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 "jfr/support/jfrIntrinsics.hpp"
31 #include "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callGenerator.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/mathexactnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/parse.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "prims/nativeLookup.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "utilities/macros.hpp"
45
46 class LibraryIntrinsic : public InlineCallGenerator {
47 // Extend the set of intrinsics known to the runtime:
48 public:
49 private:
50 bool _is_virtual;
51 bool _does_virtual_dispatch;
52 int8_t _predicates_count; // Intrinsic is predicated by several conditions
53 int8_t _last_predicate; // Last generated predicate
54 vmIntrinsics::ID _intrinsic_id;
55
56 public:
57 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
58 : InlineCallGenerator(m),
59 _is_virtual(is_virtual),
60 _does_virtual_dispatch(does_virtual_dispatch),
61 _predicates_count((int8_t)predicates_count),
62 _last_predicate((int8_t)-1),
63 _intrinsic_id(id)
64 {
65 }
66 virtual bool is_intrinsic() const { return true; }
67 virtual bool is_virtual() const { return _is_virtual; }
68 virtual bool is_predicated() const { return _predicates_count > 0; }
69 virtual int predicates_count() const { return _predicates_count; }
70 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
71 virtual JVMState* generate(JVMState* jvms);
72 virtual Node* generate_predicate(JVMState* jvms, int predicate);
73 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
74 };
75
76
77 // Local helper class for LibraryIntrinsic:
78 class LibraryCallKit : public GraphKit {
79 private:
80 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
81 Node* _result; // the result node, if any
82 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
83
84 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
85
86 public:
87 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
88 : GraphKit(jvms),
89 _intrinsic(intrinsic),
90 _result(NULL)
91 {
92 // Check if this is a root compile. In that case we don't have a caller.
93 if (!jvms->has_method()) {
94 _reexecute_sp = sp();
95 } else {
96 // Find out how many arguments the interpreter needs when deoptimizing
97 // and save the stack pointer value so it can used by uncommon_trap.
98 // We find the argument count by looking at the declared signature.
99 bool ignored_will_link;
100 ciSignature* declared_signature = NULL;
101 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
102 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
103 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
104 }
105 }
106
107 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
108
109 ciMethod* caller() const { return jvms()->method(); }
110 int bci() const { return jvms()->bci(); }
111 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
112 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
113 ciMethod* callee() const { return _intrinsic->method(); }
114
115 bool try_to_inline(int predicate);
116 Node* try_to_predicate(int predicate);
117
118 void push_result() {
119 // Push the result onto the stack.
120 if (!stopped() && result() != NULL) {
121 BasicType bt = result()->bottom_type()->basic_type();
122 push_node(bt, result());
123 }
124 }
125
126 private:
127 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
128 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
129 }
130
131 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
132 void set_result(RegionNode* region, PhiNode* value);
133 Node* result() { return _result; }
134
135 virtual int reexecute_sp() { return _reexecute_sp; }
136
137 // Helper functions to inline natives
138 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
139 Node* generate_slow_guard(Node* test, RegionNode* region);
140 Node* generate_fair_guard(Node* test, RegionNode* region);
141 Node* generate_negative_guard(Node* index, RegionNode* region,
142 // resulting CastII of index:
143 Node* *pos_index = NULL);
144 Node* generate_nonpositive_guard(Node* index, bool never_negative,
145 // resulting CastII of index:
146 Node* *pos_index = NULL);
147 Node* generate_limit_guard(Node* offset, Node* subseq_length,
148 Node* array_length,
149 RegionNode* region);
150 Node* generate_current_thread(Node* &tls_output);
151 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
152 bool disjoint_bases, const char* &name, bool dest_uninitialized);
153 Node* load_mirror_from_klass(Node* klass);
154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path,
156 int offset);
157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path) {
159 int offset = java_lang_Class::klass_offset_in_bytes();
160 return load_klass_from_mirror_common(mirror, never_see_null,
161 region, null_path,
162 offset);
163 }
164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path) {
166 int offset = java_lang_Class::array_klass_offset_in_bytes();
167 return load_klass_from_mirror_common(mirror, never_see_null,
168 region, null_path,
169 offset);
170 }
171 Node* generate_access_flags_guard(Node* kls,
172 int modifier_mask, int modifier_bits,
173 RegionNode* region);
174 Node* generate_interface_guard(Node* kls, RegionNode* region);
175 Node* generate_array_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, false, false);
177 }
178 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, true);
180 }
181 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, true, false);
183 }
184 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, true);
186 }
187 Node* generate_array_guard_common(Node* kls, RegionNode* region,
188 bool obj_array, bool not_array);
189 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
190 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
191 bool is_virtual = false, bool is_static = false);
192 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
193 return generate_method_call(method_id, false, true);
194 }
195 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, true, false);
197 }
198 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
199
200 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
201 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
202 bool inline_string_compareTo();
203 bool inline_string_indexOf();
204 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
205 bool inline_string_equals();
206 Node* round_double_node(Node* n);
207 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
208 bool inline_math_native(vmIntrinsics::ID id);
209 bool inline_trig(vmIntrinsics::ID id);
210 bool inline_math(vmIntrinsics::ID id);
211 template <typename OverflowOp>
212 bool inline_math_overflow(Node* arg1, Node* arg2);
213 void inline_math_mathExact(Node* math, Node* test);
214 bool inline_math_addExactI(bool is_increment);
215 bool inline_math_addExactL(bool is_increment);
216 bool inline_math_multiplyExactI();
217 bool inline_math_multiplyExactL();
218 bool inline_math_negateExactI();
219 bool inline_math_negateExactL();
220 bool inline_math_subtractExactI(bool is_decrement);
221 bool inline_math_subtractExactL(bool is_decrement);
222 bool inline_exp();
223 bool inline_pow();
224 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_min_max(vmIntrinsics::ID id);
226 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
227 // This returns Type::AnyPtr, RawPtr, or OopPtr.
228 int classify_unsafe_addr(Node* &base, Node* &offset);
229 Node* make_unsafe_address(Node* base, Node* offset);
230 // Helper for inline_unsafe_access.
231 // Generates the guards that check whether the result of
232 // Unsafe.getObject should be recorded in an SATB log buffer.
233 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
234 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
235 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
236 static bool klass_needs_init_guard(Node* kls);
237 bool inline_unsafe_allocate();
238 bool inline_unsafe_copyMemory();
239 bool inline_native_currentThread();
240 #ifdef JFR_HAVE_INTRINSICS
241 bool inline_native_classID();
242 bool inline_native_getEventWriter();
243 #endif
244 bool inline_native_time_funcs(address method, const char* funcName);
245 bool inline_native_isInterrupted();
246 bool inline_native_Class_query(vmIntrinsics::ID id);
247 bool inline_native_subtype_check();
248
249 bool inline_native_newArray();
250 bool inline_native_getLength();
251 bool inline_array_copyOf(bool is_copyOfRange);
252 bool inline_array_equals();
253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
254 bool inline_native_clone(bool is_virtual);
255 bool inline_native_Reflection_getCallerClass();
256 // Helper function for inlining native object hash method
257 bool inline_native_hashcode(bool is_virtual, bool is_static);
258 bool inline_native_getClass();
259
260 // Helper functions for inlining arraycopy
261 bool inline_arraycopy();
262 void generate_arraycopy(const TypePtr* adr_type,
263 BasicType basic_elem_type,
264 Node* src, Node* src_offset,
265 Node* dest, Node* dest_offset,
266 Node* copy_length,
267 bool disjoint_bases = false,
268 bool length_never_negative = false,
269 RegionNode* slow_region = NULL);
270 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
271 RegionNode* slow_region);
272 void generate_clear_array(const TypePtr* adr_type,
273 Node* dest,
274 BasicType basic_elem_type,
275 Node* slice_off,
276 Node* slice_len,
277 Node* slice_end);
278 bool generate_block_arraycopy(const TypePtr* adr_type,
279 BasicType basic_elem_type,
280 AllocateNode* alloc,
281 Node* src, Node* src_offset,
282 Node* dest, Node* dest_offset,
283 Node* dest_size, bool dest_uninitialized);
284 void generate_slow_arraycopy(const TypePtr* adr_type,
285 Node* src, Node* src_offset,
286 Node* dest, Node* dest_offset,
287 Node* copy_length, bool dest_uninitialized);
288 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
289 Node* dest_elem_klass,
290 Node* src, Node* src_offset,
291 Node* dest, Node* dest_offset,
292 Node* copy_length, bool dest_uninitialized);
293 Node* generate_generic_arraycopy(const TypePtr* adr_type,
294 Node* src, Node* src_offset,
295 Node* dest, Node* dest_offset,
296 Node* copy_length, bool dest_uninitialized);
297 void generate_unchecked_arraycopy(const TypePtr* adr_type,
298 BasicType basic_elem_type,
299 bool disjoint_bases,
300 Node* src, Node* src_offset,
301 Node* dest, Node* dest_offset,
302 Node* copy_length, bool dest_uninitialized);
303 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
304 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
305 bool inline_unsafe_ordered_store(BasicType type);
306 bool inline_unsafe_fence(vmIntrinsics::ID id);
307 bool inline_fp_conversions(vmIntrinsics::ID id);
308 bool inline_number_methods(vmIntrinsics::ID id);
309 bool inline_reference_get();
310 bool inline_aescrypt_Block(vmIntrinsics::ID id);
311 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
312 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
313 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
314 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
315 bool inline_sha_implCompress(vmIntrinsics::ID id);
316 bool inline_digestBase_implCompressMB(int predicate);
317 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
318 bool long_state, address stubAddr, const char *stubName,
319 Node* src_start, Node* ofs, Node* limit);
320 Node* get_state_from_sha_object(Node *sha_object);
321 Node* get_state_from_sha5_object(Node *sha_object);
322 Node* inline_digestBase_implCompressMB_predicate(int predicate);
323 bool inline_encodeISOArray();
324 bool inline_updateCRC32();
325 bool inline_updateBytesCRC32();
326 bool inline_updateByteBufferCRC32();
327 bool inline_multiplyToLen();
328 bool inline_squareToLen();
329 bool inline_mulAdd();
330 bool inline_montgomeryMultiply();
331 bool inline_montgomerySquare();
332
333 bool inline_profileBoolean();
334 };
335
336
337 //---------------------------make_vm_intrinsic----------------------------
338 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
339 vmIntrinsics::ID id = m->intrinsic_id();
340 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
341
342 ccstr disable_intr = NULL;
343
344 if ((DisableIntrinsic[0] != '\0'
345 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
346 (method_has_option_value("DisableIntrinsic", disable_intr)
347 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
348 // disabled by a user request on the command line:
349 // example: -XX:DisableIntrinsic=_hashCode,_getClass
350 return NULL;
351 }
352
353 if (!m->is_loaded()) {
354 // do not attempt to inline unloaded methods
355 return NULL;
356 }
357
358 // Only a few intrinsics implement a virtual dispatch.
359 // They are expensive calls which are also frequently overridden.
360 if (is_virtual) {
361 switch (id) {
362 case vmIntrinsics::_hashCode:
363 case vmIntrinsics::_clone:
364 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
365 break;
366 default:
367 return NULL;
368 }
369 }
370
371 // -XX:-InlineNatives disables nearly all intrinsics:
372 if (!InlineNatives) {
373 switch (id) {
374 case vmIntrinsics::_indexOf:
375 case vmIntrinsics::_compareTo:
376 case vmIntrinsics::_equals:
377 case vmIntrinsics::_equalsC:
378 case vmIntrinsics::_getAndAddInt:
379 case vmIntrinsics::_getAndAddLong:
380 case vmIntrinsics::_getAndSetInt:
381 case vmIntrinsics::_getAndSetLong:
382 case vmIntrinsics::_getAndSetObject:
383 case vmIntrinsics::_loadFence:
384 case vmIntrinsics::_storeFence:
385 case vmIntrinsics::_fullFence:
386 break; // InlineNatives does not control String.compareTo
387 case vmIntrinsics::_Reference_get:
388 break; // InlineNatives does not control Reference.get
389 default:
390 return NULL;
391 }
392 }
393
394 int predicates = 0;
395 bool does_virtual_dispatch = false;
396
397 switch (id) {
398 case vmIntrinsics::_compareTo:
399 if (!SpecialStringCompareTo) return NULL;
400 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
401 break;
402 case vmIntrinsics::_indexOf:
403 if (!SpecialStringIndexOf) return NULL;
404 break;
405 case vmIntrinsics::_equals:
406 if (!SpecialStringEquals) return NULL;
407 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
408 break;
409 case vmIntrinsics::_equalsC:
410 if (!SpecialArraysEquals) return NULL;
411 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
412 break;
413 case vmIntrinsics::_arraycopy:
414 if (!InlineArrayCopy) return NULL;
415 break;
416 case vmIntrinsics::_copyMemory:
417 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
418 if (!InlineArrayCopy) return NULL;
419 break;
420 case vmIntrinsics::_hashCode:
421 if (!InlineObjectHash) return NULL;
422 does_virtual_dispatch = true;
423 break;
424 case vmIntrinsics::_clone:
425 does_virtual_dispatch = true;
426 case vmIntrinsics::_copyOf:
427 case vmIntrinsics::_copyOfRange:
428 if (!InlineObjectCopy) return NULL;
429 // These also use the arraycopy intrinsic mechanism:
430 if (!InlineArrayCopy) return NULL;
431 break;
432 case vmIntrinsics::_encodeISOArray:
433 if (!SpecialEncodeISOArray) return NULL;
434 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
435 break;
436 case vmIntrinsics::_checkIndex:
437 // We do not intrinsify this. The optimizer does fine with it.
438 return NULL;
439
440 case vmIntrinsics::_getCallerClass:
441 if (!UseNewReflection) return NULL;
442 if (!InlineReflectionGetCallerClass) return NULL;
443 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
444 break;
445
446 case vmIntrinsics::_bitCount_i:
447 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
448 break;
449
450 case vmIntrinsics::_bitCount_l:
451 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
452 break;
453
454 case vmIntrinsics::_numberOfLeadingZeros_i:
455 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
456 break;
457
458 case vmIntrinsics::_numberOfLeadingZeros_l:
459 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
460 break;
461
462 case vmIntrinsics::_numberOfTrailingZeros_i:
463 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
464 break;
465
466 case vmIntrinsics::_numberOfTrailingZeros_l:
467 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
468 break;
469
470 case vmIntrinsics::_reverseBytes_c:
471 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
472 break;
473 case vmIntrinsics::_reverseBytes_s:
474 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
475 break;
476 case vmIntrinsics::_reverseBytes_i:
477 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
478 break;
479 case vmIntrinsics::_reverseBytes_l:
480 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
481 break;
482
483 case vmIntrinsics::_Reference_get:
484 // Use the intrinsic version of Reference.get() so that the value in
485 // the referent field can be registered by the G1 pre-barrier code.
486 // Also add memory barrier to prevent commoning reads from this field
487 // across safepoint since GC can change it value.
488 break;
489
490 case vmIntrinsics::_compareAndSwapObject:
491 #ifdef _LP64
492 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
493 #endif
494 break;
495
496 case vmIntrinsics::_compareAndSwapLong:
497 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
498 break;
499
500 case vmIntrinsics::_getAndAddInt:
501 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
502 break;
503
504 case vmIntrinsics::_getAndAddLong:
505 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
506 break;
507
508 case vmIntrinsics::_getAndSetInt:
509 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
510 break;
511
512 case vmIntrinsics::_getAndSetLong:
513 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
514 break;
515
516 case vmIntrinsics::_getAndSetObject:
517 #ifdef _LP64
518 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
519 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
520 break;
521 #else
522 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
523 break;
524 #endif
525
526 case vmIntrinsics::_aescrypt_encryptBlock:
527 case vmIntrinsics::_aescrypt_decryptBlock:
528 if (!UseAESIntrinsics) return NULL;
529 break;
530
531 case vmIntrinsics::_multiplyToLen:
532 if (!UseMultiplyToLenIntrinsic) return NULL;
533 break;
534
535 case vmIntrinsics::_squareToLen:
536 if (!UseSquareToLenIntrinsic) return NULL;
537 break;
538
539 case vmIntrinsics::_mulAdd:
540 if (!UseMulAddIntrinsic) return NULL;
541 break;
542
543 case vmIntrinsics::_montgomeryMultiply:
544 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
545 break;
546 case vmIntrinsics::_montgomerySquare:
547 if (!UseMontgomerySquareIntrinsic) return NULL;
548 break;
549
550 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
551 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
552 if (!UseAESIntrinsics) return NULL;
553 // these two require the predicated logic
554 predicates = 1;
555 break;
556
557 case vmIntrinsics::_sha_implCompress:
558 if (!UseSHA1Intrinsics) return NULL;
559 break;
560
561 case vmIntrinsics::_sha2_implCompress:
562 if (!UseSHA256Intrinsics) return NULL;
563 break;
564
565 case vmIntrinsics::_sha5_implCompress:
566 if (!UseSHA512Intrinsics) return NULL;
567 break;
568
569 case vmIntrinsics::_digestBase_implCompressMB:
570 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
571 predicates = 3;
572 break;
573
574 case vmIntrinsics::_updateCRC32:
575 case vmIntrinsics::_updateBytesCRC32:
576 case vmIntrinsics::_updateByteBufferCRC32:
577 if (!UseCRC32Intrinsics) return NULL;
578 break;
579
580 case vmIntrinsics::_incrementExactI:
581 case vmIntrinsics::_addExactI:
582 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
583 break;
584 case vmIntrinsics::_incrementExactL:
585 case vmIntrinsics::_addExactL:
586 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
587 break;
588 case vmIntrinsics::_decrementExactI:
589 case vmIntrinsics::_subtractExactI:
590 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
591 break;
592 case vmIntrinsics::_decrementExactL:
593 case vmIntrinsics::_subtractExactL:
594 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
595 break;
596 case vmIntrinsics::_negateExactI:
597 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
598 break;
599 case vmIntrinsics::_negateExactL:
600 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
601 break;
602 case vmIntrinsics::_multiplyExactI:
603 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
604 break;
605 case vmIntrinsics::_multiplyExactL:
606 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
607 break;
608
609 default:
610 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
611 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
612 break;
613 }
614
615 // -XX:-InlineClassNatives disables natives from the Class class.
616 // The flag applies to all reflective calls, notably Array.newArray
617 // (visible to Java programmers as Array.newInstance).
618 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
619 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
620 if (!InlineClassNatives) return NULL;
621 }
622
623 // -XX:-InlineThreadNatives disables natives from the Thread class.
624 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
625 if (!InlineThreadNatives) return NULL;
626 }
627
628 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
629 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
630 m->holder()->name() == ciSymbol::java_lang_Float() ||
631 m->holder()->name() == ciSymbol::java_lang_Double()) {
632 if (!InlineMathNatives) return NULL;
633 }
634
635 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
636 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
637 if (!InlineUnsafeOps) return NULL;
638 }
639
640 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
641 }
642
643 //----------------------register_library_intrinsics-----------------------
644 // Initialize this file's data structures, for each Compile instance.
645 void Compile::register_library_intrinsics() {
646 // Nothing to do here.
647 }
648
649 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
650 LibraryCallKit kit(jvms, this);
651 Compile* C = kit.C;
652 int nodes = C->unique();
653 #ifndef PRODUCT
654 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
655 char buf[1000];
656 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
657 tty->print_cr("Intrinsic %s", str);
658 }
659 #endif
660 ciMethod* callee = kit.callee();
661 const int bci = kit.bci();
662
663 // Try to inline the intrinsic.
664 if (kit.try_to_inline(_last_predicate)) {
665 if (C->print_intrinsics() || C->print_inlining()) {
666 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
667 }
668 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
669 if (C->log()) {
670 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
671 vmIntrinsics::name_at(intrinsic_id()),
672 (is_virtual() ? " virtual='1'" : ""),
673 C->unique() - nodes);
674 }
675 // Push the result from the inlined method onto the stack.
676 kit.push_result();
677 return kit.transfer_exceptions_into_jvms();
678 }
679
680 // The intrinsic bailed out
681 if (C->print_intrinsics() || C->print_inlining()) {
682 if (jvms->has_method()) {
683 // Not a root compile.
684 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
685 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
686 } else {
687 // Root compile
688 tty->print("Did not generate intrinsic %s%s at bci:%d in",
689 vmIntrinsics::name_at(intrinsic_id()),
690 (is_virtual() ? " (virtual)" : ""), bci);
691 }
692 }
693 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
694 return NULL;
695 }
696
697 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
698 LibraryCallKit kit(jvms, this);
699 Compile* C = kit.C;
700 int nodes = C->unique();
701 _last_predicate = predicate;
702 #ifndef PRODUCT
703 assert(is_predicated() && predicate < predicates_count(), "sanity");
704 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
705 char buf[1000];
706 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
707 tty->print_cr("Predicate for intrinsic %s", str);
708 }
709 #endif
710 ciMethod* callee = kit.callee();
711 const int bci = kit.bci();
712
713 Node* slow_ctl = kit.try_to_predicate(predicate);
714 if (!kit.failing()) {
715 if (C->print_intrinsics() || C->print_inlining()) {
716 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
717 }
718 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
719 if (C->log()) {
720 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
721 vmIntrinsics::name_at(intrinsic_id()),
722 (is_virtual() ? " virtual='1'" : ""),
723 C->unique() - nodes);
724 }
725 return slow_ctl; // Could be NULL if the check folds.
726 }
727
728 // The intrinsic bailed out
729 if (C->print_intrinsics() || C->print_inlining()) {
730 if (jvms->has_method()) {
731 // Not a root compile.
732 const char* msg = "failed to generate predicate for intrinsic";
733 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
734 } else {
735 // Root compile
736 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
737 vmIntrinsics::name_at(intrinsic_id()),
738 (is_virtual() ? " (virtual)" : ""), bci);
739 }
740 }
741 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
742 return NULL;
743 }
744
745 bool LibraryCallKit::try_to_inline(int predicate) {
746 // Handle symbolic names for otherwise undistinguished boolean switches:
747 const bool is_store = true;
748 const bool is_native_ptr = true;
749 const bool is_static = true;
750 const bool is_volatile = true;
751
752 if (!jvms()->has_method()) {
753 // Root JVMState has a null method.
754 assert(map()->memory()->Opcode() == Op_Parm, "");
755 // Insert the memory aliasing node
756 set_all_memory(reset_memory());
757 }
758 assert(merged_memory(), "");
759
760
761 switch (intrinsic_id()) {
762 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
763 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
764 case vmIntrinsics::_getClass: return inline_native_getClass();
765
766 case vmIntrinsics::_dsin:
767 case vmIntrinsics::_dcos:
768 case vmIntrinsics::_dtan:
769 case vmIntrinsics::_dabs:
770 case vmIntrinsics::_datan2:
771 case vmIntrinsics::_dsqrt:
772 case vmIntrinsics::_dexp:
773 case vmIntrinsics::_dlog:
774 case vmIntrinsics::_dlog10:
775 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
776
777 case vmIntrinsics::_min:
778 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
779
780 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
781 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
782 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
783 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
784 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
785 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
786 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
787 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
788 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
789 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
790 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
791 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
792
793 case vmIntrinsics::_arraycopy: return inline_arraycopy();
794
795 case vmIntrinsics::_compareTo: return inline_string_compareTo();
796 case vmIntrinsics::_indexOf: return inline_string_indexOf();
797 case vmIntrinsics::_equals: return inline_string_equals();
798
799 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
800 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
801 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
802 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
803 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
804 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
805 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
806 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
807 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
808
809 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
810 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
811 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
812 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
813 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
814 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
815 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
816 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
817 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
818
819 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
820 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
821 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
822 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
823 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
824 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
825 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
826 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
827
828 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
829 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
830 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
831 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
832 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
833 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
834 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
835 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
836
837 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
838 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
839 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
840 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
841 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
842 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
843 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
844 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
845 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
846
847 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
848 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
849 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
850 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
851 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
852 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
853 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
854 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
855 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
856
857 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
858 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
859 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
860 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
861
862 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
863 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
864 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
865
866 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
867 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
868 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
869
870 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
871 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
872 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
873 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
874 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
875
876 case vmIntrinsics::_loadFence:
877 case vmIntrinsics::_storeFence:
878 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
879
880 case vmIntrinsics::_currentThread: return inline_native_currentThread();
881 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
882
883 #ifdef JFR_HAVE_INTRINSICS
884 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
885 case vmIntrinsics::_getClassId: return inline_native_classID();
886 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
887 #endif
888 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
889 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
890 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
891 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
892 case vmIntrinsics::_newArray: return inline_native_newArray();
893 case vmIntrinsics::_getLength: return inline_native_getLength();
894 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
895 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
896 case vmIntrinsics::_equalsC: return inline_array_equals();
897 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
898
899 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
900
901 case vmIntrinsics::_isInstance:
902 case vmIntrinsics::_getModifiers:
903 case vmIntrinsics::_isInterface:
904 case vmIntrinsics::_isArray:
905 case vmIntrinsics::_isPrimitive:
906 case vmIntrinsics::_getSuperclass:
907 case vmIntrinsics::_getComponentType:
908 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
909
910 case vmIntrinsics::_floatToRawIntBits:
911 case vmIntrinsics::_floatToIntBits:
912 case vmIntrinsics::_intBitsToFloat:
913 case vmIntrinsics::_doubleToRawLongBits:
914 case vmIntrinsics::_doubleToLongBits:
915 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
916
917 case vmIntrinsics::_numberOfLeadingZeros_i:
918 case vmIntrinsics::_numberOfLeadingZeros_l:
919 case vmIntrinsics::_numberOfTrailingZeros_i:
920 case vmIntrinsics::_numberOfTrailingZeros_l:
921 case vmIntrinsics::_bitCount_i:
922 case vmIntrinsics::_bitCount_l:
923 case vmIntrinsics::_reverseBytes_i:
924 case vmIntrinsics::_reverseBytes_l:
925 case vmIntrinsics::_reverseBytes_s:
926 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
927
928 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
929
930 case vmIntrinsics::_Reference_get: return inline_reference_get();
931
932 case vmIntrinsics::_aescrypt_encryptBlock:
933 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
934
935 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
936 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
937 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
938
939 case vmIntrinsics::_sha_implCompress:
940 case vmIntrinsics::_sha2_implCompress:
941 case vmIntrinsics::_sha5_implCompress:
942 return inline_sha_implCompress(intrinsic_id());
943
944 case vmIntrinsics::_digestBase_implCompressMB:
945 return inline_digestBase_implCompressMB(predicate);
946
947 case vmIntrinsics::_multiplyToLen:
948 return inline_multiplyToLen();
949
950 case vmIntrinsics::_squareToLen:
951 return inline_squareToLen();
952
953 case vmIntrinsics::_mulAdd:
954 return inline_mulAdd();
955
956 case vmIntrinsics::_montgomeryMultiply:
957 return inline_montgomeryMultiply();
958 case vmIntrinsics::_montgomerySquare:
959 return inline_montgomerySquare();
960
961 case vmIntrinsics::_encodeISOArray:
962 return inline_encodeISOArray();
963
964 case vmIntrinsics::_updateCRC32:
965 return inline_updateCRC32();
966 case vmIntrinsics::_updateBytesCRC32:
967 return inline_updateBytesCRC32();
968 case vmIntrinsics::_updateByteBufferCRC32:
969 return inline_updateByteBufferCRC32();
970
971 case vmIntrinsics::_profileBoolean:
972 return inline_profileBoolean();
973
974 default:
975 // If you get here, it may be that someone has added a new intrinsic
976 // to the list in vmSymbols.hpp without implementing it here.
977 #ifndef PRODUCT
978 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
979 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
980 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
981 }
982 #endif
983 return false;
984 }
985 }
986
987 Node* LibraryCallKit::try_to_predicate(int predicate) {
988 if (!jvms()->has_method()) {
989 // Root JVMState has a null method.
990 assert(map()->memory()->Opcode() == Op_Parm, "");
991 // Insert the memory aliasing node
992 set_all_memory(reset_memory());
993 }
994 assert(merged_memory(), "");
995
996 switch (intrinsic_id()) {
997 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
998 return inline_cipherBlockChaining_AESCrypt_predicate(false);
999 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1000 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1001 case vmIntrinsics::_digestBase_implCompressMB:
1002 return inline_digestBase_implCompressMB_predicate(predicate);
1003
1004 default:
1005 // If you get here, it may be that someone has added a new intrinsic
1006 // to the list in vmSymbols.hpp without implementing it here.
1007 #ifndef PRODUCT
1008 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1009 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1010 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1011 }
1012 #endif
1013 Node* slow_ctl = control();
1014 set_control(top()); // No fast path instrinsic
1015 return slow_ctl;
1016 }
1017 }
1018
1019 //------------------------------set_result-------------------------------
1020 // Helper function for finishing intrinsics.
1021 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1022 record_for_igvn(region);
1023 set_control(_gvn.transform(region));
1024 set_result( _gvn.transform(value));
1025 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1026 }
1027
1028 //------------------------------generate_guard---------------------------
1029 // Helper function for generating guarded fast-slow graph structures.
1030 // The given 'test', if true, guards a slow path. If the test fails
1031 // then a fast path can be taken. (We generally hope it fails.)
1032 // In all cases, GraphKit::control() is updated to the fast path.
1033 // The returned value represents the control for the slow path.
1034 // The return value is never 'top'; it is either a valid control
1035 // or NULL if it is obvious that the slow path can never be taken.
1036 // Also, if region and the slow control are not NULL, the slow edge
1037 // is appended to the region.
1038 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1039 if (stopped()) {
1040 // Already short circuited.
1041 return NULL;
1042 }
1043
1044 // Build an if node and its projections.
1045 // If test is true we take the slow path, which we assume is uncommon.
1046 if (_gvn.type(test) == TypeInt::ZERO) {
1047 // The slow branch is never taken. No need to build this guard.
1048 return NULL;
1049 }
1050
1051 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1052
1053 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1054 if (if_slow == top()) {
1055 // The slow branch is never taken. No need to build this guard.
1056 return NULL;
1057 }
1058
1059 if (region != NULL)
1060 region->add_req(if_slow);
1061
1062 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1063 set_control(if_fast);
1064
1065 return if_slow;
1066 }
1067
1068 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1069 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1070 }
1071 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1072 return generate_guard(test, region, PROB_FAIR);
1073 }
1074
1075 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1076 Node* *pos_index) {
1077 if (stopped())
1078 return NULL; // already stopped
1079 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1080 return NULL; // index is already adequately typed
1081 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1082 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1083 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1084 if (is_neg != NULL && pos_index != NULL) {
1085 // Emulate effect of Parse::adjust_map_after_if.
1086 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1087 ccast->set_req(0, control());
1088 (*pos_index) = _gvn.transform(ccast);
1089 }
1090 return is_neg;
1091 }
1092
1093 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1094 Node* *pos_index) {
1095 if (stopped())
1096 return NULL; // already stopped
1097 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1098 return NULL; // index is already adequately typed
1099 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1100 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1101 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1102 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1103 if (is_notp != NULL && pos_index != NULL) {
1104 // Emulate effect of Parse::adjust_map_after_if.
1105 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1106 ccast->set_req(0, control());
1107 (*pos_index) = _gvn.transform(ccast);
1108 }
1109 return is_notp;
1110 }
1111
1112 // Make sure that 'position' is a valid limit index, in [0..length].
1113 // There are two equivalent plans for checking this:
1114 // A. (offset + copyLength) unsigned<= arrayLength
1115 // B. offset <= (arrayLength - copyLength)
1116 // We require that all of the values above, except for the sum and
1117 // difference, are already known to be non-negative.
1118 // Plan A is robust in the face of overflow, if offset and copyLength
1119 // are both hugely positive.
1120 //
1121 // Plan B is less direct and intuitive, but it does not overflow at
1122 // all, since the difference of two non-negatives is always
1123 // representable. Whenever Java methods must perform the equivalent
1124 // check they generally use Plan B instead of Plan A.
1125 // For the moment we use Plan A.
1126 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1127 Node* subseq_length,
1128 Node* array_length,
1129 RegionNode* region) {
1130 if (stopped())
1131 return NULL; // already stopped
1132 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1133 if (zero_offset && subseq_length->eqv_uncast(array_length))
1134 return NULL; // common case of whole-array copy
1135 Node* last = subseq_length;
1136 if (!zero_offset) // last += offset
1137 last = _gvn.transform(new (C) AddINode(last, offset));
1138 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1139 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1140 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1141 return is_over;
1142 }
1143
1144
1145 //--------------------------generate_current_thread--------------------
1146 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1147 ciKlass* thread_klass = env()->Thread_klass();
1148 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1149 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1150 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1151 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1152 tls_output = thread;
1153 return threadObj;
1154 }
1155
1156
1157 //------------------------------make_string_method_node------------------------
1158 // Helper method for String intrinsic functions. This version is called
1159 // with str1 and str2 pointing to String object nodes.
1160 //
1161 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1162 Node* no_ctrl = NULL;
1163
1164 // Get start addr of string
1165 Node* str1_value = load_String_value(no_ctrl, str1);
1166 Node* str1_offset = load_String_offset(no_ctrl, str1);
1167 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1168
1169 // Get length of string 1
1170 Node* str1_len = load_String_length(no_ctrl, str1);
1171
1172 Node* str2_value = load_String_value(no_ctrl, str2);
1173 Node* str2_offset = load_String_offset(no_ctrl, str2);
1174 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1175
1176 Node* str2_len = NULL;
1177 Node* result = NULL;
1178
1179 switch (opcode) {
1180 case Op_StrIndexOf:
1181 // Get length of string 2
1182 str2_len = load_String_length(no_ctrl, str2);
1183
1184 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1185 str1_start, str1_len, str2_start, str2_len);
1186 break;
1187 case Op_StrComp:
1188 // Get length of string 2
1189 str2_len = load_String_length(no_ctrl, str2);
1190
1191 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1192 str1_start, str1_len, str2_start, str2_len);
1193 break;
1194 case Op_StrEquals:
1195 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1196 str1_start, str2_start, str1_len);
1197 break;
1198 default:
1199 ShouldNotReachHere();
1200 return NULL;
1201 }
1202
1203 // All these intrinsics have checks.
1204 C->set_has_split_ifs(true); // Has chance for split-if optimization
1205
1206 return _gvn.transform(result);
1207 }
1208
1209 // Helper method for String intrinsic functions. This version is called
1210 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1211 // to Int nodes containing the lenghts of str1 and str2.
1212 //
1213 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1214 Node* result = NULL;
1215 switch (opcode) {
1216 case Op_StrIndexOf:
1217 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1218 str1_start, cnt1, str2_start, cnt2);
1219 break;
1220 case Op_StrComp:
1221 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1222 str1_start, cnt1, str2_start, cnt2);
1223 break;
1224 case Op_StrEquals:
1225 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1226 str1_start, str2_start, cnt1);
1227 break;
1228 default:
1229 ShouldNotReachHere();
1230 return NULL;
1231 }
1232
1233 // All these intrinsics have checks.
1234 C->set_has_split_ifs(true); // Has chance for split-if optimization
1235
1236 return _gvn.transform(result);
1237 }
1238
1239 //------------------------------inline_string_compareTo------------------------
1240 // public int java.lang.String.compareTo(String anotherString);
1241 bool LibraryCallKit::inline_string_compareTo() {
1242 Node* receiver = null_check(argument(0));
1243 Node* arg = null_check(argument(1));
1244 if (stopped()) {
1245 return true;
1246 }
1247 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1248 return true;
1249 }
1250
1251 //------------------------------inline_string_equals------------------------
1252 bool LibraryCallKit::inline_string_equals() {
1253 Node* receiver = null_check_receiver();
1254 // NOTE: Do not null check argument for String.equals() because spec
1255 // allows to specify NULL as argument.
1256 Node* argument = this->argument(1);
1257 if (stopped()) {
1258 return true;
1259 }
1260
1261 // paths (plus control) merge
1262 RegionNode* region = new (C) RegionNode(5);
1263 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1264
1265 // does source == target string?
1266 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1267 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1268
1269 Node* if_eq = generate_slow_guard(bol, NULL);
1270 if (if_eq != NULL) {
1271 // receiver == argument
1272 phi->init_req(2, intcon(1));
1273 region->init_req(2, if_eq);
1274 }
1275
1276 // get String klass for instanceOf
1277 ciInstanceKlass* klass = env()->String_klass();
1278
1279 if (!stopped()) {
1280 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1281 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1282 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1283
1284 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1285 //instanceOf == true, fallthrough
1286
1287 if (inst_false != NULL) {
1288 phi->init_req(3, intcon(0));
1289 region->init_req(3, inst_false);
1290 }
1291 }
1292
1293 if (!stopped()) {
1294 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1295
1296 // Properly cast the argument to String
1297 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1298 // This path is taken only when argument's type is String:NotNull.
1299 argument = cast_not_null(argument, false);
1300
1301 Node* no_ctrl = NULL;
1302
1303 // Get start addr of receiver
1304 Node* receiver_val = load_String_value(no_ctrl, receiver);
1305 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1306 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1307
1308 // Get length of receiver
1309 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1310
1311 // Get start addr of argument
1312 Node* argument_val = load_String_value(no_ctrl, argument);
1313 Node* argument_offset = load_String_offset(no_ctrl, argument);
1314 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1315
1316 // Get length of argument
1317 Node* argument_cnt = load_String_length(no_ctrl, argument);
1318
1319 // Check for receiver count != argument count
1320 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1321 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1322 Node* if_ne = generate_slow_guard(bol, NULL);
1323 if (if_ne != NULL) {
1324 phi->init_req(4, intcon(0));
1325 region->init_req(4, if_ne);
1326 }
1327
1328 // Check for count == 0 is done by assembler code for StrEquals.
1329
1330 if (!stopped()) {
1331 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1332 phi->init_req(1, equals);
1333 region->init_req(1, control());
1334 }
1335 }
1336
1337 // post merge
1338 set_control(_gvn.transform(region));
1339 record_for_igvn(region);
1340
1341 set_result(_gvn.transform(phi));
1342 return true;
1343 }
1344
1345 //------------------------------inline_array_equals----------------------------
1346 bool LibraryCallKit::inline_array_equals() {
1347 Node* arg1 = argument(0);
1348 Node* arg2 = argument(1);
1349 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1350 return true;
1351 }
1352
1353 // Java version of String.indexOf(constant string)
1354 // class StringDecl {
1355 // StringDecl(char[] ca) {
1356 // offset = 0;
1357 // count = ca.length;
1358 // value = ca;
1359 // }
1360 // int offset;
1361 // int count;
1362 // char[] value;
1363 // }
1364 //
1365 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1366 // int targetOffset, int cache_i, int md2) {
1367 // int cache = cache_i;
1368 // int sourceOffset = string_object.offset;
1369 // int sourceCount = string_object.count;
1370 // int targetCount = target_object.length;
1371 //
1372 // int targetCountLess1 = targetCount - 1;
1373 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1374 //
1375 // char[] source = string_object.value;
1376 // char[] target = target_object;
1377 // int lastChar = target[targetCountLess1];
1378 //
1379 // outer_loop:
1380 // for (int i = sourceOffset; i < sourceEnd; ) {
1381 // int src = source[i + targetCountLess1];
1382 // if (src == lastChar) {
1383 // // With random strings and a 4-character alphabet,
1384 // // reverse matching at this point sets up 0.8% fewer
1385 // // frames, but (paradoxically) makes 0.3% more probes.
1386 // // Since those probes are nearer the lastChar probe,
1387 // // there is may be a net D$ win with reverse matching.
1388 // // But, reversing loop inhibits unroll of inner loop
1389 // // for unknown reason. So, does running outer loop from
1390 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1391 // for (int j = 0; j < targetCountLess1; j++) {
1392 // if (target[targetOffset + j] != source[i+j]) {
1393 // if ((cache & (1 << source[i+j])) == 0) {
1394 // if (md2 < j+1) {
1395 // i += j+1;
1396 // continue outer_loop;
1397 // }
1398 // }
1399 // i += md2;
1400 // continue outer_loop;
1401 // }
1402 // }
1403 // return i - sourceOffset;
1404 // }
1405 // if ((cache & (1 << src)) == 0) {
1406 // i += targetCountLess1;
1407 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1408 // i++;
1409 // }
1410 // return -1;
1411 // }
1412
1413 //------------------------------string_indexOf------------------------
1414 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1415 jint cache_i, jint md2_i) {
1416
1417 Node* no_ctrl = NULL;
1418 float likely = PROB_LIKELY(0.9);
1419 float unlikely = PROB_UNLIKELY(0.9);
1420
1421 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1422
1423 Node* source = load_String_value(no_ctrl, string_object);
1424 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1425 Node* sourceCount = load_String_length(no_ctrl, string_object);
1426
1427 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1428 jint target_length = target_array->length();
1429 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1430 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1431
1432 // String.value field is known to be @Stable.
1433 if (UseImplicitStableValues) {
1434 target = cast_array_to_stable(target, target_type);
1435 }
1436
1437 IdealKit kit(this, false, true);
1438 #define __ kit.
1439 Node* zero = __ ConI(0);
1440 Node* one = __ ConI(1);
1441 Node* cache = __ ConI(cache_i);
1442 Node* md2 = __ ConI(md2_i);
1443 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1444 Node* targetCount = __ ConI(target_length);
1445 Node* targetCountLess1 = __ ConI(target_length - 1);
1446 Node* targetOffset = __ ConI(targetOffset_i);
1447 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1448
1449 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1450 Node* outer_loop = __ make_label(2 /* goto */);
1451 Node* return_ = __ make_label(1);
1452
1453 __ set(rtn,__ ConI(-1));
1454 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1455 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1456 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1457 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1458 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1459 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1460 Node* tpj = __ AddI(targetOffset, __ value(j));
1461 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1462 Node* ipj = __ AddI(__ value(i), __ value(j));
1463 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1464 __ if_then(targ, BoolTest::ne, src2); {
1465 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1466 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1467 __ increment(i, __ AddI(__ value(j), one));
1468 __ goto_(outer_loop);
1469 } __ end_if(); __ dead(j);
1470 }__ end_if(); __ dead(j);
1471 __ increment(i, md2);
1472 __ goto_(outer_loop);
1473 }__ end_if();
1474 __ increment(j, one);
1475 }__ end_loop(); __ dead(j);
1476 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1477 __ goto_(return_);
1478 }__ end_if();
1479 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1480 __ increment(i, targetCountLess1);
1481 }__ end_if();
1482 __ increment(i, one);
1483 __ bind(outer_loop);
1484 }__ end_loop(); __ dead(i);
1485 __ bind(return_);
1486
1487 // Final sync IdealKit and GraphKit.
1488 final_sync(kit);
1489 Node* result = __ value(rtn);
1490 #undef __
1491 C->set_has_loops(true);
1492 return result;
1493 }
1494
1495 //------------------------------inline_string_indexOf------------------------
1496 bool LibraryCallKit::inline_string_indexOf() {
1497 Node* receiver = argument(0);
1498 Node* arg = argument(1);
1499
1500 Node* result;
1501 // Disable the use of pcmpestri until it can be guaranteed that
1502 // the load doesn't cross into the uncommited space.
1503 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1504 UseSSE42Intrinsics) {
1505 // Generate SSE4.2 version of indexOf
1506 // We currently only have match rules that use SSE4.2
1507
1508 receiver = null_check(receiver);
1509 arg = null_check(arg);
1510 if (stopped()) {
1511 return true;
1512 }
1513
1514 ciInstanceKlass* str_klass = env()->String_klass();
1515 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1516
1517 // Make the merge point
1518 RegionNode* result_rgn = new (C) RegionNode(4);
1519 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1520 Node* no_ctrl = NULL;
1521
1522 // Get start addr of source string
1523 Node* source = load_String_value(no_ctrl, receiver);
1524 Node* source_offset = load_String_offset(no_ctrl, receiver);
1525 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1526
1527 // Get length of source string
1528 Node* source_cnt = load_String_length(no_ctrl, receiver);
1529
1530 // Get start addr of substring
1531 Node* substr = load_String_value(no_ctrl, arg);
1532 Node* substr_offset = load_String_offset(no_ctrl, arg);
1533 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1534
1535 // Get length of source string
1536 Node* substr_cnt = load_String_length(no_ctrl, arg);
1537
1538 // Check for substr count > string count
1539 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1540 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1541 Node* if_gt = generate_slow_guard(bol, NULL);
1542 if (if_gt != NULL) {
1543 result_phi->init_req(2, intcon(-1));
1544 result_rgn->init_req(2, if_gt);
1545 }
1546
1547 if (!stopped()) {
1548 // Check for substr count == 0
1549 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1550 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1551 Node* if_zero = generate_slow_guard(bol, NULL);
1552 if (if_zero != NULL) {
1553 result_phi->init_req(3, intcon(0));
1554 result_rgn->init_req(3, if_zero);
1555 }
1556 }
1557
1558 if (!stopped()) {
1559 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1560 result_phi->init_req(1, result);
1561 result_rgn->init_req(1, control());
1562 }
1563 set_control(_gvn.transform(result_rgn));
1564 record_for_igvn(result_rgn);
1565 result = _gvn.transform(result_phi);
1566
1567 } else { // Use LibraryCallKit::string_indexOf
1568 // don't intrinsify if argument isn't a constant string.
1569 if (!arg->is_Con()) {
1570 return false;
1571 }
1572 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1573 if (str_type == NULL) {
1574 return false;
1575 }
1576 ciInstanceKlass* klass = env()->String_klass();
1577 ciObject* str_const = str_type->const_oop();
1578 if (str_const == NULL || str_const->klass() != klass) {
1579 return false;
1580 }
1581 ciInstance* str = str_const->as_instance();
1582 assert(str != NULL, "must be instance");
1583
1584 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1585 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1586
1587 int o;
1588 int c;
1589 if (java_lang_String::has_offset_field()) {
1590 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1591 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1592 } else {
1593 o = 0;
1594 c = pat->length();
1595 }
1596
1597 // constant strings have no offset and count == length which
1598 // simplifies the resulting code somewhat so lets optimize for that.
1599 if (o != 0 || c != pat->length()) {
1600 return false;
1601 }
1602
1603 receiver = null_check(receiver, T_OBJECT);
1604 // NOTE: No null check on the argument is needed since it's a constant String oop.
1605 if (stopped()) {
1606 return true;
1607 }
1608
1609 // The null string as a pattern always returns 0 (match at beginning of string)
1610 if (c == 0) {
1611 set_result(intcon(0));
1612 return true;
1613 }
1614
1615 // Generate default indexOf
1616 jchar lastChar = pat->char_at(o + (c - 1));
1617 int cache = 0;
1618 int i;
1619 for (i = 0; i < c - 1; i++) {
1620 assert(i < pat->length(), "out of range");
1621 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1622 }
1623
1624 int md2 = c;
1625 for (i = 0; i < c - 1; i++) {
1626 assert(i < pat->length(), "out of range");
1627 if (pat->char_at(o + i) == lastChar) {
1628 md2 = (c - 1) - i;
1629 }
1630 }
1631
1632 result = string_indexOf(receiver, pat, o, cache, md2);
1633 }
1634 set_result(result);
1635 return true;
1636 }
1637
1638 //--------------------------round_double_node--------------------------------
1639 // Round a double node if necessary.
1640 Node* LibraryCallKit::round_double_node(Node* n) {
1641 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1642 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1643 return n;
1644 }
1645
1646 //------------------------------inline_math-----------------------------------
1647 // public static double Math.abs(double)
1648 // public static double Math.sqrt(double)
1649 // public static double Math.log(double)
1650 // public static double Math.log10(double)
1651 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1652 Node* arg = round_double_node(argument(0));
1653 Node* n = NULL;
1654 switch (id) {
1655 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1656 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1657 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1658 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1659 default: fatal_unexpected_iid(id); break;
1660 }
1661 set_result(_gvn.transform(n));
1662 return true;
1663 }
1664
1665 //------------------------------inline_trig----------------------------------
1666 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1667 // argument reduction which will turn into a fast/slow diamond.
1668 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1669 Node* arg = round_double_node(argument(0));
1670 Node* n = NULL;
1671
1672 switch (id) {
1673 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1674 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1675 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1676 default: fatal_unexpected_iid(id); break;
1677 }
1678 n = _gvn.transform(n);
1679
1680 // Rounding required? Check for argument reduction!
1681 if (Matcher::strict_fp_requires_explicit_rounding) {
1682 static const double pi_4 = 0.7853981633974483;
1683 static const double neg_pi_4 = -0.7853981633974483;
1684 // pi/2 in 80-bit extended precision
1685 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1686 // -pi/2 in 80-bit extended precision
1687 // 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};
1688 // Cutoff value for using this argument reduction technique
1689 //static const double pi_2_minus_epsilon = 1.564660403643354;
1690 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1691
1692 // Pseudocode for sin:
1693 // if (x <= Math.PI / 4.0) {
1694 // if (x >= -Math.PI / 4.0) return fsin(x);
1695 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1696 // } else {
1697 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1698 // }
1699 // return StrictMath.sin(x);
1700
1701 // Pseudocode for cos:
1702 // if (x <= Math.PI / 4.0) {
1703 // if (x >= -Math.PI / 4.0) return fcos(x);
1704 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1705 // } else {
1706 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1707 // }
1708 // return StrictMath.cos(x);
1709
1710 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1711 // requires a special machine instruction to load it. Instead we'll try
1712 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1713 // probably do the math inside the SIN encoding.
1714
1715 // Make the merge point
1716 RegionNode* r = new (C) RegionNode(3);
1717 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1718
1719 // Flatten arg so we need only 1 test
1720 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1721 // Node for PI/4 constant
1722 Node *pi4 = makecon(TypeD::make(pi_4));
1723 // Check PI/4 : abs(arg)
1724 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1725 // Check: If PI/4 < abs(arg) then go slow
1726 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1727 // Branch either way
1728 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1729 set_control(opt_iff(r,iff));
1730
1731 // Set fast path result
1732 phi->init_req(2, n);
1733
1734 // Slow path - non-blocking leaf call
1735 Node* call = NULL;
1736 switch (id) {
1737 case vmIntrinsics::_dsin:
1738 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1739 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1740 "Sin", NULL, arg, top());
1741 break;
1742 case vmIntrinsics::_dcos:
1743 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1744 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1745 "Cos", NULL, arg, top());
1746 break;
1747 case vmIntrinsics::_dtan:
1748 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1749 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1750 "Tan", NULL, arg, top());
1751 break;
1752 }
1753 assert(control()->in(0) == call, "");
1754 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1755 r->init_req(1, control());
1756 phi->init_req(1, slow_result);
1757
1758 // Post-merge
1759 set_control(_gvn.transform(r));
1760 record_for_igvn(r);
1761 n = _gvn.transform(phi);
1762
1763 C->set_has_split_ifs(true); // Has chance for split-if optimization
1764 }
1765 set_result(n);
1766 return true;
1767 }
1768
1769 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1770 //-------------------
1771 //result=(result.isNaN())? funcAddr():result;
1772 // Check: If isNaN() by checking result!=result? then either trap
1773 // or go to runtime
1774 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1775 // Build the boolean node
1776 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1777
1778 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1779 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1780 // The pow or exp intrinsic returned a NaN, which requires a call
1781 // to the runtime. Recompile with the runtime call.
1782 uncommon_trap(Deoptimization::Reason_intrinsic,
1783 Deoptimization::Action_make_not_entrant);
1784 }
1785 return result;
1786 } else {
1787 // If this inlining ever returned NaN in the past, we compile a call
1788 // to the runtime to properly handle corner cases
1789
1790 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1791 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1792 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1793
1794 if (!if_slow->is_top()) {
1795 RegionNode* result_region = new (C) RegionNode(3);
1796 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1797
1798 result_region->init_req(1, if_fast);
1799 result_val->init_req(1, result);
1800
1801 set_control(if_slow);
1802
1803 const TypePtr* no_memory_effects = NULL;
1804 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1805 no_memory_effects,
1806 x, top(), y, y ? top() : NULL);
1807 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1808 #ifdef ASSERT
1809 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1810 assert(value_top == top(), "second value must be top");
1811 #endif
1812
1813 result_region->init_req(2, control());
1814 result_val->init_req(2, value);
1815 set_control(_gvn.transform(result_region));
1816 return _gvn.transform(result_val);
1817 } else {
1818 return result;
1819 }
1820 }
1821 }
1822
1823 //------------------------------inline_exp-------------------------------------
1824 // Inline exp instructions, if possible. The Intel hardware only misses
1825 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1826 bool LibraryCallKit::inline_exp() {
1827 Node* arg = round_double_node(argument(0));
1828 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1829
1830 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1831 set_result(n);
1832
1833 C->set_has_split_ifs(true); // Has chance for split-if optimization
1834 return true;
1835 }
1836
1837 //------------------------------inline_pow-------------------------------------
1838 // Inline power instructions, if possible.
1839 bool LibraryCallKit::inline_pow() {
1840 // Pseudocode for pow
1841 // if (y == 2) {
1842 // return x * x;
1843 // } else {
1844 // if (x <= 0.0) {
1845 // long longy = (long)y;
1846 // if ((double)longy == y) { // if y is long
1847 // if (y + 1 == y) longy = 0; // huge number: even
1848 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1849 // } else {
1850 // result = NaN;
1851 // }
1852 // } else {
1853 // result = DPow(x,y);
1854 // }
1855 // if (result != result)? {
1856 // result = uncommon_trap() or runtime_call();
1857 // }
1858 // return result;
1859 // }
1860
1861 Node* x = round_double_node(argument(0));
1862 Node* y = round_double_node(argument(2));
1863
1864 Node* result = NULL;
1865
1866 Node* const_two_node = makecon(TypeD::make(2.0));
1867 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1868 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1869 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1870 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1871 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1872
1873 RegionNode* region_node = new (C) RegionNode(3);
1874 region_node->init_req(1, if_true);
1875
1876 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1877 // special case for x^y where y == 2, we can convert it to x * x
1878 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1879
1880 // set control to if_false since we will now process the false branch
1881 set_control(if_false);
1882
1883 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1884 // Short form: skip the fancy tests and just check for NaN result.
1885 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1886 } else {
1887 // If this inlining ever returned NaN in the past, include all
1888 // checks + call to the runtime.
1889
1890 // Set the merge point for If node with condition of (x <= 0.0)
1891 // There are four possible paths to region node and phi node
1892 RegionNode *r = new (C) RegionNode(4);
1893 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1894
1895 // Build the first if node: if (x <= 0.0)
1896 // Node for 0 constant
1897 Node *zeronode = makecon(TypeD::ZERO);
1898 // Check x:0
1899 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1900 // Check: If (x<=0) then go complex path
1901 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1902 // Branch either way
1903 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1904 // Fast path taken; set region slot 3
1905 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1906 r->init_req(3,fast_taken); // Capture fast-control
1907
1908 // Fast path not-taken, i.e. slow path
1909 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1910
1911 // Set fast path result
1912 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1913 phi->init_req(3, fast_result);
1914
1915 // Complex path
1916 // Build the second if node (if y is long)
1917 // Node for (long)y
1918 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1919 // Node for (double)((long) y)
1920 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1921 // Check (double)((long) y) : y
1922 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1923 // Check if (y isn't long) then go to slow path
1924
1925 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1926 // Branch either way
1927 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1928 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1929
1930 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1931
1932 // Calculate DPow(abs(x), y)*(1 & (long)y)
1933 // Node for constant 1
1934 Node *conone = longcon(1);
1935 // 1& (long)y
1936 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1937
1938 // A huge number is always even. Detect a huge number by checking
1939 // if y + 1 == y and set integer to be tested for parity to 0.
1940 // Required for corner case:
1941 // (long)9.223372036854776E18 = max_jlong
1942 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1943 // max_jlong is odd but 9.223372036854776E18 is even
1944 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1945 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1946 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1947 Node* correctedsign = NULL;
1948 if (ConditionalMoveLimit != 0) {
1949 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1950 } else {
1951 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1952 RegionNode *r = new (C) RegionNode(3);
1953 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1954 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1955 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1956 phi->init_req(1, signnode);
1957 phi->init_req(2, longcon(0));
1958 correctedsign = _gvn.transform(phi);
1959 ylong_path = _gvn.transform(r);
1960 record_for_igvn(r);
1961 }
1962
1963 // zero node
1964 Node *conzero = longcon(0);
1965 // Check (1&(long)y)==0?
1966 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1967 // Check if (1&(long)y)!=0?, if so the result is negative
1968 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1969 // abs(x)
1970 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1971 // abs(x)^y
1972 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1973 // -abs(x)^y
1974 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1975 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1976 Node *signresult = NULL;
1977 if (ConditionalMoveLimit != 0) {
1978 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1979 } else {
1980 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1981 RegionNode *r = new (C) RegionNode(3);
1982 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1983 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1984 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1985 phi->init_req(1, absxpowy);
1986 phi->init_req(2, negabsxpowy);
1987 signresult = _gvn.transform(phi);
1988 ylong_path = _gvn.transform(r);
1989 record_for_igvn(r);
1990 }
1991 // Set complex path fast result
1992 r->init_req(2, ylong_path);
1993 phi->init_req(2, signresult);
1994
1995 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1996 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1997 r->init_req(1,slow_path);
1998 phi->init_req(1,slow_result);
1999
2000 // Post merge
2001 set_control(_gvn.transform(r));
2002 record_for_igvn(r);
2003 result = _gvn.transform(phi);
2004 }
2005
2006 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2007
2008 // control from finish_pow_exp is now input to the region node
2009 region_node->set_req(2, control());
2010 // the result from finish_pow_exp is now input to the phi node
2011 phi_node->init_req(2, result);
2012 set_control(_gvn.transform(region_node));
2013 record_for_igvn(region_node);
2014 set_result(_gvn.transform(phi_node));
2015
2016 C->set_has_split_ifs(true); // Has chance for split-if optimization
2017 return true;
2018 }
2019
2020 //------------------------------runtime_math-----------------------------
2021 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2022 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2023 "must be (DD)D or (D)D type");
2024
2025 // Inputs
2026 Node* a = round_double_node(argument(0));
2027 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2028
2029 const TypePtr* no_memory_effects = NULL;
2030 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2031 no_memory_effects,
2032 a, top(), b, b ? top() : NULL);
2033 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2034 #ifdef ASSERT
2035 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2036 assert(value_top == top(), "second value must be top");
2037 #endif
2038
2039 set_result(value);
2040 return true;
2041 }
2042
2043 //------------------------------inline_math_native-----------------------------
2044 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2045 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2046 switch (id) {
2047 // These intrinsics are not properly supported on all hardware
2048 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2049 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2050 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2051 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2052 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2053 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2054
2055 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2056 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2057 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2058 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2059
2060 // These intrinsics are supported on all hardware
2061 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2062 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2063
2064 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2065 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2066 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2067 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2068 #undef FN_PTR
2069
2070 // These intrinsics are not yet correctly implemented
2071 case vmIntrinsics::_datan2:
2072 return false;
2073
2074 default:
2075 fatal_unexpected_iid(id);
2076 return false;
2077 }
2078 }
2079
2080 static bool is_simple_name(Node* n) {
2081 return (n->req() == 1 // constant
2082 || (n->is_Type() && n->as_Type()->type()->singleton())
2083 || n->is_Proj() // parameter or return value
2084 || n->is_Phi() // local of some sort
2085 );
2086 }
2087
2088 //----------------------------inline_min_max-----------------------------------
2089 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2090 set_result(generate_min_max(id, argument(0), argument(1)));
2091 return true;
2092 }
2093
2094 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2095 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2096 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2097 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2098 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2099
2100 {
2101 PreserveJVMState pjvms(this);
2102 PreserveReexecuteState preexecs(this);
2103 jvms()->set_should_reexecute(true);
2104
2105 set_control(slow_path);
2106 set_i_o(i_o());
2107
2108 uncommon_trap(Deoptimization::Reason_intrinsic,
2109 Deoptimization::Action_none);
2110 }
2111
2112 set_control(fast_path);
2113 set_result(math);
2114 }
2115
2116 template <typename OverflowOp>
2117 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2118 typedef typename OverflowOp::MathOp MathOp;
2119
2120 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2121 Node* operation = _gvn.transform( mathOp );
2122 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2123 inline_math_mathExact(operation, ofcheck);
2124 return true;
2125 }
2126
2127 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2128 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2129 }
2130
2131 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2132 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2133 }
2134
2135 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2136 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2137 }
2138
2139 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2140 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2141 }
2142
2143 bool LibraryCallKit::inline_math_negateExactI() {
2144 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2145 }
2146
2147 bool LibraryCallKit::inline_math_negateExactL() {
2148 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2149 }
2150
2151 bool LibraryCallKit::inline_math_multiplyExactI() {
2152 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2153 }
2154
2155 bool LibraryCallKit::inline_math_multiplyExactL() {
2156 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2157 }
2158
2159 Node*
2160 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2161 // These are the candidate return value:
2162 Node* xvalue = x0;
2163 Node* yvalue = y0;
2164
2165 if (xvalue == yvalue) {
2166 return xvalue;
2167 }
2168
2169 bool want_max = (id == vmIntrinsics::_max);
2170
2171 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2172 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2173 if (txvalue == NULL || tyvalue == NULL) return top();
2174 // This is not really necessary, but it is consistent with a
2175 // hypothetical MaxINode::Value method:
2176 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2177
2178 // %%% This folding logic should (ideally) be in a different place.
2179 // Some should be inside IfNode, and there to be a more reliable
2180 // transformation of ?: style patterns into cmoves. We also want
2181 // more powerful optimizations around cmove and min/max.
2182
2183 // Try to find a dominating comparison of these guys.
2184 // It can simplify the index computation for Arrays.copyOf
2185 // and similar uses of System.arraycopy.
2186 // First, compute the normalized version of CmpI(x, y).
2187 int cmp_op = Op_CmpI;
2188 Node* xkey = xvalue;
2189 Node* ykey = yvalue;
2190 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2191 if (ideal_cmpxy->is_Cmp()) {
2192 // E.g., if we have CmpI(length - offset, count),
2193 // it might idealize to CmpI(length, count + offset)
2194 cmp_op = ideal_cmpxy->Opcode();
2195 xkey = ideal_cmpxy->in(1);
2196 ykey = ideal_cmpxy->in(2);
2197 }
2198
2199 // Start by locating any relevant comparisons.
2200 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2201 Node* cmpxy = NULL;
2202 Node* cmpyx = NULL;
2203 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2204 Node* cmp = start_from->fast_out(k);
2205 if (cmp->outcnt() > 0 && // must have prior uses
2206 cmp->in(0) == NULL && // must be context-independent
2207 cmp->Opcode() == cmp_op) { // right kind of compare
2208 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2209 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2210 }
2211 }
2212
2213 const int NCMPS = 2;
2214 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2215 int cmpn;
2216 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2217 if (cmps[cmpn] != NULL) break; // find a result
2218 }
2219 if (cmpn < NCMPS) {
2220 // Look for a dominating test that tells us the min and max.
2221 int depth = 0; // Limit search depth for speed
2222 Node* dom = control();
2223 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2224 if (++depth >= 100) break;
2225 Node* ifproj = dom;
2226 if (!ifproj->is_Proj()) continue;
2227 Node* iff = ifproj->in(0);
2228 if (!iff->is_If()) continue;
2229 Node* bol = iff->in(1);
2230 if (!bol->is_Bool()) continue;
2231 Node* cmp = bol->in(1);
2232 if (cmp == NULL) continue;
2233 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2234 if (cmps[cmpn] == cmp) break;
2235 if (cmpn == NCMPS) continue;
2236 BoolTest::mask btest = bol->as_Bool()->_test._test;
2237 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2238 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2239 // At this point, we know that 'x btest y' is true.
2240 switch (btest) {
2241 case BoolTest::eq:
2242 // They are proven equal, so we can collapse the min/max.
2243 // Either value is the answer. Choose the simpler.
2244 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2245 return yvalue;
2246 return xvalue;
2247 case BoolTest::lt: // x < y
2248 case BoolTest::le: // x <= y
2249 return (want_max ? yvalue : xvalue);
2250 case BoolTest::gt: // x > y
2251 case BoolTest::ge: // x >= y
2252 return (want_max ? xvalue : yvalue);
2253 }
2254 }
2255 }
2256
2257 // We failed to find a dominating test.
2258 // Let's pick a test that might GVN with prior tests.
2259 Node* best_bol = NULL;
2260 BoolTest::mask best_btest = BoolTest::illegal;
2261 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2262 Node* cmp = cmps[cmpn];
2263 if (cmp == NULL) continue;
2264 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2265 Node* bol = cmp->fast_out(j);
2266 if (!bol->is_Bool()) continue;
2267 BoolTest::mask btest = bol->as_Bool()->_test._test;
2268 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2269 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2270 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2271 best_bol = bol->as_Bool();
2272 best_btest = btest;
2273 }
2274 }
2275 }
2276
2277 Node* answer_if_true = NULL;
2278 Node* answer_if_false = NULL;
2279 switch (best_btest) {
2280 default:
2281 if (cmpxy == NULL)
2282 cmpxy = ideal_cmpxy;
2283 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2284 // and fall through:
2285 case BoolTest::lt: // x < y
2286 case BoolTest::le: // x <= y
2287 answer_if_true = (want_max ? yvalue : xvalue);
2288 answer_if_false = (want_max ? xvalue : yvalue);
2289 break;
2290 case BoolTest::gt: // x > y
2291 case BoolTest::ge: // x >= y
2292 answer_if_true = (want_max ? xvalue : yvalue);
2293 answer_if_false = (want_max ? yvalue : xvalue);
2294 break;
2295 }
2296
2297 jint hi, lo;
2298 if (want_max) {
2299 // We can sharpen the minimum.
2300 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2301 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2302 } else {
2303 // We can sharpen the maximum.
2304 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2305 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2306 }
2307
2308 // Use a flow-free graph structure, to avoid creating excess control edges
2309 // which could hinder other optimizations.
2310 // Since Math.min/max is often used with arraycopy, we want
2311 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2312 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2313 answer_if_false, answer_if_true,
2314 TypeInt::make(lo, hi, widen));
2315
2316 return _gvn.transform(cmov);
2317
2318 /*
2319 // This is not as desirable as it may seem, since Min and Max
2320 // nodes do not have a full set of optimizations.
2321 // And they would interfere, anyway, with 'if' optimizations
2322 // and with CMoveI canonical forms.
2323 switch (id) {
2324 case vmIntrinsics::_min:
2325 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2326 case vmIntrinsics::_max:
2327 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2328 default:
2329 ShouldNotReachHere();
2330 }
2331 */
2332 }
2333
2334 inline int
2335 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2336 const TypePtr* base_type = TypePtr::NULL_PTR;
2337 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2338 if (base_type == NULL) {
2339 // Unknown type.
2340 return Type::AnyPtr;
2341 } else if (base_type == TypePtr::NULL_PTR) {
2342 // Since this is a NULL+long form, we have to switch to a rawptr.
2343 base = _gvn.transform(new (C) CastX2PNode(offset));
2344 offset = MakeConX(0);
2345 return Type::RawPtr;
2346 } else if (base_type->base() == Type::RawPtr) {
2347 return Type::RawPtr;
2348 } else if (base_type->isa_oopptr()) {
2349 // Base is never null => always a heap address.
2350 if (base_type->ptr() == TypePtr::NotNull) {
2351 return Type::OopPtr;
2352 }
2353 // Offset is small => always a heap address.
2354 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2355 if (offset_type != NULL &&
2356 base_type->offset() == 0 && // (should always be?)
2357 offset_type->_lo >= 0 &&
2358 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2359 return Type::OopPtr;
2360 }
2361 // Otherwise, it might either be oop+off or NULL+addr.
2362 return Type::AnyPtr;
2363 } else {
2364 // No information:
2365 return Type::AnyPtr;
2366 }
2367 }
2368
2369 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2370 int kind = classify_unsafe_addr(base, offset);
2371 if (kind == Type::RawPtr) {
2372 return basic_plus_adr(top(), base, offset);
2373 } else {
2374 return basic_plus_adr(base, offset);
2375 }
2376 }
2377
2378 //--------------------------inline_number_methods-----------------------------
2379 // inline int Integer.numberOfLeadingZeros(int)
2380 // inline int Long.numberOfLeadingZeros(long)
2381 //
2382 // inline int Integer.numberOfTrailingZeros(int)
2383 // inline int Long.numberOfTrailingZeros(long)
2384 //
2385 // inline int Integer.bitCount(int)
2386 // inline int Long.bitCount(long)
2387 //
2388 // inline char Character.reverseBytes(char)
2389 // inline short Short.reverseBytes(short)
2390 // inline int Integer.reverseBytes(int)
2391 // inline long Long.reverseBytes(long)
2392 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2393 Node* arg = argument(0);
2394 Node* n = NULL;
2395 switch (id) {
2396 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2397 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2398 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2399 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2400 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2401 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2402 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2403 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2404 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2405 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2406 default: fatal_unexpected_iid(id); break;
2407 }
2408 set_result(_gvn.transform(n));
2409 return true;
2410 }
2411
2412 //----------------------------inline_unsafe_access----------------------------
2413
2414 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2415
2416 // Helper that guards and inserts a pre-barrier.
2417 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2418 Node* pre_val, bool need_mem_bar) {
2419 // We could be accessing the referent field of a reference object. If so, when G1
2420 // is enabled, we need to log the value in the referent field in an SATB buffer.
2421 // This routine performs some compile time filters and generates suitable
2422 // runtime filters that guard the pre-barrier code.
2423 // Also add memory barrier for non volatile load from the referent field
2424 // to prevent commoning of loads across safepoint.
2425 if (!UseG1GC && !need_mem_bar)
2426 return;
2427
2428 // Some compile time checks.
2429
2430 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2431 const TypeX* otype = offset->find_intptr_t_type();
2432 if (otype != NULL && otype->is_con() &&
2433 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2434 // Constant offset but not the reference_offset so just return
2435 return;
2436 }
2437
2438 // We only need to generate the runtime guards for instances.
2439 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2440 if (btype != NULL) {
2441 if (btype->isa_aryptr()) {
2442 // Array type so nothing to do
2443 return;
2444 }
2445
2446 const TypeInstPtr* itype = btype->isa_instptr();
2447 if (itype != NULL) {
2448 // Can the klass of base_oop be statically determined to be
2449 // _not_ a sub-class of Reference and _not_ Object?
2450 ciKlass* klass = itype->klass();
2451 if ( klass->is_loaded() &&
2452 !klass->is_subtype_of(env()->Reference_klass()) &&
2453 !env()->Object_klass()->is_subtype_of(klass)) {
2454 return;
2455 }
2456 }
2457 }
2458
2459 // The compile time filters did not reject base_oop/offset so
2460 // we need to generate the following runtime filters
2461 //
2462 // if (offset == java_lang_ref_Reference::_reference_offset) {
2463 // if (instance_of(base, java.lang.ref.Reference)) {
2464 // pre_barrier(_, pre_val, ...);
2465 // }
2466 // }
2467
2468 float likely = PROB_LIKELY( 0.999);
2469 float unlikely = PROB_UNLIKELY(0.999);
2470
2471 IdealKit ideal(this);
2472 #define __ ideal.
2473
2474 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2475
2476 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2477 // Update graphKit memory and control from IdealKit.
2478 sync_kit(ideal);
2479
2480 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2481 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2482
2483 // Update IdealKit memory and control from graphKit.
2484 __ sync_kit(this);
2485
2486 Node* one = __ ConI(1);
2487 // is_instof == 0 if base_oop == NULL
2488 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2489
2490 // Update graphKit from IdeakKit.
2491 sync_kit(ideal);
2492
2493 // Use the pre-barrier to record the value in the referent field
2494 pre_barrier(false /* do_load */,
2495 __ ctrl(),
2496 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2497 pre_val /* pre_val */,
2498 T_OBJECT);
2499 if (need_mem_bar) {
2500 // Add memory barrier to prevent commoning reads from this field
2501 // across safepoint since GC can change its value.
2502 insert_mem_bar(Op_MemBarCPUOrder);
2503 }
2504 // Update IdealKit from graphKit.
2505 __ sync_kit(this);
2506
2507 } __ end_if(); // _ref_type != ref_none
2508 } __ end_if(); // offset == referent_offset
2509
2510 // Final sync IdealKit and GraphKit.
2511 final_sync(ideal);
2512 #undef __
2513 }
2514
2515
2516 // Interpret Unsafe.fieldOffset cookies correctly:
2517 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2518
2519 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2520 // Attempt to infer a sharper value type from the offset and base type.
2521 ciKlass* sharpened_klass = NULL;
2522
2523 // See if it is an instance field, with an object type.
2524 if (alias_type->field() != NULL) {
2525 assert(!is_native_ptr, "native pointer op cannot use a java address");
2526 if (alias_type->field()->type()->is_klass()) {
2527 sharpened_klass = alias_type->field()->type()->as_klass();
2528 }
2529 }
2530
2531 // See if it is a narrow oop array.
2532 if (adr_type->isa_aryptr()) {
2533 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2534 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2535 if (elem_type != NULL) {
2536 sharpened_klass = elem_type->klass();
2537 }
2538 }
2539 }
2540
2541 // The sharpened class might be unloaded if there is no class loader
2542 // contraint in place.
2543 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2544 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2545
2546 #ifndef PRODUCT
2547 if (C->print_intrinsics() || C->print_inlining()) {
2548 tty->print(" from base type: "); adr_type->dump();
2549 tty->print(" sharpened value: "); tjp->dump();
2550 }
2551 #endif
2552 // Sharpen the value type.
2553 return tjp;
2554 }
2555 return NULL;
2556 }
2557
2558 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2559 if (callee()->is_static()) return false; // caller must have the capability!
2560 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2561
2562 #ifndef PRODUCT
2563 {
2564 ResourceMark rm;
2565 // Check the signatures.
2566 ciSignature* sig = callee()->signature();
2567 #ifdef ASSERT
2568 if (!is_store) {
2569 // Object getObject(Object base, int/long offset), etc.
2570 BasicType rtype = sig->return_type()->basic_type();
2571 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2572 rtype = T_ADDRESS; // it is really a C void*
2573 assert(rtype == type, "getter must return the expected value");
2574 if (!is_native_ptr) {
2575 assert(sig->count() == 2, "oop getter has 2 arguments");
2576 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2577 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2578 } else {
2579 assert(sig->count() == 1, "native getter has 1 argument");
2580 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2581 }
2582 } else {
2583 // void putObject(Object base, int/long offset, Object x), etc.
2584 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2585 if (!is_native_ptr) {
2586 assert(sig->count() == 3, "oop putter has 3 arguments");
2587 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2588 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2589 } else {
2590 assert(sig->count() == 2, "native putter has 2 arguments");
2591 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2592 }
2593 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2594 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2595 vtype = T_ADDRESS; // it is really a C void*
2596 assert(vtype == type, "putter must accept the expected value");
2597 }
2598 #endif // ASSERT
2599 }
2600 #endif //PRODUCT
2601
2602 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2603
2604 Node* receiver = argument(0); // type: oop
2605
2606 // Build address expression. See the code in inline_unsafe_prefetch.
2607 Node* adr;
2608 Node* heap_base_oop = top();
2609 Node* offset = top();
2610 Node* val;
2611
2612 if (!is_native_ptr) {
2613 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2614 Node* base = argument(1); // type: oop
2615 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2616 offset = argument(2); // type: long
2617 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2618 // to be plain byte offsets, which are also the same as those accepted
2619 // by oopDesc::field_base.
2620 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2621 "fieldOffset must be byte-scaled");
2622 // 32-bit machines ignore the high half!
2623 offset = ConvL2X(offset);
2624 adr = make_unsafe_address(base, offset);
2625 heap_base_oop = base;
2626 val = is_store ? argument(4) : NULL;
2627 } else {
2628 Node* ptr = argument(1); // type: long
2629 ptr = ConvL2X(ptr); // adjust Java long to machine word
2630 adr = make_unsafe_address(NULL, ptr);
2631 val = is_store ? argument(3) : NULL;
2632 }
2633
2634 // Can base be NULL? Otherwise, always on-heap access.
2635 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2636
2637 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2638
2639 // Try to categorize the address.
2640 Compile::AliasType* alias_type = C->alias_type(adr_type);
2641 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2642
2643 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2644 alias_type->adr_type() == TypeAryPtr::RANGE) {
2645 return false; // not supported
2646 }
2647
2648 bool mismatched = false;
2649 BasicType bt = alias_type->basic_type();
2650 if (bt != T_ILLEGAL) {
2651 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2652 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2653 // Alias type doesn't differentiate between byte[] and boolean[]).
2654 // Use address type to get the element type.
2655 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2656 }
2657 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2658 // accessing an array field with getObject is not a mismatch
2659 bt = T_OBJECT;
2660 }
2661 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2662 // Don't intrinsify mismatched object accesses
2663 return false;
2664 }
2665 mismatched = (bt != type);
2666 }
2667
2668 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2669
2670 // First guess at the value type.
2671 const Type *value_type = Type::get_const_basic_type(type);
2672
2673 // We will need memory barriers unless we can determine a unique
2674 // alias category for this reference. (Note: If for some reason
2675 // the barriers get omitted and the unsafe reference begins to "pollute"
2676 // the alias analysis of the rest of the graph, either Compile::can_alias
2677 // or Compile::must_alias will throw a diagnostic assert.)
2678 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2679
2680 // If we are reading the value of the referent field of a Reference
2681 // object (either by using Unsafe directly or through reflection)
2682 // then, if G1 is enabled, we need to record the referent in an
2683 // SATB log buffer using the pre-barrier mechanism.
2684 // Also we need to add memory barrier to prevent commoning reads
2685 // from this field across safepoint since GC can change its value.
2686 bool need_read_barrier = !is_native_ptr && !is_store &&
2687 offset != top() && heap_base_oop != top();
2688
2689 if (!is_store && type == T_OBJECT) {
2690 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2691 if (tjp != NULL) {
2692 value_type = tjp;
2693 }
2694 }
2695
2696 receiver = null_check(receiver);
2697 if (stopped()) {
2698 return true;
2699 }
2700 // Heap pointers get a null-check from the interpreter,
2701 // as a courtesy. However, this is not guaranteed by Unsafe,
2702 // and it is not possible to fully distinguish unintended nulls
2703 // from intended ones in this API.
2704
2705 if (is_volatile) {
2706 // We need to emit leading and trailing CPU membars (see below) in
2707 // addition to memory membars when is_volatile. This is a little
2708 // too strong, but avoids the need to insert per-alias-type
2709 // volatile membars (for stores; compare Parse::do_put_xxx), which
2710 // we cannot do effectively here because we probably only have a
2711 // rough approximation of type.
2712 need_mem_bar = true;
2713 // For Stores, place a memory ordering barrier now.
2714 if (is_store) {
2715 insert_mem_bar(Op_MemBarRelease);
2716 } else {
2717 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2718 insert_mem_bar(Op_MemBarVolatile);
2719 }
2720 }
2721 }
2722
2723 // Memory barrier to prevent normal and 'unsafe' accesses from
2724 // bypassing each other. Happens after null checks, so the
2725 // exception paths do not take memory state from the memory barrier,
2726 // so there's no problems making a strong assert about mixing users
2727 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2728 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2729 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2730
2731 if (!is_store) {
2732 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2733 // To be valid, unsafe loads may depend on other conditions than
2734 // the one that guards them: pin the Load node
2735 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2736 // load value
2737 switch (type) {
2738 case T_BOOLEAN:
2739 case T_CHAR:
2740 case T_BYTE:
2741 case T_SHORT:
2742 case T_INT:
2743 case T_LONG:
2744 case T_FLOAT:
2745 case T_DOUBLE:
2746 break;
2747 case T_OBJECT:
2748 if (need_read_barrier) {
2749 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2750 }
2751 break;
2752 case T_ADDRESS:
2753 // Cast to an int type.
2754 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2755 p = ConvX2UL(p);
2756 break;
2757 default:
2758 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2759 break;
2760 }
2761 // The load node has the control of the preceding MemBarCPUOrder. All
2762 // following nodes will have the control of the MemBarCPUOrder inserted at
2763 // the end of this method. So, pushing the load onto the stack at a later
2764 // point is fine.
2765 set_result(p);
2766 } else {
2767 // place effect of store into memory
2768 switch (type) {
2769 case T_DOUBLE:
2770 val = dstore_rounding(val);
2771 break;
2772 case T_ADDRESS:
2773 // Repackage the long as a pointer.
2774 val = ConvL2X(val);
2775 val = _gvn.transform(new (C) CastX2PNode(val));
2776 break;
2777 }
2778
2779 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2780 if (type != T_OBJECT ) {
2781 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2782 } else {
2783 // Possibly an oop being stored to Java heap or native memory
2784 if (!can_access_non_heap) {
2785 // oop to Java heap.
2786 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2787 } else {
2788 // We can't tell at compile time if we are storing in the Java heap or outside
2789 // of it. So we need to emit code to conditionally do the proper type of
2790 // store.
2791
2792 IdealKit ideal(this);
2793 #define __ ideal.
2794 // QQQ who knows what probability is here??
2795 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2796 // Sync IdealKit and graphKit.
2797 sync_kit(ideal);
2798 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2799 // Update IdealKit memory.
2800 __ sync_kit(this);
2801 } __ else_(); {
2802 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched);
2803 } __ end_if();
2804 // Final sync IdealKit and GraphKit.
2805 final_sync(ideal);
2806 #undef __
2807 }
2808 }
2809 }
2810
2811 if (is_volatile) {
2812 if (!is_store) {
2813 insert_mem_bar(Op_MemBarAcquire);
2814 } else {
2815 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2816 insert_mem_bar(Op_MemBarVolatile);
2817 }
2818 }
2819 }
2820
2821 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2822
2823 return true;
2824 }
2825
2826 //----------------------------inline_unsafe_prefetch----------------------------
2827
2828 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2829 #ifndef PRODUCT
2830 {
2831 ResourceMark rm;
2832 // Check the signatures.
2833 ciSignature* sig = callee()->signature();
2834 #ifdef ASSERT
2835 // Object getObject(Object base, int/long offset), etc.
2836 BasicType rtype = sig->return_type()->basic_type();
2837 if (!is_native_ptr) {
2838 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2839 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2840 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2841 } else {
2842 assert(sig->count() == 1, "native prefetch has 1 argument");
2843 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2844 }
2845 #endif // ASSERT
2846 }
2847 #endif // !PRODUCT
2848
2849 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2850
2851 const int idx = is_static ? 0 : 1;
2852 if (!is_static) {
2853 null_check_receiver();
2854 if (stopped()) {
2855 return true;
2856 }
2857 }
2858
2859 // Build address expression. See the code in inline_unsafe_access.
2860 Node *adr;
2861 if (!is_native_ptr) {
2862 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2863 Node* base = argument(idx + 0); // type: oop
2864 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2865 Node* offset = argument(idx + 1); // type: long
2866 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2867 // to be plain byte offsets, which are also the same as those accepted
2868 // by oopDesc::field_base.
2869 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2870 "fieldOffset must be byte-scaled");
2871 // 32-bit machines ignore the high half!
2872 offset = ConvL2X(offset);
2873 adr = make_unsafe_address(base, offset);
2874 } else {
2875 Node* ptr = argument(idx + 0); // type: long
2876 ptr = ConvL2X(ptr); // adjust Java long to machine word
2877 adr = make_unsafe_address(NULL, ptr);
2878 }
2879
2880 // Generate the read or write prefetch
2881 Node *prefetch;
2882 if (is_store) {
2883 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2884 } else {
2885 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2886 }
2887 prefetch->init_req(0, control());
2888 set_i_o(_gvn.transform(prefetch));
2889
2890 return true;
2891 }
2892
2893 //----------------------------inline_unsafe_load_store----------------------------
2894 // This method serves a couple of different customers (depending on LoadStoreKind):
2895 //
2896 // LS_cmpxchg:
2897 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2898 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2899 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2900 //
2901 // LS_xadd:
2902 // public int getAndAddInt( Object o, long offset, int delta)
2903 // public long getAndAddLong(Object o, long offset, long delta)
2904 //
2905 // LS_xchg:
2906 // int getAndSet(Object o, long offset, int newValue)
2907 // long getAndSet(Object o, long offset, long newValue)
2908 // Object getAndSet(Object o, long offset, Object newValue)
2909 //
2910 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2911 // This basic scheme here is the same as inline_unsafe_access, but
2912 // differs in enough details that combining them would make the code
2913 // overly confusing. (This is a true fact! I originally combined
2914 // them, but even I was confused by it!) As much code/comments as
2915 // possible are retained from inline_unsafe_access though to make
2916 // the correspondences clearer. - dl
2917
2918 if (callee()->is_static()) return false; // caller must have the capability!
2919
2920 #ifndef PRODUCT
2921 BasicType rtype;
2922 {
2923 ResourceMark rm;
2924 // Check the signatures.
2925 ciSignature* sig = callee()->signature();
2926 rtype = sig->return_type()->basic_type();
2927 if (kind == LS_xadd || kind == LS_xchg) {
2928 // Check the signatures.
2929 #ifdef ASSERT
2930 assert(rtype == type, "get and set must return the expected type");
2931 assert(sig->count() == 3, "get and set has 3 arguments");
2932 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2933 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2934 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2935 #endif // ASSERT
2936 } else if (kind == LS_cmpxchg) {
2937 // Check the signatures.
2938 #ifdef ASSERT
2939 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2940 assert(sig->count() == 4, "CAS has 4 arguments");
2941 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2942 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2943 #endif // ASSERT
2944 } else {
2945 ShouldNotReachHere();
2946 }
2947 }
2948 #endif //PRODUCT
2949
2950 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2951
2952 // Get arguments:
2953 Node* receiver = NULL;
2954 Node* base = NULL;
2955 Node* offset = NULL;
2956 Node* oldval = NULL;
2957 Node* newval = NULL;
2958 if (kind == LS_cmpxchg) {
2959 const bool two_slot_type = type2size[type] == 2;
2960 receiver = argument(0); // type: oop
2961 base = argument(1); // type: oop
2962 offset = argument(2); // type: long
2963 oldval = argument(4); // type: oop, int, or long
2964 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2965 } else if (kind == LS_xadd || kind == LS_xchg){
2966 receiver = argument(0); // type: oop
2967 base = argument(1); // type: oop
2968 offset = argument(2); // type: long
2969 oldval = NULL;
2970 newval = argument(4); // type: oop, int, or long
2971 }
2972
2973 // Build field offset expression.
2974 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2975 // to be plain byte offsets, which are also the same as those accepted
2976 // by oopDesc::field_base.
2977 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2978 // 32-bit machines ignore the high half of long offsets
2979 offset = ConvL2X(offset);
2980 Node* adr = make_unsafe_address(base, offset);
2981 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2982
2983 Compile::AliasType* alias_type = C->alias_type(adr_type);
2984 BasicType bt = alias_type->basic_type();
2985 if (bt != T_ILLEGAL &&
2986 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2987 // Don't intrinsify mismatched object accesses.
2988 return false;
2989 }
2990
2991 // For CAS, unlike inline_unsafe_access, there seems no point in
2992 // trying to refine types. Just use the coarse types here.
2993 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2994 const Type *value_type = Type::get_const_basic_type(type);
2995
2996 if (kind == LS_xchg && type == T_OBJECT) {
2997 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2998 if (tjp != NULL) {
2999 value_type = tjp;
3000 }
3001 }
3002
3003 // Null check receiver.
3004 receiver = null_check(receiver);
3005 if (stopped()) {
3006 return true;
3007 }
3008
3009 int alias_idx = C->get_alias_index(adr_type);
3010
3011 // Memory-model-wise, a LoadStore acts like a little synchronized
3012 // block, so needs barriers on each side. These don't translate
3013 // into actual barriers on most machines, but we still need rest of
3014 // compiler to respect ordering.
3015
3016 insert_mem_bar(Op_MemBarRelease);
3017 insert_mem_bar(Op_MemBarCPUOrder);
3018
3019 // 4984716: MemBars must be inserted before this
3020 // memory node in order to avoid a false
3021 // dependency which will confuse the scheduler.
3022 Node *mem = memory(alias_idx);
3023
3024 // For now, we handle only those cases that actually exist: ints,
3025 // longs, and Object. Adding others should be straightforward.
3026 Node* load_store = NULL;
3027 switch(type) {
3028 case T_INT:
3029 if (kind == LS_xadd) {
3030 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3031 } else if (kind == LS_xchg) {
3032 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3033 } else if (kind == LS_cmpxchg) {
3034 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3035 } else {
3036 ShouldNotReachHere();
3037 }
3038 break;
3039 case T_LONG:
3040 if (kind == LS_xadd) {
3041 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3042 } else if (kind == LS_xchg) {
3043 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3044 } else if (kind == LS_cmpxchg) {
3045 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3046 } else {
3047 ShouldNotReachHere();
3048 }
3049 break;
3050 case T_OBJECT:
3051 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3052 // could be delayed during Parse (for example, in adjust_map_after_if()).
3053 // Execute transformation here to avoid barrier generation in such case.
3054 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3055 newval = _gvn.makecon(TypePtr::NULL_PTR);
3056
3057 // Reference stores need a store barrier.
3058 if (kind == LS_xchg) {
3059 // If pre-barrier must execute before the oop store, old value will require do_load here.
3060 if (!can_move_pre_barrier()) {
3061 pre_barrier(true /* do_load*/,
3062 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3063 NULL /* pre_val*/,
3064 T_OBJECT);
3065 } // Else move pre_barrier to use load_store value, see below.
3066 } else if (kind == LS_cmpxchg) {
3067 // Same as for newval above:
3068 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3069 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3070 }
3071 // The only known value which might get overwritten is oldval.
3072 pre_barrier(false /* do_load */,
3073 control(), NULL, NULL, max_juint, NULL, NULL,
3074 oldval /* pre_val */,
3075 T_OBJECT);
3076 } else {
3077 ShouldNotReachHere();
3078 }
3079
3080 #ifdef _LP64
3081 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3082 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3083 if (kind == LS_xchg) {
3084 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3085 newval_enc, adr_type, value_type->make_narrowoop()));
3086 } else {
3087 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3088 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3089 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3090 newval_enc, oldval_enc));
3091 }
3092 } else
3093 #endif
3094 {
3095 if (kind == LS_xchg) {
3096 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3097 } else {
3098 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3099 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3100 }
3101 }
3102 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3103 break;
3104 default:
3105 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3106 break;
3107 }
3108
3109 // SCMemProjNodes represent the memory state of a LoadStore. Their
3110 // main role is to prevent LoadStore nodes from being optimized away
3111 // when their results aren't used.
3112 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3113 set_memory(proj, alias_idx);
3114
3115 if (type == T_OBJECT && kind == LS_xchg) {
3116 #ifdef _LP64
3117 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3118 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3119 }
3120 #endif
3121 if (can_move_pre_barrier()) {
3122 // Don't need to load pre_val. The old value is returned by load_store.
3123 // The pre_barrier can execute after the xchg as long as no safepoint
3124 // gets inserted between them.
3125 pre_barrier(false /* do_load */,
3126 control(), NULL, NULL, max_juint, NULL, NULL,
3127 load_store /* pre_val */,
3128 T_OBJECT);
3129 }
3130 }
3131
3132 // Add the trailing membar surrounding the access
3133 insert_mem_bar(Op_MemBarCPUOrder);
3134 insert_mem_bar(Op_MemBarAcquire);
3135
3136 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3137 set_result(load_store);
3138 return true;
3139 }
3140
3141 //----------------------------inline_unsafe_ordered_store----------------------
3142 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3143 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3144 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3145 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3146 // This is another variant of inline_unsafe_access, differing in
3147 // that it always issues store-store ("release") barrier and ensures
3148 // store-atomicity (which only matters for "long").
3149
3150 if (callee()->is_static()) return false; // caller must have the capability!
3151
3152 #ifndef PRODUCT
3153 {
3154 ResourceMark rm;
3155 // Check the signatures.
3156 ciSignature* sig = callee()->signature();
3157 #ifdef ASSERT
3158 BasicType rtype = sig->return_type()->basic_type();
3159 assert(rtype == T_VOID, "must return void");
3160 assert(sig->count() == 3, "has 3 arguments");
3161 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3162 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3163 #endif // ASSERT
3164 }
3165 #endif //PRODUCT
3166
3167 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3168
3169 // Get arguments:
3170 Node* receiver = argument(0); // type: oop
3171 Node* base = argument(1); // type: oop
3172 Node* offset = argument(2); // type: long
3173 Node* val = argument(4); // type: oop, int, or long
3174
3175 // Null check receiver.
3176 receiver = null_check(receiver);
3177 if (stopped()) {
3178 return true;
3179 }
3180
3181 // Build field offset expression.
3182 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3183 // 32-bit machines ignore the high half of long offsets
3184 offset = ConvL2X(offset);
3185 Node* adr = make_unsafe_address(base, offset);
3186 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3187 const Type *value_type = Type::get_const_basic_type(type);
3188 Compile::AliasType* alias_type = C->alias_type(adr_type);
3189
3190 insert_mem_bar(Op_MemBarRelease);
3191 insert_mem_bar(Op_MemBarCPUOrder);
3192 // Ensure that the store is atomic for longs:
3193 const bool require_atomic_access = true;
3194 Node* store;
3195 if (type == T_OBJECT) // reference stores need a store barrier.
3196 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3197 else {
3198 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3199 }
3200 insert_mem_bar(Op_MemBarCPUOrder);
3201 return true;
3202 }
3203
3204 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3205 // Regardless of form, don't allow previous ld/st to move down,
3206 // then issue acquire, release, or volatile mem_bar.
3207 insert_mem_bar(Op_MemBarCPUOrder);
3208 switch(id) {
3209 case vmIntrinsics::_loadFence:
3210 insert_mem_bar(Op_LoadFence);
3211 return true;
3212 case vmIntrinsics::_storeFence:
3213 insert_mem_bar(Op_StoreFence);
3214 return true;
3215 case vmIntrinsics::_fullFence:
3216 insert_mem_bar(Op_MemBarVolatile);
3217 return true;
3218 default:
3219 fatal_unexpected_iid(id);
3220 return false;
3221 }
3222 }
3223
3224 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3225 if (!kls->is_Con()) {
3226 return true;
3227 }
3228 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3229 if (klsptr == NULL) {
3230 return true;
3231 }
3232 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3233 // don't need a guard for a klass that is already initialized
3234 return !ik->is_initialized();
3235 }
3236
3237 //----------------------------inline_unsafe_allocate---------------------------
3238 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3239 bool LibraryCallKit::inline_unsafe_allocate() {
3240 if (callee()->is_static()) return false; // caller must have the capability!
3241
3242 null_check_receiver(); // null-check, then ignore
3243 Node* cls = null_check(argument(1));
3244 if (stopped()) return true;
3245
3246 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3247 kls = null_check(kls);
3248 if (stopped()) return true; // argument was like int.class
3249
3250 Node* test = NULL;
3251 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3252 // Note: The argument might still be an illegal value like
3253 // Serializable.class or Object[].class. The runtime will handle it.
3254 // But we must make an explicit check for initialization.
3255 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3256 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3257 // can generate code to load it as unsigned byte.
3258 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3259 Node* bits = intcon(InstanceKlass::fully_initialized);
3260 test = _gvn.transform(new (C) SubINode(inst, bits));
3261 // The 'test' is non-zero if we need to take a slow path.
3262 }
3263
3264 Node* obj = new_instance(kls, test);
3265 set_result(obj);
3266 return true;
3267 }
3268
3269 #ifdef JFR_HAVE_INTRINSICS
3270 /*
3271 * oop -> myklass
3272 * myklass->trace_id |= USED
3273 * return myklass->trace_id & ~0x3
3274 */
3275 bool LibraryCallKit::inline_native_classID() {
3276 Node* cls = null_check(argument(0), T_OBJECT);
3277 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3278 kls = null_check(kls, T_OBJECT);
3279
3280 ByteSize offset = KLASS_TRACE_ID_OFFSET;
3281 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3282 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3283
3284 Node* clsused = longcon(0x01l); // set the class bit
3285 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3286 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3287 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3288
3289 #ifdef TRACE_ID_META_BITS
3290 Node* mbits = longcon(~TRACE_ID_META_BITS);
3291 tvalue = _gvn.transform(new (C) AndLNode(tvalue, mbits));
3292 #endif
3293 #ifdef TRACE_ID_SHIFT
3294 Node* cbits = intcon(TRACE_ID_SHIFT);
3295 tvalue = _gvn.transform(new (C) URShiftLNode(tvalue, cbits));
3296 #endif
3297
3298 set_result(tvalue);
3299 return true;
3300 }
3301
3302 bool LibraryCallKit::inline_native_getEventWriter() {
3303 Node* tls_ptr = _gvn.transform(new (C) ThreadLocalNode());
3304
3305 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3306 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)
3307 );
3308
3309 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3310
3311 Node* jobj_cmp_null = _gvn.transform( new (C) CmpPNode(jobj, null()) );
3312 Node* test_jobj_eq_null = _gvn.transform( new (C) BoolNode(jobj_cmp_null, BoolTest::eq) );
3313
3314 IfNode* iff_jobj_null =
3315 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3316
3317 enum { _normal_path = 1,
3318 _null_path = 2,
3319 PATH_LIMIT };
3320
3321 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3322 PhiNode* result_val = new (C) PhiNode(result_rgn, TypePtr::BOTTOM);
3323
3324 Node* jobj_is_null = _gvn.transform(new (C) IfTrueNode(iff_jobj_null));
3325 result_rgn->init_req(_null_path, jobj_is_null);
3326 result_val->init_req(_null_path, null());
3327
3328 Node* jobj_is_not_null = _gvn.transform(new (C) IfFalseNode(iff_jobj_null));
3329 result_rgn->init_req(_normal_path, jobj_is_not_null);
3330
3331 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3332 result_val->init_req(_normal_path, res);
3333
3334 set_result(result_rgn, result_val);
3335
3336 return true;
3337 }
3338 #endif // JFR_HAVE_INTRINSICS
3339
3340 //------------------------inline_native_time_funcs--------------
3341 // inline code for System.currentTimeMillis() and System.nanoTime()
3342 // these have the same type and signature
3343 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3344 const TypeFunc* tf = OptoRuntime::void_long_Type();
3345 const TypePtr* no_memory_effects = NULL;
3346 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3347 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3348 #ifdef ASSERT
3349 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3350 assert(value_top == top(), "second value must be top");
3351 #endif
3352 set_result(value);
3353 return true;
3354 }
3355
3356 //------------------------inline_native_currentThread------------------
3357 bool LibraryCallKit::inline_native_currentThread() {
3358 Node* junk = NULL;
3359 set_result(generate_current_thread(junk));
3360 return true;
3361 }
3362
3363 //------------------------inline_native_isInterrupted------------------
3364 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3365 bool LibraryCallKit::inline_native_isInterrupted() {
3366 // Add a fast path to t.isInterrupted(clear_int):
3367 // (t == Thread.current() &&
3368 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3369 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3370 // So, in the common case that the interrupt bit is false,
3371 // we avoid making a call into the VM. Even if the interrupt bit
3372 // is true, if the clear_int argument is false, we avoid the VM call.
3373 // However, if the receiver is not currentThread, we must call the VM,
3374 // because there must be some locking done around the operation.
3375
3376 // We only go to the fast case code if we pass two guards.
3377 // Paths which do not pass are accumulated in the slow_region.
3378
3379 enum {
3380 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3381 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3382 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3383 PATH_LIMIT
3384 };
3385
3386 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3387 // out of the function.
3388 insert_mem_bar(Op_MemBarCPUOrder);
3389
3390 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3391 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3392
3393 RegionNode* slow_region = new (C) RegionNode(1);
3394 record_for_igvn(slow_region);
3395
3396 // (a) Receiving thread must be the current thread.
3397 Node* rec_thr = argument(0);
3398 Node* tls_ptr = NULL;
3399 Node* cur_thr = generate_current_thread(tls_ptr);
3400 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3401 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3402
3403 generate_slow_guard(bol_thr, slow_region);
3404
3405 // (b) Interrupt bit on TLS must be false.
3406 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3407 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3408 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3409
3410 // Set the control input on the field _interrupted read to prevent it floating up.
3411 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3412 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3413 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3414
3415 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3416
3417 // First fast path: if (!TLS._interrupted) return false;
3418 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3419 result_rgn->init_req(no_int_result_path, false_bit);
3420 result_val->init_req(no_int_result_path, intcon(0));
3421
3422 // drop through to next case
3423 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3424
3425 #ifndef TARGET_OS_FAMILY_windows
3426 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3427 Node* clr_arg = argument(1);
3428 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3429 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3430 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3431
3432 // Second fast path: ... else if (!clear_int) return true;
3433 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3434 result_rgn->init_req(no_clear_result_path, false_arg);
3435 result_val->init_req(no_clear_result_path, intcon(1));
3436
3437 // drop through to next case
3438 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3439 #else
3440 // To return true on Windows you must read the _interrupted field
3441 // and check the the event state i.e. take the slow path.
3442 #endif // TARGET_OS_FAMILY_windows
3443
3444 // (d) Otherwise, go to the slow path.
3445 slow_region->add_req(control());
3446 set_control( _gvn.transform(slow_region));
3447
3448 if (stopped()) {
3449 // There is no slow path.
3450 result_rgn->init_req(slow_result_path, top());
3451 result_val->init_req(slow_result_path, top());
3452 } else {
3453 // non-virtual because it is a private non-static
3454 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3455
3456 Node* slow_val = set_results_for_java_call(slow_call);
3457 // this->control() comes from set_results_for_java_call
3458
3459 Node* fast_io = slow_call->in(TypeFunc::I_O);
3460 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3461
3462 // These two phis are pre-filled with copies of of the fast IO and Memory
3463 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3464 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3465
3466 result_rgn->init_req(slow_result_path, control());
3467 result_io ->init_req(slow_result_path, i_o());
3468 result_mem->init_req(slow_result_path, reset_memory());
3469 result_val->init_req(slow_result_path, slow_val);
3470
3471 set_all_memory(_gvn.transform(result_mem));
3472 set_i_o( _gvn.transform(result_io));
3473 }
3474
3475 C->set_has_split_ifs(true); // Has chance for split-if optimization
3476 set_result(result_rgn, result_val);
3477 return true;
3478 }
3479
3480 //---------------------------load_mirror_from_klass----------------------------
3481 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3482 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3483 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3484 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3485 }
3486
3487 //-----------------------load_klass_from_mirror_common-------------------------
3488 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3489 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3490 // and branch to the given path on the region.
3491 // If never_see_null, take an uncommon trap on null, so we can optimistically
3492 // compile for the non-null case.
3493 // If the region is NULL, force never_see_null = true.
3494 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3495 bool never_see_null,
3496 RegionNode* region,
3497 int null_path,
3498 int offset) {
3499 if (region == NULL) never_see_null = true;
3500 Node* p = basic_plus_adr(mirror, offset);
3501 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3502 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3503 Node* null_ctl = top();
3504 kls = null_check_oop(kls, &null_ctl, never_see_null);
3505 if (region != NULL) {
3506 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3507 region->init_req(null_path, null_ctl);
3508 } else {
3509 assert(null_ctl == top(), "no loose ends");
3510 }
3511 return kls;
3512 }
3513
3514 //--------------------(inline_native_Class_query helpers)---------------------
3515 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3516 // Fall through if (mods & mask) == bits, take the guard otherwise.
3517 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3518 // Branch around if the given klass has the given modifier bit set.
3519 // Like generate_guard, adds a new path onto the region.
3520 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3521 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3522 Node* mask = intcon(modifier_mask);
3523 Node* bits = intcon(modifier_bits);
3524 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3525 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3526 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3527 return generate_fair_guard(bol, region);
3528 }
3529 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3530 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3531 }
3532
3533 //-------------------------inline_native_Class_query-------------------
3534 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3535 const Type* return_type = TypeInt::BOOL;
3536 Node* prim_return_value = top(); // what happens if it's a primitive class?
3537 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3538 bool expect_prim = false; // most of these guys expect to work on refs
3539
3540 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3541
3542 Node* mirror = argument(0);
3543 Node* obj = top();
3544
3545 switch (id) {
3546 case vmIntrinsics::_isInstance:
3547 // nothing is an instance of a primitive type
3548 prim_return_value = intcon(0);
3549 obj = argument(1);
3550 break;
3551 case vmIntrinsics::_getModifiers:
3552 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3553 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3554 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3555 break;
3556 case vmIntrinsics::_isInterface:
3557 prim_return_value = intcon(0);
3558 break;
3559 case vmIntrinsics::_isArray:
3560 prim_return_value = intcon(0);
3561 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3562 break;
3563 case vmIntrinsics::_isPrimitive:
3564 prim_return_value = intcon(1);
3565 expect_prim = true; // obviously
3566 break;
3567 case vmIntrinsics::_getSuperclass:
3568 prim_return_value = null();
3569 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3570 break;
3571 case vmIntrinsics::_getComponentType:
3572 prim_return_value = null();
3573 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3574 break;
3575 case vmIntrinsics::_getClassAccessFlags:
3576 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3577 return_type = TypeInt::INT; // not bool! 6297094
3578 break;
3579 default:
3580 fatal_unexpected_iid(id);
3581 break;
3582 }
3583
3584 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3585 if (mirror_con == NULL) return false; // cannot happen?
3586
3587 #ifndef PRODUCT
3588 if (C->print_intrinsics() || C->print_inlining()) {
3589 ciType* k = mirror_con->java_mirror_type();
3590 if (k) {
3591 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3592 k->print_name();
3593 tty->cr();
3594 }
3595 }
3596 #endif
3597
3598 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3599 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3600 record_for_igvn(region);
3601 PhiNode* phi = new (C) PhiNode(region, return_type);
3602
3603 // The mirror will never be null of Reflection.getClassAccessFlags, however
3604 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3605 // if it is. See bug 4774291.
3606
3607 // For Reflection.getClassAccessFlags(), the null check occurs in
3608 // the wrong place; see inline_unsafe_access(), above, for a similar
3609 // situation.
3610 mirror = null_check(mirror);
3611 // If mirror or obj is dead, only null-path is taken.
3612 if (stopped()) return true;
3613
3614 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3615
3616 // Now load the mirror's klass metaobject, and null-check it.
3617 // Side-effects region with the control path if the klass is null.
3618 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3619 // If kls is null, we have a primitive mirror.
3620 phi->init_req(_prim_path, prim_return_value);
3621 if (stopped()) { set_result(region, phi); return true; }
3622 bool safe_for_replace = (region->in(_prim_path) == top());
3623
3624 Node* p; // handy temp
3625 Node* null_ctl;
3626
3627 // Now that we have the non-null klass, we can perform the real query.
3628 // For constant classes, the query will constant-fold in LoadNode::Value.
3629 Node* query_value = top();
3630 switch (id) {
3631 case vmIntrinsics::_isInstance:
3632 // nothing is an instance of a primitive type
3633 query_value = gen_instanceof(obj, kls, safe_for_replace);
3634 break;
3635
3636 case vmIntrinsics::_getModifiers:
3637 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3638 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3639 break;
3640
3641 case vmIntrinsics::_isInterface:
3642 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3643 if (generate_interface_guard(kls, region) != NULL)
3644 // A guard was added. If the guard is taken, it was an interface.
3645 phi->add_req(intcon(1));
3646 // If we fall through, it's a plain class.
3647 query_value = intcon(0);
3648 break;
3649
3650 case vmIntrinsics::_isArray:
3651 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3652 if (generate_array_guard(kls, region) != NULL)
3653 // A guard was added. If the guard is taken, it was an array.
3654 phi->add_req(intcon(1));
3655 // If we fall through, it's a plain class.
3656 query_value = intcon(0);
3657 break;
3658
3659 case vmIntrinsics::_isPrimitive:
3660 query_value = intcon(0); // "normal" path produces false
3661 break;
3662
3663 case vmIntrinsics::_getSuperclass:
3664 // The rules here are somewhat unfortunate, but we can still do better
3665 // with random logic than with a JNI call.
3666 // Interfaces store null or Object as _super, but must report null.
3667 // Arrays store an intermediate super as _super, but must report Object.
3668 // Other types can report the actual _super.
3669 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3670 if (generate_interface_guard(kls, region) != NULL)
3671 // A guard was added. If the guard is taken, it was an interface.
3672 phi->add_req(null());
3673 if (generate_array_guard(kls, region) != NULL)
3674 // A guard was added. If the guard is taken, it was an array.
3675 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3676 // If we fall through, it's a plain class. Get its _super.
3677 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3678 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3679 null_ctl = top();
3680 kls = null_check_oop(kls, &null_ctl);
3681 if (null_ctl != top()) {
3682 // If the guard is taken, Object.superClass is null (both klass and mirror).
3683 region->add_req(null_ctl);
3684 phi ->add_req(null());
3685 }
3686 if (!stopped()) {
3687 query_value = load_mirror_from_klass(kls);
3688 }
3689 break;
3690
3691 case vmIntrinsics::_getComponentType:
3692 if (generate_array_guard(kls, region) != NULL) {
3693 // Be sure to pin the oop load to the guard edge just created:
3694 Node* is_array_ctrl = region->in(region->req()-1);
3695 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3696 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3697 phi->add_req(cmo);
3698 }
3699 query_value = null(); // non-array case is null
3700 break;
3701
3702 case vmIntrinsics::_getClassAccessFlags:
3703 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3704 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3705 break;
3706
3707 default:
3708 fatal_unexpected_iid(id);
3709 break;
3710 }
3711
3712 // Fall-through is the normal case of a query to a real class.
3713 phi->init_req(1, query_value);
3714 region->init_req(1, control());
3715
3716 C->set_has_split_ifs(true); // Has chance for split-if optimization
3717 set_result(region, phi);
3718 return true;
3719 }
3720
3721 //--------------------------inline_native_subtype_check------------------------
3722 // This intrinsic takes the JNI calls out of the heart of
3723 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3724 bool LibraryCallKit::inline_native_subtype_check() {
3725 // Pull both arguments off the stack.
3726 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3727 args[0] = argument(0);
3728 args[1] = argument(1);
3729 Node* klasses[2]; // corresponding Klasses: superk, subk
3730 klasses[0] = klasses[1] = top();
3731
3732 enum {
3733 // A full decision tree on {superc is prim, subc is prim}:
3734 _prim_0_path = 1, // {P,N} => false
3735 // {P,P} & superc!=subc => false
3736 _prim_same_path, // {P,P} & superc==subc => true
3737 _prim_1_path, // {N,P} => false
3738 _ref_subtype_path, // {N,N} & subtype check wins => true
3739 _both_ref_path, // {N,N} & subtype check loses => false
3740 PATH_LIMIT
3741 };
3742
3743 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3744 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3745 record_for_igvn(region);
3746
3747 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3748 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3749 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3750
3751 // First null-check both mirrors and load each mirror's klass metaobject.
3752 int which_arg;
3753 for (which_arg = 0; which_arg <= 1; which_arg++) {
3754 Node* arg = args[which_arg];
3755 arg = null_check(arg);
3756 if (stopped()) break;
3757 args[which_arg] = arg;
3758
3759 Node* p = basic_plus_adr(arg, class_klass_offset);
3760 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3761 klasses[which_arg] = _gvn.transform(kls);
3762 }
3763
3764 // Having loaded both klasses, test each for null.
3765 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3766 for (which_arg = 0; which_arg <= 1; which_arg++) {
3767 Node* kls = klasses[which_arg];
3768 Node* null_ctl = top();
3769 kls = null_check_oop(kls, &null_ctl, never_see_null);
3770 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3771 region->init_req(prim_path, null_ctl);
3772 if (stopped()) break;
3773 klasses[which_arg] = kls;
3774 }
3775
3776 if (!stopped()) {
3777 // now we have two reference types, in klasses[0..1]
3778 Node* subk = klasses[1]; // the argument to isAssignableFrom
3779 Node* superk = klasses[0]; // the receiver
3780 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3781 // now we have a successful reference subtype check
3782 region->set_req(_ref_subtype_path, control());
3783 }
3784
3785 // If both operands are primitive (both klasses null), then
3786 // we must return true when they are identical primitives.
3787 // It is convenient to test this after the first null klass check.
3788 set_control(region->in(_prim_0_path)); // go back to first null check
3789 if (!stopped()) {
3790 // Since superc is primitive, make a guard for the superc==subc case.
3791 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3792 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3793 generate_guard(bol_eq, region, PROB_FAIR);
3794 if (region->req() == PATH_LIMIT+1) {
3795 // A guard was added. If the added guard is taken, superc==subc.
3796 region->swap_edges(PATH_LIMIT, _prim_same_path);
3797 region->del_req(PATH_LIMIT);
3798 }
3799 region->set_req(_prim_0_path, control()); // Not equal after all.
3800 }
3801
3802 // these are the only paths that produce 'true':
3803 phi->set_req(_prim_same_path, intcon(1));
3804 phi->set_req(_ref_subtype_path, intcon(1));
3805
3806 // pull together the cases:
3807 assert(region->req() == PATH_LIMIT, "sane region");
3808 for (uint i = 1; i < region->req(); i++) {
3809 Node* ctl = region->in(i);
3810 if (ctl == NULL || ctl == top()) {
3811 region->set_req(i, top());
3812 phi ->set_req(i, top());
3813 } else if (phi->in(i) == NULL) {
3814 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3815 }
3816 }
3817
3818 set_control(_gvn.transform(region));
3819 set_result(_gvn.transform(phi));
3820 return true;
3821 }
3822
3823 //---------------------generate_array_guard_common------------------------
3824 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3825 bool obj_array, bool not_array) {
3826 // If obj_array/non_array==false/false:
3827 // Branch around if the given klass is in fact an array (either obj or prim).
3828 // If obj_array/non_array==false/true:
3829 // Branch around if the given klass is not an array klass of any kind.
3830 // If obj_array/non_array==true/true:
3831 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3832 // If obj_array/non_array==true/false:
3833 // Branch around if the kls is an oop array (Object[] or subtype)
3834 //
3835 // Like generate_guard, adds a new path onto the region.
3836 jint layout_con = 0;
3837 Node* layout_val = get_layout_helper(kls, layout_con);
3838 if (layout_val == NULL) {
3839 bool query = (obj_array
3840 ? Klass::layout_helper_is_objArray(layout_con)
3841 : Klass::layout_helper_is_array(layout_con));
3842 if (query == not_array) {
3843 return NULL; // never a branch
3844 } else { // always a branch
3845 Node* always_branch = control();
3846 if (region != NULL)
3847 region->add_req(always_branch);
3848 set_control(top());
3849 return always_branch;
3850 }
3851 }
3852 // Now test the correct condition.
3853 jint nval = (obj_array
3854 ? ((jint)Klass::_lh_array_tag_type_value
3855 << Klass::_lh_array_tag_shift)
3856 : Klass::_lh_neutral_value);
3857 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3858 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3859 // invert the test if we are looking for a non-array
3860 if (not_array) btest = BoolTest(btest).negate();
3861 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3862 return generate_fair_guard(bol, region);
3863 }
3864
3865
3866 //-----------------------inline_native_newArray--------------------------
3867 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3868 bool LibraryCallKit::inline_native_newArray() {
3869 Node* mirror = argument(0);
3870 Node* count_val = argument(1);
3871
3872 mirror = null_check(mirror);
3873 // If mirror or obj is dead, only null-path is taken.
3874 if (stopped()) return true;
3875
3876 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3877 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3878 PhiNode* result_val = new(C) PhiNode(result_reg,
3879 TypeInstPtr::NOTNULL);
3880 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3881 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3882 TypePtr::BOTTOM);
3883
3884 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3885 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3886 result_reg, _slow_path);
3887 Node* normal_ctl = control();
3888 Node* no_array_ctl = result_reg->in(_slow_path);
3889
3890 // Generate code for the slow case. We make a call to newArray().
3891 set_control(no_array_ctl);
3892 if (!stopped()) {
3893 // Either the input type is void.class, or else the
3894 // array klass has not yet been cached. Either the
3895 // ensuing call will throw an exception, or else it
3896 // will cache the array klass for next time.
3897 PreserveJVMState pjvms(this);
3898 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3899 Node* slow_result = set_results_for_java_call(slow_call);
3900 // this->control() comes from set_results_for_java_call
3901 result_reg->set_req(_slow_path, control());
3902 result_val->set_req(_slow_path, slow_result);
3903 result_io ->set_req(_slow_path, i_o());
3904 result_mem->set_req(_slow_path, reset_memory());
3905 }
3906
3907 set_control(normal_ctl);
3908 if (!stopped()) {
3909 // Normal case: The array type has been cached in the java.lang.Class.
3910 // The following call works fine even if the array type is polymorphic.
3911 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3912 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3913 result_reg->init_req(_normal_path, control());
3914 result_val->init_req(_normal_path, obj);
3915 result_io ->init_req(_normal_path, i_o());
3916 result_mem->init_req(_normal_path, reset_memory());
3917 }
3918
3919 // Return the combined state.
3920 set_i_o( _gvn.transform(result_io) );
3921 set_all_memory( _gvn.transform(result_mem));
3922
3923 C->set_has_split_ifs(true); // Has chance for split-if optimization
3924 set_result(result_reg, result_val);
3925 return true;
3926 }
3927
3928 //----------------------inline_native_getLength--------------------------
3929 // public static native int java.lang.reflect.Array.getLength(Object array);
3930 bool LibraryCallKit::inline_native_getLength() {
3931 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3932
3933 Node* array = null_check(argument(0));
3934 // If array is dead, only null-path is taken.
3935 if (stopped()) return true;
3936
3937 // Deoptimize if it is a non-array.
3938 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3939
3940 if (non_array != NULL) {
3941 PreserveJVMState pjvms(this);
3942 set_control(non_array);
3943 uncommon_trap(Deoptimization::Reason_intrinsic,
3944 Deoptimization::Action_maybe_recompile);
3945 }
3946
3947 // If control is dead, only non-array-path is taken.
3948 if (stopped()) return true;
3949
3950 // The works fine even if the array type is polymorphic.
3951 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3952 Node* result = load_array_length(array);
3953
3954 C->set_has_split_ifs(true); // Has chance for split-if optimization
3955 set_result(result);
3956 return true;
3957 }
3958
3959 //------------------------inline_array_copyOf----------------------------
3960 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3961 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3962 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3963 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3964
3965 // Get the arguments.
3966 Node* original = argument(0);
3967 Node* start = is_copyOfRange? argument(1): intcon(0);
3968 Node* end = is_copyOfRange? argument(2): argument(1);
3969 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3970
3971 Node* newcopy = NULL;
3972
3973 // Set the original stack and the reexecute bit for the interpreter to reexecute
3974 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3975 { PreserveReexecuteState preexecs(this);
3976 jvms()->set_should_reexecute(true);
3977
3978 array_type_mirror = null_check(array_type_mirror);
3979 original = null_check(original);
3980
3981 // Check if a null path was taken unconditionally.
3982 if (stopped()) return true;
3983
3984 Node* orig_length = load_array_length(original);
3985
3986 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3987 klass_node = null_check(klass_node);
3988
3989 RegionNode* bailout = new (C) RegionNode(1);
3990 record_for_igvn(bailout);
3991
3992 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3993 // Bail out if that is so.
3994 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3995 if (not_objArray != NULL) {
3996 // Improve the klass node's type from the new optimistic assumption:
3997 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3998 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3999 Node* cast = new (C) CastPPNode(klass_node, akls);
4000 cast->init_req(0, control());
4001 klass_node = _gvn.transform(cast);
4002 }
4003
4004 // Bail out if either start or end is negative.
4005 generate_negative_guard(start, bailout, &start);
4006 generate_negative_guard(end, bailout, &end);
4007
4008 Node* length = end;
4009 if (_gvn.type(start) != TypeInt::ZERO) {
4010 length = _gvn.transform(new (C) SubINode(end, start));
4011 }
4012
4013 // Bail out if length is negative.
4014 // Without this the new_array would throw
4015 // NegativeArraySizeException but IllegalArgumentException is what
4016 // should be thrown
4017 generate_negative_guard(length, bailout, &length);
4018
4019 if (bailout->req() > 1) {
4020 PreserveJVMState pjvms(this);
4021 set_control(_gvn.transform(bailout));
4022 uncommon_trap(Deoptimization::Reason_intrinsic,
4023 Deoptimization::Action_maybe_recompile);
4024 }
4025
4026 if (!stopped()) {
4027 // How many elements will we copy from the original?
4028 // The answer is MinI(orig_length - start, length).
4029 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
4030 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4031
4032 newcopy = new_array(klass_node, length, 0); // no argments to push
4033
4034 // Generate a direct call to the right arraycopy function(s).
4035 // We know the copy is disjoint but we might not know if the
4036 // oop stores need checking.
4037 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4038 // This will fail a store-check if x contains any non-nulls.
4039 bool disjoint_bases = true;
4040 // if start > orig_length then the length of the copy may be
4041 // negative.
4042 bool length_never_negative = !is_copyOfRange;
4043 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4044 original, start, newcopy, intcon(0), moved,
4045 disjoint_bases, length_never_negative);
4046 }
4047 } // original reexecute is set back here
4048
4049 C->set_has_split_ifs(true); // Has chance for split-if optimization
4050 if (!stopped()) {
4051 set_result(newcopy);
4052 }
4053 return true;
4054 }
4055
4056
4057 //----------------------generate_virtual_guard---------------------------
4058 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4059 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4060 RegionNode* slow_region) {
4061 ciMethod* method = callee();
4062 int vtable_index = method->vtable_index();
4063 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4064 err_msg_res("bad index %d", vtable_index));
4065 // Get the Method* out of the appropriate vtable entry.
4066 int entry_offset = (InstanceKlass::vtable_start_offset() +
4067 vtable_index*vtableEntry::size()) * wordSize +
4068 vtableEntry::method_offset_in_bytes();
4069 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4070 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4071
4072 // Compare the target method with the expected method (e.g., Object.hashCode).
4073 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4074
4075 Node* native_call = makecon(native_call_addr);
4076 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4077 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4078
4079 return generate_slow_guard(test_native, slow_region);
4080 }
4081
4082 //-----------------------generate_method_call----------------------------
4083 // Use generate_method_call to make a slow-call to the real
4084 // method if the fast path fails. An alternative would be to
4085 // use a stub like OptoRuntime::slow_arraycopy_Java.
4086 // This only works for expanding the current library call,
4087 // not another intrinsic. (E.g., don't use this for making an
4088 // arraycopy call inside of the copyOf intrinsic.)
4089 CallJavaNode*
4090 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4091 // When compiling the intrinsic method itself, do not use this technique.
4092 guarantee(callee() != C->method(), "cannot make slow-call to self");
4093
4094 ciMethod* method = callee();
4095 // ensure the JVMS we have will be correct for this call
4096 guarantee(method_id == method->intrinsic_id(), "must match");
4097
4098 const TypeFunc* tf = TypeFunc::make(method);
4099 CallJavaNode* slow_call;
4100 if (is_static) {
4101 assert(!is_virtual, "");
4102 slow_call = new(C) CallStaticJavaNode(C, tf,
4103 SharedRuntime::get_resolve_static_call_stub(),
4104 method, bci());
4105 } else if (is_virtual) {
4106 null_check_receiver();
4107 int vtable_index = Method::invalid_vtable_index;
4108 if (UseInlineCaches) {
4109 // Suppress the vtable call
4110 } else {
4111 // hashCode and clone are not a miranda methods,
4112 // so the vtable index is fixed.
4113 // No need to use the linkResolver to get it.
4114 vtable_index = method->vtable_index();
4115 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4116 err_msg_res("bad index %d", vtable_index));
4117 }
4118 slow_call = new(C) CallDynamicJavaNode(tf,
4119 SharedRuntime::get_resolve_virtual_call_stub(),
4120 method, vtable_index, bci());
4121 } else { // neither virtual nor static: opt_virtual
4122 null_check_receiver();
4123 slow_call = new(C) CallStaticJavaNode(C, tf,
4124 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4125 method, bci());
4126 slow_call->set_optimized_virtual(true);
4127 }
4128 set_arguments_for_java_call(slow_call);
4129 set_edges_for_java_call(slow_call);
4130 return slow_call;
4131 }
4132
4133
4134 /**
4135 * Build special case code for calls to hashCode on an object. This call may
4136 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4137 * slightly different code.
4138 */
4139 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4140 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4141 assert(!(is_virtual && is_static), "either virtual, special, or static");
4142
4143 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4144
4145 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4146 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4147 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4148 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4149 Node* obj = NULL;
4150 if (!is_static) {
4151 // Check for hashing null object
4152 obj = null_check_receiver();
4153 if (stopped()) return true; // unconditionally null
4154 result_reg->init_req(_null_path, top());
4155 result_val->init_req(_null_path, top());
4156 } else {
4157 // Do a null check, and return zero if null.
4158 // System.identityHashCode(null) == 0
4159 obj = argument(0);
4160 Node* null_ctl = top();
4161 obj = null_check_oop(obj, &null_ctl);
4162 result_reg->init_req(_null_path, null_ctl);
4163 result_val->init_req(_null_path, _gvn.intcon(0));
4164 }
4165
4166 // Unconditionally null? Then return right away.
4167 if (stopped()) {
4168 set_control( result_reg->in(_null_path));
4169 if (!stopped())
4170 set_result(result_val->in(_null_path));
4171 return true;
4172 }
4173
4174 // We only go to the fast case code if we pass a number of guards. The
4175 // paths which do not pass are accumulated in the slow_region.
4176 RegionNode* slow_region = new (C) RegionNode(1);
4177 record_for_igvn(slow_region);
4178
4179 // If this is a virtual call, we generate a funny guard. We pull out
4180 // the vtable entry corresponding to hashCode() from the target object.
4181 // If the target method which we are calling happens to be the native
4182 // Object hashCode() method, we pass the guard. We do not need this
4183 // guard for non-virtual calls -- the caller is known to be the native
4184 // Object hashCode().
4185 if (is_virtual) {
4186 // After null check, get the object's klass.
4187 Node* obj_klass = load_object_klass(obj);
4188 generate_virtual_guard(obj_klass, slow_region);
4189 }
4190
4191 // Get the header out of the object, use LoadMarkNode when available
4192 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4193 // The control of the load must be NULL. Otherwise, the load can move before
4194 // the null check after castPP removal.
4195 Node* no_ctrl = NULL;
4196 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4197
4198 // Test the header to see if it is unlocked.
4199 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4200 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4201 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4202 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4203 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4204
4205 generate_slow_guard(test_unlocked, slow_region);
4206
4207 // Get the hash value and check to see that it has been properly assigned.
4208 // We depend on hash_mask being at most 32 bits and avoid the use of
4209 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4210 // vm: see markOop.hpp.
4211 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4212 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4213 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4214 // This hack lets the hash bits live anywhere in the mark object now, as long
4215 // as the shift drops the relevant bits into the low 32 bits. Note that
4216 // Java spec says that HashCode is an int so there's no point in capturing
4217 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4218 hshifted_header = ConvX2I(hshifted_header);
4219 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4220
4221 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4222 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4223 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4224
4225 generate_slow_guard(test_assigned, slow_region);
4226
4227 Node* init_mem = reset_memory();
4228 // fill in the rest of the null path:
4229 result_io ->init_req(_null_path, i_o());
4230 result_mem->init_req(_null_path, init_mem);
4231
4232 result_val->init_req(_fast_path, hash_val);
4233 result_reg->init_req(_fast_path, control());
4234 result_io ->init_req(_fast_path, i_o());
4235 result_mem->init_req(_fast_path, init_mem);
4236
4237 // Generate code for the slow case. We make a call to hashCode().
4238 set_control(_gvn.transform(slow_region));
4239 if (!stopped()) {
4240 // No need for PreserveJVMState, because we're using up the present state.
4241 set_all_memory(init_mem);
4242 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4243 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4244 Node* slow_result = set_results_for_java_call(slow_call);
4245 // this->control() comes from set_results_for_java_call
4246 result_reg->init_req(_slow_path, control());
4247 result_val->init_req(_slow_path, slow_result);
4248 result_io ->set_req(_slow_path, i_o());
4249 result_mem ->set_req(_slow_path, reset_memory());
4250 }
4251
4252 // Return the combined state.
4253 set_i_o( _gvn.transform(result_io) );
4254 set_all_memory( _gvn.transform(result_mem));
4255
4256 set_result(result_reg, result_val);
4257 return true;
4258 }
4259
4260 //---------------------------inline_native_getClass----------------------------
4261 // public final native Class<?> java.lang.Object.getClass();
4262 //
4263 // Build special case code for calls to getClass on an object.
4264 bool LibraryCallKit::inline_native_getClass() {
4265 Node* obj = null_check_receiver();
4266 if (stopped()) return true;
4267 set_result(load_mirror_from_klass(load_object_klass(obj)));
4268 return true;
4269 }
4270
4271 //-----------------inline_native_Reflection_getCallerClass---------------------
4272 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4273 //
4274 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4275 //
4276 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4277 // in that it must skip particular security frames and checks for
4278 // caller sensitive methods.
4279 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4280 #ifndef PRODUCT
4281 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4282 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4283 }
4284 #endif
4285
4286 if (!jvms()->has_method()) {
4287 #ifndef PRODUCT
4288 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4289 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4290 }
4291 #endif
4292 return false;
4293 }
4294
4295 // Walk back up the JVM state to find the caller at the required
4296 // depth.
4297 JVMState* caller_jvms = jvms();
4298
4299 // Cf. JVM_GetCallerClass
4300 // NOTE: Start the loop at depth 1 because the current JVM state does
4301 // not include the Reflection.getCallerClass() frame.
4302 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4303 ciMethod* m = caller_jvms->method();
4304 switch (n) {
4305 case 0:
4306 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4307 break;
4308 case 1:
4309 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4310 if (!m->caller_sensitive()) {
4311 #ifndef PRODUCT
4312 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4313 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4314 }
4315 #endif
4316 return false; // bail-out; let JVM_GetCallerClass do the work
4317 }
4318 break;
4319 default:
4320 if (!m->is_ignored_by_security_stack_walk()) {
4321 // We have reached the desired frame; return the holder class.
4322 // Acquire method holder as java.lang.Class and push as constant.
4323 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4324 ciInstance* caller_mirror = caller_klass->java_mirror();
4325 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4326
4327 #ifndef PRODUCT
4328 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4329 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());
4330 tty->print_cr(" JVM state at this point:");
4331 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4332 ciMethod* m = jvms()->of_depth(i)->method();
4333 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4334 }
4335 }
4336 #endif
4337 return true;
4338 }
4339 break;
4340 }
4341 }
4342
4343 #ifndef PRODUCT
4344 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4345 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4346 tty->print_cr(" JVM state at this point:");
4347 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4348 ciMethod* m = jvms()->of_depth(i)->method();
4349 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4350 }
4351 }
4352 #endif
4353
4354 return false; // bail-out; let JVM_GetCallerClass do the work
4355 }
4356
4357 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4358 Node* arg = argument(0);
4359 Node* result = NULL;
4360
4361 switch (id) {
4362 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4363 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4364 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4365 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4366
4367 case vmIntrinsics::_doubleToLongBits: {
4368 // two paths (plus control) merge in a wood
4369 RegionNode *r = new (C) RegionNode(3);
4370 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4371
4372 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4373 // Build the boolean node
4374 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4375
4376 // Branch either way.
4377 // NaN case is less traveled, which makes all the difference.
4378 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4379 Node *opt_isnan = _gvn.transform(ifisnan);
4380 assert( opt_isnan->is_If(), "Expect an IfNode");
4381 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4382 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4383
4384 set_control(iftrue);
4385
4386 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4387 Node *slow_result = longcon(nan_bits); // return NaN
4388 phi->init_req(1, _gvn.transform( slow_result ));
4389 r->init_req(1, iftrue);
4390
4391 // Else fall through
4392 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4393 set_control(iffalse);
4394
4395 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4396 r->init_req(2, iffalse);
4397
4398 // Post merge
4399 set_control(_gvn.transform(r));
4400 record_for_igvn(r);
4401
4402 C->set_has_split_ifs(true); // Has chance for split-if optimization
4403 result = phi;
4404 assert(result->bottom_type()->isa_long(), "must be");
4405 break;
4406 }
4407
4408 case vmIntrinsics::_floatToIntBits: {
4409 // two paths (plus control) merge in a wood
4410 RegionNode *r = new (C) RegionNode(3);
4411 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4412
4413 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4414 // Build the boolean node
4415 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4416
4417 // Branch either way.
4418 // NaN case is less traveled, which makes all the difference.
4419 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4420 Node *opt_isnan = _gvn.transform(ifisnan);
4421 assert( opt_isnan->is_If(), "Expect an IfNode");
4422 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4423 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4424
4425 set_control(iftrue);
4426
4427 static const jint nan_bits = 0x7fc00000;
4428 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4429 phi->init_req(1, _gvn.transform( slow_result ));
4430 r->init_req(1, iftrue);
4431
4432 // Else fall through
4433 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4434 set_control(iffalse);
4435
4436 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4437 r->init_req(2, iffalse);
4438
4439 // Post merge
4440 set_control(_gvn.transform(r));
4441 record_for_igvn(r);
4442
4443 C->set_has_split_ifs(true); // Has chance for split-if optimization
4444 result = phi;
4445 assert(result->bottom_type()->isa_int(), "must be");
4446 break;
4447 }
4448
4449 default:
4450 fatal_unexpected_iid(id);
4451 break;
4452 }
4453 set_result(_gvn.transform(result));
4454 return true;
4455 }
4456
4457 #ifdef _LP64
4458 #define XTOP ,top() /*additional argument*/
4459 #else //_LP64
4460 #define XTOP /*no additional argument*/
4461 #endif //_LP64
4462
4463 //----------------------inline_unsafe_copyMemory-------------------------
4464 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4465 bool LibraryCallKit::inline_unsafe_copyMemory() {
4466 if (callee()->is_static()) return false; // caller must have the capability!
4467 null_check_receiver(); // null-check receiver
4468 if (stopped()) return true;
4469
4470 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4471
4472 Node* src_ptr = argument(1); // type: oop
4473 Node* src_off = ConvL2X(argument(2)); // type: long
4474 Node* dst_ptr = argument(4); // type: oop
4475 Node* dst_off = ConvL2X(argument(5)); // type: long
4476 Node* size = ConvL2X(argument(7)); // type: long
4477
4478 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4479 "fieldOffset must be byte-scaled");
4480
4481 Node* src = make_unsafe_address(src_ptr, src_off);
4482 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4483
4484 // Conservatively insert a memory barrier on all memory slices.
4485 // Do not let writes of the copy source or destination float below the copy.
4486 insert_mem_bar(Op_MemBarCPUOrder);
4487
4488 // Call it. Note that the length argument is not scaled.
4489 make_runtime_call(RC_LEAF|RC_NO_FP,
4490 OptoRuntime::fast_arraycopy_Type(),
4491 StubRoutines::unsafe_arraycopy(),
4492 "unsafe_arraycopy",
4493 TypeRawPtr::BOTTOM,
4494 src, dst, size XTOP);
4495
4496 // Do not let reads of the copy destination float above the copy.
4497 insert_mem_bar(Op_MemBarCPUOrder);
4498
4499 return true;
4500 }
4501
4502 //------------------------clone_coping-----------------------------------
4503 // Helper function for inline_native_clone.
4504 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4505 assert(obj_size != NULL, "");
4506 Node* raw_obj = alloc_obj->in(1);
4507 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4508
4509 AllocateNode* alloc = NULL;
4510 if (ReduceBulkZeroing) {
4511 // We will be completely responsible for initializing this object -
4512 // mark Initialize node as complete.
4513 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4514 // The object was just allocated - there should be no any stores!
4515 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4516 // Mark as complete_with_arraycopy so that on AllocateNode
4517 // expansion, we know this AllocateNode is initialized by an array
4518 // copy and a StoreStore barrier exists after the array copy.
4519 alloc->initialization()->set_complete_with_arraycopy();
4520 }
4521
4522 // Copy the fastest available way.
4523 // TODO: generate fields copies for small objects instead.
4524 Node* src = obj;
4525 Node* dest = alloc_obj;
4526 Node* size = _gvn.transform(obj_size);
4527
4528 // Exclude the header but include array length to copy by 8 bytes words.
4529 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4530 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4531 instanceOopDesc::base_offset_in_bytes();
4532 // base_off:
4533 // 8 - 32-bit VM
4534 // 12 - 64-bit VM, compressed klass
4535 // 16 - 64-bit VM, normal klass
4536 if (base_off % BytesPerLong != 0) {
4537 assert(UseCompressedClassPointers, "");
4538 if (is_array) {
4539 // Exclude length to copy by 8 bytes words.
4540 base_off += sizeof(int);
4541 } else {
4542 // Include klass to copy by 8 bytes words.
4543 base_off = instanceOopDesc::klass_offset_in_bytes();
4544 }
4545 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4546 }
4547 src = basic_plus_adr(src, base_off);
4548 dest = basic_plus_adr(dest, base_off);
4549
4550 // Compute the length also, if needed:
4551 Node* countx = size;
4552 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4553 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4554
4555 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4556 bool disjoint_bases = true;
4557 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4558 src, NULL, dest, NULL, countx,
4559 /*dest_uninitialized*/true);
4560
4561 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4562 if (card_mark) {
4563 assert(!is_array, "");
4564 // Put in store barrier for any and all oops we are sticking
4565 // into this object. (We could avoid this if we could prove
4566 // that the object type contains no oop fields at all.)
4567 Node* no_particular_value = NULL;
4568 Node* no_particular_field = NULL;
4569 int raw_adr_idx = Compile::AliasIdxRaw;
4570 post_barrier(control(),
4571 memory(raw_adr_type),
4572 alloc_obj,
4573 no_particular_field,
4574 raw_adr_idx,
4575 no_particular_value,
4576 T_OBJECT,
4577 false);
4578 }
4579
4580 // Do not let reads from the cloned object float above the arraycopy.
4581 if (alloc != NULL) {
4582 // Do not let stores that initialize this object be reordered with
4583 // a subsequent store that would make this object accessible by
4584 // other threads.
4585 // Record what AllocateNode this StoreStore protects so that
4586 // escape analysis can go from the MemBarStoreStoreNode to the
4587 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4588 // based on the escape status of the AllocateNode.
4589 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4590 } else {
4591 insert_mem_bar(Op_MemBarCPUOrder);
4592 }
4593 }
4594
4595 //------------------------inline_native_clone----------------------------
4596 // protected native Object java.lang.Object.clone();
4597 //
4598 // Here are the simple edge cases:
4599 // null receiver => normal trap
4600 // virtual and clone was overridden => slow path to out-of-line clone
4601 // not cloneable or finalizer => slow path to out-of-line Object.clone
4602 //
4603 // The general case has two steps, allocation and copying.
4604 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4605 //
4606 // Copying also has two cases, oop arrays and everything else.
4607 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4608 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4609 //
4610 // These steps fold up nicely if and when the cloned object's klass
4611 // can be sharply typed as an object array, a type array, or an instance.
4612 //
4613 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4614 PhiNode* result_val;
4615
4616 // Set the reexecute bit for the interpreter to reexecute
4617 // the bytecode that invokes Object.clone if deoptimization happens.
4618 { PreserveReexecuteState preexecs(this);
4619 jvms()->set_should_reexecute(true);
4620
4621 Node* obj = null_check_receiver();
4622 if (stopped()) return true;
4623
4624 Node* obj_klass = load_object_klass(obj);
4625 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4626 const TypeOopPtr* toop = ((tklass != NULL)
4627 ? tklass->as_instance_type()
4628 : TypeInstPtr::NOTNULL);
4629
4630 // Conservatively insert a memory barrier on all memory slices.
4631 // Do not let writes into the original float below the clone.
4632 insert_mem_bar(Op_MemBarCPUOrder);
4633
4634 // paths into result_reg:
4635 enum {
4636 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4637 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4638 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4639 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4640 PATH_LIMIT
4641 };
4642 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4643 result_val = new(C) PhiNode(result_reg,
4644 TypeInstPtr::NOTNULL);
4645 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4646 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4647 TypePtr::BOTTOM);
4648 record_for_igvn(result_reg);
4649
4650 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4651 int raw_adr_idx = Compile::AliasIdxRaw;
4652
4653 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4654 if (array_ctl != NULL) {
4655 // It's an array.
4656 PreserveJVMState pjvms(this);
4657 set_control(array_ctl);
4658 Node* obj_length = load_array_length(obj);
4659 Node* obj_size = NULL;
4660 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4661
4662 if (!use_ReduceInitialCardMarks()) {
4663 // If it is an oop array, it requires very special treatment,
4664 // because card marking is required on each card of the array.
4665 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4666 if (is_obja != NULL) {
4667 PreserveJVMState pjvms2(this);
4668 set_control(is_obja);
4669 // Generate a direct call to the right arraycopy function(s).
4670 bool disjoint_bases = true;
4671 bool length_never_negative = true;
4672 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4673 obj, intcon(0), alloc_obj, intcon(0),
4674 obj_length,
4675 disjoint_bases, length_never_negative);
4676 result_reg->init_req(_objArray_path, control());
4677 result_val->init_req(_objArray_path, alloc_obj);
4678 result_i_o ->set_req(_objArray_path, i_o());
4679 result_mem ->set_req(_objArray_path, reset_memory());
4680 }
4681 }
4682 // Otherwise, there are no card marks to worry about.
4683 // (We can dispense with card marks if we know the allocation
4684 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4685 // causes the non-eden paths to take compensating steps to
4686 // simulate a fresh allocation, so that no further
4687 // card marks are required in compiled code to initialize
4688 // the object.)
4689
4690 if (!stopped()) {
4691 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4692
4693 // Present the results of the copy.
4694 result_reg->init_req(_array_path, control());
4695 result_val->init_req(_array_path, alloc_obj);
4696 result_i_o ->set_req(_array_path, i_o());
4697 result_mem ->set_req(_array_path, reset_memory());
4698 }
4699 }
4700
4701 // We only go to the instance fast case code if we pass a number of guards.
4702 // The paths which do not pass are accumulated in the slow_region.
4703 RegionNode* slow_region = new (C) RegionNode(1);
4704 record_for_igvn(slow_region);
4705 if (!stopped()) {
4706 // It's an instance (we did array above). Make the slow-path tests.
4707 // If this is a virtual call, we generate a funny guard. We grab
4708 // the vtable entry corresponding to clone() from the target object.
4709 // If the target method which we are calling happens to be the
4710 // Object clone() method, we pass the guard. We do not need this
4711 // guard for non-virtual calls; the caller is known to be the native
4712 // Object clone().
4713 if (is_virtual) {
4714 generate_virtual_guard(obj_klass, slow_region);
4715 }
4716
4717 // The object must be cloneable and must not have a finalizer.
4718 // Both of these conditions may be checked in a single test.
4719 // We could optimize the cloneable test further, but we don't care.
4720 generate_access_flags_guard(obj_klass,
4721 // Test both conditions:
4722 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4723 // Must be cloneable but not finalizer:
4724 JVM_ACC_IS_CLONEABLE,
4725 slow_region);
4726 }
4727
4728 if (!stopped()) {
4729 // It's an instance, and it passed the slow-path tests.
4730 PreserveJVMState pjvms(this);
4731 Node* obj_size = NULL;
4732 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4733 // is reexecuted if deoptimization occurs and there could be problems when merging
4734 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4735 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4736
4737 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4738
4739 // Present the results of the slow call.
4740 result_reg->init_req(_instance_path, control());
4741 result_val->init_req(_instance_path, alloc_obj);
4742 result_i_o ->set_req(_instance_path, i_o());
4743 result_mem ->set_req(_instance_path, reset_memory());
4744 }
4745
4746 // Generate code for the slow case. We make a call to clone().
4747 set_control(_gvn.transform(slow_region));
4748 if (!stopped()) {
4749 PreserveJVMState pjvms(this);
4750 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4751 Node* slow_result = set_results_for_java_call(slow_call);
4752 // this->control() comes from set_results_for_java_call
4753 result_reg->init_req(_slow_path, control());
4754 result_val->init_req(_slow_path, slow_result);
4755 result_i_o ->set_req(_slow_path, i_o());
4756 result_mem ->set_req(_slow_path, reset_memory());
4757 }
4758
4759 // Return the combined state.
4760 set_control( _gvn.transform(result_reg));
4761 set_i_o( _gvn.transform(result_i_o));
4762 set_all_memory( _gvn.transform(result_mem));
4763 } // original reexecute is set back here
4764
4765 set_result(_gvn.transform(result_val));
4766 return true;
4767 }
4768
4769 //------------------------------basictype2arraycopy----------------------------
4770 address LibraryCallKit::basictype2arraycopy(BasicType t,
4771 Node* src_offset,
4772 Node* dest_offset,
4773 bool disjoint_bases,
4774 const char* &name,
4775 bool dest_uninitialized) {
4776 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4777 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4778
4779 bool aligned = false;
4780 bool disjoint = disjoint_bases;
4781
4782 // if the offsets are the same, we can treat the memory regions as
4783 // disjoint, because either the memory regions are in different arrays,
4784 // or they are identical (which we can treat as disjoint.) We can also
4785 // treat a copy with a destination index less that the source index
4786 // as disjoint since a low->high copy will work correctly in this case.
4787 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4788 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4789 // both indices are constants
4790 int s_offs = src_offset_inttype->get_con();
4791 int d_offs = dest_offset_inttype->get_con();
4792 int element_size = type2aelembytes(t);
4793 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4794 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4795 if (s_offs >= d_offs) disjoint = true;
4796 } else if (src_offset == dest_offset && src_offset != NULL) {
4797 // This can occur if the offsets are identical non-constants.
4798 disjoint = true;
4799 }
4800
4801 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4802 }
4803
4804
4805 //------------------------------inline_arraycopy-----------------------
4806 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4807 // Object dest, int destPos,
4808 // int length);
4809 bool LibraryCallKit::inline_arraycopy() {
4810 // Get the arguments.
4811 Node* src = argument(0); // type: oop
4812 Node* src_offset = argument(1); // type: int
4813 Node* dest = argument(2); // type: oop
4814 Node* dest_offset = argument(3); // type: int
4815 Node* length = argument(4); // type: int
4816
4817 // Compile time checks. If any of these checks cannot be verified at compile time,
4818 // we do not make a fast path for this call. Instead, we let the call remain as it
4819 // is. The checks we choose to mandate at compile time are:
4820 //
4821 // (1) src and dest are arrays.
4822 const Type* src_type = src->Value(&_gvn);
4823 const Type* dest_type = dest->Value(&_gvn);
4824 const TypeAryPtr* top_src = src_type->isa_aryptr();
4825 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4826
4827 // Do we have the type of src?
4828 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4829 // Do we have the type of dest?
4830 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4831 // Is the type for src from speculation?
4832 bool src_spec = false;
4833 // Is the type for dest from speculation?
4834 bool dest_spec = false;
4835
4836 if (!has_src || !has_dest) {
4837 // We don't have sufficient type information, let's see if
4838 // speculative types can help. We need to have types for both src
4839 // and dest so that it pays off.
4840
4841 // Do we already have or could we have type information for src
4842 bool could_have_src = has_src;
4843 // Do we already have or could we have type information for dest
4844 bool could_have_dest = has_dest;
4845
4846 ciKlass* src_k = NULL;
4847 if (!has_src) {
4848 src_k = src_type->speculative_type();
4849 if (src_k != NULL && src_k->is_array_klass()) {
4850 could_have_src = true;
4851 }
4852 }
4853
4854 ciKlass* dest_k = NULL;
4855 if (!has_dest) {
4856 dest_k = dest_type->speculative_type();
4857 if (dest_k != NULL && dest_k->is_array_klass()) {
4858 could_have_dest = true;
4859 }
4860 }
4861
4862 if (could_have_src && could_have_dest) {
4863 // This is going to pay off so emit the required guards
4864 if (!has_src) {
4865 src = maybe_cast_profiled_obj(src, src_k);
4866 src_type = _gvn.type(src);
4867 top_src = src_type->isa_aryptr();
4868 has_src = (top_src != NULL && top_src->klass() != NULL);
4869 src_spec = true;
4870 }
4871 if (!has_dest) {
4872 dest = maybe_cast_profiled_obj(dest, dest_k);
4873 dest_type = _gvn.type(dest);
4874 top_dest = dest_type->isa_aryptr();
4875 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4876 dest_spec = true;
4877 }
4878 }
4879 }
4880
4881 if (!has_src || !has_dest) {
4882 // Conservatively insert a memory barrier on all memory slices.
4883 // Do not let writes into the source float below the arraycopy.
4884 insert_mem_bar(Op_MemBarCPUOrder);
4885
4886 // Call StubRoutines::generic_arraycopy stub.
4887 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4888 src, src_offset, dest, dest_offset, length);
4889
4890 // Do not let reads from the destination float above the arraycopy.
4891 // Since we cannot type the arrays, we don't know which slices
4892 // might be affected. We could restrict this barrier only to those
4893 // memory slices which pertain to array elements--but don't bother.
4894 if (!InsertMemBarAfterArraycopy)
4895 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4896 insert_mem_bar(Op_MemBarCPUOrder);
4897 return true;
4898 }
4899
4900 // (2) src and dest arrays must have elements of the same BasicType
4901 // Figure out the size and type of the elements we will be copying.
4902 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4903 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4904 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4905 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4906
4907 if (src_elem != dest_elem || dest_elem == T_VOID) {
4908 // The component types are not the same or are not recognized. Punt.
4909 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4910 generate_slow_arraycopy(TypePtr::BOTTOM,
4911 src, src_offset, dest, dest_offset, length,
4912 /*dest_uninitialized*/false);
4913 return true;
4914 }
4915
4916 if (src_elem == T_OBJECT) {
4917 // If both arrays are object arrays then having the exact types
4918 // for both will remove the need for a subtype check at runtime
4919 // before the call and may make it possible to pick a faster copy
4920 // routine (without a subtype check on every element)
4921 // Do we have the exact type of src?
4922 bool could_have_src = src_spec;
4923 // Do we have the exact type of dest?
4924 bool could_have_dest = dest_spec;
4925 ciKlass* src_k = top_src->klass();
4926 ciKlass* dest_k = top_dest->klass();
4927 if (!src_spec) {
4928 src_k = src_type->speculative_type();
4929 if (src_k != NULL && src_k->is_array_klass()) {
4930 could_have_src = true;
4931 }
4932 }
4933 if (!dest_spec) {
4934 dest_k = dest_type->speculative_type();
4935 if (dest_k != NULL && dest_k->is_array_klass()) {
4936 could_have_dest = true;
4937 }
4938 }
4939 if (could_have_src && could_have_dest) {
4940 // If we can have both exact types, emit the missing guards
4941 if (could_have_src && !src_spec) {
4942 src = maybe_cast_profiled_obj(src, src_k);
4943 }
4944 if (could_have_dest && !dest_spec) {
4945 dest = maybe_cast_profiled_obj(dest, dest_k);
4946 }
4947 }
4948 }
4949
4950 //---------------------------------------------------------------------------
4951 // We will make a fast path for this call to arraycopy.
4952
4953 // We have the following tests left to perform:
4954 //
4955 // (3) src and dest must not be null.
4956 // (4) src_offset must not be negative.
4957 // (5) dest_offset must not be negative.
4958 // (6) length must not be negative.
4959 // (7) src_offset + length must not exceed length of src.
4960 // (8) dest_offset + length must not exceed length of dest.
4961 // (9) each element of an oop array must be assignable
4962
4963 RegionNode* slow_region = new (C) RegionNode(1);
4964 record_for_igvn(slow_region);
4965
4966 // (3) operands must not be null
4967 // We currently perform our null checks with the null_check routine.
4968 // This means that the null exceptions will be reported in the caller
4969 // rather than (correctly) reported inside of the native arraycopy call.
4970 // This should be corrected, given time. We do our null check with the
4971 // stack pointer restored.
4972 src = null_check(src, T_ARRAY);
4973 dest = null_check(dest, T_ARRAY);
4974
4975 // (4) src_offset must not be negative.
4976 generate_negative_guard(src_offset, slow_region);
4977
4978 // (5) dest_offset must not be negative.
4979 generate_negative_guard(dest_offset, slow_region);
4980
4981 // (6) length must not be negative (moved to generate_arraycopy()).
4982 // generate_negative_guard(length, slow_region);
4983
4984 // (7) src_offset + length must not exceed length of src.
4985 generate_limit_guard(src_offset, length,
4986 load_array_length(src),
4987 slow_region);
4988
4989 // (8) dest_offset + length must not exceed length of dest.
4990 generate_limit_guard(dest_offset, length,
4991 load_array_length(dest),
4992 slow_region);
4993
4994 // (9) each element of an oop array must be assignable
4995 // The generate_arraycopy subroutine checks this.
4996
4997 // This is where the memory effects are placed:
4998 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4999 generate_arraycopy(adr_type, dest_elem,
5000 src, src_offset, dest, dest_offset, length,
5001 false, false, slow_region);
5002
5003 return true;
5004 }
5005
5006 //-----------------------------generate_arraycopy----------------------
5007 // Generate an optimized call to arraycopy.
5008 // Caller must guard against non-arrays.
5009 // Caller must determine a common array basic-type for both arrays.
5010 // Caller must validate offsets against array bounds.
5011 // The slow_region has already collected guard failure paths
5012 // (such as out of bounds length or non-conformable array types).
5013 // The generated code has this shape, in general:
5014 //
5015 // if (length == 0) return // via zero_path
5016 // slowval = -1
5017 // if (types unknown) {
5018 // slowval = call generic copy loop
5019 // if (slowval == 0) return // via checked_path
5020 // } else if (indexes in bounds) {
5021 // if ((is object array) && !(array type check)) {
5022 // slowval = call checked copy loop
5023 // if (slowval == 0) return // via checked_path
5024 // } else {
5025 // call bulk copy loop
5026 // return // via fast_path
5027 // }
5028 // }
5029 // // adjust params for remaining work:
5030 // if (slowval != -1) {
5031 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
5032 // }
5033 // slow_region:
5034 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
5035 // return // via slow_call_path
5036 //
5037 // This routine is used from several intrinsics: System.arraycopy,
5038 // Object.clone (the array subcase), and Arrays.copyOf[Range].
5039 //
5040 void
5041 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
5042 BasicType basic_elem_type,
5043 Node* src, Node* src_offset,
5044 Node* dest, Node* dest_offset,
5045 Node* copy_length,
5046 bool disjoint_bases,
5047 bool length_never_negative,
5048 RegionNode* slow_region) {
5049
5050 if (slow_region == NULL) {
5051 slow_region = new(C) RegionNode(1);
5052 record_for_igvn(slow_region);
5053 }
5054
5055 Node* original_dest = dest;
5056 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5057 bool dest_uninitialized = false;
5058
5059 // See if this is the initialization of a newly-allocated array.
5060 // If so, we will take responsibility here for initializing it to zero.
5061 // (Note: Because tightly_coupled_allocation performs checks on the
5062 // out-edges of the dest, we need to avoid making derived pointers
5063 // from it until we have checked its uses.)
5064 if (ReduceBulkZeroing
5065 && !ZeroTLAB // pointless if already zeroed
5066 && basic_elem_type != T_CONFLICT // avoid corner case
5067 && !src->eqv_uncast(dest)
5068 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5069 != NULL)
5070 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5071 && alloc->maybe_set_complete(&_gvn)) {
5072 // "You break it, you buy it."
5073 InitializeNode* init = alloc->initialization();
5074 assert(init->is_complete(), "we just did this");
5075 init->set_complete_with_arraycopy();
5076 assert(dest->is_CheckCastPP(), "sanity");
5077 assert(dest->in(0)->in(0) == init, "dest pinned");
5078 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5079 // From this point on, every exit path is responsible for
5080 // initializing any non-copied parts of the object to zero.
5081 // Also, if this flag is set we make sure that arraycopy interacts properly
5082 // with G1, eliding pre-barriers. See CR 6627983.
5083 dest_uninitialized = true;
5084 } else {
5085 // No zeroing elimination here.
5086 alloc = NULL;
5087 //original_dest = dest;
5088 //dest_uninitialized = false;
5089 }
5090
5091 // Results are placed here:
5092 enum { fast_path = 1, // normal void-returning assembly stub
5093 checked_path = 2, // special assembly stub with cleanup
5094 slow_call_path = 3, // something went wrong; call the VM
5095 zero_path = 4, // bypass when length of copy is zero
5096 bcopy_path = 5, // copy primitive array by 64-bit blocks
5097 PATH_LIMIT = 6
5098 };
5099 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5100 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5101 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5102 record_for_igvn(result_region);
5103 _gvn.set_type_bottom(result_i_o);
5104 _gvn.set_type_bottom(result_memory);
5105 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5106
5107 // The slow_control path:
5108 Node* slow_control;
5109 Node* slow_i_o = i_o();
5110 Node* slow_mem = memory(adr_type);
5111 debug_only(slow_control = (Node*) badAddress);
5112
5113 // Checked control path:
5114 Node* checked_control = top();
5115 Node* checked_mem = NULL;
5116 Node* checked_i_o = NULL;
5117 Node* checked_value = NULL;
5118
5119 if (basic_elem_type == T_CONFLICT) {
5120 assert(!dest_uninitialized, "");
5121 Node* cv = generate_generic_arraycopy(adr_type,
5122 src, src_offset, dest, dest_offset,
5123 copy_length, dest_uninitialized);
5124 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5125 checked_control = control();
5126 checked_i_o = i_o();
5127 checked_mem = memory(adr_type);
5128 checked_value = cv;
5129 set_control(top()); // no fast path
5130 }
5131
5132 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5133 if (not_pos != NULL) {
5134 PreserveJVMState pjvms(this);
5135 set_control(not_pos);
5136
5137 // (6) length must not be negative.
5138 if (!length_never_negative) {
5139 generate_negative_guard(copy_length, slow_region);
5140 }
5141
5142 // copy_length is 0.
5143 if (!stopped() && dest_uninitialized) {
5144 Node* dest_length = alloc->in(AllocateNode::ALength);
5145 if (copy_length->eqv_uncast(dest_length)
5146 || _gvn.find_int_con(dest_length, 1) <= 0) {
5147 // There is no zeroing to do. No need for a secondary raw memory barrier.
5148 } else {
5149 // Clear the whole thing since there are no source elements to copy.
5150 generate_clear_array(adr_type, dest, basic_elem_type,
5151 intcon(0), NULL,
5152 alloc->in(AllocateNode::AllocSize));
5153 // Use a secondary InitializeNode as raw memory barrier.
5154 // Currently it is needed only on this path since other
5155 // paths have stub or runtime calls as raw memory barriers.
5156 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5157 Compile::AliasIdxRaw,
5158 top())->as_Initialize();
5159 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5160 }
5161 }
5162
5163 // Present the results of the fast call.
5164 result_region->init_req(zero_path, control());
5165 result_i_o ->init_req(zero_path, i_o());
5166 result_memory->init_req(zero_path, memory(adr_type));
5167 }
5168
5169 if (!stopped() && dest_uninitialized) {
5170 // We have to initialize the *uncopied* part of the array to zero.
5171 // The copy destination is the slice dest[off..off+len]. The other slices
5172 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5173 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5174 Node* dest_length = alloc->in(AllocateNode::ALength);
5175 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5176 copy_length));
5177
5178 // If there is a head section that needs zeroing, do it now.
5179 if (find_int_con(dest_offset, -1) != 0) {
5180 generate_clear_array(adr_type, dest, basic_elem_type,
5181 intcon(0), dest_offset,
5182 NULL);
5183 }
5184
5185 // Next, perform a dynamic check on the tail length.
5186 // It is often zero, and we can win big if we prove this.
5187 // There are two wins: Avoid generating the ClearArray
5188 // with its attendant messy index arithmetic, and upgrade
5189 // the copy to a more hardware-friendly word size of 64 bits.
5190 Node* tail_ctl = NULL;
5191 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5192 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5193 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5194 tail_ctl = generate_slow_guard(bol_lt, NULL);
5195 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5196 }
5197
5198 // At this point, let's assume there is no tail.
5199 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5200 // There is no tail. Try an upgrade to a 64-bit copy.
5201 bool didit = false;
5202 { PreserveJVMState pjvms(this);
5203 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5204 src, src_offset, dest, dest_offset,
5205 dest_size, dest_uninitialized);
5206 if (didit) {
5207 // Present the results of the block-copying fast call.
5208 result_region->init_req(bcopy_path, control());
5209 result_i_o ->init_req(bcopy_path, i_o());
5210 result_memory->init_req(bcopy_path, memory(adr_type));
5211 }
5212 }
5213 if (didit)
5214 set_control(top()); // no regular fast path
5215 }
5216
5217 // Clear the tail, if any.
5218 if (tail_ctl != NULL) {
5219 Node* notail_ctl = stopped() ? NULL : control();
5220 set_control(tail_ctl);
5221 if (notail_ctl == NULL) {
5222 generate_clear_array(adr_type, dest, basic_elem_type,
5223 dest_tail, NULL,
5224 dest_size);
5225 } else {
5226 // Make a local merge.
5227 Node* done_ctl = new(C) RegionNode(3);
5228 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5229 done_ctl->init_req(1, notail_ctl);
5230 done_mem->init_req(1, memory(adr_type));
5231 generate_clear_array(adr_type, dest, basic_elem_type,
5232 dest_tail, NULL,
5233 dest_size);
5234 done_ctl->init_req(2, control());
5235 done_mem->init_req(2, memory(adr_type));
5236 set_control( _gvn.transform(done_ctl));
5237 set_memory( _gvn.transform(done_mem), adr_type );
5238 }
5239 }
5240 }
5241
5242 BasicType copy_type = basic_elem_type;
5243 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5244 if (!stopped() && copy_type == T_OBJECT) {
5245 // If src and dest have compatible element types, we can copy bits.
5246 // Types S[] and D[] are compatible if D is a supertype of S.
5247 //
5248 // If they are not, we will use checked_oop_disjoint_arraycopy,
5249 // which performs a fast optimistic per-oop check, and backs off
5250 // further to JVM_ArrayCopy on the first per-oop check that fails.
5251 // (Actually, we don't move raw bits only; the GC requires card marks.)
5252
5253 // Get the Klass* for both src and dest
5254 Node* src_klass = load_object_klass(src);
5255 Node* dest_klass = load_object_klass(dest);
5256
5257 // Generate the subtype check.
5258 // This might fold up statically, or then again it might not.
5259 //
5260 // Non-static example: Copying List<String>.elements to a new String[].
5261 // The backing store for a List<String> is always an Object[],
5262 // but its elements are always type String, if the generic types
5263 // are correct at the source level.
5264 //
5265 // Test S[] against D[], not S against D, because (probably)
5266 // the secondary supertype cache is less busy for S[] than S.
5267 // This usually only matters when D is an interface.
5268 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5269 // Plug failing path into checked_oop_disjoint_arraycopy
5270 if (not_subtype_ctrl != top()) {
5271 PreserveJVMState pjvms(this);
5272 set_control(not_subtype_ctrl);
5273 // (At this point we can assume disjoint_bases, since types differ.)
5274 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5275 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5276 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5277 Node* dest_elem_klass = _gvn.transform(n1);
5278 Node* cv = generate_checkcast_arraycopy(adr_type,
5279 dest_elem_klass,
5280 src, src_offset, dest, dest_offset,
5281 ConvI2X(copy_length), dest_uninitialized);
5282 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5283 checked_control = control();
5284 checked_i_o = i_o();
5285 checked_mem = memory(adr_type);
5286 checked_value = cv;
5287 }
5288 // At this point we know we do not need type checks on oop stores.
5289
5290 // Let's see if we need card marks:
5291 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5292 // If we do not need card marks, copy using the jint or jlong stub.
5293 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5294 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5295 "sizes agree");
5296 }
5297 }
5298
5299 if (!stopped()) {
5300 // Generate the fast path, if possible.
5301 PreserveJVMState pjvms(this);
5302 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5303 src, src_offset, dest, dest_offset,
5304 ConvI2X(copy_length), dest_uninitialized);
5305
5306 // Present the results of the fast call.
5307 result_region->init_req(fast_path, control());
5308 result_i_o ->init_req(fast_path, i_o());
5309 result_memory->init_req(fast_path, memory(adr_type));
5310 }
5311
5312 // Here are all the slow paths up to this point, in one bundle:
5313 slow_control = top();
5314 if (slow_region != NULL)
5315 slow_control = _gvn.transform(slow_region);
5316 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5317
5318 set_control(checked_control);
5319 if (!stopped()) {
5320 // Clean up after the checked call.
5321 // The returned value is either 0 or -1^K,
5322 // where K = number of partially transferred array elements.
5323 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5324 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5325 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5326
5327 // If it is 0, we are done, so transfer to the end.
5328 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5329 result_region->init_req(checked_path, checks_done);
5330 result_i_o ->init_req(checked_path, checked_i_o);
5331 result_memory->init_req(checked_path, checked_mem);
5332
5333 // If it is not zero, merge into the slow call.
5334 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5335 RegionNode* slow_reg2 = new(C) RegionNode(3);
5336 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5337 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5338 record_for_igvn(slow_reg2);
5339 slow_reg2 ->init_req(1, slow_control);
5340 slow_i_o2 ->init_req(1, slow_i_o);
5341 slow_mem2 ->init_req(1, slow_mem);
5342 slow_reg2 ->init_req(2, control());
5343 slow_i_o2 ->init_req(2, checked_i_o);
5344 slow_mem2 ->init_req(2, checked_mem);
5345
5346 slow_control = _gvn.transform(slow_reg2);
5347 slow_i_o = _gvn.transform(slow_i_o2);
5348 slow_mem = _gvn.transform(slow_mem2);
5349
5350 if (alloc != NULL) {
5351 // We'll restart from the very beginning, after zeroing the whole thing.
5352 // This can cause double writes, but that's OK since dest is brand new.
5353 // So we ignore the low 31 bits of the value returned from the stub.
5354 } else {
5355 // We must continue the copy exactly where it failed, or else
5356 // another thread might see the wrong number of writes to dest.
5357 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5358 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5359 slow_offset->init_req(1, intcon(0));
5360 slow_offset->init_req(2, checked_offset);
5361 slow_offset = _gvn.transform(slow_offset);
5362
5363 // Adjust the arguments by the conditionally incoming offset.
5364 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5365 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5366 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5367
5368 // Tweak the node variables to adjust the code produced below:
5369 src_offset = src_off_plus;
5370 dest_offset = dest_off_plus;
5371 copy_length = length_minus;
5372 }
5373 }
5374
5375 set_control(slow_control);
5376 if (!stopped()) {
5377 // Generate the slow path, if needed.
5378 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5379
5380 set_memory(slow_mem, adr_type);
5381 set_i_o(slow_i_o);
5382
5383 if (dest_uninitialized) {
5384 generate_clear_array(adr_type, dest, basic_elem_type,
5385 intcon(0), NULL,
5386 alloc->in(AllocateNode::AllocSize));
5387 }
5388
5389 generate_slow_arraycopy(adr_type,
5390 src, src_offset, dest, dest_offset,
5391 copy_length, /*dest_uninitialized*/false);
5392
5393 result_region->init_req(slow_call_path, control());
5394 result_i_o ->init_req(slow_call_path, i_o());
5395 result_memory->init_req(slow_call_path, memory(adr_type));
5396 }
5397
5398 // Remove unused edges.
5399 for (uint i = 1; i < result_region->req(); i++) {
5400 if (result_region->in(i) == NULL)
5401 result_region->init_req(i, top());
5402 }
5403
5404 // Finished; return the combined state.
5405 set_control( _gvn.transform(result_region));
5406 set_i_o( _gvn.transform(result_i_o) );
5407 set_memory( _gvn.transform(result_memory), adr_type );
5408
5409 // The memory edges above are precise in order to model effects around
5410 // array copies accurately to allow value numbering of field loads around
5411 // arraycopy. Such field loads, both before and after, are common in Java
5412 // collections and similar classes involving header/array data structures.
5413 //
5414 // But with low number of register or when some registers are used or killed
5415 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5416 // The next memory barrier is added to avoid it. If the arraycopy can be
5417 // optimized away (which it can, sometimes) then we can manually remove
5418 // the membar also.
5419 //
5420 // Do not let reads from the cloned object float above the arraycopy.
5421 if (alloc != NULL) {
5422 // Do not let stores that initialize this object be reordered with
5423 // a subsequent store that would make this object accessible by
5424 // other threads.
5425 // Record what AllocateNode this StoreStore protects so that
5426 // escape analysis can go from the MemBarStoreStoreNode to the
5427 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5428 // based on the escape status of the AllocateNode.
5429 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5430 } else if (InsertMemBarAfterArraycopy)
5431 insert_mem_bar(Op_MemBarCPUOrder);
5432 }
5433
5434
5435 // Helper function which determines if an arraycopy immediately follows
5436 // an allocation, with no intervening tests or other escapes for the object.
5437 AllocateArrayNode*
5438 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5439 RegionNode* slow_region) {
5440 if (stopped()) return NULL; // no fast path
5441 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5442
5443 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5444 if (alloc == NULL) return NULL;
5445
5446 Node* rawmem = memory(Compile::AliasIdxRaw);
5447 // Is the allocation's memory state untouched?
5448 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5449 // Bail out if there have been raw-memory effects since the allocation.
5450 // (Example: There might have been a call or safepoint.)
5451 return NULL;
5452 }
5453 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5454 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5455 return NULL;
5456 }
5457
5458 // There must be no unexpected observers of this allocation.
5459 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5460 Node* obs = ptr->fast_out(i);
5461 if (obs != this->map()) {
5462 return NULL;
5463 }
5464 }
5465
5466 // This arraycopy must unconditionally follow the allocation of the ptr.
5467 Node* alloc_ctl = ptr->in(0);
5468 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5469
5470 Node* ctl = control();
5471 while (ctl != alloc_ctl) {
5472 // There may be guards which feed into the slow_region.
5473 // Any other control flow means that we might not get a chance
5474 // to finish initializing the allocated object.
5475 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5476 IfNode* iff = ctl->in(0)->as_If();
5477 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5478 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5479 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5480 ctl = iff->in(0); // This test feeds the known slow_region.
5481 continue;
5482 }
5483 // One more try: Various low-level checks bottom out in
5484 // uncommon traps. If the debug-info of the trap omits
5485 // any reference to the allocation, as we've already
5486 // observed, then there can be no objection to the trap.
5487 bool found_trap = false;
5488 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5489 Node* obs = not_ctl->fast_out(j);
5490 if (obs->in(0) == not_ctl && obs->is_Call() &&
5491 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5492 found_trap = true; break;
5493 }
5494 }
5495 if (found_trap) {
5496 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5497 continue;
5498 }
5499 }
5500 return NULL;
5501 }
5502
5503 // If we get this far, we have an allocation which immediately
5504 // precedes the arraycopy, and we can take over zeroing the new object.
5505 // The arraycopy will finish the initialization, and provide
5506 // a new control state to which we will anchor the destination pointer.
5507
5508 return alloc;
5509 }
5510
5511 // Helper for initialization of arrays, creating a ClearArray.
5512 // It writes zero bits in [start..end), within the body of an array object.
5513 // The memory effects are all chained onto the 'adr_type' alias category.
5514 //
5515 // Since the object is otherwise uninitialized, we are free
5516 // to put a little "slop" around the edges of the cleared area,
5517 // as long as it does not go back into the array's header,
5518 // or beyond the array end within the heap.
5519 //
5520 // The lower edge can be rounded down to the nearest jint and the
5521 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5522 //
5523 // Arguments:
5524 // adr_type memory slice where writes are generated
5525 // dest oop of the destination array
5526 // basic_elem_type element type of the destination
5527 // slice_idx array index of first element to store
5528 // slice_len number of elements to store (or NULL)
5529 // dest_size total size in bytes of the array object
5530 //
5531 // Exactly one of slice_len or dest_size must be non-NULL.
5532 // If dest_size is non-NULL, zeroing extends to the end of the object.
5533 // If slice_len is non-NULL, the slice_idx value must be a constant.
5534 void
5535 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5536 Node* dest,
5537 BasicType basic_elem_type,
5538 Node* slice_idx,
5539 Node* slice_len,
5540 Node* dest_size) {
5541 // one or the other but not both of slice_len and dest_size:
5542 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5543 if (slice_len == NULL) slice_len = top();
5544 if (dest_size == NULL) dest_size = top();
5545
5546 // operate on this memory slice:
5547 Node* mem = memory(adr_type); // memory slice to operate on
5548
5549 // scaling and rounding of indexes:
5550 int scale = exact_log2(type2aelembytes(basic_elem_type));
5551 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5552 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5553 int bump_bit = (-1 << scale) & BytesPerInt;
5554
5555 // determine constant starts and ends
5556 const intptr_t BIG_NEG = -128;
5557 assert(BIG_NEG + 2*abase < 0, "neg enough");
5558 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5559 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5560 if (slice_len_con == 0) {
5561 return; // nothing to do here
5562 }
5563 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5564 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5565 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5566 assert(end_con < 0, "not two cons");
5567 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5568 BytesPerLong);
5569 }
5570
5571 if (start_con >= 0 && end_con >= 0) {
5572 // Constant start and end. Simple.
5573 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5574 start_con, end_con, &_gvn);
5575 } else if (start_con >= 0 && dest_size != top()) {
5576 // Constant start, pre-rounded end after the tail of the array.
5577 Node* end = dest_size;
5578 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5579 start_con, end, &_gvn);
5580 } else if (start_con >= 0 && slice_len != top()) {
5581 // Constant start, non-constant end. End needs rounding up.
5582 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5583 intptr_t end_base = abase + (slice_idx_con << scale);
5584 int end_round = (-1 << scale) & (BytesPerLong - 1);
5585 Node* end = ConvI2X(slice_len);
5586 if (scale != 0)
5587 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5588 end_base += end_round;
5589 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5590 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5591 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5592 start_con, end, &_gvn);
5593 } else if (start_con < 0 && dest_size != top()) {
5594 // Non-constant start, pre-rounded end after the tail of the array.
5595 // This is almost certainly a "round-to-end" operation.
5596 Node* start = slice_idx;
5597 start = ConvI2X(start);
5598 if (scale != 0)
5599 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5600 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5601 if ((bump_bit | clear_low) != 0) {
5602 int to_clear = (bump_bit | clear_low);
5603 // Align up mod 8, then store a jint zero unconditionally
5604 // just before the mod-8 boundary.
5605 if (((abase + bump_bit) & ~to_clear) - bump_bit
5606 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5607 bump_bit = 0;
5608 assert((abase & to_clear) == 0, "array base must be long-aligned");
5609 } else {
5610 // Bump 'start' up to (or past) the next jint boundary:
5611 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5612 assert((abase & clear_low) == 0, "array base must be int-aligned");
5613 }
5614 // Round bumped 'start' down to jlong boundary in body of array.
5615 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5616 if (bump_bit != 0) {
5617 // Store a zero to the immediately preceding jint:
5618 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5619 Node* p1 = basic_plus_adr(dest, x1);
5620 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5621 mem = _gvn.transform(mem);
5622 }
5623 }
5624 Node* end = dest_size; // pre-rounded
5625 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5626 start, end, &_gvn);
5627 } else {
5628 // Non-constant start, unrounded non-constant end.
5629 // (Nobody zeroes a random midsection of an array using this routine.)
5630 ShouldNotReachHere(); // fix caller
5631 }
5632
5633 // Done.
5634 set_memory(mem, adr_type);
5635 }
5636
5637
5638 bool
5639 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5640 BasicType basic_elem_type,
5641 AllocateNode* alloc,
5642 Node* src, Node* src_offset,
5643 Node* dest, Node* dest_offset,
5644 Node* dest_size, bool dest_uninitialized) {
5645 // See if there is an advantage from block transfer.
5646 int scale = exact_log2(type2aelembytes(basic_elem_type));
5647 if (scale >= LogBytesPerLong)
5648 return false; // it is already a block transfer
5649
5650 // Look at the alignment of the starting offsets.
5651 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5652
5653 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5654 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5655 if (src_off_con < 0 || dest_off_con < 0)
5656 // At present, we can only understand constants.
5657 return false;
5658
5659 intptr_t src_off = abase + (src_off_con << scale);
5660 intptr_t dest_off = abase + (dest_off_con << scale);
5661
5662 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5663 // Non-aligned; too bad.
5664 // One more chance: Pick off an initial 32-bit word.
5665 // This is a common case, since abase can be odd mod 8.
5666 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5667 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5668 Node* sptr = basic_plus_adr(src, src_off);
5669 Node* dptr = basic_plus_adr(dest, dest_off);
5670 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5671 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5672 src_off += BytesPerInt;
5673 dest_off += BytesPerInt;
5674 } else {
5675 return false;
5676 }
5677 }
5678 assert(src_off % BytesPerLong == 0, "");
5679 assert(dest_off % BytesPerLong == 0, "");
5680
5681 // Do this copy by giant steps.
5682 Node* sptr = basic_plus_adr(src, src_off);
5683 Node* dptr = basic_plus_adr(dest, dest_off);
5684 Node* countx = dest_size;
5685 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5686 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5687
5688 bool disjoint_bases = true; // since alloc != NULL
5689 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5690 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5691
5692 return true;
5693 }
5694
5695
5696 // Helper function; generates code for the slow case.
5697 // We make a call to a runtime method which emulates the native method,
5698 // but without the native wrapper overhead.
5699 void
5700 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5701 Node* src, Node* src_offset,
5702 Node* dest, Node* dest_offset,
5703 Node* copy_length, bool dest_uninitialized) {
5704 assert(!dest_uninitialized, "Invariant");
5705 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5706 OptoRuntime::slow_arraycopy_Type(),
5707 OptoRuntime::slow_arraycopy_Java(),
5708 "slow_arraycopy", adr_type,
5709 src, src_offset, dest, dest_offset,
5710 copy_length);
5711
5712 // Handle exceptions thrown by this fellow:
5713 make_slow_call_ex(call, env()->Throwable_klass(), false);
5714 }
5715
5716 // Helper function; generates code for cases requiring runtime checks.
5717 Node*
5718 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5719 Node* dest_elem_klass,
5720 Node* src, Node* src_offset,
5721 Node* dest, Node* dest_offset,
5722 Node* copy_length, bool dest_uninitialized) {
5723 if (stopped()) return NULL;
5724
5725 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5726 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5727 return NULL;
5728 }
5729
5730 // Pick out the parameters required to perform a store-check
5731 // for the target array. This is an optimistic check. It will
5732 // look in each non-null element's class, at the desired klass's
5733 // super_check_offset, for the desired klass.
5734 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5735 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5736 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5737 Node* check_offset = ConvI2X(_gvn.transform(n3));
5738 Node* check_value = dest_elem_klass;
5739
5740 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5741 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5742
5743 // (We know the arrays are never conjoint, because their types differ.)
5744 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5745 OptoRuntime::checkcast_arraycopy_Type(),
5746 copyfunc_addr, "checkcast_arraycopy", adr_type,
5747 // five arguments, of which two are
5748 // intptr_t (jlong in LP64)
5749 src_start, dest_start,
5750 copy_length XTOP,
5751 check_offset XTOP,
5752 check_value);
5753
5754 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5755 }
5756
5757
5758 // Helper function; generates code for cases requiring runtime checks.
5759 Node*
5760 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5761 Node* src, Node* src_offset,
5762 Node* dest, Node* dest_offset,
5763 Node* copy_length, bool dest_uninitialized) {
5764 assert(!dest_uninitialized, "Invariant");
5765 if (stopped()) return NULL;
5766 address copyfunc_addr = StubRoutines::generic_arraycopy();
5767 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5768 return NULL;
5769 }
5770
5771 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5772 OptoRuntime::generic_arraycopy_Type(),
5773 copyfunc_addr, "generic_arraycopy", adr_type,
5774 src, src_offset, dest, dest_offset, copy_length);
5775
5776 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5777 }
5778
5779 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5780 void
5781 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5782 BasicType basic_elem_type,
5783 bool disjoint_bases,
5784 Node* src, Node* src_offset,
5785 Node* dest, Node* dest_offset,
5786 Node* copy_length, bool dest_uninitialized) {
5787 if (stopped()) return; // nothing to do
5788
5789 Node* src_start = src;
5790 Node* dest_start = dest;
5791 if (src_offset != NULL || dest_offset != NULL) {
5792 assert(src_offset != NULL && dest_offset != NULL, "");
5793 src_start = array_element_address(src, src_offset, basic_elem_type);
5794 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5795 }
5796
5797 // Figure out which arraycopy runtime method to call.
5798 const char* copyfunc_name = "arraycopy";
5799 address copyfunc_addr =
5800 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5801 disjoint_bases, copyfunc_name, dest_uninitialized);
5802
5803 // Call it. Note that the count_ix value is not scaled to a byte-size.
5804 make_runtime_call(RC_LEAF|RC_NO_FP,
5805 OptoRuntime::fast_arraycopy_Type(),
5806 copyfunc_addr, copyfunc_name, adr_type,
5807 src_start, dest_start, copy_length XTOP);
5808 }
5809
5810 //-------------inline_encodeISOArray-----------------------------------
5811 // encode char[] to byte[] in ISO_8859_1
5812 bool LibraryCallKit::inline_encodeISOArray() {
5813 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5814 // no receiver since it is static method
5815 Node *src = argument(0);
5816 Node *src_offset = argument(1);
5817 Node *dst = argument(2);
5818 Node *dst_offset = argument(3);
5819 Node *length = argument(4);
5820
5821 const Type* src_type = src->Value(&_gvn);
5822 const Type* dst_type = dst->Value(&_gvn);
5823 const TypeAryPtr* top_src = src_type->isa_aryptr();
5824 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5825 if (top_src == NULL || top_src->klass() == NULL ||
5826 top_dest == NULL || top_dest->klass() == NULL) {
5827 // failed array check
5828 return false;
5829 }
5830
5831 // Figure out the size and type of the elements we will be copying.
5832 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5833 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5834 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5835 return false;
5836 }
5837 Node* src_start = array_element_address(src, src_offset, src_elem);
5838 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5839 // 'src_start' points to src array + scaled offset
5840 // 'dst_start' points to dst array + scaled offset
5841
5842 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5843 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5844 enc = _gvn.transform(enc);
5845 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5846 set_memory(res_mem, mtype);
5847 set_result(enc);
5848 return true;
5849 }
5850
5851 //-------------inline_multiplyToLen-----------------------------------
5852 bool LibraryCallKit::inline_multiplyToLen() {
5853 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5854
5855 address stubAddr = StubRoutines::multiplyToLen();
5856 if (stubAddr == NULL) {
5857 return false; // Intrinsic's stub is not implemented on this platform
5858 }
5859 const char* stubName = "multiplyToLen";
5860
5861 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5862
5863 // no receiver because it is a static method
5864 Node* x = argument(0);
5865 Node* xlen = argument(1);
5866 Node* y = argument(2);
5867 Node* ylen = argument(3);
5868 Node* z = argument(4);
5869
5870 const Type* x_type = x->Value(&_gvn);
5871 const Type* y_type = y->Value(&_gvn);
5872 const TypeAryPtr* top_x = x_type->isa_aryptr();
5873 const TypeAryPtr* top_y = y_type->isa_aryptr();
5874 if (top_x == NULL || top_x->klass() == NULL ||
5875 top_y == NULL || top_y->klass() == NULL) {
5876 // failed array check
5877 return false;
5878 }
5879
5880 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5881 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5882 if (x_elem != T_INT || y_elem != T_INT) {
5883 return false;
5884 }
5885
5886 // Set the original stack and the reexecute bit for the interpreter to reexecute
5887 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5888 // on the return from z array allocation in runtime.
5889 { PreserveReexecuteState preexecs(this);
5890 jvms()->set_should_reexecute(true);
5891
5892 Node* x_start = array_element_address(x, intcon(0), x_elem);
5893 Node* y_start = array_element_address(y, intcon(0), y_elem);
5894 // 'x_start' points to x array + scaled xlen
5895 // 'y_start' points to y array + scaled ylen
5896
5897 // Allocate the result array
5898 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5899 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5900 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5901
5902 IdealKit ideal(this);
5903
5904 #define __ ideal.
5905 Node* one = __ ConI(1);
5906 Node* zero = __ ConI(0);
5907 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5908 __ set(need_alloc, zero);
5909 __ set(z_alloc, z);
5910 __ if_then(z, BoolTest::eq, null()); {
5911 __ increment (need_alloc, one);
5912 } __ else_(); {
5913 // Update graphKit memory and control from IdealKit.
5914 sync_kit(ideal);
5915 Node* zlen_arg = load_array_length(z);
5916 // Update IdealKit memory and control from graphKit.
5917 __ sync_kit(this);
5918 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5919 __ increment (need_alloc, one);
5920 } __ end_if();
5921 } __ end_if();
5922
5923 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5924 // Update graphKit memory and control from IdealKit.
5925 sync_kit(ideal);
5926 Node * narr = new_array(klass_node, zlen, 1);
5927 // Update IdealKit memory and control from graphKit.
5928 __ sync_kit(this);
5929 __ set(z_alloc, narr);
5930 } __ end_if();
5931
5932 sync_kit(ideal);
5933 z = __ value(z_alloc);
5934 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5935 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5936 // Final sync IdealKit and GraphKit.
5937 final_sync(ideal);
5938 #undef __
5939
5940 Node* z_start = array_element_address(z, intcon(0), T_INT);
5941
5942 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5943 OptoRuntime::multiplyToLen_Type(),
5944 stubAddr, stubName, TypePtr::BOTTOM,
5945 x_start, xlen, y_start, ylen, z_start, zlen);
5946 } // original reexecute is set back here
5947
5948 C->set_has_split_ifs(true); // Has chance for split-if optimization
5949 set_result(z);
5950 return true;
5951 }
5952
5953 //-------------inline_squareToLen------------------------------------
5954 bool LibraryCallKit::inline_squareToLen() {
5955 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5956
5957 address stubAddr = StubRoutines::squareToLen();
5958 if (stubAddr == NULL) {
5959 return false; // Intrinsic's stub is not implemented on this platform
5960 }
5961 const char* stubName = "squareToLen";
5962
5963 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5964
5965 Node* x = argument(0);
5966 Node* len = argument(1);
5967 Node* z = argument(2);
5968 Node* zlen = argument(3);
5969
5970 const Type* x_type = x->Value(&_gvn);
5971 const Type* z_type = z->Value(&_gvn);
5972 const TypeAryPtr* top_x = x_type->isa_aryptr();
5973 const TypeAryPtr* top_z = z_type->isa_aryptr();
5974 if (top_x == NULL || top_x->klass() == NULL ||
5975 top_z == NULL || top_z->klass() == NULL) {
5976 // failed array check
5977 return false;
5978 }
5979
5980 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5981 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5982 if (x_elem != T_INT || z_elem != T_INT) {
5983 return false;
5984 }
5985
5986
5987 Node* x_start = array_element_address(x, intcon(0), x_elem);
5988 Node* z_start = array_element_address(z, intcon(0), z_elem);
5989
5990 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5991 OptoRuntime::squareToLen_Type(),
5992 stubAddr, stubName, TypePtr::BOTTOM,
5993 x_start, len, z_start, zlen);
5994
5995 set_result(z);
5996 return true;
5997 }
5998
5999 //-------------inline_mulAdd------------------------------------------
6000 bool LibraryCallKit::inline_mulAdd() {
6001 assert(UseMulAddIntrinsic, "not implementated on this platform");
6002
6003 address stubAddr = StubRoutines::mulAdd();
6004 if (stubAddr == NULL) {
6005 return false; // Intrinsic's stub is not implemented on this platform
6006 }
6007 const char* stubName = "mulAdd";
6008
6009 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
6010
6011 Node* out = argument(0);
6012 Node* in = argument(1);
6013 Node* offset = argument(2);
6014 Node* len = argument(3);
6015 Node* k = argument(4);
6016
6017 const Type* out_type = out->Value(&_gvn);
6018 const Type* in_type = in->Value(&_gvn);
6019 const TypeAryPtr* top_out = out_type->isa_aryptr();
6020 const TypeAryPtr* top_in = in_type->isa_aryptr();
6021 if (top_out == NULL || top_out->klass() == NULL ||
6022 top_in == NULL || top_in->klass() == NULL) {
6023 // failed array check
6024 return false;
6025 }
6026
6027 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6028 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6029 if (out_elem != T_INT || in_elem != T_INT) {
6030 return false;
6031 }
6032
6033 Node* outlen = load_array_length(out);
6034 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
6035 Node* out_start = array_element_address(out, intcon(0), out_elem);
6036 Node* in_start = array_element_address(in, intcon(0), in_elem);
6037
6038 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6039 OptoRuntime::mulAdd_Type(),
6040 stubAddr, stubName, TypePtr::BOTTOM,
6041 out_start,in_start, new_offset, len, k);
6042 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6043 set_result(result);
6044 return true;
6045 }
6046
6047 //-------------inline_montgomeryMultiply-----------------------------------
6048 bool LibraryCallKit::inline_montgomeryMultiply() {
6049 address stubAddr = StubRoutines::montgomeryMultiply();
6050 if (stubAddr == NULL) {
6051 return false; // Intrinsic's stub is not implemented on this platform
6052 }
6053
6054 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6055 const char* stubName = "montgomery_multiply";
6056
6057 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6058
6059 Node* a = argument(0);
6060 Node* b = argument(1);
6061 Node* n = argument(2);
6062 Node* len = argument(3);
6063 Node* inv = argument(4);
6064 Node* m = argument(6);
6065
6066 const Type* a_type = a->Value(&_gvn);
6067 const TypeAryPtr* top_a = a_type->isa_aryptr();
6068 const Type* b_type = b->Value(&_gvn);
6069 const TypeAryPtr* top_b = b_type->isa_aryptr();
6070 const Type* n_type = a->Value(&_gvn);
6071 const TypeAryPtr* top_n = n_type->isa_aryptr();
6072 const Type* m_type = a->Value(&_gvn);
6073 const TypeAryPtr* top_m = m_type->isa_aryptr();
6074 if (top_a == NULL || top_a->klass() == NULL ||
6075 top_b == NULL || top_b->klass() == NULL ||
6076 top_n == NULL || top_n->klass() == NULL ||
6077 top_m == NULL || top_m->klass() == NULL) {
6078 // failed array check
6079 return false;
6080 }
6081
6082 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6083 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6084 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6085 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6086 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6087 return false;
6088 }
6089
6090 // Make the call
6091 {
6092 Node* a_start = array_element_address(a, intcon(0), a_elem);
6093 Node* b_start = array_element_address(b, intcon(0), b_elem);
6094 Node* n_start = array_element_address(n, intcon(0), n_elem);
6095 Node* m_start = array_element_address(m, intcon(0), m_elem);
6096
6097 Node* call = NULL;
6098 if (CCallingConventionRequiresIntsAsLongs) {
6099 Node* len_I2L = ConvI2L(len);
6100 call = make_runtime_call(RC_LEAF,
6101 OptoRuntime::montgomeryMultiply_Type(),
6102 stubAddr, stubName, TypePtr::BOTTOM,
6103 a_start, b_start, n_start, len_I2L XTOP, inv,
6104 top(), m_start);
6105 } else {
6106 call = make_runtime_call(RC_LEAF,
6107 OptoRuntime::montgomeryMultiply_Type(),
6108 stubAddr, stubName, TypePtr::BOTTOM,
6109 a_start, b_start, n_start, len, inv, top(),
6110 m_start);
6111 }
6112 set_result(m);
6113 }
6114
6115 return true;
6116 }
6117
6118 bool LibraryCallKit::inline_montgomerySquare() {
6119 address stubAddr = StubRoutines::montgomerySquare();
6120 if (stubAddr == NULL) {
6121 return false; // Intrinsic's stub is not implemented on this platform
6122 }
6123
6124 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6125 const char* stubName = "montgomery_square";
6126
6127 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6128
6129 Node* a = argument(0);
6130 Node* n = argument(1);
6131 Node* len = argument(2);
6132 Node* inv = argument(3);
6133 Node* m = argument(5);
6134
6135 const Type* a_type = a->Value(&_gvn);
6136 const TypeAryPtr* top_a = a_type->isa_aryptr();
6137 const Type* n_type = a->Value(&_gvn);
6138 const TypeAryPtr* top_n = n_type->isa_aryptr();
6139 const Type* m_type = a->Value(&_gvn);
6140 const TypeAryPtr* top_m = m_type->isa_aryptr();
6141 if (top_a == NULL || top_a->klass() == NULL ||
6142 top_n == NULL || top_n->klass() == NULL ||
6143 top_m == NULL || top_m->klass() == NULL) {
6144 // failed array check
6145 return false;
6146 }
6147
6148 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6149 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6150 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6151 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6152 return false;
6153 }
6154
6155 // Make the call
6156 {
6157 Node* a_start = array_element_address(a, intcon(0), a_elem);
6158 Node* n_start = array_element_address(n, intcon(0), n_elem);
6159 Node* m_start = array_element_address(m, intcon(0), m_elem);
6160
6161 Node* call = NULL;
6162 if (CCallingConventionRequiresIntsAsLongs) {
6163 Node* len_I2L = ConvI2L(len);
6164 call = make_runtime_call(RC_LEAF,
6165 OptoRuntime::montgomerySquare_Type(),
6166 stubAddr, stubName, TypePtr::BOTTOM,
6167 a_start, n_start, len_I2L XTOP, inv, top(),
6168 m_start);
6169 } else {
6170 call = make_runtime_call(RC_LEAF,
6171 OptoRuntime::montgomerySquare_Type(),
6172 stubAddr, stubName, TypePtr::BOTTOM,
6173 a_start, n_start, len, inv, top(),
6174 m_start);
6175 }
6176
6177 set_result(m);
6178 }
6179
6180 return true;
6181 }
6182
6183
6184 /**
6185 * Calculate CRC32 for byte.
6186 * int java.util.zip.CRC32.update(int crc, int b)
6187 */
6188 bool LibraryCallKit::inline_updateCRC32() {
6189 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6190 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6191 // no receiver since it is static method
6192 Node* crc = argument(0); // type: int
6193 Node* b = argument(1); // type: int
6194
6195 /*
6196 * int c = ~ crc;
6197 * b = timesXtoThe32[(b ^ c) & 0xFF];
6198 * b = b ^ (c >>> 8);
6199 * crc = ~b;
6200 */
6201
6202 Node* M1 = intcon(-1);
6203 crc = _gvn.transform(new (C) XorINode(crc, M1));
6204 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6205 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6206
6207 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6208 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6209 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6210 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6211
6212 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6213 result = _gvn.transform(new (C) XorINode(crc, result));
6214 result = _gvn.transform(new (C) XorINode(result, M1));
6215 set_result(result);
6216 return true;
6217 }
6218
6219 /**
6220 * Calculate CRC32 for byte[] array.
6221 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6222 */
6223 bool LibraryCallKit::inline_updateBytesCRC32() {
6224 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6225 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6226 // no receiver since it is static method
6227 Node* crc = argument(0); // type: int
6228 Node* src = argument(1); // type: oop
6229 Node* offset = argument(2); // type: int
6230 Node* length = argument(3); // type: int
6231
6232 const Type* src_type = src->Value(&_gvn);
6233 const TypeAryPtr* top_src = src_type->isa_aryptr();
6234 if (top_src == NULL || top_src->klass() == NULL) {
6235 // failed array check
6236 return false;
6237 }
6238
6239 // Figure out the size and type of the elements we will be copying.
6240 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6241 if (src_elem != T_BYTE) {
6242 return false;
6243 }
6244
6245 // 'src_start' points to src array + scaled offset
6246 Node* src_start = array_element_address(src, offset, src_elem);
6247
6248 // We assume that range check is done by caller.
6249 // TODO: generate range check (offset+length < src.length) in debug VM.
6250
6251 // Call the stub.
6252 address stubAddr = StubRoutines::updateBytesCRC32();
6253 const char *stubName = "updateBytesCRC32";
6254 Node* call;
6255 if (CCallingConventionRequiresIntsAsLongs) {
6256 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6257 stubAddr, stubName, TypePtr::BOTTOM,
6258 crc XTOP, src_start, length XTOP);
6259 } else {
6260 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6261 stubAddr, stubName, TypePtr::BOTTOM,
6262 crc, src_start, length);
6263 }
6264 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6265 set_result(result);
6266 return true;
6267 }
6268
6269 /**
6270 * Calculate CRC32 for ByteBuffer.
6271 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6272 */
6273 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6274 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6275 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6276 // no receiver since it is static method
6277 Node* crc = argument(0); // type: int
6278 Node* src = argument(1); // type: long
6279 Node* offset = argument(3); // type: int
6280 Node* length = argument(4); // type: int
6281
6282 src = ConvL2X(src); // adjust Java long to machine word
6283 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6284 offset = ConvI2X(offset);
6285
6286 // 'src_start' points to src array + scaled offset
6287 Node* src_start = basic_plus_adr(top(), base, offset);
6288
6289 // Call the stub.
6290 address stubAddr = StubRoutines::updateBytesCRC32();
6291 const char *stubName = "updateBytesCRC32";
6292 Node* call;
6293 if (CCallingConventionRequiresIntsAsLongs) {
6294 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6295 stubAddr, stubName, TypePtr::BOTTOM,
6296 crc XTOP, src_start, length XTOP);
6297 } else {
6298 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6299 stubAddr, stubName, TypePtr::BOTTOM,
6300 crc, src_start, length);
6301 }
6302 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6303 set_result(result);
6304 return true;
6305 }
6306
6307 //----------------------------inline_reference_get----------------------------
6308 // public T java.lang.ref.Reference.get();
6309 bool LibraryCallKit::inline_reference_get() {
6310 const int referent_offset = java_lang_ref_Reference::referent_offset;
6311 guarantee(referent_offset > 0, "should have already been set");
6312
6313 // Get the argument:
6314 Node* reference_obj = null_check_receiver();
6315 if (stopped()) return true;
6316
6317 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6318
6319 ciInstanceKlass* klass = env()->Object_klass();
6320 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6321
6322 Node* no_ctrl = NULL;
6323 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6324
6325 // Use the pre-barrier to record the value in the referent field
6326 pre_barrier(false /* do_load */,
6327 control(),
6328 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6329 result /* pre_val */,
6330 T_OBJECT);
6331
6332 // Add memory barrier to prevent commoning reads from this field
6333 // across safepoint since GC can change its value.
6334 insert_mem_bar(Op_MemBarCPUOrder);
6335
6336 set_result(result);
6337 return true;
6338 }
6339
6340
6341 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6342 bool is_exact=true, bool is_static=false) {
6343
6344 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6345 assert(tinst != NULL, "obj is null");
6346 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6347 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6348
6349 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6350 ciSymbol::make(fieldTypeString),
6351 is_static);
6352 if (field == NULL) return (Node *) NULL;
6353 assert (field != NULL, "undefined field");
6354
6355 // Next code copied from Parse::do_get_xxx():
6356
6357 // Compute address and memory type.
6358 int offset = field->offset_in_bytes();
6359 bool is_vol = field->is_volatile();
6360 ciType* field_klass = field->type();
6361 assert(field_klass->is_loaded(), "should be loaded");
6362 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6363 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6364 BasicType bt = field->layout_type();
6365
6366 // Build the resultant type of the load
6367 const Type *type;
6368 if (bt == T_OBJECT) {
6369 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6370 } else {
6371 type = Type::get_const_basic_type(bt);
6372 }
6373
6374 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6375 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6376 }
6377 // Build the load.
6378 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6379 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6380 // If reference is volatile, prevent following memory ops from
6381 // floating up past the volatile read. Also prevents commoning
6382 // another volatile read.
6383 if (is_vol) {
6384 // Memory barrier includes bogus read of value to force load BEFORE membar
6385 insert_mem_bar(Op_MemBarAcquire, loadedField);
6386 }
6387 return loadedField;
6388 }
6389
6390
6391 //------------------------------inline_aescrypt_Block-----------------------
6392 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6393 address stubAddr = NULL;
6394 const char *stubName;
6395 assert(UseAES, "need AES instruction support");
6396
6397 switch(id) {
6398 case vmIntrinsics::_aescrypt_encryptBlock:
6399 stubAddr = StubRoutines::aescrypt_encryptBlock();
6400 stubName = "aescrypt_encryptBlock";
6401 break;
6402 case vmIntrinsics::_aescrypt_decryptBlock:
6403 stubAddr = StubRoutines::aescrypt_decryptBlock();
6404 stubName = "aescrypt_decryptBlock";
6405 break;
6406 }
6407 if (stubAddr == NULL) return false;
6408
6409 Node* aescrypt_object = argument(0);
6410 Node* src = argument(1);
6411 Node* src_offset = argument(2);
6412 Node* dest = argument(3);
6413 Node* dest_offset = argument(4);
6414
6415 // (1) src and dest are arrays.
6416 const Type* src_type = src->Value(&_gvn);
6417 const Type* dest_type = dest->Value(&_gvn);
6418 const TypeAryPtr* top_src = src_type->isa_aryptr();
6419 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6420 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6421
6422 // for the quick and dirty code we will skip all the checks.
6423 // we are just trying to get the call to be generated.
6424 Node* src_start = src;
6425 Node* dest_start = dest;
6426 if (src_offset != NULL || dest_offset != NULL) {
6427 assert(src_offset != NULL && dest_offset != NULL, "");
6428 src_start = array_element_address(src, src_offset, T_BYTE);
6429 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6430 }
6431
6432 // now need to get the start of its expanded key array
6433 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6434 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6435 if (k_start == NULL) return false;
6436
6437 if (Matcher::pass_original_key_for_aes()) {
6438 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6439 // compatibility issues between Java key expansion and SPARC crypto instructions
6440 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6441 if (original_k_start == NULL) return false;
6442
6443 // Call the stub.
6444 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6445 stubAddr, stubName, TypePtr::BOTTOM,
6446 src_start, dest_start, k_start, original_k_start);
6447 } else {
6448 // Call the stub.
6449 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6450 stubAddr, stubName, TypePtr::BOTTOM,
6451 src_start, dest_start, k_start);
6452 }
6453
6454 return true;
6455 }
6456
6457 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6458 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6459 address stubAddr = NULL;
6460 const char *stubName = NULL;
6461
6462 assert(UseAES, "need AES instruction support");
6463
6464 switch(id) {
6465 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6466 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6467 stubName = "cipherBlockChaining_encryptAESCrypt";
6468 break;
6469 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6470 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6471 stubName = "cipherBlockChaining_decryptAESCrypt";
6472 break;
6473 }
6474 if (stubAddr == NULL) return false;
6475
6476 Node* cipherBlockChaining_object = argument(0);
6477 Node* src = argument(1);
6478 Node* src_offset = argument(2);
6479 Node* len = argument(3);
6480 Node* dest = argument(4);
6481 Node* dest_offset = argument(5);
6482
6483 // (1) src and dest are arrays.
6484 const Type* src_type = src->Value(&_gvn);
6485 const Type* dest_type = dest->Value(&_gvn);
6486 const TypeAryPtr* top_src = src_type->isa_aryptr();
6487 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6488 assert (top_src != NULL && top_src->klass() != NULL
6489 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6490
6491 // checks are the responsibility of the caller
6492 Node* src_start = src;
6493 Node* dest_start = dest;
6494 if (src_offset != NULL || dest_offset != NULL) {
6495 assert(src_offset != NULL && dest_offset != NULL, "");
6496 src_start = array_element_address(src, src_offset, T_BYTE);
6497 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6498 }
6499
6500 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6501 // (because of the predicated logic executed earlier).
6502 // so we cast it here safely.
6503 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6504
6505 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6506 if (embeddedCipherObj == NULL) return false;
6507
6508 // cast it to what we know it will be at runtime
6509 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6510 assert(tinst != NULL, "CBC obj is null");
6511 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6512 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6513 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6514
6515 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6516 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6517 const TypeOopPtr* xtype = aklass->as_instance_type();
6518 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6519 aescrypt_object = _gvn.transform(aescrypt_object);
6520
6521 // we need to get the start of the aescrypt_object's expanded key array
6522 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6523 if (k_start == NULL) return false;
6524
6525 // similarly, get the start address of the r vector
6526 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6527 if (objRvec == NULL) return false;
6528 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6529
6530 Node* cbcCrypt;
6531 if (Matcher::pass_original_key_for_aes()) {
6532 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6533 // compatibility issues between Java key expansion and SPARC crypto instructions
6534 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6535 if (original_k_start == NULL) return false;
6536
6537 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6538 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6539 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6540 stubAddr, stubName, TypePtr::BOTTOM,
6541 src_start, dest_start, k_start, r_start, len, original_k_start);
6542 } else {
6543 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6544 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6545 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6546 stubAddr, stubName, TypePtr::BOTTOM,
6547 src_start, dest_start, k_start, r_start, len);
6548 }
6549
6550 // return cipher length (int)
6551 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6552 set_result(retvalue);
6553 return true;
6554 }
6555
6556 //------------------------------get_key_start_from_aescrypt_object-----------------------
6557 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6558 #ifdef PPC64
6559 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6560 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6561 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6562 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6563 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6564 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6565 if (objSessionK == NULL) {
6566 return (Node *) NULL;
6567 }
6568 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6569 #else
6570 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6571 #endif // PPC64
6572 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6573 if (objAESCryptKey == NULL) return (Node *) NULL;
6574
6575 // now have the array, need to get the start address of the K array
6576 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6577 return k_start;
6578 }
6579
6580 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6581 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6582 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6583 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6584 if (objAESCryptKey == NULL) return (Node *) NULL;
6585
6586 // now have the array, need to get the start address of the lastKey array
6587 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6588 return original_k_start;
6589 }
6590
6591 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6592 // Return node representing slow path of predicate check.
6593 // the pseudo code we want to emulate with this predicate is:
6594 // for encryption:
6595 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6596 // for decryption:
6597 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6598 // note cipher==plain is more conservative than the original java code but that's OK
6599 //
6600 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6601 // The receiver was checked for NULL already.
6602 Node* objCBC = argument(0);
6603
6604 // Load embeddedCipher field of CipherBlockChaining object.
6605 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6606
6607 // get AESCrypt klass for instanceOf check
6608 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6609 // will have same classloader as CipherBlockChaining object
6610 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6611 assert(tinst != NULL, "CBCobj is null");
6612 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6613
6614 // we want to do an instanceof comparison against the AESCrypt class
6615 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6616 if (!klass_AESCrypt->is_loaded()) {
6617 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6618 Node* ctrl = control();
6619 set_control(top()); // no regular fast path
6620 return ctrl;
6621 }
6622 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6623
6624 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6625 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6626 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6627
6628 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6629
6630 // for encryption, we are done
6631 if (!decrypting)
6632 return instof_false; // even if it is NULL
6633
6634 // for decryption, we need to add a further check to avoid
6635 // taking the intrinsic path when cipher and plain are the same
6636 // see the original java code for why.
6637 RegionNode* region = new(C) RegionNode(3);
6638 region->init_req(1, instof_false);
6639 Node* src = argument(1);
6640 Node* dest = argument(4);
6641 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6642 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6643 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6644 region->init_req(2, src_dest_conjoint);
6645
6646 record_for_igvn(region);
6647 return _gvn.transform(region);
6648 }
6649
6650 //------------------------------inline_sha_implCompress-----------------------
6651 //
6652 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6653 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6654 //
6655 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6656 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6657 //
6658 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6659 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6660 //
6661 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6662 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6663
6664 Node* sha_obj = argument(0);
6665 Node* src = argument(1); // type oop
6666 Node* ofs = argument(2); // type int
6667
6668 const Type* src_type = src->Value(&_gvn);
6669 const TypeAryPtr* top_src = src_type->isa_aryptr();
6670 if (top_src == NULL || top_src->klass() == NULL) {
6671 // failed array check
6672 return false;
6673 }
6674 // Figure out the size and type of the elements we will be copying.
6675 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6676 if (src_elem != T_BYTE) {
6677 return false;
6678 }
6679 // 'src_start' points to src array + offset
6680 Node* src_start = array_element_address(src, ofs, src_elem);
6681 Node* state = NULL;
6682 address stubAddr;
6683 const char *stubName;
6684
6685 switch(id) {
6686 case vmIntrinsics::_sha_implCompress:
6687 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6688 state = get_state_from_sha_object(sha_obj);
6689 stubAddr = StubRoutines::sha1_implCompress();
6690 stubName = "sha1_implCompress";
6691 break;
6692 case vmIntrinsics::_sha2_implCompress:
6693 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6694 state = get_state_from_sha_object(sha_obj);
6695 stubAddr = StubRoutines::sha256_implCompress();
6696 stubName = "sha256_implCompress";
6697 break;
6698 case vmIntrinsics::_sha5_implCompress:
6699 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6700 state = get_state_from_sha5_object(sha_obj);
6701 stubAddr = StubRoutines::sha512_implCompress();
6702 stubName = "sha512_implCompress";
6703 break;
6704 default:
6705 fatal_unexpected_iid(id);
6706 return false;
6707 }
6708 if (state == NULL) return false;
6709
6710 // Call the stub.
6711 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6712 stubAddr, stubName, TypePtr::BOTTOM,
6713 src_start, state);
6714
6715 return true;
6716 }
6717
6718 //------------------------------inline_digestBase_implCompressMB-----------------------
6719 //
6720 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6721 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6722 //
6723 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6724 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6725 "need SHA1/SHA256/SHA512 instruction support");
6726 assert((uint)predicate < 3, "sanity");
6727 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6728
6729 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6730 Node* src = argument(1); // byte[] array
6731 Node* ofs = argument(2); // type int
6732 Node* limit = argument(3); // type int
6733
6734 const Type* src_type = src->Value(&_gvn);
6735 const TypeAryPtr* top_src = src_type->isa_aryptr();
6736 if (top_src == NULL || top_src->klass() == NULL) {
6737 // failed array check
6738 return false;
6739 }
6740 // Figure out the size and type of the elements we will be copying.
6741 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6742 if (src_elem != T_BYTE) {
6743 return false;
6744 }
6745 // 'src_start' points to src array + offset
6746 Node* src_start = array_element_address(src, ofs, src_elem);
6747
6748 const char* klass_SHA_name = NULL;
6749 const char* stub_name = NULL;
6750 address stub_addr = NULL;
6751 bool long_state = false;
6752
6753 switch (predicate) {
6754 case 0:
6755 if (UseSHA1Intrinsics) {
6756 klass_SHA_name = "sun/security/provider/SHA";
6757 stub_name = "sha1_implCompressMB";
6758 stub_addr = StubRoutines::sha1_implCompressMB();
6759 }
6760 break;
6761 case 1:
6762 if (UseSHA256Intrinsics) {
6763 klass_SHA_name = "sun/security/provider/SHA2";
6764 stub_name = "sha256_implCompressMB";
6765 stub_addr = StubRoutines::sha256_implCompressMB();
6766 }
6767 break;
6768 case 2:
6769 if (UseSHA512Intrinsics) {
6770 klass_SHA_name = "sun/security/provider/SHA5";
6771 stub_name = "sha512_implCompressMB";
6772 stub_addr = StubRoutines::sha512_implCompressMB();
6773 long_state = true;
6774 }
6775 break;
6776 default:
6777 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6778 }
6779 if (klass_SHA_name != NULL) {
6780 // get DigestBase klass to lookup for SHA klass
6781 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6782 assert(tinst != NULL, "digestBase_obj is not instance???");
6783 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6784
6785 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6786 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6787 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6788 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6789 }
6790 return false;
6791 }
6792 //------------------------------inline_sha_implCompressMB-----------------------
6793 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6794 bool long_state, address stubAddr, const char *stubName,
6795 Node* src_start, Node* ofs, Node* limit) {
6796 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6797 const TypeOopPtr* xtype = aklass->as_instance_type();
6798 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6799 sha_obj = _gvn.transform(sha_obj);
6800
6801 Node* state;
6802 if (long_state) {
6803 state = get_state_from_sha5_object(sha_obj);
6804 } else {
6805 state = get_state_from_sha_object(sha_obj);
6806 }
6807 if (state == NULL) return false;
6808
6809 // Call the stub.
6810 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6811 OptoRuntime::digestBase_implCompressMB_Type(),
6812 stubAddr, stubName, TypePtr::BOTTOM,
6813 src_start, state, ofs, limit);
6814 // return ofs (int)
6815 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6816 set_result(result);
6817
6818 return true;
6819 }
6820
6821 //------------------------------get_state_from_sha_object-----------------------
6822 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6823 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6824 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6825 if (sha_state == NULL) return (Node *) NULL;
6826
6827 // now have the array, need to get the start address of the state array
6828 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6829 return state;
6830 }
6831
6832 //------------------------------get_state_from_sha5_object-----------------------
6833 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6834 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6835 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6836 if (sha_state == NULL) return (Node *) NULL;
6837
6838 // now have the array, need to get the start address of the state array
6839 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6840 return state;
6841 }
6842
6843 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6844 // Return node representing slow path of predicate check.
6845 // the pseudo code we want to emulate with this predicate is:
6846 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6847 //
6848 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6849 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6850 "need SHA1/SHA256/SHA512 instruction support");
6851 assert((uint)predicate < 3, "sanity");
6852
6853 // The receiver was checked for NULL already.
6854 Node* digestBaseObj = argument(0);
6855
6856 // get DigestBase klass for instanceOf check
6857 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6858 assert(tinst != NULL, "digestBaseObj is null");
6859 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6860
6861 const char* klass_SHA_name = NULL;
6862 switch (predicate) {
6863 case 0:
6864 if (UseSHA1Intrinsics) {
6865 // we want to do an instanceof comparison against the SHA class
6866 klass_SHA_name = "sun/security/provider/SHA";
6867 }
6868 break;
6869 case 1:
6870 if (UseSHA256Intrinsics) {
6871 // we want to do an instanceof comparison against the SHA2 class
6872 klass_SHA_name = "sun/security/provider/SHA2";
6873 }
6874 break;
6875 case 2:
6876 if (UseSHA512Intrinsics) {
6877 // we want to do an instanceof comparison against the SHA5 class
6878 klass_SHA_name = "sun/security/provider/SHA5";
6879 }
6880 break;
6881 default:
6882 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6883 }
6884
6885 ciKlass* klass_SHA = NULL;
6886 if (klass_SHA_name != NULL) {
6887 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6888 }
6889 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6890 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6891 Node* ctrl = control();
6892 set_control(top()); // no intrinsic path
6893 return ctrl;
6894 }
6895 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6896
6897 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6898 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6899 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6900 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6901
6902 return instof_false; // even if it is NULL
6903 }
6904
6905 bool LibraryCallKit::inline_profileBoolean() {
6906 Node* counts = argument(1);
6907 const TypeAryPtr* ary = NULL;
6908 ciArray* aobj = NULL;
6909 if (counts->is_Con()
6910 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6911 && (aobj = ary->const_oop()->as_array()) != NULL
6912 && (aobj->length() == 2)) {
6913 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6914 jint false_cnt = aobj->element_value(0).as_int();
6915 jint true_cnt = aobj->element_value(1).as_int();
6916
6917 if (C->log() != NULL) {
6918 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6919 false_cnt, true_cnt);
6920 }
6921
6922 if (false_cnt + true_cnt == 0) {
6923 // According to profile, never executed.
6924 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6925 Deoptimization::Action_reinterpret);
6926 return true;
6927 }
6928
6929 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6930 // is a number of each value occurrences.
6931 Node* result = argument(0);
6932 if (false_cnt == 0 || true_cnt == 0) {
6933 // According to profile, one value has been never seen.
6934 int expected_val = (false_cnt == 0) ? 1 : 0;
6935
6936 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6937 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6938
6939 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6940 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6941 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6942
6943 { // Slow path: uncommon trap for never seen value and then reexecute
6944 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6945 // the value has been seen at least once.
6946 PreserveJVMState pjvms(this);
6947 PreserveReexecuteState preexecs(this);
6948 jvms()->set_should_reexecute(true);
6949
6950 set_control(slow_path);
6951 set_i_o(i_o());
6952
6953 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6954 Deoptimization::Action_reinterpret);
6955 }
6956 // The guard for never seen value enables sharpening of the result and
6957 // returning a constant. It allows to eliminate branches on the same value
6958 // later on.
6959 set_control(fast_path);
6960 result = intcon(expected_val);
6961 }
6962 // Stop profiling.
6963 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6964 // By replacing method body with profile data (represented as ProfileBooleanNode
6965 // on IR level) we effectively disable profiling.
6966 // It enables full speed execution once optimized code is generated.
6967 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6968 C->record_for_igvn(profile);
6969 set_result(profile);
6970 return true;
6971 } else {
6972 // Continue profiling.
6973 // Profile data isn't available at the moment. So, execute method's bytecode version.
6974 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6975 // is compiled and counters aren't available since corresponding MethodHandle
6976 // isn't a compile-time constant.
6977 return false;
6978 }
6979 }