized) {
5579 if (stopped()) return; // nothing to do
5580
5581 Node* src_start = src;
5582 Node* dest_start = dest;
5583 if (src_offset != NULL || dest_offset != NULL) {
5584 assert(src_offset != NULL && dest_offset != NULL, "");
5585 src_start = array_element_address(src, src_offset, basic_elem_type);
5586 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5587 }
5588
5589 // Figure out which arraycopy runtime method to call.
5590 const char* copyfunc_name = "arraycopy";
5591 address copyfunc_addr =
5592 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5593 disjoint_bases, copyfunc_name, dest_uninitialized);
5594
5595 // Call it. Note that the count_ix value is not scaled to a byte-size.
5596 make_runtime_call(RC_LEAF|RC_NO_FP,
5597 OptoRuntime::fast_arraycopy_Type(),
5598 copyfunc_addr, copyfunc_name, adr_type,
5599 src_start, dest_start, copy_length XTOP);
5600 }
5601
5602 //-------------inline_encodeISOArray-----------------------------------
5603 // encode char[] to byte[] in ISO_8859_1
5604 bool LibraryCallKit::inline_encodeISOArray() {
5605 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5606 // no receiver since it is static method
5607 Node *src = argument(0);
5608 Node *src_offset = argument(1);
5609 Node *dst = argument(2);
5610 Node *dst_offset = argument(3);
5611 Node *length = argument(4);
5612
5613 const Type* src_type = src->Value(&_gvn);
5614 const Type* dst_type = dst->Value(&_gvn);
5615 const TypeAryPtr* top_src = src_type->isa_aryptr();
5616 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5617 if (top_src == NULL || top_src->klass() == NULL ||
5618 top_dest == NULL || top_dest->klass() == NULL) {
5619 // failed array check
5620 return false;
5621 }
5622
5623 // Figure out the size and type of the elements we will be copying.
5624 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5625 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5626 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5627 return false;
5628 }
5629 Node* src_start = array_element_address(src, src_offset, src_elem);
5630 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5631 // 'src_start' points to src array + scaled offset
5632 // 'dst_start' points to dst array + scaled offset
5633
5634 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5635 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5636 enc = _gvn.transform(enc);
5637 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5638 set_memory(res_mem, mtype);
5639 set_result(enc);
5640 return true;
5641 }
5642
5643 /**
5644 * Calculate CRC32 for byte.
5645 * int java.util.zip.CRC32.update(int crc, int b)
5646 */
5647 bool LibraryCallKit::inline_updateCRC32() {
5648 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5649 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5650 // no receiver since it is static method
5651 Node* crc = argument(0); // type: int
5652 Node* b = argument(1); // type: int
5653
5654 /*
5655 * int c = ~ crc;
5656 * b = timesXtoThe32[(b ^ c) & 0xFF];
5657 * b = b ^ (c >>> 8);
5658 * crc = ~b;
5659 */
5660
5661 Node* M1 = intcon(-1);
5662 crc = _gvn.transform(new (C) XorINode(crc, M1));
5663 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5664 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5665
5666 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5667 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5668 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5669 result = make_load(control(), adr, TypeInt::INT, T_INT);
5670
5671 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5672 result = _gvn.transform(new (C) XorINode(crc, result));
5673 result = _gvn.transform(new (C) XorINode(result, M1));
5674 set_result(result);
5675 return true;
5676 }
5677
5678 /**
5679 * Calculate CRC32 for byte[] array.
5680 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5681 */
5682 bool LibraryCallKit::inline_updateBytesCRC32() {
5683 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5684 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5685 // no receiver since it is static method
5686 Node* crc = argument(0); // type: int
5687 Node* src = argument(1); // type: oop
5688 Node* offset = argument(2); // type: int
5689 Node* length = argument(3); // type: int
5690
5691 const Type* src_type = src->Value(&_gvn);
5692 const TypeAryPtr* top_src = src_type->isa_aryptr();
5693 if (top_src == NULL || top_src->klass() == NULL) {
5694 // failed array check
5695 return false;
5696 }
5697
5698 // Figure out the size and type of the elements we will be copying.
5699 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5700 if (src_elem != T_BYTE) {
5701 return false;
5702 }
5703
5704 // 'src_start' points to src array + scaled offset
5705 Node* src_start = array_element_address(src, offset, src_elem);
5706
5707 // We assume that range check is done by caller.
5708 // TODO: generate range check (offset+length < src.length) in debug VM.
5709
5710 // Call the stub.
5711 address stubAddr = StubRoutines::updateBytesCRC32();
5712 const char *stubName = "updateBytesCRC32";
5713
5714 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5715 stubAddr, stubName, TypePtr::BOTTOM,
5716 crc, src_start, length);
5717 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5718 set_result(result);
5719 return true;
5720 }
5721
5722 /**
5723 * Calculate CRC32 for ByteBuffer.
5724 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5725 */
5726 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5727 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5728 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5729 // no receiver since it is static method
5730 Node* crc = argument(0); // type: int
5731 Node* src = argument(1); // type: long
5732 Node* offset = argument(3); // type: int
5733 Node* length = argument(4); // type: int
5734
5735 src = ConvL2X(src); // adjust Java long to machine word
5736 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5737 offset = ConvI2X(offset);
5738
5739 // 'src_start' points to src array + scaled offset
5740 Node* src_start = basic_plus_adr(top(), base, offset);
5741
5742 // Call the stub.
5743 address stubAddr = StubRoutines::updateBytesCRC32();
5744 const char *stubName = "updateBytesCRC32";
5745
5746 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5747 stubAddr, stubName, TypePtr::BOTTOM,
5748 crc, src_start, length);
5749 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5750 set_result(result);
5751 return true;
5752 }
5753
5754 //----------------------------inline_reference_get----------------------------
5755 // public T java.lang.ref.Reference.get();
5756 bool LibraryCallKit::inline_reference_get() {
5757 const int referent_offset = java_lang_ref_Reference::referent_offset;
5758 guarantee(referent_offset > 0, "should have already been set");
5759
5760 // Get the argument:
5761 Node* reference_obj = null_check_receiver();
5762 if (stopped()) return true;
5763
5764 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5765
5766 ciInstanceKlass* klass = env()->Object_klass();
5767 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5768
5769 Node* no_ctrl = NULL;
5770 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5771
5772 // Use the pre-barrier to record the value in the referent field
5773 pre_barrier(false /* do_load */,
5774 control(),
5775 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5776 result /* pre_val */,
5777 T_OBJECT);
5778
5779 // Add memory barrier to prevent commoning reads from this field
5780 // across safepoint since GC can change its value.
5781 insert_mem_bar(Op_MemBarCPUOrder);
5782
5783 set_result(result);
5784 return true;
5785 }
5786
5787
5788 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5789 bool is_exact=true, bool is_static=false) {
5790
5791 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5792 assert(tinst != NULL, "obj is null");
5793 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5794 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5795
5796 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5797 ciSymbol::make(fieldTypeString),
5798 is_static);
5799 if (field == NULL) return (Node *) NULL;
5800 assert (field != NULL, "undefined field");
5801
5802 // Next code copied from Parse::do_get_xxx():
5803
5804 // Compute address and memory type.
5805 int offset = field->offset_in_bytes();
5806 bool is_vol = field->is_volatile();
5807 ciType* field_klass = field->type();
5808 assert(field_klass->is_loaded(), "should be loaded");
5809 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5810 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5811 BasicType bt = field->layout_type();
5812
5813 // Build the resultant type of the load
5814 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5815
5816 // Build the load.
5817 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5818 return loadedField;
5819 }
5820
5821
5822 //------------------------------inline_aescrypt_Block-----------------------
5823 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5824 address stubAddr;
5825 const char *stubName;
5826 assert(UseAES, "need AES instruction support");
5827
5828 switch(id) {
5829 case vmIntrinsics::_aescrypt_encryptBlock:
5830 stubAddr = StubRoutines::aescrypt_encryptBlock();
5831 stubName = "aescrypt_encryptBlock";
5832 break;
5833 case vmIntrinsics::_aescrypt_decryptBlock:
5834 stubAddr = StubRoutines::aescrypt_decryptBlock();
5835 stubName = "aescrypt_decryptBlock";
5836 break;
5837 }
5838 if (stubAddr == NULL) return false;
5839
5840 Node* aescrypt_object = argument(0);
5841 Node* src = argument(1);
5842 Node* src_offset = argument(2);
5843 Node* dest = argument(3);
5844 Node* dest_offset = argument(4);
5845
5846 // (1) src and dest are arrays.
5847 const Type* src_type = src->Value(&_gvn);
5848 const Type* dest_type = dest->Value(&_gvn);
5849 const TypeAryPtr* top_src = src_type->isa_aryptr();
5850 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5851 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5852
5853 // for the quick and dirty code we will skip all the checks.
5854 // we are just trying to get the call to be generated.
5855 Node* src_start = src;
5856 Node* dest_start = dest;
5857 if (src_offset != NULL || dest_offset != NULL) {
5858 assert(src_offset != NULL && dest_offset != NULL, "");
5859 src_start = array_element_address(src, src_offset, T_BYTE);
5860 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5861 }
5862
5863 // now need to get the start of its expanded key array
5864 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5865 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5866 if (k_start == NULL) return false;
5867
5868 // Call the stub.
5869 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5870 stubAddr, stubName, TypePtr::BOTTOM,
5871 src_start, dest_start, k_start);
5872
5873 return true;
5874 }
5875
5876 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5877 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5878 address stubAddr;
5879 const char *stubName;
5880
5881 assert(UseAES, "need AES instruction support");
5882
5883 switch(id) {
5884 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5885 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5886 stubName = "cipherBlockChaining_encryptAESCrypt";
5887 break;
5888 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5889 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5890 stubName = "cipherBlockChaining_decryptAESCrypt";
5891 break;
5892 }
5893 if (stubAddr == NULL) return false;
5894
5895 Node* cipherBlockChaining_object = argument(0);
5896 Node* src = argument(1);
5897 Node* src_offset = argument(2);
5898 Node* len = argument(3);
5899 Node* dest = argument(4);
5900 Node* dest_offset = argument(5);
5901
5902 // (1) src and dest are arrays.
5903 const Type* src_type = src->Value(&_gvn);
5904 const Type* dest_type = dest->Value(&_gvn);
5905 const TypeAryPtr* top_src = src_type->isa_aryptr();
5906 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5907 assert (top_src != NULL && top_src->klass() != NULL
5908 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5909
5910 // checks are the responsibility of the caller
5911 Node* src_start = src;
5912 Node* dest_start = dest;
5913 if (src_offset != NULL || dest_offset != NULL) {
5914 assert(src_offset != NULL && dest_offset != NULL, "");
5915 src_start = array_element_address(src, src_offset, T_BYTE);
5916 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5917 }
5918
5919 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5920 // (because of the predicated logic executed earlier).
5921 // so we cast it here safely.
5922 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5923
5924 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5925 if (embeddedCipherObj == NULL) return false;
5926
5927 // cast it to what we know it will be at runtime
5928 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5929 assert(tinst != NULL, "CBC obj is null");
5930 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5931 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5932 if (!klass_AESCrypt->is_loaded()) return false;
5933
5934 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5935 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5936 const TypeOopPtr* xtype = aklass->as_instance_type();
5937 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5938 aescrypt_object = _gvn.transform(aescrypt_object);
5939
5940 // we need to get the start of the aescrypt_object's expanded key array
5941 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5942 if (k_start == NULL) return false;
5943
5944 // similarly, get the start address of the r vector
5945 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5946 if (objRvec == NULL) return false;
5947 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5948
5949 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5950 make_runtime_call(RC_LEAF|RC_NO_FP,
5951 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5952 stubAddr, stubName, TypePtr::BOTTOM,
5953 src_start, dest_start, k_start, r_start, len);
5954
5955 // return is void so no result needs to be pushed
5956
5957 return true;
5958 }
5959
5960 //------------------------------get_key_start_from_aescrypt_object-----------------------
5961 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5962 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5963 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5964 if (objAESCryptKey == NULL) return (Node *) NULL;
5965
5966 // now have the array, need to get the start address of the K array
5967 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5968 return k_start;
5969 }
5970
5971 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5972 // Return node representing slow path of predicate check.
5973 // the pseudo code we want to emulate with this predicate is:
5974 // for encryption:
5975 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5976 // for decryption:
5977 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5978 // note cipher==plain is more conservative than the original java code but that's OK
5979 //
5980 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5981 // First, check receiver for NULL since it is virtual method.
5982 Node* objCBC = argument(0);
5983 objCBC = null_check(objCBC);
5984
5985 if (stopped()) return NULL; // Always NULL
5986
5987 // Load embeddedCipher field of CipherBlockChaining object.
5988 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5989
5990 // get AESCrypt klass for instanceOf check
5991 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5992 // will have same classloader as CipherBlockChaining object
5993 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5994 assert(tinst != NULL, "CBCobj is null");
5995 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5996
5997 // we want to do an instanceof comparison against the AESCrypt class
5998 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5999 if (!klass_AESCrypt->is_loaded()) {
6000 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6001 Node* ctrl = control();
6002 set_control(top()); // no regular fast path
6003 return ctrl;
6004 }
6005 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6006
6007 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6008 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6009 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6010
6011 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6012
6013 // for encryption, we are done
6014 if (!decrypting)
6015 return instof_false; // even if it is NULL
6016
6017 // for decryption, we need to add a further check to avoid
6018 // taking the intrinsic path when cipher and plain are the same
6019 // see the original java code for why.
6020 RegionNode* region = new(C) RegionNode(3);
6021 region->init_req(1, instof_false);
6022 Node* src = argument(1);
6023 Node* dest = argument(4);
6024 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6025 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6026 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6027 region->init_req(2, src_dest_conjoint);
6028
6029 record_for_igvn(region);
6030 return _gvn.transform(region);
6031 }