1 /*
2 * Copyright (c) 1999, 2019, 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(); tty->cr();
2549 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
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 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2613 Node* base = argument(1); // type: oop
2614
2615 if (!is_native_ptr) {
2616 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2617 offset = argument(2); // type: long
2618 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2619 // to be plain byte offsets, which are also the same as those accepted
2620 // by oopDesc::field_base.
2621 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2622 "fieldOffset must be byte-scaled");
2623 // 32-bit machines ignore the high half!
2624 offset = ConvL2X(offset);
2625 adr = make_unsafe_address(base, offset);
2626 heap_base_oop = base;
2627 val = is_store ? argument(4) : NULL;
2628 } else {
2629 Node* ptr = argument(1); // type: long
2630 ptr = ConvL2X(ptr); // adjust Java long to machine word
2631 adr = make_unsafe_address(NULL, ptr);
2632 val = is_store ? argument(3) : NULL;
2633 }
2634
2635 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) {
2636 return false; // off-heap oop accesses are not supported
2637 }
2638
2639 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2640
2641 // Try to categorize the address.
2642 Compile::AliasType* alias_type = C->alias_type(adr_type);
2643 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2644
2645 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2646 alias_type->adr_type() == TypeAryPtr::RANGE) {
2647 return false; // not supported
2648 }
2649
2650 bool mismatched = false;
2651 BasicType bt = alias_type->basic_type();
2652 if (bt != T_ILLEGAL) {
2653 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2654 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2655 // Alias type doesn't differentiate between byte[] and boolean[]).
2656 // Use address type to get the element type.
2657 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2658 }
2659 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2660 // accessing an array field with getObject is not a mismatch
2661 bt = T_OBJECT;
2662 }
2663 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2664 // Don't intrinsify mismatched object accesses
2665 return false;
2666 }
2667 mismatched = (bt != type);
2668 }
2669
2670 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2671
2672 // First guess at the value type.
2673 const Type *value_type = Type::get_const_basic_type(type);
2674
2675 // We will need memory barriers unless we can determine a unique
2676 // alias category for this reference. (Note: If for some reason
2677 // the barriers get omitted and the unsafe reference begins to "pollute"
2678 // the alias analysis of the rest of the graph, either Compile::can_alias
2679 // or Compile::must_alias will throw a diagnostic assert.)
2680 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2681
2682 // If we are reading the value of the referent field of a Reference
2683 // object (either by using Unsafe directly or through reflection)
2684 // then, if G1 is enabled, we need to record the referent in an
2685 // SATB log buffer using the pre-barrier mechanism.
2686 // Also we need to add memory barrier to prevent commoning reads
2687 // from this field across safepoint since GC can change its value.
2688 bool need_read_barrier = !is_native_ptr && !is_store &&
2689 offset != top() && heap_base_oop != top();
2690
2691 if (!is_store && type == T_OBJECT) {
2692 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2693 if (tjp != NULL) {
2694 value_type = tjp;
2695 }
2696 }
2697
2698 receiver = null_check(receiver);
2699 if (stopped()) {
2700 return true;
2701 }
2702 // Heap pointers get a null-check from the interpreter,
2703 // as a courtesy. However, this is not guaranteed by Unsafe,
2704 // and it is not possible to fully distinguish unintended nulls
2705 // from intended ones in this API.
2706
2707 if (is_volatile) {
2708 // We need to emit leading and trailing CPU membars (see below) in
2709 // addition to memory membars when is_volatile. This is a little
2710 // too strong, but avoids the need to insert per-alias-type
2711 // volatile membars (for stores; compare Parse::do_put_xxx), which
2712 // we cannot do effectively here because we probably only have a
2713 // rough approximation of type.
2714 need_mem_bar = true;
2715 // For Stores, place a memory ordering barrier now.
2716 if (is_store) {
2717 insert_mem_bar(Op_MemBarRelease);
2718 } else {
2719 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2720 insert_mem_bar(Op_MemBarVolatile);
2721 }
2722 }
2723 }
2724
2725 // Memory barrier to prevent normal and 'unsafe' accesses from
2726 // bypassing each other. Happens after null checks, so the
2727 // exception paths do not take memory state from the memory barrier,
2728 // so there's no problems making a strong assert about mixing users
2729 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2730 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2731 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2732
2733 if (!is_store) {
2734 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2735 // To be valid, unsafe loads may depend on other conditions than
2736 // the one that guards them: pin the Load node
2737 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2738 // load value
2739 switch (type) {
2740 case T_BOOLEAN:
2741 case T_CHAR:
2742 case T_BYTE:
2743 case T_SHORT:
2744 case T_INT:
2745 case T_LONG:
2746 case T_FLOAT:
2747 case T_DOUBLE:
2748 break;
2749 case T_OBJECT:
2750 if (need_read_barrier) {
2751 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2752 }
2753 break;
2754 case T_ADDRESS:
2755 // Cast to an int type.
2756 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2757 p = ConvX2UL(p);
2758 break;
2759 default:
2760 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2761 break;
2762 }
2763 // The load node has the control of the preceding MemBarCPUOrder. All
2764 // following nodes will have the control of the MemBarCPUOrder inserted at
2765 // the end of this method. So, pushing the load onto the stack at a later
2766 // point is fine.
2767 set_result(p);
2768 } else {
2769 // place effect of store into memory
2770 switch (type) {
2771 case T_DOUBLE:
2772 val = dstore_rounding(val);
2773 break;
2774 case T_ADDRESS:
2775 // Repackage the long as a pointer.
2776 val = ConvL2X(val);
2777 val = _gvn.transform(new (C) CastX2PNode(val));
2778 break;
2779 }
2780
2781 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2782 if (type == T_OBJECT ) {
2783 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2784 } else {
2785 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2786 }
2787 }
2788
2789 if (is_volatile) {
2790 if (!is_store) {
2791 insert_mem_bar(Op_MemBarAcquire);
2792 } else {
2793 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2794 insert_mem_bar(Op_MemBarVolatile);
2795 }
2796 }
2797 }
2798
2799 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2800
2801 return true;
2802 }
2803
2804 //----------------------------inline_unsafe_prefetch----------------------------
2805
2806 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2807 #ifndef PRODUCT
2808 {
2809 ResourceMark rm;
2810 // Check the signatures.
2811 ciSignature* sig = callee()->signature();
2812 #ifdef ASSERT
2813 // Object getObject(Object base, int/long offset), etc.
2814 BasicType rtype = sig->return_type()->basic_type();
2815 if (!is_native_ptr) {
2816 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2817 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2818 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2819 } else {
2820 assert(sig->count() == 1, "native prefetch has 1 argument");
2821 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2822 }
2823 #endif // ASSERT
2824 }
2825 #endif // !PRODUCT
2826
2827 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2828
2829 const int idx = is_static ? 0 : 1;
2830 if (!is_static) {
2831 null_check_receiver();
2832 if (stopped()) {
2833 return true;
2834 }
2835 }
2836
2837 // Build address expression. See the code in inline_unsafe_access.
2838 Node *adr;
2839 if (!is_native_ptr) {
2840 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2841 Node* base = argument(idx + 0); // type: oop
2842 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2843 Node* offset = argument(idx + 1); // type: long
2844 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2845 // to be plain byte offsets, which are also the same as those accepted
2846 // by oopDesc::field_base.
2847 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2848 "fieldOffset must be byte-scaled");
2849 // 32-bit machines ignore the high half!
2850 offset = ConvL2X(offset);
2851 adr = make_unsafe_address(base, offset);
2852 } else {
2853 Node* ptr = argument(idx + 0); // type: long
2854 ptr = ConvL2X(ptr); // adjust Java long to machine word
2855 adr = make_unsafe_address(NULL, ptr);
2856 }
2857
2858 // Generate the read or write prefetch
2859 Node *prefetch;
2860 if (is_store) {
2861 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2862 } else {
2863 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2864 }
2865 prefetch->init_req(0, control());
2866 set_i_o(_gvn.transform(prefetch));
2867
2868 return true;
2869 }
2870
2871 //----------------------------inline_unsafe_load_store----------------------------
2872 // This method serves a couple of different customers (depending on LoadStoreKind):
2873 //
2874 // LS_cmpxchg:
2875 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2876 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2877 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2878 //
2879 // LS_xadd:
2880 // public int getAndAddInt( Object o, long offset, int delta)
2881 // public long getAndAddLong(Object o, long offset, long delta)
2882 //
2883 // LS_xchg:
2884 // int getAndSet(Object o, long offset, int newValue)
2885 // long getAndSet(Object o, long offset, long newValue)
2886 // Object getAndSet(Object o, long offset, Object newValue)
2887 //
2888 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2889 // This basic scheme here is the same as inline_unsafe_access, but
2890 // differs in enough details that combining them would make the code
2891 // overly confusing. (This is a true fact! I originally combined
2892 // them, but even I was confused by it!) As much code/comments as
2893 // possible are retained from inline_unsafe_access though to make
2894 // the correspondences clearer. - dl
2895
2896 if (callee()->is_static()) return false; // caller must have the capability!
2897
2898 #ifndef PRODUCT
2899 BasicType rtype;
2900 {
2901 ResourceMark rm;
2902 // Check the signatures.
2903 ciSignature* sig = callee()->signature();
2904 rtype = sig->return_type()->basic_type();
2905 if (kind == LS_xadd || kind == LS_xchg) {
2906 // Check the signatures.
2907 #ifdef ASSERT
2908 assert(rtype == type, "get and set must return the expected type");
2909 assert(sig->count() == 3, "get and set has 3 arguments");
2910 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2911 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2912 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2913 #endif // ASSERT
2914 } else if (kind == LS_cmpxchg) {
2915 // Check the signatures.
2916 #ifdef ASSERT
2917 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2918 assert(sig->count() == 4, "CAS has 4 arguments");
2919 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2920 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2921 #endif // ASSERT
2922 } else {
2923 ShouldNotReachHere();
2924 }
2925 }
2926 #endif //PRODUCT
2927
2928 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2929
2930 // Get arguments:
2931 Node* receiver = NULL;
2932 Node* base = NULL;
2933 Node* offset = NULL;
2934 Node* oldval = NULL;
2935 Node* newval = NULL;
2936 if (kind == LS_cmpxchg) {
2937 const bool two_slot_type = type2size[type] == 2;
2938 receiver = argument(0); // type: oop
2939 base = argument(1); // type: oop
2940 offset = argument(2); // type: long
2941 oldval = argument(4); // type: oop, int, or long
2942 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2943 } else if (kind == LS_xadd || kind == LS_xchg){
2944 receiver = argument(0); // type: oop
2945 base = argument(1); // type: oop
2946 offset = argument(2); // type: long
2947 oldval = NULL;
2948 newval = argument(4); // type: oop, int, or long
2949 }
2950
2951 // Build field offset expression.
2952 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2953 // to be plain byte offsets, which are also the same as those accepted
2954 // by oopDesc::field_base.
2955 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2956 // 32-bit machines ignore the high half of long offsets
2957 offset = ConvL2X(offset);
2958 Node* adr = make_unsafe_address(base, offset);
2959 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2960
2961 Compile::AliasType* alias_type = C->alias_type(adr_type);
2962 BasicType bt = alias_type->basic_type();
2963 if (bt != T_ILLEGAL &&
2964 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2965 // Don't intrinsify mismatched object accesses.
2966 return false;
2967 }
2968
2969 // For CAS, unlike inline_unsafe_access, there seems no point in
2970 // trying to refine types. Just use the coarse types here.
2971 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2972 const Type *value_type = Type::get_const_basic_type(type);
2973
2974 if (kind == LS_xchg && type == T_OBJECT) {
2975 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2976 if (tjp != NULL) {
2977 value_type = tjp;
2978 }
2979 }
2980
2981 // Null check receiver.
2982 receiver = null_check(receiver);
2983 if (stopped()) {
2984 return true;
2985 }
2986
2987 int alias_idx = C->get_alias_index(adr_type);
2988
2989 // Memory-model-wise, a LoadStore acts like a little synchronized
2990 // block, so needs barriers on each side. These don't translate
2991 // into actual barriers on most machines, but we still need rest of
2992 // compiler to respect ordering.
2993
2994 insert_mem_bar(Op_MemBarRelease);
2995 insert_mem_bar(Op_MemBarCPUOrder);
2996
2997 // 4984716: MemBars must be inserted before this
2998 // memory node in order to avoid a false
2999 // dependency which will confuse the scheduler.
3000 Node *mem = memory(alias_idx);
3001
3002 // For now, we handle only those cases that actually exist: ints,
3003 // longs, and Object. Adding others should be straightforward.
3004 Node* load_store = NULL;
3005 switch(type) {
3006 case T_INT:
3007 if (kind == LS_xadd) {
3008 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3009 } else if (kind == LS_xchg) {
3010 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3011 } else if (kind == LS_cmpxchg) {
3012 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3013 } else {
3014 ShouldNotReachHere();
3015 }
3016 break;
3017 case T_LONG:
3018 if (kind == LS_xadd) {
3019 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3020 } else if (kind == LS_xchg) {
3021 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3022 } else if (kind == LS_cmpxchg) {
3023 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3024 } else {
3025 ShouldNotReachHere();
3026 }
3027 break;
3028 case T_OBJECT:
3029 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3030 // could be delayed during Parse (for example, in adjust_map_after_if()).
3031 // Execute transformation here to avoid barrier generation in such case.
3032 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3033 newval = _gvn.makecon(TypePtr::NULL_PTR);
3034
3035 // Reference stores need a store barrier.
3036 if (kind == LS_xchg) {
3037 // If pre-barrier must execute before the oop store, old value will require do_load here.
3038 if (!can_move_pre_barrier()) {
3039 pre_barrier(true /* do_load*/,
3040 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3041 NULL /* pre_val*/,
3042 T_OBJECT);
3043 } // Else move pre_barrier to use load_store value, see below.
3044 } else if (kind == LS_cmpxchg) {
3045 // Same as for newval above:
3046 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3047 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3048 }
3049 // The only known value which might get overwritten is oldval.
3050 pre_barrier(false /* do_load */,
3051 control(), NULL, NULL, max_juint, NULL, NULL,
3052 oldval /* pre_val */,
3053 T_OBJECT);
3054 } else {
3055 ShouldNotReachHere();
3056 }
3057
3058 #ifdef _LP64
3059 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3060 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3061 if (kind == LS_xchg) {
3062 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3063 newval_enc, adr_type, value_type->make_narrowoop()));
3064 } else {
3065 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3066 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3067 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3068 newval_enc, oldval_enc));
3069 }
3070 } else
3071 #endif
3072 {
3073 if (kind == LS_xchg) {
3074 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3075 } else {
3076 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3077 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3078 }
3079 }
3080 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3081 break;
3082 default:
3083 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3084 break;
3085 }
3086
3087 // SCMemProjNodes represent the memory state of a LoadStore. Their
3088 // main role is to prevent LoadStore nodes from being optimized away
3089 // when their results aren't used.
3090 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3091 set_memory(proj, alias_idx);
3092
3093 if (type == T_OBJECT && kind == LS_xchg) {
3094 #ifdef _LP64
3095 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3096 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3097 }
3098 #endif
3099 if (can_move_pre_barrier()) {
3100 // Don't need to load pre_val. The old value is returned by load_store.
3101 // The pre_barrier can execute after the xchg as long as no safepoint
3102 // gets inserted between them.
3103 pre_barrier(false /* do_load */,
3104 control(), NULL, NULL, max_juint, NULL, NULL,
3105 load_store /* pre_val */,
3106 T_OBJECT);
3107 }
3108 }
3109
3110 // Add the trailing membar surrounding the access
3111 insert_mem_bar(Op_MemBarCPUOrder);
3112 insert_mem_bar(Op_MemBarAcquire);
3113
3114 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3115 set_result(load_store);
3116 return true;
3117 }
3118
3119 //----------------------------inline_unsafe_ordered_store----------------------
3120 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3121 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3122 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3123 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3124 // This is another variant of inline_unsafe_access, differing in
3125 // that it always issues store-store ("release") barrier and ensures
3126 // store-atomicity (which only matters for "long").
3127
3128 if (callee()->is_static()) return false; // caller must have the capability!
3129
3130 #ifndef PRODUCT
3131 {
3132 ResourceMark rm;
3133 // Check the signatures.
3134 ciSignature* sig = callee()->signature();
3135 #ifdef ASSERT
3136 BasicType rtype = sig->return_type()->basic_type();
3137 assert(rtype == T_VOID, "must return void");
3138 assert(sig->count() == 3, "has 3 arguments");
3139 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3140 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3141 #endif // ASSERT
3142 }
3143 #endif //PRODUCT
3144
3145 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3146
3147 // Get arguments:
3148 Node* receiver = argument(0); // type: oop
3149 Node* base = argument(1); // type: oop
3150 Node* offset = argument(2); // type: long
3151 Node* val = argument(4); // type: oop, int, or long
3152
3153 // Null check receiver.
3154 receiver = null_check(receiver);
3155 if (stopped()) {
3156 return true;
3157 }
3158
3159 // Build field offset expression.
3160 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3161 // 32-bit machines ignore the high half of long offsets
3162 offset = ConvL2X(offset);
3163 Node* adr = make_unsafe_address(base, offset);
3164 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3165 const Type *value_type = Type::get_const_basic_type(type);
3166 Compile::AliasType* alias_type = C->alias_type(adr_type);
3167
3168 insert_mem_bar(Op_MemBarRelease);
3169 insert_mem_bar(Op_MemBarCPUOrder);
3170 // Ensure that the store is atomic for longs:
3171 const bool require_atomic_access = true;
3172 Node* store;
3173 if (type == T_OBJECT) // reference stores need a store barrier.
3174 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3175 else {
3176 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3177 }
3178 insert_mem_bar(Op_MemBarCPUOrder);
3179 return true;
3180 }
3181
3182 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3183 // Regardless of form, don't allow previous ld/st to move down,
3184 // then issue acquire, release, or volatile mem_bar.
3185 insert_mem_bar(Op_MemBarCPUOrder);
3186 switch(id) {
3187 case vmIntrinsics::_loadFence:
3188 insert_mem_bar(Op_LoadFence);
3189 return true;
3190 case vmIntrinsics::_storeFence:
3191 insert_mem_bar(Op_StoreFence);
3192 return true;
3193 case vmIntrinsics::_fullFence:
3194 insert_mem_bar(Op_MemBarVolatile);
3195 return true;
3196 default:
3197 fatal_unexpected_iid(id);
3198 return false;
3199 }
3200 }
3201
3202 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3203 if (!kls->is_Con()) {
3204 return true;
3205 }
3206 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3207 if (klsptr == NULL) {
3208 return true;
3209 }
3210 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3211 // don't need a guard for a klass that is already initialized
3212 return !ik->is_initialized();
3213 }
3214
3215 //----------------------------inline_unsafe_allocate---------------------------
3216 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3217 bool LibraryCallKit::inline_unsafe_allocate() {
3218 if (callee()->is_static()) return false; // caller must have the capability!
3219
3220 null_check_receiver(); // null-check, then ignore
3221 Node* cls = null_check(argument(1));
3222 if (stopped()) return true;
3223
3224 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3225 kls = null_check(kls);
3226 if (stopped()) return true; // argument was like int.class
3227
3228 Node* test = NULL;
3229 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3230 // Note: The argument might still be an illegal value like
3231 // Serializable.class or Object[].class. The runtime will handle it.
3232 // But we must make an explicit check for initialization.
3233 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3234 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3235 // can generate code to load it as unsigned byte.
3236 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3237 Node* bits = intcon(InstanceKlass::fully_initialized);
3238 test = _gvn.transform(new (C) SubINode(inst, bits));
3239 // The 'test' is non-zero if we need to take a slow path.
3240 }
3241
3242 Node* obj = new_instance(kls, test);
3243 set_result(obj);
3244 return true;
3245 }
3246
3247 #ifdef JFR_HAVE_INTRINSICS
3248 /*
3249 * oop -> myklass
3250 * myklass->trace_id |= USED
3251 * return myklass->trace_id & ~0x3
3252 */
3253 bool LibraryCallKit::inline_native_classID() {
3254 Node* cls = null_check(argument(0), T_OBJECT);
3255 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3256 kls = null_check(kls, T_OBJECT);
3257
3258 ByteSize offset = KLASS_TRACE_ID_OFFSET;
3259 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3260 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3261
3262 Node* clsused = longcon(0x01l); // set the class bit
3263 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3264 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3265 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3266
3267 #ifdef TRACE_ID_META_BITS
3268 Node* mbits = longcon(~TRACE_ID_META_BITS);
3269 tvalue = _gvn.transform(new (C) AndLNode(tvalue, mbits));
3270 #endif
3271 #ifdef TRACE_ID_SHIFT
3272 Node* cbits = intcon(TRACE_ID_SHIFT);
3273 tvalue = _gvn.transform(new (C) URShiftLNode(tvalue, cbits));
3274 #endif
3275
3276 set_result(tvalue);
3277 return true;
3278 }
3279
3280 bool LibraryCallKit::inline_native_getEventWriter() {
3281 Node* tls_ptr = _gvn.transform(new (C) ThreadLocalNode());
3282
3283 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3284 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)
3285 );
3286
3287 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3288
3289 Node* jobj_cmp_null = _gvn.transform( new (C) CmpPNode(jobj, null()) );
3290 Node* test_jobj_eq_null = _gvn.transform( new (C) BoolNode(jobj_cmp_null, BoolTest::eq) );
3291
3292 IfNode* iff_jobj_null =
3293 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3294
3295 enum { _normal_path = 1,
3296 _null_path = 2,
3297 PATH_LIMIT };
3298
3299 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3300 PhiNode* result_val = new (C) PhiNode(result_rgn, TypePtr::BOTTOM);
3301
3302 Node* jobj_is_null = _gvn.transform(new (C) IfTrueNode(iff_jobj_null));
3303 result_rgn->init_req(_null_path, jobj_is_null);
3304 result_val->init_req(_null_path, null());
3305
3306 Node* jobj_is_not_null = _gvn.transform(new (C) IfFalseNode(iff_jobj_null));
3307 result_rgn->init_req(_normal_path, jobj_is_not_null);
3308
3309 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3310 result_val->init_req(_normal_path, res);
3311
3312 set_result(result_rgn, result_val);
3313
3314 return true;
3315 }
3316 #endif // JFR_HAVE_INTRINSICS
3317
3318 //------------------------inline_native_time_funcs--------------
3319 // inline code for System.currentTimeMillis() and System.nanoTime()
3320 // these have the same type and signature
3321 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3322 const TypeFunc* tf = OptoRuntime::void_long_Type();
3323 const TypePtr* no_memory_effects = NULL;
3324 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3325 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3326 #ifdef ASSERT
3327 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3328 assert(value_top == top(), "second value must be top");
3329 #endif
3330 set_result(value);
3331 return true;
3332 }
3333
3334 //------------------------inline_native_currentThread------------------
3335 bool LibraryCallKit::inline_native_currentThread() {
3336 Node* junk = NULL;
3337 set_result(generate_current_thread(junk));
3338 return true;
3339 }
3340
3341 //------------------------inline_native_isInterrupted------------------
3342 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3343 bool LibraryCallKit::inline_native_isInterrupted() {
3344 // Add a fast path to t.isInterrupted(clear_int):
3345 // (t == Thread.current() &&
3346 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3347 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3348 // So, in the common case that the interrupt bit is false,
3349 // we avoid making a call into the VM. Even if the interrupt bit
3350 // is true, if the clear_int argument is false, we avoid the VM call.
3351 // However, if the receiver is not currentThread, we must call the VM,
3352 // because there must be some locking done around the operation.
3353
3354 // We only go to the fast case code if we pass two guards.
3355 // Paths which do not pass are accumulated in the slow_region.
3356
3357 enum {
3358 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3359 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3360 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3361 PATH_LIMIT
3362 };
3363
3364 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3365 // out of the function.
3366 insert_mem_bar(Op_MemBarCPUOrder);
3367
3368 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3369 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3370
3371 RegionNode* slow_region = new (C) RegionNode(1);
3372 record_for_igvn(slow_region);
3373
3374 // (a) Receiving thread must be the current thread.
3375 Node* rec_thr = argument(0);
3376 Node* tls_ptr = NULL;
3377 Node* cur_thr = generate_current_thread(tls_ptr);
3378 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3379 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3380
3381 generate_slow_guard(bol_thr, slow_region);
3382
3383 // (b) Interrupt bit on TLS must be false.
3384 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3385 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3386 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3387
3388 // Set the control input on the field _interrupted read to prevent it floating up.
3389 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3390 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3391 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3392
3393 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3394
3395 // First fast path: if (!TLS._interrupted) return false;
3396 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3397 result_rgn->init_req(no_int_result_path, false_bit);
3398 result_val->init_req(no_int_result_path, intcon(0));
3399
3400 // drop through to next case
3401 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3402
3403 #ifndef TARGET_OS_FAMILY_windows
3404 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3405 Node* clr_arg = argument(1);
3406 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3407 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3408 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3409
3410 // Second fast path: ... else if (!clear_int) return true;
3411 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3412 result_rgn->init_req(no_clear_result_path, false_arg);
3413 result_val->init_req(no_clear_result_path, intcon(1));
3414
3415 // drop through to next case
3416 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3417 #else
3418 // To return true on Windows you must read the _interrupted field
3419 // and check the the event state i.e. take the slow path.
3420 #endif // TARGET_OS_FAMILY_windows
3421
3422 // (d) Otherwise, go to the slow path.
3423 slow_region->add_req(control());
3424 set_control( _gvn.transform(slow_region));
3425
3426 if (stopped()) {
3427 // There is no slow path.
3428 result_rgn->init_req(slow_result_path, top());
3429 result_val->init_req(slow_result_path, top());
3430 } else {
3431 // non-virtual because it is a private non-static
3432 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3433
3434 Node* slow_val = set_results_for_java_call(slow_call);
3435 // this->control() comes from set_results_for_java_call
3436
3437 Node* fast_io = slow_call->in(TypeFunc::I_O);
3438 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3439
3440 // These two phis are pre-filled with copies of of the fast IO and Memory
3441 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3442 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3443
3444 result_rgn->init_req(slow_result_path, control());
3445 result_io ->init_req(slow_result_path, i_o());
3446 result_mem->init_req(slow_result_path, reset_memory());
3447 result_val->init_req(slow_result_path, slow_val);
3448
3449 set_all_memory(_gvn.transform(result_mem));
3450 set_i_o( _gvn.transform(result_io));
3451 }
3452
3453 C->set_has_split_ifs(true); // Has chance for split-if optimization
3454 set_result(result_rgn, result_val);
3455 return true;
3456 }
3457
3458 //---------------------------load_mirror_from_klass----------------------------
3459 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3460 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3461 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3462 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3463 }
3464
3465 //-----------------------load_klass_from_mirror_common-------------------------
3466 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3467 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3468 // and branch to the given path on the region.
3469 // If never_see_null, take an uncommon trap on null, so we can optimistically
3470 // compile for the non-null case.
3471 // If the region is NULL, force never_see_null = true.
3472 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3473 bool never_see_null,
3474 RegionNode* region,
3475 int null_path,
3476 int offset) {
3477 if (region == NULL) never_see_null = true;
3478 Node* p = basic_plus_adr(mirror, offset);
3479 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3480 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3481 Node* null_ctl = top();
3482 kls = null_check_oop(kls, &null_ctl, never_see_null);
3483 if (region != NULL) {
3484 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3485 region->init_req(null_path, null_ctl);
3486 } else {
3487 assert(null_ctl == top(), "no loose ends");
3488 }
3489 return kls;
3490 }
3491
3492 //--------------------(inline_native_Class_query helpers)---------------------
3493 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3494 // Fall through if (mods & mask) == bits, take the guard otherwise.
3495 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3496 // Branch around if the given klass has the given modifier bit set.
3497 // Like generate_guard, adds a new path onto the region.
3498 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3499 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3500 Node* mask = intcon(modifier_mask);
3501 Node* bits = intcon(modifier_bits);
3502 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3503 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3504 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3505 return generate_fair_guard(bol, region);
3506 }
3507 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3508 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3509 }
3510
3511 //-------------------------inline_native_Class_query-------------------
3512 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3513 const Type* return_type = TypeInt::BOOL;
3514 Node* prim_return_value = top(); // what happens if it's a primitive class?
3515 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3516 bool expect_prim = false; // most of these guys expect to work on refs
3517
3518 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3519
3520 Node* mirror = argument(0);
3521 Node* obj = top();
3522
3523 switch (id) {
3524 case vmIntrinsics::_isInstance:
3525 // nothing is an instance of a primitive type
3526 prim_return_value = intcon(0);
3527 obj = argument(1);
3528 break;
3529 case vmIntrinsics::_getModifiers:
3530 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3531 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3532 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3533 break;
3534 case vmIntrinsics::_isInterface:
3535 prim_return_value = intcon(0);
3536 break;
3537 case vmIntrinsics::_isArray:
3538 prim_return_value = intcon(0);
3539 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3540 break;
3541 case vmIntrinsics::_isPrimitive:
3542 prim_return_value = intcon(1);
3543 expect_prim = true; // obviously
3544 break;
3545 case vmIntrinsics::_getSuperclass:
3546 prim_return_value = null();
3547 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3548 break;
3549 case vmIntrinsics::_getComponentType:
3550 prim_return_value = null();
3551 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3552 break;
3553 case vmIntrinsics::_getClassAccessFlags:
3554 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3555 return_type = TypeInt::INT; // not bool! 6297094
3556 break;
3557 default:
3558 fatal_unexpected_iid(id);
3559 break;
3560 }
3561
3562 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3563 if (mirror_con == NULL) return false; // cannot happen?
3564
3565 #ifndef PRODUCT
3566 if (C->print_intrinsics() || C->print_inlining()) {
3567 ciType* k = mirror_con->java_mirror_type();
3568 if (k) {
3569 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3570 k->print_name();
3571 tty->cr();
3572 }
3573 }
3574 #endif
3575
3576 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3577 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3578 record_for_igvn(region);
3579 PhiNode* phi = new (C) PhiNode(region, return_type);
3580
3581 // The mirror will never be null of Reflection.getClassAccessFlags, however
3582 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3583 // if it is. See bug 4774291.
3584
3585 // For Reflection.getClassAccessFlags(), the null check occurs in
3586 // the wrong place; see inline_unsafe_access(), above, for a similar
3587 // situation.
3588 mirror = null_check(mirror);
3589 // If mirror or obj is dead, only null-path is taken.
3590 if (stopped()) return true;
3591
3592 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3593
3594 // Now load the mirror's klass metaobject, and null-check it.
3595 // Side-effects region with the control path if the klass is null.
3596 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3597 // If kls is null, we have a primitive mirror.
3598 phi->init_req(_prim_path, prim_return_value);
3599 if (stopped()) { set_result(region, phi); return true; }
3600 bool safe_for_replace = (region->in(_prim_path) == top());
3601
3602 Node* p; // handy temp
3603 Node* null_ctl;
3604
3605 // Now that we have the non-null klass, we can perform the real query.
3606 // For constant classes, the query will constant-fold in LoadNode::Value.
3607 Node* query_value = top();
3608 switch (id) {
3609 case vmIntrinsics::_isInstance:
3610 // nothing is an instance of a primitive type
3611 query_value = gen_instanceof(obj, kls, safe_for_replace);
3612 break;
3613
3614 case vmIntrinsics::_getModifiers:
3615 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3616 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3617 break;
3618
3619 case vmIntrinsics::_isInterface:
3620 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3621 if (generate_interface_guard(kls, region) != NULL)
3622 // A guard was added. If the guard is taken, it was an interface.
3623 phi->add_req(intcon(1));
3624 // If we fall through, it's a plain class.
3625 query_value = intcon(0);
3626 break;
3627
3628 case vmIntrinsics::_isArray:
3629 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3630 if (generate_array_guard(kls, region) != NULL)
3631 // A guard was added. If the guard is taken, it was an array.
3632 phi->add_req(intcon(1));
3633 // If we fall through, it's a plain class.
3634 query_value = intcon(0);
3635 break;
3636
3637 case vmIntrinsics::_isPrimitive:
3638 query_value = intcon(0); // "normal" path produces false
3639 break;
3640
3641 case vmIntrinsics::_getSuperclass:
3642 // The rules here are somewhat unfortunate, but we can still do better
3643 // with random logic than with a JNI call.
3644 // Interfaces store null or Object as _super, but must report null.
3645 // Arrays store an intermediate super as _super, but must report Object.
3646 // Other types can report the actual _super.
3647 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3648 if (generate_interface_guard(kls, region) != NULL)
3649 // A guard was added. If the guard is taken, it was an interface.
3650 phi->add_req(null());
3651 if (generate_array_guard(kls, region) != NULL)
3652 // A guard was added. If the guard is taken, it was an array.
3653 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3654 // If we fall through, it's a plain class. Get its _super.
3655 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3656 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3657 null_ctl = top();
3658 kls = null_check_oop(kls, &null_ctl);
3659 if (null_ctl != top()) {
3660 // If the guard is taken, Object.superClass is null (both klass and mirror).
3661 region->add_req(null_ctl);
3662 phi ->add_req(null());
3663 }
3664 if (!stopped()) {
3665 query_value = load_mirror_from_klass(kls);
3666 }
3667 break;
3668
3669 case vmIntrinsics::_getComponentType:
3670 if (generate_array_guard(kls, region) != NULL) {
3671 // Be sure to pin the oop load to the guard edge just created:
3672 Node* is_array_ctrl = region->in(region->req()-1);
3673 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3674 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3675 phi->add_req(cmo);
3676 }
3677 query_value = null(); // non-array case is null
3678 break;
3679
3680 case vmIntrinsics::_getClassAccessFlags:
3681 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3682 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3683 break;
3684
3685 default:
3686 fatal_unexpected_iid(id);
3687 break;
3688 }
3689
3690 // Fall-through is the normal case of a query to a real class.
3691 phi->init_req(1, query_value);
3692 region->init_req(1, control());
3693
3694 C->set_has_split_ifs(true); // Has chance for split-if optimization
3695 set_result(region, phi);
3696 return true;
3697 }
3698
3699 //--------------------------inline_native_subtype_check------------------------
3700 // This intrinsic takes the JNI calls out of the heart of
3701 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3702 bool LibraryCallKit::inline_native_subtype_check() {
3703 // Pull both arguments off the stack.
3704 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3705 args[0] = argument(0);
3706 args[1] = argument(1);
3707 Node* klasses[2]; // corresponding Klasses: superk, subk
3708 klasses[0] = klasses[1] = top();
3709
3710 enum {
3711 // A full decision tree on {superc is prim, subc is prim}:
3712 _prim_0_path = 1, // {P,N} => false
3713 // {P,P} & superc!=subc => false
3714 _prim_same_path, // {P,P} & superc==subc => true
3715 _prim_1_path, // {N,P} => false
3716 _ref_subtype_path, // {N,N} & subtype check wins => true
3717 _both_ref_path, // {N,N} & subtype check loses => false
3718 PATH_LIMIT
3719 };
3720
3721 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3722 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3723 record_for_igvn(region);
3724
3725 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3726 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3727 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3728
3729 // First null-check both mirrors and load each mirror's klass metaobject.
3730 int which_arg;
3731 for (which_arg = 0; which_arg <= 1; which_arg++) {
3732 Node* arg = args[which_arg];
3733 arg = null_check(arg);
3734 if (stopped()) break;
3735 args[which_arg] = arg;
3736
3737 Node* p = basic_plus_adr(arg, class_klass_offset);
3738 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3739 klasses[which_arg] = _gvn.transform(kls);
3740 }
3741
3742 // Having loaded both klasses, test each for null.
3743 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3744 for (which_arg = 0; which_arg <= 1; which_arg++) {
3745 Node* kls = klasses[which_arg];
3746 Node* null_ctl = top();
3747 kls = null_check_oop(kls, &null_ctl, never_see_null);
3748 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3749 region->init_req(prim_path, null_ctl);
3750 if (stopped()) break;
3751 klasses[which_arg] = kls;
3752 }
3753
3754 if (!stopped()) {
3755 // now we have two reference types, in klasses[0..1]
3756 Node* subk = klasses[1]; // the argument to isAssignableFrom
3757 Node* superk = klasses[0]; // the receiver
3758 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3759 // now we have a successful reference subtype check
3760 region->set_req(_ref_subtype_path, control());
3761 }
3762
3763 // If both operands are primitive (both klasses null), then
3764 // we must return true when they are identical primitives.
3765 // It is convenient to test this after the first null klass check.
3766 set_control(region->in(_prim_0_path)); // go back to first null check
3767 if (!stopped()) {
3768 // Since superc is primitive, make a guard for the superc==subc case.
3769 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3770 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3771 generate_guard(bol_eq, region, PROB_FAIR);
3772 if (region->req() == PATH_LIMIT+1) {
3773 // A guard was added. If the added guard is taken, superc==subc.
3774 region->swap_edges(PATH_LIMIT, _prim_same_path);
3775 region->del_req(PATH_LIMIT);
3776 }
3777 region->set_req(_prim_0_path, control()); // Not equal after all.
3778 }
3779
3780 // these are the only paths that produce 'true':
3781 phi->set_req(_prim_same_path, intcon(1));
3782 phi->set_req(_ref_subtype_path, intcon(1));
3783
3784 // pull together the cases:
3785 assert(region->req() == PATH_LIMIT, "sane region");
3786 for (uint i = 1; i < region->req(); i++) {
3787 Node* ctl = region->in(i);
3788 if (ctl == NULL || ctl == top()) {
3789 region->set_req(i, top());
3790 phi ->set_req(i, top());
3791 } else if (phi->in(i) == NULL) {
3792 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3793 }
3794 }
3795
3796 set_control(_gvn.transform(region));
3797 set_result(_gvn.transform(phi));
3798 return true;
3799 }
3800
3801 //---------------------generate_array_guard_common------------------------
3802 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3803 bool obj_array, bool not_array) {
3804 // If obj_array/non_array==false/false:
3805 // Branch around if the given klass is in fact an array (either obj or prim).
3806 // If obj_array/non_array==false/true:
3807 // Branch around if the given klass is not an array klass of any kind.
3808 // If obj_array/non_array==true/true:
3809 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3810 // If obj_array/non_array==true/false:
3811 // Branch around if the kls is an oop array (Object[] or subtype)
3812 //
3813 // Like generate_guard, adds a new path onto the region.
3814 jint layout_con = 0;
3815 Node* layout_val = get_layout_helper(kls, layout_con);
3816 if (layout_val == NULL) {
3817 bool query = (obj_array
3818 ? Klass::layout_helper_is_objArray(layout_con)
3819 : Klass::layout_helper_is_array(layout_con));
3820 if (query == not_array) {
3821 return NULL; // never a branch
3822 } else { // always a branch
3823 Node* always_branch = control();
3824 if (region != NULL)
3825 region->add_req(always_branch);
3826 set_control(top());
3827 return always_branch;
3828 }
3829 }
3830 // Now test the correct condition.
3831 jint nval = (obj_array
3832 ? (jint)(Klass::_lh_array_tag_type_value
3833 << Klass::_lh_array_tag_shift)
3834 : Klass::_lh_neutral_value);
3835 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3836 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3837 // invert the test if we are looking for a non-array
3838 if (not_array) btest = BoolTest(btest).negate();
3839 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3840 return generate_fair_guard(bol, region);
3841 }
3842
3843
3844 //-----------------------inline_native_newArray--------------------------
3845 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3846 bool LibraryCallKit::inline_native_newArray() {
3847 Node* mirror = argument(0);
3848 Node* count_val = argument(1);
3849
3850 mirror = null_check(mirror);
3851 // If mirror or obj is dead, only null-path is taken.
3852 if (stopped()) return true;
3853
3854 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3855 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3856 PhiNode* result_val = new(C) PhiNode(result_reg,
3857 TypeInstPtr::NOTNULL);
3858 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3859 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3860 TypePtr::BOTTOM);
3861
3862 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3863 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3864 result_reg, _slow_path);
3865 Node* normal_ctl = control();
3866 Node* no_array_ctl = result_reg->in(_slow_path);
3867
3868 // Generate code for the slow case. We make a call to newArray().
3869 set_control(no_array_ctl);
3870 if (!stopped()) {
3871 // Either the input type is void.class, or else the
3872 // array klass has not yet been cached. Either the
3873 // ensuing call will throw an exception, or else it
3874 // will cache the array klass for next time.
3875 PreserveJVMState pjvms(this);
3876 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3877 Node* slow_result = set_results_for_java_call(slow_call);
3878 // this->control() comes from set_results_for_java_call
3879 result_reg->set_req(_slow_path, control());
3880 result_val->set_req(_slow_path, slow_result);
3881 result_io ->set_req(_slow_path, i_o());
3882 result_mem->set_req(_slow_path, reset_memory());
3883 }
3884
3885 set_control(normal_ctl);
3886 if (!stopped()) {
3887 // Normal case: The array type has been cached in the java.lang.Class.
3888 // The following call works fine even if the array type is polymorphic.
3889 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3890 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3891 result_reg->init_req(_normal_path, control());
3892 result_val->init_req(_normal_path, obj);
3893 result_io ->init_req(_normal_path, i_o());
3894 result_mem->init_req(_normal_path, reset_memory());
3895 }
3896
3897 // Return the combined state.
3898 set_i_o( _gvn.transform(result_io) );
3899 set_all_memory( _gvn.transform(result_mem));
3900
3901 C->set_has_split_ifs(true); // Has chance for split-if optimization
3902 set_result(result_reg, result_val);
3903 return true;
3904 }
3905
3906 //----------------------inline_native_getLength--------------------------
3907 // public static native int java.lang.reflect.Array.getLength(Object array);
3908 bool LibraryCallKit::inline_native_getLength() {
3909 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3910
3911 Node* array = null_check(argument(0));
3912 // If array is dead, only null-path is taken.
3913 if (stopped()) return true;
3914
3915 // Deoptimize if it is a non-array.
3916 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3917
3918 if (non_array != NULL) {
3919 PreserveJVMState pjvms(this);
3920 set_control(non_array);
3921 uncommon_trap(Deoptimization::Reason_intrinsic,
3922 Deoptimization::Action_maybe_recompile);
3923 }
3924
3925 // If control is dead, only non-array-path is taken.
3926 if (stopped()) return true;
3927
3928 // The works fine even if the array type is polymorphic.
3929 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3930 Node* result = load_array_length(array);
3931
3932 C->set_has_split_ifs(true); // Has chance for split-if optimization
3933 set_result(result);
3934 return true;
3935 }
3936
3937 //------------------------inline_array_copyOf----------------------------
3938 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3939 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3940 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3941 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3942
3943 // Get the arguments.
3944 Node* original = argument(0);
3945 Node* start = is_copyOfRange? argument(1): intcon(0);
3946 Node* end = is_copyOfRange? argument(2): argument(1);
3947 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3948
3949 Node* newcopy = NULL;
3950
3951 // Set the original stack and the reexecute bit for the interpreter to reexecute
3952 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3953 { PreserveReexecuteState preexecs(this);
3954 jvms()->set_should_reexecute(true);
3955
3956 array_type_mirror = null_check(array_type_mirror);
3957 original = null_check(original);
3958
3959 // Check if a null path was taken unconditionally.
3960 if (stopped()) return true;
3961
3962 Node* orig_length = load_array_length(original);
3963
3964 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3965 klass_node = null_check(klass_node);
3966
3967 RegionNode* bailout = new (C) RegionNode(1);
3968 record_for_igvn(bailout);
3969
3970 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3971 // Bail out if that is so.
3972 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3973 if (not_objArray != NULL) {
3974 // Improve the klass node's type from the new optimistic assumption:
3975 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3976 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3977 Node* cast = new (C) CastPPNode(klass_node, akls);
3978 cast->init_req(0, control());
3979 klass_node = _gvn.transform(cast);
3980 }
3981
3982 // Bail out if either start or end is negative.
3983 generate_negative_guard(start, bailout, &start);
3984 generate_negative_guard(end, bailout, &end);
3985
3986 Node* length = end;
3987 if (_gvn.type(start) != TypeInt::ZERO) {
3988 length = _gvn.transform(new (C) SubINode(end, start));
3989 }
3990
3991 // Bail out if length is negative.
3992 // Without this the new_array would throw
3993 // NegativeArraySizeException but IllegalArgumentException is what
3994 // should be thrown
3995 generate_negative_guard(length, bailout, &length);
3996
3997 if (bailout->req() > 1) {
3998 PreserveJVMState pjvms(this);
3999 set_control(_gvn.transform(bailout));
4000 uncommon_trap(Deoptimization::Reason_intrinsic,
4001 Deoptimization::Action_maybe_recompile);
4002 }
4003
4004 if (!stopped()) {
4005 // How many elements will we copy from the original?
4006 // The answer is MinI(orig_length - start, length).
4007 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
4008 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4009
4010 newcopy = new_array(klass_node, length, 0); // no argments to push
4011
4012 // Generate a direct call to the right arraycopy function(s).
4013 // We know the copy is disjoint but we might not know if the
4014 // oop stores need checking.
4015 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4016 // This will fail a store-check if x contains any non-nulls.
4017 bool disjoint_bases = true;
4018 // if start > orig_length then the length of the copy may be
4019 // negative.
4020 bool length_never_negative = !is_copyOfRange;
4021 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4022 original, start, newcopy, intcon(0), moved,
4023 disjoint_bases, length_never_negative);
4024 }
4025 } // original reexecute is set back here
4026
4027 C->set_has_split_ifs(true); // Has chance for split-if optimization
4028 if (!stopped()) {
4029 set_result(newcopy);
4030 }
4031 return true;
4032 }
4033
4034
4035 //----------------------generate_virtual_guard---------------------------
4036 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4037 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4038 RegionNode* slow_region) {
4039 ciMethod* method = callee();
4040 int vtable_index = method->vtable_index();
4041 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4042 err_msg_res("bad index %d", vtable_index));
4043 // Get the Method* out of the appropriate vtable entry.
4044 int entry_offset = (InstanceKlass::vtable_start_offset() +
4045 vtable_index*vtableEntry::size()) * wordSize +
4046 vtableEntry::method_offset_in_bytes();
4047 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4048 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4049
4050 // Compare the target method with the expected method (e.g., Object.hashCode).
4051 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4052
4053 Node* native_call = makecon(native_call_addr);
4054 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4055 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4056
4057 return generate_slow_guard(test_native, slow_region);
4058 }
4059
4060 //-----------------------generate_method_call----------------------------
4061 // Use generate_method_call to make a slow-call to the real
4062 // method if the fast path fails. An alternative would be to
4063 // use a stub like OptoRuntime::slow_arraycopy_Java.
4064 // This only works for expanding the current library call,
4065 // not another intrinsic. (E.g., don't use this for making an
4066 // arraycopy call inside of the copyOf intrinsic.)
4067 CallJavaNode*
4068 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4069 // When compiling the intrinsic method itself, do not use this technique.
4070 guarantee(callee() != C->method(), "cannot make slow-call to self");
4071
4072 ciMethod* method = callee();
4073 // ensure the JVMS we have will be correct for this call
4074 guarantee(method_id == method->intrinsic_id(), "must match");
4075
4076 const TypeFunc* tf = TypeFunc::make(method);
4077 CallJavaNode* slow_call;
4078 if (is_static) {
4079 assert(!is_virtual, "");
4080 slow_call = new(C) CallStaticJavaNode(C, tf,
4081 SharedRuntime::get_resolve_static_call_stub(),
4082 method, bci());
4083 } else if (is_virtual) {
4084 null_check_receiver();
4085 int vtable_index = Method::invalid_vtable_index;
4086 if (UseInlineCaches) {
4087 // Suppress the vtable call
4088 } else {
4089 // hashCode and clone are not a miranda methods,
4090 // so the vtable index is fixed.
4091 // No need to use the linkResolver to get it.
4092 vtable_index = method->vtable_index();
4093 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4094 err_msg_res("bad index %d", vtable_index));
4095 }
4096 slow_call = new(C) CallDynamicJavaNode(tf,
4097 SharedRuntime::get_resolve_virtual_call_stub(),
4098 method, vtable_index, bci());
4099 } else { // neither virtual nor static: opt_virtual
4100 null_check_receiver();
4101 slow_call = new(C) CallStaticJavaNode(C, tf,
4102 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4103 method, bci());
4104 slow_call->set_optimized_virtual(true);
4105 }
4106 set_arguments_for_java_call(slow_call);
4107 set_edges_for_java_call(slow_call);
4108 return slow_call;
4109 }
4110
4111
4112 /**
4113 * Build special case code for calls to hashCode on an object. This call may
4114 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4115 * slightly different code.
4116 */
4117 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4118 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4119 assert(!(is_virtual && is_static), "either virtual, special, or static");
4120
4121 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4122
4123 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4124 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4125 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4126 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4127 Node* obj = NULL;
4128 if (!is_static) {
4129 // Check for hashing null object
4130 obj = null_check_receiver();
4131 if (stopped()) return true; // unconditionally null
4132 result_reg->init_req(_null_path, top());
4133 result_val->init_req(_null_path, top());
4134 } else {
4135 // Do a null check, and return zero if null.
4136 // System.identityHashCode(null) == 0
4137 obj = argument(0);
4138 Node* null_ctl = top();
4139 obj = null_check_oop(obj, &null_ctl);
4140 result_reg->init_req(_null_path, null_ctl);
4141 result_val->init_req(_null_path, _gvn.intcon(0));
4142 }
4143
4144 // Unconditionally null? Then return right away.
4145 if (stopped()) {
4146 set_control( result_reg->in(_null_path));
4147 if (!stopped())
4148 set_result(result_val->in(_null_path));
4149 return true;
4150 }
4151
4152 // We only go to the fast case code if we pass a number of guards. The
4153 // paths which do not pass are accumulated in the slow_region.
4154 RegionNode* slow_region = new (C) RegionNode(1);
4155 record_for_igvn(slow_region);
4156
4157 // If this is a virtual call, we generate a funny guard. We pull out
4158 // the vtable entry corresponding to hashCode() from the target object.
4159 // If the target method which we are calling happens to be the native
4160 // Object hashCode() method, we pass the guard. We do not need this
4161 // guard for non-virtual calls -- the caller is known to be the native
4162 // Object hashCode().
4163 if (is_virtual) {
4164 // After null check, get the object's klass.
4165 Node* obj_klass = load_object_klass(obj);
4166 generate_virtual_guard(obj_klass, slow_region);
4167 }
4168
4169 // Get the header out of the object, use LoadMarkNode when available
4170 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4171 // The control of the load must be NULL. Otherwise, the load can move before
4172 // the null check after castPP removal.
4173 Node* no_ctrl = NULL;
4174 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4175
4176 // Test the header to see if it is unlocked.
4177 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4178 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4179 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4180 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4181 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4182
4183 generate_slow_guard(test_unlocked, slow_region);
4184
4185 // Get the hash value and check to see that it has been properly assigned.
4186 // We depend on hash_mask being at most 32 bits and avoid the use of
4187 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4188 // vm: see markOop.hpp.
4189 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4190 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4191 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4192 // This hack lets the hash bits live anywhere in the mark object now, as long
4193 // as the shift drops the relevant bits into the low 32 bits. Note that
4194 // Java spec says that HashCode is an int so there's no point in capturing
4195 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4196 hshifted_header = ConvX2I(hshifted_header);
4197 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4198
4199 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4200 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4201 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4202
4203 generate_slow_guard(test_assigned, slow_region);
4204
4205 Node* init_mem = reset_memory();
4206 // fill in the rest of the null path:
4207 result_io ->init_req(_null_path, i_o());
4208 result_mem->init_req(_null_path, init_mem);
4209
4210 result_val->init_req(_fast_path, hash_val);
4211 result_reg->init_req(_fast_path, control());
4212 result_io ->init_req(_fast_path, i_o());
4213 result_mem->init_req(_fast_path, init_mem);
4214
4215 // Generate code for the slow case. We make a call to hashCode().
4216 set_control(_gvn.transform(slow_region));
4217 if (!stopped()) {
4218 // No need for PreserveJVMState, because we're using up the present state.
4219 set_all_memory(init_mem);
4220 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4221 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4222 Node* slow_result = set_results_for_java_call(slow_call);
4223 // this->control() comes from set_results_for_java_call
4224 result_reg->init_req(_slow_path, control());
4225 result_val->init_req(_slow_path, slow_result);
4226 result_io ->set_req(_slow_path, i_o());
4227 result_mem ->set_req(_slow_path, reset_memory());
4228 }
4229
4230 // Return the combined state.
4231 set_i_o( _gvn.transform(result_io) );
4232 set_all_memory( _gvn.transform(result_mem));
4233
4234 set_result(result_reg, result_val);
4235 return true;
4236 }
4237
4238 //---------------------------inline_native_getClass----------------------------
4239 // public final native Class<?> java.lang.Object.getClass();
4240 //
4241 // Build special case code for calls to getClass on an object.
4242 bool LibraryCallKit::inline_native_getClass() {
4243 Node* obj = null_check_receiver();
4244 if (stopped()) return true;
4245 set_result(load_mirror_from_klass(load_object_klass(obj)));
4246 return true;
4247 }
4248
4249 //-----------------inline_native_Reflection_getCallerClass---------------------
4250 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4251 //
4252 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4253 //
4254 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4255 // in that it must skip particular security frames and checks for
4256 // caller sensitive methods.
4257 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4258 #ifndef PRODUCT
4259 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4260 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4261 }
4262 #endif
4263
4264 if (!jvms()->has_method()) {
4265 #ifndef PRODUCT
4266 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4267 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4268 }
4269 #endif
4270 return false;
4271 }
4272
4273 // Walk back up the JVM state to find the caller at the required
4274 // depth.
4275 JVMState* caller_jvms = jvms();
4276
4277 // Cf. JVM_GetCallerClass
4278 // NOTE: Start the loop at depth 1 because the current JVM state does
4279 // not include the Reflection.getCallerClass() frame.
4280 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4281 ciMethod* m = caller_jvms->method();
4282 switch (n) {
4283 case 0:
4284 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4285 break;
4286 case 1:
4287 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4288 if (!m->caller_sensitive()) {
4289 #ifndef PRODUCT
4290 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4291 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4292 }
4293 #endif
4294 return false; // bail-out; let JVM_GetCallerClass do the work
4295 }
4296 break;
4297 default:
4298 if (!m->is_ignored_by_security_stack_walk()) {
4299 // We have reached the desired frame; return the holder class.
4300 // Acquire method holder as java.lang.Class and push as constant.
4301 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4302 ciInstance* caller_mirror = caller_klass->java_mirror();
4303 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4304
4305 #ifndef PRODUCT
4306 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4307 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());
4308 tty->print_cr(" JVM state at this point:");
4309 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4310 ciMethod* m = jvms()->of_depth(i)->method();
4311 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4312 }
4313 }
4314 #endif
4315 return true;
4316 }
4317 break;
4318 }
4319 }
4320
4321 #ifndef PRODUCT
4322 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4323 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4324 tty->print_cr(" JVM state at this point:");
4325 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4326 ciMethod* m = jvms()->of_depth(i)->method();
4327 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4328 }
4329 }
4330 #endif
4331
4332 return false; // bail-out; let JVM_GetCallerClass do the work
4333 }
4334
4335 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4336 Node* arg = argument(0);
4337 Node* result = NULL;
4338
4339 switch (id) {
4340 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4341 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4342 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4343 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4344
4345 case vmIntrinsics::_doubleToLongBits: {
4346 // two paths (plus control) merge in a wood
4347 RegionNode *r = new (C) RegionNode(3);
4348 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4349
4350 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4351 // Build the boolean node
4352 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4353
4354 // Branch either way.
4355 // NaN case is less traveled, which makes all the difference.
4356 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4357 Node *opt_isnan = _gvn.transform(ifisnan);
4358 assert( opt_isnan->is_If(), "Expect an IfNode");
4359 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4360 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4361
4362 set_control(iftrue);
4363
4364 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4365 Node *slow_result = longcon(nan_bits); // return NaN
4366 phi->init_req(1, _gvn.transform( slow_result ));
4367 r->init_req(1, iftrue);
4368
4369 // Else fall through
4370 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4371 set_control(iffalse);
4372
4373 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4374 r->init_req(2, iffalse);
4375
4376 // Post merge
4377 set_control(_gvn.transform(r));
4378 record_for_igvn(r);
4379
4380 C->set_has_split_ifs(true); // Has chance for split-if optimization
4381 result = phi;
4382 assert(result->bottom_type()->isa_long(), "must be");
4383 break;
4384 }
4385
4386 case vmIntrinsics::_floatToIntBits: {
4387 // two paths (plus control) merge in a wood
4388 RegionNode *r = new (C) RegionNode(3);
4389 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4390
4391 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4392 // Build the boolean node
4393 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4394
4395 // Branch either way.
4396 // NaN case is less traveled, which makes all the difference.
4397 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4398 Node *opt_isnan = _gvn.transform(ifisnan);
4399 assert( opt_isnan->is_If(), "Expect an IfNode");
4400 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4401 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4402
4403 set_control(iftrue);
4404
4405 static const jint nan_bits = 0x7fc00000;
4406 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4407 phi->init_req(1, _gvn.transform( slow_result ));
4408 r->init_req(1, iftrue);
4409
4410 // Else fall through
4411 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4412 set_control(iffalse);
4413
4414 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4415 r->init_req(2, iffalse);
4416
4417 // Post merge
4418 set_control(_gvn.transform(r));
4419 record_for_igvn(r);
4420
4421 C->set_has_split_ifs(true); // Has chance for split-if optimization
4422 result = phi;
4423 assert(result->bottom_type()->isa_int(), "must be");
4424 break;
4425 }
4426
4427 default:
4428 fatal_unexpected_iid(id);
4429 break;
4430 }
4431 set_result(_gvn.transform(result));
4432 return true;
4433 }
4434
4435 #ifdef _LP64
4436 #define XTOP ,top() /*additional argument*/
4437 #else //_LP64
4438 #define XTOP /*no additional argument*/
4439 #endif //_LP64
4440
4441 //----------------------inline_unsafe_copyMemory-------------------------
4442 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4443 bool LibraryCallKit::inline_unsafe_copyMemory() {
4444 if (callee()->is_static()) return false; // caller must have the capability!
4445 null_check_receiver(); // null-check receiver
4446 if (stopped()) return true;
4447
4448 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4449
4450 Node* src_ptr = argument(1); // type: oop
4451 Node* src_off = ConvL2X(argument(2)); // type: long
4452 Node* dst_ptr = argument(4); // type: oop
4453 Node* dst_off = ConvL2X(argument(5)); // type: long
4454 Node* size = ConvL2X(argument(7)); // type: long
4455
4456 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4457 "fieldOffset must be byte-scaled");
4458
4459 Node* src = make_unsafe_address(src_ptr, src_off);
4460 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4461
4462 // Conservatively insert a memory barrier on all memory slices.
4463 // Do not let writes of the copy source or destination float below the copy.
4464 insert_mem_bar(Op_MemBarCPUOrder);
4465
4466 // Call it. Note that the length argument is not scaled.
4467 make_runtime_call(RC_LEAF|RC_NO_FP,
4468 OptoRuntime::fast_arraycopy_Type(),
4469 StubRoutines::unsafe_arraycopy(),
4470 "unsafe_arraycopy",
4471 TypeRawPtr::BOTTOM,
4472 src, dst, size XTOP);
4473
4474 // Do not let reads of the copy destination float above the copy.
4475 insert_mem_bar(Op_MemBarCPUOrder);
4476
4477 return true;
4478 }
4479
4480 //------------------------clone_coping-----------------------------------
4481 // Helper function for inline_native_clone.
4482 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4483 assert(obj_size != NULL, "");
4484 Node* raw_obj = alloc_obj->in(1);
4485 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4486
4487 AllocateNode* alloc = NULL;
4488 if (ReduceBulkZeroing) {
4489 // We will be completely responsible for initializing this object -
4490 // mark Initialize node as complete.
4491 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4492 // The object was just allocated - there should be no any stores!
4493 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4494 // Mark as complete_with_arraycopy so that on AllocateNode
4495 // expansion, we know this AllocateNode is initialized by an array
4496 // copy and a StoreStore barrier exists after the array copy.
4497 alloc->initialization()->set_complete_with_arraycopy();
4498 }
4499
4500 // Copy the fastest available way.
4501 // TODO: generate fields copies for small objects instead.
4502 Node* src = obj;
4503 Node* dest = alloc_obj;
4504 Node* size = _gvn.transform(obj_size);
4505
4506 // Exclude the header but include array length to copy by 8 bytes words.
4507 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4508 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4509 instanceOopDesc::base_offset_in_bytes();
4510 // base_off:
4511 // 8 - 32-bit VM
4512 // 12 - 64-bit VM, compressed klass
4513 // 16 - 64-bit VM, normal klass
4514 if (base_off % BytesPerLong != 0) {
4515 assert(UseCompressedClassPointers, "");
4516 if (is_array) {
4517 // Exclude length to copy by 8 bytes words.
4518 base_off += sizeof(int);
4519 } else {
4520 // Include klass to copy by 8 bytes words.
4521 base_off = instanceOopDesc::klass_offset_in_bytes();
4522 }
4523 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4524 }
4525 src = basic_plus_adr(src, base_off);
4526 dest = basic_plus_adr(dest, base_off);
4527
4528 // Compute the length also, if needed:
4529 Node* countx = size;
4530 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4531 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4532
4533 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4534 bool disjoint_bases = true;
4535 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4536 src, NULL, dest, NULL, countx,
4537 /*dest_uninitialized*/true);
4538
4539 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4540 if (card_mark) {
4541 assert(!is_array, "");
4542 // Put in store barrier for any and all oops we are sticking
4543 // into this object. (We could avoid this if we could prove
4544 // that the object type contains no oop fields at all.)
4545 Node* no_particular_value = NULL;
4546 Node* no_particular_field = NULL;
4547 int raw_adr_idx = Compile::AliasIdxRaw;
4548 post_barrier(control(),
4549 memory(raw_adr_type),
4550 alloc_obj,
4551 no_particular_field,
4552 raw_adr_idx,
4553 no_particular_value,
4554 T_OBJECT,
4555 false);
4556 }
4557
4558 // Do not let reads from the cloned object float above the arraycopy.
4559 if (alloc != NULL) {
4560 // Do not let stores that initialize this object be reordered with
4561 // a subsequent store that would make this object accessible by
4562 // other threads.
4563 // Record what AllocateNode this StoreStore protects so that
4564 // escape analysis can go from the MemBarStoreStoreNode to the
4565 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4566 // based on the escape status of the AllocateNode.
4567 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4568 } else {
4569 insert_mem_bar(Op_MemBarCPUOrder);
4570 }
4571 }
4572
4573 //------------------------inline_native_clone----------------------------
4574 // protected native Object java.lang.Object.clone();
4575 //
4576 // Here are the simple edge cases:
4577 // null receiver => normal trap
4578 // virtual and clone was overridden => slow path to out-of-line clone
4579 // not cloneable or finalizer => slow path to out-of-line Object.clone
4580 //
4581 // The general case has two steps, allocation and copying.
4582 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4583 //
4584 // Copying also has two cases, oop arrays and everything else.
4585 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4586 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4587 //
4588 // These steps fold up nicely if and when the cloned object's klass
4589 // can be sharply typed as an object array, a type array, or an instance.
4590 //
4591 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4592 PhiNode* result_val;
4593
4594 // Set the reexecute bit for the interpreter to reexecute
4595 // the bytecode that invokes Object.clone if deoptimization happens.
4596 { PreserveReexecuteState preexecs(this);
4597 jvms()->set_should_reexecute(true);
4598
4599 Node* obj = null_check_receiver();
4600 if (stopped()) return true;
4601
4602 Node* obj_klass = load_object_klass(obj);
4603 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4604 const TypeOopPtr* toop = ((tklass != NULL)
4605 ? tklass->as_instance_type()
4606 : TypeInstPtr::NOTNULL);
4607
4608 // Conservatively insert a memory barrier on all memory slices.
4609 // Do not let writes into the original float below the clone.
4610 insert_mem_bar(Op_MemBarCPUOrder);
4611
4612 // paths into result_reg:
4613 enum {
4614 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4615 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4616 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4617 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4618 PATH_LIMIT
4619 };
4620 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4621 result_val = new(C) PhiNode(result_reg,
4622 TypeInstPtr::NOTNULL);
4623 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4624 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4625 TypePtr::BOTTOM);
4626 record_for_igvn(result_reg);
4627
4628 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4629 int raw_adr_idx = Compile::AliasIdxRaw;
4630
4631 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4632 if (array_ctl != NULL) {
4633 // It's an array.
4634 PreserveJVMState pjvms(this);
4635 set_control(array_ctl);
4636 Node* obj_length = load_array_length(obj);
4637 Node* obj_size = NULL;
4638 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4639
4640 if (!use_ReduceInitialCardMarks()) {
4641 // If it is an oop array, it requires very special treatment,
4642 // because card marking is required on each card of the array.
4643 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4644 if (is_obja != NULL) {
4645 PreserveJVMState pjvms2(this);
4646 set_control(is_obja);
4647 // Generate a direct call to the right arraycopy function(s).
4648 bool disjoint_bases = true;
4649 bool length_never_negative = true;
4650 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4651 obj, intcon(0), alloc_obj, intcon(0),
4652 obj_length,
4653 disjoint_bases, length_never_negative);
4654 result_reg->init_req(_objArray_path, control());
4655 result_val->init_req(_objArray_path, alloc_obj);
4656 result_i_o ->set_req(_objArray_path, i_o());
4657 result_mem ->set_req(_objArray_path, reset_memory());
4658 }
4659 }
4660 // Otherwise, there are no card marks to worry about.
4661 // (We can dispense with card marks if we know the allocation
4662 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4663 // causes the non-eden paths to take compensating steps to
4664 // simulate a fresh allocation, so that no further
4665 // card marks are required in compiled code to initialize
4666 // the object.)
4667
4668 if (!stopped()) {
4669 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4670
4671 // Present the results of the copy.
4672 result_reg->init_req(_array_path, control());
4673 result_val->init_req(_array_path, alloc_obj);
4674 result_i_o ->set_req(_array_path, i_o());
4675 result_mem ->set_req(_array_path, reset_memory());
4676 }
4677 }
4678
4679 // We only go to the instance fast case code if we pass a number of guards.
4680 // The paths which do not pass are accumulated in the slow_region.
4681 RegionNode* slow_region = new (C) RegionNode(1);
4682 record_for_igvn(slow_region);
4683 if (!stopped()) {
4684 // It's an instance (we did array above). Make the slow-path tests.
4685 // If this is a virtual call, we generate a funny guard. We grab
4686 // the vtable entry corresponding to clone() from the target object.
4687 // If the target method which we are calling happens to be the
4688 // Object clone() method, we pass the guard. We do not need this
4689 // guard for non-virtual calls; the caller is known to be the native
4690 // Object clone().
4691 if (is_virtual) {
4692 generate_virtual_guard(obj_klass, slow_region);
4693 }
4694
4695 // The object must be cloneable and must not have a finalizer.
4696 // Both of these conditions may be checked in a single test.
4697 // We could optimize the cloneable test further, but we don't care.
4698 generate_access_flags_guard(obj_klass,
4699 // Test both conditions:
4700 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4701 // Must be cloneable but not finalizer:
4702 JVM_ACC_IS_CLONEABLE,
4703 slow_region);
4704 }
4705
4706 if (!stopped()) {
4707 // It's an instance, and it passed the slow-path tests.
4708 PreserveJVMState pjvms(this);
4709 Node* obj_size = NULL;
4710 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4711 // is reexecuted if deoptimization occurs and there could be problems when merging
4712 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4713 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4714
4715 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4716
4717 // Present the results of the slow call.
4718 result_reg->init_req(_instance_path, control());
4719 result_val->init_req(_instance_path, alloc_obj);
4720 result_i_o ->set_req(_instance_path, i_o());
4721 result_mem ->set_req(_instance_path, reset_memory());
4722 }
4723
4724 // Generate code for the slow case. We make a call to clone().
4725 set_control(_gvn.transform(slow_region));
4726 if (!stopped()) {
4727 PreserveJVMState pjvms(this);
4728 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4729 Node* slow_result = set_results_for_java_call(slow_call);
4730 // this->control() comes from set_results_for_java_call
4731 result_reg->init_req(_slow_path, control());
4732 result_val->init_req(_slow_path, slow_result);
4733 result_i_o ->set_req(_slow_path, i_o());
4734 result_mem ->set_req(_slow_path, reset_memory());
4735 }
4736
4737 // Return the combined state.
4738 set_control( _gvn.transform(result_reg));
4739 set_i_o( _gvn.transform(result_i_o));
4740 set_all_memory( _gvn.transform(result_mem));
4741 } // original reexecute is set back here
4742
4743 set_result(_gvn.transform(result_val));
4744 return true;
4745 }
4746
4747 //------------------------------basictype2arraycopy----------------------------
4748 address LibraryCallKit::basictype2arraycopy(BasicType t,
4749 Node* src_offset,
4750 Node* dest_offset,
4751 bool disjoint_bases,
4752 const char* &name,
4753 bool dest_uninitialized) {
4754 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4755 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4756
4757 bool aligned = false;
4758 bool disjoint = disjoint_bases;
4759
4760 // if the offsets are the same, we can treat the memory regions as
4761 // disjoint, because either the memory regions are in different arrays,
4762 // or they are identical (which we can treat as disjoint.) We can also
4763 // treat a copy with a destination index less that the source index
4764 // as disjoint since a low->high copy will work correctly in this case.
4765 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4766 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4767 // both indices are constants
4768 int s_offs = src_offset_inttype->get_con();
4769 int d_offs = dest_offset_inttype->get_con();
4770 int element_size = type2aelembytes(t);
4771 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4772 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4773 if (s_offs >= d_offs) disjoint = true;
4774 } else if (src_offset == dest_offset && src_offset != NULL) {
4775 // This can occur if the offsets are identical non-constants.
4776 disjoint = true;
4777 }
4778
4779 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4780 }
4781
4782
4783 //------------------------------inline_arraycopy-----------------------
4784 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4785 // Object dest, int destPos,
4786 // int length);
4787 bool LibraryCallKit::inline_arraycopy() {
4788 // Get the arguments.
4789 Node* src = argument(0); // type: oop
4790 Node* src_offset = argument(1); // type: int
4791 Node* dest = argument(2); // type: oop
4792 Node* dest_offset = argument(3); // type: int
4793 Node* length = argument(4); // type: int
4794
4795 // Compile time checks. If any of these checks cannot be verified at compile time,
4796 // we do not make a fast path for this call. Instead, we let the call remain as it
4797 // is. The checks we choose to mandate at compile time are:
4798 //
4799 // (1) src and dest are arrays.
4800 const Type* src_type = src->Value(&_gvn);
4801 const Type* dest_type = dest->Value(&_gvn);
4802 const TypeAryPtr* top_src = src_type->isa_aryptr();
4803 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4804
4805 // Do we have the type of src?
4806 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4807 // Do we have the type of dest?
4808 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4809 // Is the type for src from speculation?
4810 bool src_spec = false;
4811 // Is the type for dest from speculation?
4812 bool dest_spec = false;
4813
4814 if (!has_src || !has_dest) {
4815 // We don't have sufficient type information, let's see if
4816 // speculative types can help. We need to have types for both src
4817 // and dest so that it pays off.
4818
4819 // Do we already have or could we have type information for src
4820 bool could_have_src = has_src;
4821 // Do we already have or could we have type information for dest
4822 bool could_have_dest = has_dest;
4823
4824 ciKlass* src_k = NULL;
4825 if (!has_src) {
4826 src_k = src_type->speculative_type();
4827 if (src_k != NULL && src_k->is_array_klass()) {
4828 could_have_src = true;
4829 }
4830 }
4831
4832 ciKlass* dest_k = NULL;
4833 if (!has_dest) {
4834 dest_k = dest_type->speculative_type();
4835 if (dest_k != NULL && dest_k->is_array_klass()) {
4836 could_have_dest = true;
4837 }
4838 }
4839
4840 if (could_have_src && could_have_dest) {
4841 // This is going to pay off so emit the required guards
4842 if (!has_src) {
4843 src = maybe_cast_profiled_obj(src, src_k);
4844 src_type = _gvn.type(src);
4845 top_src = src_type->isa_aryptr();
4846 has_src = (top_src != NULL && top_src->klass() != NULL);
4847 src_spec = true;
4848 }
4849 if (!has_dest) {
4850 dest = maybe_cast_profiled_obj(dest, dest_k);
4851 dest_type = _gvn.type(dest);
4852 top_dest = dest_type->isa_aryptr();
4853 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4854 dest_spec = true;
4855 }
4856 }
4857 }
4858
4859 if (!has_src || !has_dest) {
4860 // Conservatively insert a memory barrier on all memory slices.
4861 // Do not let writes into the source float below the arraycopy.
4862 insert_mem_bar(Op_MemBarCPUOrder);
4863
4864 // Call StubRoutines::generic_arraycopy stub.
4865 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4866 src, src_offset, dest, dest_offset, length);
4867
4868 // Do not let reads from the destination float above the arraycopy.
4869 // Since we cannot type the arrays, we don't know which slices
4870 // might be affected. We could restrict this barrier only to those
4871 // memory slices which pertain to array elements--but don't bother.
4872 if (!InsertMemBarAfterArraycopy)
4873 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4874 insert_mem_bar(Op_MemBarCPUOrder);
4875 return true;
4876 }
4877
4878 // (2) src and dest arrays must have elements of the same BasicType
4879 // Figure out the size and type of the elements we will be copying.
4880 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4881 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4882 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4883 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4884
4885 if (src_elem != dest_elem || dest_elem == T_VOID) {
4886 // The component types are not the same or are not recognized. Punt.
4887 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4888 generate_slow_arraycopy(TypePtr::BOTTOM,
4889 src, src_offset, dest, dest_offset, length,
4890 /*dest_uninitialized*/false);
4891 return true;
4892 }
4893
4894 if (src_elem == T_OBJECT) {
4895 // If both arrays are object arrays then having the exact types
4896 // for both will remove the need for a subtype check at runtime
4897 // before the call and may make it possible to pick a faster copy
4898 // routine (without a subtype check on every element)
4899 // Do we have the exact type of src?
4900 bool could_have_src = src_spec;
4901 // Do we have the exact type of dest?
4902 bool could_have_dest = dest_spec;
4903 ciKlass* src_k = top_src->klass();
4904 ciKlass* dest_k = top_dest->klass();
4905 if (!src_spec) {
4906 src_k = src_type->speculative_type();
4907 if (src_k != NULL && src_k->is_array_klass()) {
4908 could_have_src = true;
4909 }
4910 }
4911 if (!dest_spec) {
4912 dest_k = dest_type->speculative_type();
4913 if (dest_k != NULL && dest_k->is_array_klass()) {
4914 could_have_dest = true;
4915 }
4916 }
4917 if (could_have_src && could_have_dest) {
4918 // If we can have both exact types, emit the missing guards
4919 if (could_have_src && !src_spec) {
4920 src = maybe_cast_profiled_obj(src, src_k);
4921 }
4922 if (could_have_dest && !dest_spec) {
4923 dest = maybe_cast_profiled_obj(dest, dest_k);
4924 }
4925 }
4926 }
4927
4928 //---------------------------------------------------------------------------
4929 // We will make a fast path for this call to arraycopy.
4930
4931 // We have the following tests left to perform:
4932 //
4933 // (3) src and dest must not be null.
4934 // (4) src_offset must not be negative.
4935 // (5) dest_offset must not be negative.
4936 // (6) length must not be negative.
4937 // (7) src_offset + length must not exceed length of src.
4938 // (8) dest_offset + length must not exceed length of dest.
4939 // (9) each element of an oop array must be assignable
4940
4941 RegionNode* slow_region = new (C) RegionNode(1);
4942 record_for_igvn(slow_region);
4943
4944 // (3) operands must not be null
4945 // We currently perform our null checks with the null_check routine.
4946 // This means that the null exceptions will be reported in the caller
4947 // rather than (correctly) reported inside of the native arraycopy call.
4948 // This should be corrected, given time. We do our null check with the
4949 // stack pointer restored.
4950 src = null_check(src, T_ARRAY);
4951 dest = null_check(dest, T_ARRAY);
4952
4953 // (4) src_offset must not be negative.
4954 generate_negative_guard(src_offset, slow_region);
4955
4956 // (5) dest_offset must not be negative.
4957 generate_negative_guard(dest_offset, slow_region);
4958
4959 // (6) length must not be negative (moved to generate_arraycopy()).
4960 // generate_negative_guard(length, slow_region);
4961
4962 // (7) src_offset + length must not exceed length of src.
4963 generate_limit_guard(src_offset, length,
4964 load_array_length(src),
4965 slow_region);
4966
4967 // (8) dest_offset + length must not exceed length of dest.
4968 generate_limit_guard(dest_offset, length,
4969 load_array_length(dest),
4970 slow_region);
4971
4972 // (9) each element of an oop array must be assignable
4973 // The generate_arraycopy subroutine checks this.
4974
4975 // This is where the memory effects are placed:
4976 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4977 generate_arraycopy(adr_type, dest_elem,
4978 src, src_offset, dest, dest_offset, length,
4979 false, false, slow_region);
4980
4981 return true;
4982 }
4983
4984 //-----------------------------generate_arraycopy----------------------
4985 // Generate an optimized call to arraycopy.
4986 // Caller must guard against non-arrays.
4987 // Caller must determine a common array basic-type for both arrays.
4988 // Caller must validate offsets against array bounds.
4989 // The slow_region has already collected guard failure paths
4990 // (such as out of bounds length or non-conformable array types).
4991 // The generated code has this shape, in general:
4992 //
4993 // if (length == 0) return // via zero_path
4994 // slowval = -1
4995 // if (types unknown) {
4996 // slowval = call generic copy loop
4997 // if (slowval == 0) return // via checked_path
4998 // } else if (indexes in bounds) {
4999 // if ((is object array) && !(array type check)) {
5000 // slowval = call checked copy loop
5001 // if (slowval == 0) return // via checked_path
5002 // } else {
5003 // call bulk copy loop
5004 // return // via fast_path
5005 // }
5006 // }
5007 // // adjust params for remaining work:
5008 // if (slowval != -1) {
5009 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
5010 // }
5011 // slow_region:
5012 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
5013 // return // via slow_call_path
5014 //
5015 // This routine is used from several intrinsics: System.arraycopy,
5016 // Object.clone (the array subcase), and Arrays.copyOf[Range].
5017 //
5018 void
5019 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
5020 BasicType basic_elem_type,
5021 Node* src, Node* src_offset,
5022 Node* dest, Node* dest_offset,
5023 Node* copy_length,
5024 bool disjoint_bases,
5025 bool length_never_negative,
5026 RegionNode* slow_region) {
5027
5028 if (slow_region == NULL) {
5029 slow_region = new(C) RegionNode(1);
5030 record_for_igvn(slow_region);
5031 }
5032
5033 Node* original_dest = dest;
5034 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5035 bool dest_uninitialized = false;
5036
5037 // See if this is the initialization of a newly-allocated array.
5038 // If so, we will take responsibility here for initializing it to zero.
5039 // (Note: Because tightly_coupled_allocation performs checks on the
5040 // out-edges of the dest, we need to avoid making derived pointers
5041 // from it until we have checked its uses.)
5042 if (ReduceBulkZeroing
5043 && !ZeroTLAB // pointless if already zeroed
5044 && basic_elem_type != T_CONFLICT // avoid corner case
5045 && !src->eqv_uncast(dest)
5046 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5047 != NULL)
5048 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5049 && alloc->maybe_set_complete(&_gvn)) {
5050 // "You break it, you buy it."
5051 InitializeNode* init = alloc->initialization();
5052 assert(init->is_complete(), "we just did this");
5053 init->set_complete_with_arraycopy();
5054 assert(dest->is_CheckCastPP(), "sanity");
5055 assert(dest->in(0)->in(0) == init, "dest pinned");
5056 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5057 // From this point on, every exit path is responsible for
5058 // initializing any non-copied parts of the object to zero.
5059 // Also, if this flag is set we make sure that arraycopy interacts properly
5060 // with G1, eliding pre-barriers. See CR 6627983.
5061 dest_uninitialized = true;
5062 } else {
5063 // No zeroing elimination here.
5064 alloc = NULL;
5065 //original_dest = dest;
5066 //dest_uninitialized = false;
5067 }
5068
5069 // Results are placed here:
5070 enum { fast_path = 1, // normal void-returning assembly stub
5071 checked_path = 2, // special assembly stub with cleanup
5072 slow_call_path = 3, // something went wrong; call the VM
5073 zero_path = 4, // bypass when length of copy is zero
5074 bcopy_path = 5, // copy primitive array by 64-bit blocks
5075 PATH_LIMIT = 6
5076 };
5077 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5078 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5079 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5080 record_for_igvn(result_region);
5081 _gvn.set_type_bottom(result_i_o);
5082 _gvn.set_type_bottom(result_memory);
5083 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5084
5085 // The slow_control path:
5086 Node* slow_control;
5087 Node* slow_i_o = i_o();
5088 Node* slow_mem = memory(adr_type);
5089 debug_only(slow_control = (Node*) badAddress);
5090
5091 // Checked control path:
5092 Node* checked_control = top();
5093 Node* checked_mem = NULL;
5094 Node* checked_i_o = NULL;
5095 Node* checked_value = NULL;
5096
5097 if (basic_elem_type == T_CONFLICT) {
5098 assert(!dest_uninitialized, "");
5099 Node* cv = generate_generic_arraycopy(adr_type,
5100 src, src_offset, dest, dest_offset,
5101 copy_length, dest_uninitialized);
5102 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5103 checked_control = control();
5104 checked_i_o = i_o();
5105 checked_mem = memory(adr_type);
5106 checked_value = cv;
5107 set_control(top()); // no fast path
5108 }
5109
5110 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5111 if (not_pos != NULL) {
5112 PreserveJVMState pjvms(this);
5113 set_control(not_pos);
5114
5115 // (6) length must not be negative.
5116 if (!length_never_negative) {
5117 generate_negative_guard(copy_length, slow_region);
5118 }
5119
5120 // copy_length is 0.
5121 if (!stopped() && dest_uninitialized) {
5122 Node* dest_length = alloc->in(AllocateNode::ALength);
5123 if (copy_length->eqv_uncast(dest_length)
5124 || _gvn.find_int_con(dest_length, 1) <= 0) {
5125 // There is no zeroing to do. No need for a secondary raw memory barrier.
5126 } else {
5127 // Clear the whole thing since there are no source elements to copy.
5128 generate_clear_array(adr_type, dest, basic_elem_type,
5129 intcon(0), NULL,
5130 alloc->in(AllocateNode::AllocSize));
5131 // Use a secondary InitializeNode as raw memory barrier.
5132 // Currently it is needed only on this path since other
5133 // paths have stub or runtime calls as raw memory barriers.
5134 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5135 Compile::AliasIdxRaw,
5136 top())->as_Initialize();
5137 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5138 }
5139 }
5140
5141 // Present the results of the fast call.
5142 result_region->init_req(zero_path, control());
5143 result_i_o ->init_req(zero_path, i_o());
5144 result_memory->init_req(zero_path, memory(adr_type));
5145 }
5146
5147 if (!stopped() && dest_uninitialized) {
5148 // We have to initialize the *uncopied* part of the array to zero.
5149 // The copy destination is the slice dest[off..off+len]. The other slices
5150 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5151 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5152 Node* dest_length = alloc->in(AllocateNode::ALength);
5153 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5154 copy_length));
5155
5156 // If there is a head section that needs zeroing, do it now.
5157 if (find_int_con(dest_offset, -1) != 0) {
5158 generate_clear_array(adr_type, dest, basic_elem_type,
5159 intcon(0), dest_offset,
5160 NULL);
5161 }
5162
5163 // Next, perform a dynamic check on the tail length.
5164 // It is often zero, and we can win big if we prove this.
5165 // There are two wins: Avoid generating the ClearArray
5166 // with its attendant messy index arithmetic, and upgrade
5167 // the copy to a more hardware-friendly word size of 64 bits.
5168 Node* tail_ctl = NULL;
5169 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5170 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5171 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5172 tail_ctl = generate_slow_guard(bol_lt, NULL);
5173 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5174 }
5175
5176 // At this point, let's assume there is no tail.
5177 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5178 // There is no tail. Try an upgrade to a 64-bit copy.
5179 bool didit = false;
5180 { PreserveJVMState pjvms(this);
5181 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5182 src, src_offset, dest, dest_offset,
5183 dest_size, dest_uninitialized);
5184 if (didit) {
5185 // Present the results of the block-copying fast call.
5186 result_region->init_req(bcopy_path, control());
5187 result_i_o ->init_req(bcopy_path, i_o());
5188 result_memory->init_req(bcopy_path, memory(adr_type));
5189 }
5190 }
5191 if (didit)
5192 set_control(top()); // no regular fast path
5193 }
5194
5195 // Clear the tail, if any.
5196 if (tail_ctl != NULL) {
5197 Node* notail_ctl = stopped() ? NULL : control();
5198 set_control(tail_ctl);
5199 if (notail_ctl == NULL) {
5200 generate_clear_array(adr_type, dest, basic_elem_type,
5201 dest_tail, NULL,
5202 dest_size);
5203 } else {
5204 // Make a local merge.
5205 Node* done_ctl = new(C) RegionNode(3);
5206 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5207 done_ctl->init_req(1, notail_ctl);
5208 done_mem->init_req(1, memory(adr_type));
5209 generate_clear_array(adr_type, dest, basic_elem_type,
5210 dest_tail, NULL,
5211 dest_size);
5212 done_ctl->init_req(2, control());
5213 done_mem->init_req(2, memory(adr_type));
5214 set_control( _gvn.transform(done_ctl));
5215 set_memory( _gvn.transform(done_mem), adr_type );
5216 }
5217 }
5218 }
5219
5220 BasicType copy_type = basic_elem_type;
5221 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5222 if (!stopped() && copy_type == T_OBJECT) {
5223 // If src and dest have compatible element types, we can copy bits.
5224 // Types S[] and D[] are compatible if D is a supertype of S.
5225 //
5226 // If they are not, we will use checked_oop_disjoint_arraycopy,
5227 // which performs a fast optimistic per-oop check, and backs off
5228 // further to JVM_ArrayCopy on the first per-oop check that fails.
5229 // (Actually, we don't move raw bits only; the GC requires card marks.)
5230
5231 // Get the Klass* for both src and dest
5232 Node* src_klass = load_object_klass(src);
5233 Node* dest_klass = load_object_klass(dest);
5234
5235 // Generate the subtype check.
5236 // This might fold up statically, or then again it might not.
5237 //
5238 // Non-static example: Copying List<String>.elements to a new String[].
5239 // The backing store for a List<String> is always an Object[],
5240 // but its elements are always type String, if the generic types
5241 // are correct at the source level.
5242 //
5243 // Test S[] against D[], not S against D, because (probably)
5244 // the secondary supertype cache is less busy for S[] than S.
5245 // This usually only matters when D is an interface.
5246 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5247 // Plug failing path into checked_oop_disjoint_arraycopy
5248 if (not_subtype_ctrl != top()) {
5249 PreserveJVMState pjvms(this);
5250 set_control(not_subtype_ctrl);
5251 // (At this point we can assume disjoint_bases, since types differ.)
5252 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5253 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5254 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5255 Node* dest_elem_klass = _gvn.transform(n1);
5256 Node* cv = generate_checkcast_arraycopy(adr_type,
5257 dest_elem_klass,
5258 src, src_offset, dest, dest_offset,
5259 ConvI2X(copy_length), dest_uninitialized);
5260 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5261 checked_control = control();
5262 checked_i_o = i_o();
5263 checked_mem = memory(adr_type);
5264 checked_value = cv;
5265 }
5266 // At this point we know we do not need type checks on oop stores.
5267
5268 // Let's see if we need card marks:
5269 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5270 // If we do not need card marks, copy using the jint or jlong stub.
5271 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5272 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5273 "sizes agree");
5274 }
5275 }
5276
5277 if (!stopped()) {
5278 // Generate the fast path, if possible.
5279 PreserveJVMState pjvms(this);
5280 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5281 src, src_offset, dest, dest_offset,
5282 ConvI2X(copy_length), dest_uninitialized);
5283
5284 // Present the results of the fast call.
5285 result_region->init_req(fast_path, control());
5286 result_i_o ->init_req(fast_path, i_o());
5287 result_memory->init_req(fast_path, memory(adr_type));
5288 }
5289
5290 // Here are all the slow paths up to this point, in one bundle:
5291 slow_control = top();
5292 if (slow_region != NULL)
5293 slow_control = _gvn.transform(slow_region);
5294 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5295
5296 set_control(checked_control);
5297 if (!stopped()) {
5298 // Clean up after the checked call.
5299 // The returned value is either 0 or -1^K,
5300 // where K = number of partially transferred array elements.
5301 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5302 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5303 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5304
5305 // If it is 0, we are done, so transfer to the end.
5306 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5307 result_region->init_req(checked_path, checks_done);
5308 result_i_o ->init_req(checked_path, checked_i_o);
5309 result_memory->init_req(checked_path, checked_mem);
5310
5311 // If it is not zero, merge into the slow call.
5312 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5313 RegionNode* slow_reg2 = new(C) RegionNode(3);
5314 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5315 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5316 record_for_igvn(slow_reg2);
5317 slow_reg2 ->init_req(1, slow_control);
5318 slow_i_o2 ->init_req(1, slow_i_o);
5319 slow_mem2 ->init_req(1, slow_mem);
5320 slow_reg2 ->init_req(2, control());
5321 slow_i_o2 ->init_req(2, checked_i_o);
5322 slow_mem2 ->init_req(2, checked_mem);
5323
5324 slow_control = _gvn.transform(slow_reg2);
5325 slow_i_o = _gvn.transform(slow_i_o2);
5326 slow_mem = _gvn.transform(slow_mem2);
5327
5328 if (alloc != NULL) {
5329 // We'll restart from the very beginning, after zeroing the whole thing.
5330 // This can cause double writes, but that's OK since dest is brand new.
5331 // So we ignore the low 31 bits of the value returned from the stub.
5332 } else {
5333 // We must continue the copy exactly where it failed, or else
5334 // another thread might see the wrong number of writes to dest.
5335 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5336 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5337 slow_offset->init_req(1, intcon(0));
5338 slow_offset->init_req(2, checked_offset);
5339 slow_offset = _gvn.transform(slow_offset);
5340
5341 // Adjust the arguments by the conditionally incoming offset.
5342 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5343 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5344 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5345
5346 // Tweak the node variables to adjust the code produced below:
5347 src_offset = src_off_plus;
5348 dest_offset = dest_off_plus;
5349 copy_length = length_minus;
5350 }
5351 }
5352
5353 set_control(slow_control);
5354 if (!stopped()) {
5355 // Generate the slow path, if needed.
5356 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5357
5358 set_memory(slow_mem, adr_type);
5359 set_i_o(slow_i_o);
5360
5361 if (dest_uninitialized) {
5362 generate_clear_array(adr_type, dest, basic_elem_type,
5363 intcon(0), NULL,
5364 alloc->in(AllocateNode::AllocSize));
5365 }
5366
5367 generate_slow_arraycopy(adr_type,
5368 src, src_offset, dest, dest_offset,
5369 copy_length, /*dest_uninitialized*/false);
5370
5371 result_region->init_req(slow_call_path, control());
5372 result_i_o ->init_req(slow_call_path, i_o());
5373 result_memory->init_req(slow_call_path, memory(adr_type));
5374 }
5375
5376 // Remove unused edges.
5377 for (uint i = 1; i < result_region->req(); i++) {
5378 if (result_region->in(i) == NULL)
5379 result_region->init_req(i, top());
5380 }
5381
5382 // Finished; return the combined state.
5383 set_control( _gvn.transform(result_region));
5384 set_i_o( _gvn.transform(result_i_o) );
5385 set_memory( _gvn.transform(result_memory), adr_type );
5386
5387 // The memory edges above are precise in order to model effects around
5388 // array copies accurately to allow value numbering of field loads around
5389 // arraycopy. Such field loads, both before and after, are common in Java
5390 // collections and similar classes involving header/array data structures.
5391 //
5392 // But with low number of register or when some registers are used or killed
5393 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5394 // The next memory barrier is added to avoid it. If the arraycopy can be
5395 // optimized away (which it can, sometimes) then we can manually remove
5396 // the membar also.
5397 //
5398 // Do not let reads from the cloned object float above the arraycopy.
5399 if (alloc != NULL) {
5400 // Do not let stores that initialize this object be reordered with
5401 // a subsequent store that would make this object accessible by
5402 // other threads.
5403 // Record what AllocateNode this StoreStore protects so that
5404 // escape analysis can go from the MemBarStoreStoreNode to the
5405 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5406 // based on the escape status of the AllocateNode.
5407 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5408 } else if (InsertMemBarAfterArraycopy)
5409 insert_mem_bar(Op_MemBarCPUOrder);
5410 }
5411
5412
5413 // Helper function which determines if an arraycopy immediately follows
5414 // an allocation, with no intervening tests or other escapes for the object.
5415 AllocateArrayNode*
5416 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5417 RegionNode* slow_region) {
5418 if (stopped()) return NULL; // no fast path
5419 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5420
5421 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5422 if (alloc == NULL) return NULL;
5423
5424 Node* rawmem = memory(Compile::AliasIdxRaw);
5425 // Is the allocation's memory state untouched?
5426 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5427 // Bail out if there have been raw-memory effects since the allocation.
5428 // (Example: There might have been a call or safepoint.)
5429 return NULL;
5430 }
5431 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5432 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5433 return NULL;
5434 }
5435
5436 // There must be no unexpected observers of this allocation.
5437 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5438 Node* obs = ptr->fast_out(i);
5439 if (obs != this->map()) {
5440 return NULL;
5441 }
5442 }
5443
5444 // This arraycopy must unconditionally follow the allocation of the ptr.
5445 Node* alloc_ctl = ptr->in(0);
5446 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5447
5448 Node* ctl = control();
5449 while (ctl != alloc_ctl) {
5450 // There may be guards which feed into the slow_region.
5451 // Any other control flow means that we might not get a chance
5452 // to finish initializing the allocated object.
5453 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5454 IfNode* iff = ctl->in(0)->as_If();
5455 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5456 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5457 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5458 ctl = iff->in(0); // This test feeds the known slow_region.
5459 continue;
5460 }
5461 // One more try: Various low-level checks bottom out in
5462 // uncommon traps. If the debug-info of the trap omits
5463 // any reference to the allocation, as we've already
5464 // observed, then there can be no objection to the trap.
5465 bool found_trap = false;
5466 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5467 Node* obs = not_ctl->fast_out(j);
5468 if (obs->in(0) == not_ctl && obs->is_Call() &&
5469 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5470 found_trap = true; break;
5471 }
5472 }
5473 if (found_trap) {
5474 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5475 continue;
5476 }
5477 }
5478 return NULL;
5479 }
5480
5481 // If we get this far, we have an allocation which immediately
5482 // precedes the arraycopy, and we can take over zeroing the new object.
5483 // The arraycopy will finish the initialization, and provide
5484 // a new control state to which we will anchor the destination pointer.
5485
5486 return alloc;
5487 }
5488
5489 // Helper for initialization of arrays, creating a ClearArray.
5490 // It writes zero bits in [start..end), within the body of an array object.
5491 // The memory effects are all chained onto the 'adr_type' alias category.
5492 //
5493 // Since the object is otherwise uninitialized, we are free
5494 // to put a little "slop" around the edges of the cleared area,
5495 // as long as it does not go back into the array's header,
5496 // or beyond the array end within the heap.
5497 //
5498 // The lower edge can be rounded down to the nearest jint and the
5499 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5500 //
5501 // Arguments:
5502 // adr_type memory slice where writes are generated
5503 // dest oop of the destination array
5504 // basic_elem_type element type of the destination
5505 // slice_idx array index of first element to store
5506 // slice_len number of elements to store (or NULL)
5507 // dest_size total size in bytes of the array object
5508 //
5509 // Exactly one of slice_len or dest_size must be non-NULL.
5510 // If dest_size is non-NULL, zeroing extends to the end of the object.
5511 // If slice_len is non-NULL, the slice_idx value must be a constant.
5512 void
5513 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5514 Node* dest,
5515 BasicType basic_elem_type,
5516 Node* slice_idx,
5517 Node* slice_len,
5518 Node* dest_size) {
5519 // one or the other but not both of slice_len and dest_size:
5520 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5521 if (slice_len == NULL) slice_len = top();
5522 if (dest_size == NULL) dest_size = top();
5523
5524 // operate on this memory slice:
5525 Node* mem = memory(adr_type); // memory slice to operate on
5526
5527 // scaling and rounding of indexes:
5528 int scale = exact_log2(type2aelembytes(basic_elem_type));
5529 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5530 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5531 int bump_bit = (-1 << scale) & BytesPerInt;
5532
5533 // determine constant starts and ends
5534 const intptr_t BIG_NEG = -128;
5535 assert(BIG_NEG + 2*abase < 0, "neg enough");
5536 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5537 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5538 if (slice_len_con == 0) {
5539 return; // nothing to do here
5540 }
5541 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5542 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5543 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5544 assert(end_con < 0, "not two cons");
5545 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5546 BytesPerLong);
5547 }
5548
5549 if (start_con >= 0 && end_con >= 0) {
5550 // Constant start and end. Simple.
5551 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5552 start_con, end_con, &_gvn);
5553 } else if (start_con >= 0 && dest_size != top()) {
5554 // Constant start, pre-rounded end after the tail of the array.
5555 Node* end = dest_size;
5556 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5557 start_con, end, &_gvn);
5558 } else if (start_con >= 0 && slice_len != top()) {
5559 // Constant start, non-constant end. End needs rounding up.
5560 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5561 intptr_t end_base = abase + (slice_idx_con << scale);
5562 int end_round = (-1 << scale) & (BytesPerLong - 1);
5563 Node* end = ConvI2X(slice_len);
5564 if (scale != 0)
5565 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5566 end_base += end_round;
5567 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5568 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5569 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5570 start_con, end, &_gvn);
5571 } else if (start_con < 0 && dest_size != top()) {
5572 // Non-constant start, pre-rounded end after the tail of the array.
5573 // This is almost certainly a "round-to-end" operation.
5574 Node* start = slice_idx;
5575 start = ConvI2X(start);
5576 if (scale != 0)
5577 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5578 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5579 if ((bump_bit | clear_low) != 0) {
5580 int to_clear = (bump_bit | clear_low);
5581 // Align up mod 8, then store a jint zero unconditionally
5582 // just before the mod-8 boundary.
5583 if (((abase + bump_bit) & ~to_clear) - bump_bit
5584 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5585 bump_bit = 0;
5586 assert((abase & to_clear) == 0, "array base must be long-aligned");
5587 } else {
5588 // Bump 'start' up to (or past) the next jint boundary:
5589 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5590 assert((abase & clear_low) == 0, "array base must be int-aligned");
5591 }
5592 // Round bumped 'start' down to jlong boundary in body of array.
5593 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5594 if (bump_bit != 0) {
5595 // Store a zero to the immediately preceding jint:
5596 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5597 Node* p1 = basic_plus_adr(dest, x1);
5598 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5599 mem = _gvn.transform(mem);
5600 }
5601 }
5602 Node* end = dest_size; // pre-rounded
5603 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5604 start, end, &_gvn);
5605 } else {
5606 // Non-constant start, unrounded non-constant end.
5607 // (Nobody zeroes a random midsection of an array using this routine.)
5608 ShouldNotReachHere(); // fix caller
5609 }
5610
5611 // Done.
5612 set_memory(mem, adr_type);
5613 }
5614
5615
5616 bool
5617 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5618 BasicType basic_elem_type,
5619 AllocateNode* alloc,
5620 Node* src, Node* src_offset,
5621 Node* dest, Node* dest_offset,
5622 Node* dest_size, bool dest_uninitialized) {
5623 // See if there is an advantage from block transfer.
5624 int scale = exact_log2(type2aelembytes(basic_elem_type));
5625 if (scale >= LogBytesPerLong)
5626 return false; // it is already a block transfer
5627
5628 // Look at the alignment of the starting offsets.
5629 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5630
5631 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5632 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5633 if (src_off_con < 0 || dest_off_con < 0)
5634 // At present, we can only understand constants.
5635 return false;
5636
5637 intptr_t src_off = abase + (src_off_con << scale);
5638 intptr_t dest_off = abase + (dest_off_con << scale);
5639
5640 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5641 // Non-aligned; too bad.
5642 // One more chance: Pick off an initial 32-bit word.
5643 // This is a common case, since abase can be odd mod 8.
5644 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5645 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5646 Node* sptr = basic_plus_adr(src, src_off);
5647 Node* dptr = basic_plus_adr(dest, dest_off);
5648 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5649 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5650 src_off += BytesPerInt;
5651 dest_off += BytesPerInt;
5652 } else {
5653 return false;
5654 }
5655 }
5656 assert(src_off % BytesPerLong == 0, "");
5657 assert(dest_off % BytesPerLong == 0, "");
5658
5659 // Do this copy by giant steps.
5660 Node* sptr = basic_plus_adr(src, src_off);
5661 Node* dptr = basic_plus_adr(dest, dest_off);
5662 Node* countx = dest_size;
5663 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5664 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5665
5666 bool disjoint_bases = true; // since alloc != NULL
5667 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5668 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5669
5670 return true;
5671 }
5672
5673
5674 // Helper function; generates code for the slow case.
5675 // We make a call to a runtime method which emulates the native method,
5676 // but without the native wrapper overhead.
5677 void
5678 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5679 Node* src, Node* src_offset,
5680 Node* dest, Node* dest_offset,
5681 Node* copy_length, bool dest_uninitialized) {
5682 assert(!dest_uninitialized, "Invariant");
5683 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5684 OptoRuntime::slow_arraycopy_Type(),
5685 OptoRuntime::slow_arraycopy_Java(),
5686 "slow_arraycopy", adr_type,
5687 src, src_offset, dest, dest_offset,
5688 copy_length);
5689
5690 // Handle exceptions thrown by this fellow:
5691 make_slow_call_ex(call, env()->Throwable_klass(), false);
5692 }
5693
5694 // Helper function; generates code for cases requiring runtime checks.
5695 Node*
5696 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5697 Node* dest_elem_klass,
5698 Node* src, Node* src_offset,
5699 Node* dest, Node* dest_offset,
5700 Node* copy_length, bool dest_uninitialized) {
5701 if (stopped()) return NULL;
5702
5703 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5704 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5705 return NULL;
5706 }
5707
5708 // Pick out the parameters required to perform a store-check
5709 // for the target array. This is an optimistic check. It will
5710 // look in each non-null element's class, at the desired klass's
5711 // super_check_offset, for the desired klass.
5712 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5713 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5714 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5715 Node* check_offset = ConvI2X(_gvn.transform(n3));
5716 Node* check_value = dest_elem_klass;
5717
5718 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5719 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5720
5721 // (We know the arrays are never conjoint, because their types differ.)
5722 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5723 OptoRuntime::checkcast_arraycopy_Type(),
5724 copyfunc_addr, "checkcast_arraycopy", adr_type,
5725 // five arguments, of which two are
5726 // intptr_t (jlong in LP64)
5727 src_start, dest_start,
5728 copy_length XTOP,
5729 check_offset XTOP,
5730 check_value);
5731
5732 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5733 }
5734
5735
5736 // Helper function; generates code for cases requiring runtime checks.
5737 Node*
5738 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5739 Node* src, Node* src_offset,
5740 Node* dest, Node* dest_offset,
5741 Node* copy_length, bool dest_uninitialized) {
5742 assert(!dest_uninitialized, "Invariant");
5743 if (stopped()) return NULL;
5744 address copyfunc_addr = StubRoutines::generic_arraycopy();
5745 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5746 return NULL;
5747 }
5748
5749 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5750 OptoRuntime::generic_arraycopy_Type(),
5751 copyfunc_addr, "generic_arraycopy", adr_type,
5752 src, src_offset, dest, dest_offset, copy_length);
5753
5754 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5755 }
5756
5757 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5758 void
5759 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5760 BasicType basic_elem_type,
5761 bool disjoint_bases,
5762 Node* src, Node* src_offset,
5763 Node* dest, Node* dest_offset,
5764 Node* copy_length, bool dest_uninitialized) {
5765 if (stopped()) return; // nothing to do
5766
5767 Node* src_start = src;
5768 Node* dest_start = dest;
5769 if (src_offset != NULL || dest_offset != NULL) {
5770 assert(src_offset != NULL && dest_offset != NULL, "");
5771 src_start = array_element_address(src, src_offset, basic_elem_type);
5772 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5773 }
5774
5775 // Figure out which arraycopy runtime method to call.
5776 const char* copyfunc_name = "arraycopy";
5777 address copyfunc_addr =
5778 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5779 disjoint_bases, copyfunc_name, dest_uninitialized);
5780
5781 // Call it. Note that the count_ix value is not scaled to a byte-size.
5782 make_runtime_call(RC_LEAF|RC_NO_FP,
5783 OptoRuntime::fast_arraycopy_Type(),
5784 copyfunc_addr, copyfunc_name, adr_type,
5785 src_start, dest_start, copy_length XTOP);
5786 }
5787
5788 //-------------inline_encodeISOArray-----------------------------------
5789 // encode char[] to byte[] in ISO_8859_1
5790 bool LibraryCallKit::inline_encodeISOArray() {
5791 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5792 // no receiver since it is static method
5793 Node *src = argument(0);
5794 Node *src_offset = argument(1);
5795 Node *dst = argument(2);
5796 Node *dst_offset = argument(3);
5797 Node *length = argument(4);
5798
5799 const Type* src_type = src->Value(&_gvn);
5800 const Type* dst_type = dst->Value(&_gvn);
5801 const TypeAryPtr* top_src = src_type->isa_aryptr();
5802 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5803 if (top_src == NULL || top_src->klass() == NULL ||
5804 top_dest == NULL || top_dest->klass() == NULL) {
5805 // failed array check
5806 return false;
5807 }
5808
5809 // Figure out the size and type of the elements we will be copying.
5810 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5811 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5812 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5813 return false;
5814 }
5815 Node* src_start = array_element_address(src, src_offset, src_elem);
5816 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5817 // 'src_start' points to src array + scaled offset
5818 // 'dst_start' points to dst array + scaled offset
5819
5820 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5821 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5822 enc = _gvn.transform(enc);
5823 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5824 set_memory(res_mem, mtype);
5825 set_result(enc);
5826 return true;
5827 }
5828
5829 //-------------inline_multiplyToLen-----------------------------------
5830 bool LibraryCallKit::inline_multiplyToLen() {
5831 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5832
5833 address stubAddr = StubRoutines::multiplyToLen();
5834 if (stubAddr == NULL) {
5835 return false; // Intrinsic's stub is not implemented on this platform
5836 }
5837 const char* stubName = "multiplyToLen";
5838
5839 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5840
5841 // no receiver because it is a static method
5842 Node* x = argument(0);
5843 Node* xlen = argument(1);
5844 Node* y = argument(2);
5845 Node* ylen = argument(3);
5846 Node* z = argument(4);
5847
5848 const Type* x_type = x->Value(&_gvn);
5849 const Type* y_type = y->Value(&_gvn);
5850 const TypeAryPtr* top_x = x_type->isa_aryptr();
5851 const TypeAryPtr* top_y = y_type->isa_aryptr();
5852 if (top_x == NULL || top_x->klass() == NULL ||
5853 top_y == NULL || top_y->klass() == NULL) {
5854 // failed array check
5855 return false;
5856 }
5857
5858 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5859 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5860 if (x_elem != T_INT || y_elem != T_INT) {
5861 return false;
5862 }
5863
5864 // Set the original stack and the reexecute bit for the interpreter to reexecute
5865 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5866 // on the return from z array allocation in runtime.
5867 { PreserveReexecuteState preexecs(this);
5868 jvms()->set_should_reexecute(true);
5869
5870 Node* x_start = array_element_address(x, intcon(0), x_elem);
5871 Node* y_start = array_element_address(y, intcon(0), y_elem);
5872 // 'x_start' points to x array + scaled xlen
5873 // 'y_start' points to y array + scaled ylen
5874
5875 // Allocate the result array
5876 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5877 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5878 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5879
5880 IdealKit ideal(this);
5881
5882 #define __ ideal.
5883 Node* one = __ ConI(1);
5884 Node* zero = __ ConI(0);
5885 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5886 __ set(need_alloc, zero);
5887 __ set(z_alloc, z);
5888 __ if_then(z, BoolTest::eq, null()); {
5889 __ increment (need_alloc, one);
5890 } __ else_(); {
5891 // Update graphKit memory and control from IdealKit.
5892 sync_kit(ideal);
5893 Node* zlen_arg = load_array_length(z);
5894 // Update IdealKit memory and control from graphKit.
5895 __ sync_kit(this);
5896 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5897 __ increment (need_alloc, one);
5898 } __ end_if();
5899 } __ end_if();
5900
5901 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5902 // Update graphKit memory and control from IdealKit.
5903 sync_kit(ideal);
5904 Node * narr = new_array(klass_node, zlen, 1);
5905 // Update IdealKit memory and control from graphKit.
5906 __ sync_kit(this);
5907 __ set(z_alloc, narr);
5908 } __ end_if();
5909
5910 sync_kit(ideal);
5911 z = __ value(z_alloc);
5912 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5913 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5914 // Final sync IdealKit and GraphKit.
5915 final_sync(ideal);
5916 #undef __
5917
5918 Node* z_start = array_element_address(z, intcon(0), T_INT);
5919
5920 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5921 OptoRuntime::multiplyToLen_Type(),
5922 stubAddr, stubName, TypePtr::BOTTOM,
5923 x_start, xlen, y_start, ylen, z_start, zlen);
5924 } // original reexecute is set back here
5925
5926 C->set_has_split_ifs(true); // Has chance for split-if optimization
5927 set_result(z);
5928 return true;
5929 }
5930
5931 //-------------inline_squareToLen------------------------------------
5932 bool LibraryCallKit::inline_squareToLen() {
5933 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5934
5935 address stubAddr = StubRoutines::squareToLen();
5936 if (stubAddr == NULL) {
5937 return false; // Intrinsic's stub is not implemented on this platform
5938 }
5939 const char* stubName = "squareToLen";
5940
5941 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5942
5943 Node* x = argument(0);
5944 Node* len = argument(1);
5945 Node* z = argument(2);
5946 Node* zlen = argument(3);
5947
5948 const Type* x_type = x->Value(&_gvn);
5949 const Type* z_type = z->Value(&_gvn);
5950 const TypeAryPtr* top_x = x_type->isa_aryptr();
5951 const TypeAryPtr* top_z = z_type->isa_aryptr();
5952 if (top_x == NULL || top_x->klass() == NULL ||
5953 top_z == NULL || top_z->klass() == NULL) {
5954 // failed array check
5955 return false;
5956 }
5957
5958 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5959 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5960 if (x_elem != T_INT || z_elem != T_INT) {
5961 return false;
5962 }
5963
5964
5965 Node* x_start = array_element_address(x, intcon(0), x_elem);
5966 Node* z_start = array_element_address(z, intcon(0), z_elem);
5967
5968 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5969 OptoRuntime::squareToLen_Type(),
5970 stubAddr, stubName, TypePtr::BOTTOM,
5971 x_start, len, z_start, zlen);
5972
5973 set_result(z);
5974 return true;
5975 }
5976
5977 //-------------inline_mulAdd------------------------------------------
5978 bool LibraryCallKit::inline_mulAdd() {
5979 assert(UseMulAddIntrinsic, "not implementated on this platform");
5980
5981 address stubAddr = StubRoutines::mulAdd();
5982 if (stubAddr == NULL) {
5983 return false; // Intrinsic's stub is not implemented on this platform
5984 }
5985 const char* stubName = "mulAdd";
5986
5987 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5988
5989 Node* out = argument(0);
5990 Node* in = argument(1);
5991 Node* offset = argument(2);
5992 Node* len = argument(3);
5993 Node* k = argument(4);
5994
5995 const Type* out_type = out->Value(&_gvn);
5996 const Type* in_type = in->Value(&_gvn);
5997 const TypeAryPtr* top_out = out_type->isa_aryptr();
5998 const TypeAryPtr* top_in = in_type->isa_aryptr();
5999 if (top_out == NULL || top_out->klass() == NULL ||
6000 top_in == NULL || top_in->klass() == NULL) {
6001 // failed array check
6002 return false;
6003 }
6004
6005 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6006 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6007 if (out_elem != T_INT || in_elem != T_INT) {
6008 return false;
6009 }
6010
6011 Node* outlen = load_array_length(out);
6012 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
6013 Node* out_start = array_element_address(out, intcon(0), out_elem);
6014 Node* in_start = array_element_address(in, intcon(0), in_elem);
6015
6016 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6017 OptoRuntime::mulAdd_Type(),
6018 stubAddr, stubName, TypePtr::BOTTOM,
6019 out_start,in_start, new_offset, len, k);
6020 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6021 set_result(result);
6022 return true;
6023 }
6024
6025 //-------------inline_montgomeryMultiply-----------------------------------
6026 bool LibraryCallKit::inline_montgomeryMultiply() {
6027 address stubAddr = StubRoutines::montgomeryMultiply();
6028 if (stubAddr == NULL) {
6029 return false; // Intrinsic's stub is not implemented on this platform
6030 }
6031
6032 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6033 const char* stubName = "montgomery_multiply";
6034
6035 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6036
6037 Node* a = argument(0);
6038 Node* b = argument(1);
6039 Node* n = argument(2);
6040 Node* len = argument(3);
6041 Node* inv = argument(4);
6042 Node* m = argument(6);
6043
6044 const Type* a_type = a->Value(&_gvn);
6045 const TypeAryPtr* top_a = a_type->isa_aryptr();
6046 const Type* b_type = b->Value(&_gvn);
6047 const TypeAryPtr* top_b = b_type->isa_aryptr();
6048 const Type* n_type = a->Value(&_gvn);
6049 const TypeAryPtr* top_n = n_type->isa_aryptr();
6050 const Type* m_type = a->Value(&_gvn);
6051 const TypeAryPtr* top_m = m_type->isa_aryptr();
6052 if (top_a == NULL || top_a->klass() == NULL ||
6053 top_b == NULL || top_b->klass() == NULL ||
6054 top_n == NULL || top_n->klass() == NULL ||
6055 top_m == NULL || top_m->klass() == NULL) {
6056 // failed array check
6057 return false;
6058 }
6059
6060 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6061 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6062 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6063 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6064 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6065 return false;
6066 }
6067
6068 // Make the call
6069 {
6070 Node* a_start = array_element_address(a, intcon(0), a_elem);
6071 Node* b_start = array_element_address(b, intcon(0), b_elem);
6072 Node* n_start = array_element_address(n, intcon(0), n_elem);
6073 Node* m_start = array_element_address(m, intcon(0), m_elem);
6074
6075 Node* call = NULL;
6076 if (CCallingConventionRequiresIntsAsLongs) {
6077 Node* len_I2L = ConvI2L(len);
6078 call = make_runtime_call(RC_LEAF,
6079 OptoRuntime::montgomeryMultiply_Type(),
6080 stubAddr, stubName, TypePtr::BOTTOM,
6081 a_start, b_start, n_start, len_I2L XTOP, inv,
6082 top(), m_start);
6083 } else {
6084 call = make_runtime_call(RC_LEAF,
6085 OptoRuntime::montgomeryMultiply_Type(),
6086 stubAddr, stubName, TypePtr::BOTTOM,
6087 a_start, b_start, n_start, len, inv, top(),
6088 m_start);
6089 }
6090 set_result(m);
6091 }
6092
6093 return true;
6094 }
6095
6096 bool LibraryCallKit::inline_montgomerySquare() {
6097 address stubAddr = StubRoutines::montgomerySquare();
6098 if (stubAddr == NULL) {
6099 return false; // Intrinsic's stub is not implemented on this platform
6100 }
6101
6102 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6103 const char* stubName = "montgomery_square";
6104
6105 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6106
6107 Node* a = argument(0);
6108 Node* n = argument(1);
6109 Node* len = argument(2);
6110 Node* inv = argument(3);
6111 Node* m = argument(5);
6112
6113 const Type* a_type = a->Value(&_gvn);
6114 const TypeAryPtr* top_a = a_type->isa_aryptr();
6115 const Type* n_type = a->Value(&_gvn);
6116 const TypeAryPtr* top_n = n_type->isa_aryptr();
6117 const Type* m_type = a->Value(&_gvn);
6118 const TypeAryPtr* top_m = m_type->isa_aryptr();
6119 if (top_a == NULL || top_a->klass() == NULL ||
6120 top_n == NULL || top_n->klass() == NULL ||
6121 top_m == NULL || top_m->klass() == NULL) {
6122 // failed array check
6123 return false;
6124 }
6125
6126 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6127 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6128 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6129 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6130 return false;
6131 }
6132
6133 // Make the call
6134 {
6135 Node* a_start = array_element_address(a, intcon(0), a_elem);
6136 Node* n_start = array_element_address(n, intcon(0), n_elem);
6137 Node* m_start = array_element_address(m, intcon(0), m_elem);
6138
6139 Node* call = NULL;
6140 if (CCallingConventionRequiresIntsAsLongs) {
6141 Node* len_I2L = ConvI2L(len);
6142 call = make_runtime_call(RC_LEAF,
6143 OptoRuntime::montgomerySquare_Type(),
6144 stubAddr, stubName, TypePtr::BOTTOM,
6145 a_start, n_start, len_I2L XTOP, inv, top(),
6146 m_start);
6147 } else {
6148 call = make_runtime_call(RC_LEAF,
6149 OptoRuntime::montgomerySquare_Type(),
6150 stubAddr, stubName, TypePtr::BOTTOM,
6151 a_start, n_start, len, inv, top(),
6152 m_start);
6153 }
6154
6155 set_result(m);
6156 }
6157
6158 return true;
6159 }
6160
6161
6162 /**
6163 * Calculate CRC32 for byte.
6164 * int java.util.zip.CRC32.update(int crc, int b)
6165 */
6166 bool LibraryCallKit::inline_updateCRC32() {
6167 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6168 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6169 // no receiver since it is static method
6170 Node* crc = argument(0); // type: int
6171 Node* b = argument(1); // type: int
6172
6173 /*
6174 * int c = ~ crc;
6175 * b = timesXtoThe32[(b ^ c) & 0xFF];
6176 * b = b ^ (c >>> 8);
6177 * crc = ~b;
6178 */
6179
6180 Node* M1 = intcon(-1);
6181 crc = _gvn.transform(new (C) XorINode(crc, M1));
6182 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6183 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6184
6185 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6186 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6187 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6188 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6189
6190 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6191 result = _gvn.transform(new (C) XorINode(crc, result));
6192 result = _gvn.transform(new (C) XorINode(result, M1));
6193 set_result(result);
6194 return true;
6195 }
6196
6197 /**
6198 * Calculate CRC32 for byte[] array.
6199 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6200 */
6201 bool LibraryCallKit::inline_updateBytesCRC32() {
6202 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6203 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6204 // no receiver since it is static method
6205 Node* crc = argument(0); // type: int
6206 Node* src = argument(1); // type: oop
6207 Node* offset = argument(2); // type: int
6208 Node* length = argument(3); // type: int
6209
6210 const Type* src_type = src->Value(&_gvn);
6211 const TypeAryPtr* top_src = src_type->isa_aryptr();
6212 if (top_src == NULL || top_src->klass() == NULL) {
6213 // failed array check
6214 return false;
6215 }
6216
6217 // Figure out the size and type of the elements we will be copying.
6218 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6219 if (src_elem != T_BYTE) {
6220 return false;
6221 }
6222
6223 // 'src_start' points to src array + scaled offset
6224 Node* src_start = array_element_address(src, offset, src_elem);
6225
6226 // We assume that range check is done by caller.
6227 // TODO: generate range check (offset+length < src.length) in debug VM.
6228
6229 // Call the stub.
6230 address stubAddr = StubRoutines::updateBytesCRC32();
6231 const char *stubName = "updateBytesCRC32";
6232 Node* call;
6233 if (CCallingConventionRequiresIntsAsLongs) {
6234 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6235 stubAddr, stubName, TypePtr::BOTTOM,
6236 crc XTOP, src_start, length XTOP);
6237 } else {
6238 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6239 stubAddr, stubName, TypePtr::BOTTOM,
6240 crc, src_start, length);
6241 }
6242 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6243 set_result(result);
6244 return true;
6245 }
6246
6247 /**
6248 * Calculate CRC32 for ByteBuffer.
6249 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6250 */
6251 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6252 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6253 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6254 // no receiver since it is static method
6255 Node* crc = argument(0); // type: int
6256 Node* src = argument(1); // type: long
6257 Node* offset = argument(3); // type: int
6258 Node* length = argument(4); // type: int
6259
6260 src = ConvL2X(src); // adjust Java long to machine word
6261 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6262 offset = ConvI2X(offset);
6263
6264 // 'src_start' points to src array + scaled offset
6265 Node* src_start = basic_plus_adr(top(), base, offset);
6266
6267 // Call the stub.
6268 address stubAddr = StubRoutines::updateBytesCRC32();
6269 const char *stubName = "updateBytesCRC32";
6270 Node* call;
6271 if (CCallingConventionRequiresIntsAsLongs) {
6272 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6273 stubAddr, stubName, TypePtr::BOTTOM,
6274 crc XTOP, src_start, length XTOP);
6275 } else {
6276 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6277 stubAddr, stubName, TypePtr::BOTTOM,
6278 crc, src_start, length);
6279 }
6280 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6281 set_result(result);
6282 return true;
6283 }
6284
6285 //----------------------------inline_reference_get----------------------------
6286 // public T java.lang.ref.Reference.get();
6287 bool LibraryCallKit::inline_reference_get() {
6288 const int referent_offset = java_lang_ref_Reference::referent_offset;
6289 guarantee(referent_offset > 0, "should have already been set");
6290
6291 // Get the argument:
6292 Node* reference_obj = null_check_receiver();
6293 if (stopped()) return true;
6294
6295 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6296
6297 ciInstanceKlass* klass = env()->Object_klass();
6298 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6299
6300 Node* no_ctrl = NULL;
6301 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6302
6303 // Use the pre-barrier to record the value in the referent field
6304 pre_barrier(false /* do_load */,
6305 control(),
6306 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6307 result /* pre_val */,
6308 T_OBJECT);
6309
6310 // Add memory barrier to prevent commoning reads from this field
6311 // across safepoint since GC can change its value.
6312 insert_mem_bar(Op_MemBarCPUOrder);
6313
6314 set_result(result);
6315 return true;
6316 }
6317
6318
6319 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6320 bool is_exact=true, bool is_static=false) {
6321
6322 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6323 assert(tinst != NULL, "obj is null");
6324 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6325 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6326
6327 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6328 ciSymbol::make(fieldTypeString),
6329 is_static);
6330 if (field == NULL) return (Node *) NULL;
6331 assert (field != NULL, "undefined field");
6332
6333 // Next code copied from Parse::do_get_xxx():
6334
6335 // Compute address and memory type.
6336 int offset = field->offset_in_bytes();
6337 bool is_vol = field->is_volatile();
6338 ciType* field_klass = field->type();
6339 assert(field_klass->is_loaded(), "should be loaded");
6340 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6341 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6342 BasicType bt = field->layout_type();
6343
6344 // Build the resultant type of the load
6345 const Type *type;
6346 if (bt == T_OBJECT) {
6347 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6348 } else {
6349 type = Type::get_const_basic_type(bt);
6350 }
6351
6352 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6353 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6354 }
6355 // Build the load.
6356 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6357 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6358 // If reference is volatile, prevent following memory ops from
6359 // floating up past the volatile read. Also prevents commoning
6360 // another volatile read.
6361 if (is_vol) {
6362 // Memory barrier includes bogus read of value to force load BEFORE membar
6363 insert_mem_bar(Op_MemBarAcquire, loadedField);
6364 }
6365 return loadedField;
6366 }
6367
6368
6369 //------------------------------inline_aescrypt_Block-----------------------
6370 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6371 address stubAddr = NULL;
6372 const char *stubName;
6373 assert(UseAES, "need AES instruction support");
6374
6375 switch(id) {
6376 case vmIntrinsics::_aescrypt_encryptBlock:
6377 stubAddr = StubRoutines::aescrypt_encryptBlock();
6378 stubName = "aescrypt_encryptBlock";
6379 break;
6380 case vmIntrinsics::_aescrypt_decryptBlock:
6381 stubAddr = StubRoutines::aescrypt_decryptBlock();
6382 stubName = "aescrypt_decryptBlock";
6383 break;
6384 }
6385 if (stubAddr == NULL) return false;
6386
6387 Node* aescrypt_object = argument(0);
6388 Node* src = argument(1);
6389 Node* src_offset = argument(2);
6390 Node* dest = argument(3);
6391 Node* dest_offset = argument(4);
6392
6393 // (1) src and dest are arrays.
6394 const Type* src_type = src->Value(&_gvn);
6395 const Type* dest_type = dest->Value(&_gvn);
6396 const TypeAryPtr* top_src = src_type->isa_aryptr();
6397 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6398 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6399
6400 // for the quick and dirty code we will skip all the checks.
6401 // we are just trying to get the call to be generated.
6402 Node* src_start = src;
6403 Node* dest_start = dest;
6404 if (src_offset != NULL || dest_offset != NULL) {
6405 assert(src_offset != NULL && dest_offset != NULL, "");
6406 src_start = array_element_address(src, src_offset, T_BYTE);
6407 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6408 }
6409
6410 // now need to get the start of its expanded key array
6411 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6412 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6413 if (k_start == NULL) return false;
6414
6415 if (Matcher::pass_original_key_for_aes()) {
6416 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6417 // compatibility issues between Java key expansion and SPARC crypto instructions
6418 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6419 if (original_k_start == NULL) return false;
6420
6421 // Call the stub.
6422 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6423 stubAddr, stubName, TypePtr::BOTTOM,
6424 src_start, dest_start, k_start, original_k_start);
6425 } else {
6426 // Call the stub.
6427 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6428 stubAddr, stubName, TypePtr::BOTTOM,
6429 src_start, dest_start, k_start);
6430 }
6431
6432 return true;
6433 }
6434
6435 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6436 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6437 address stubAddr = NULL;
6438 const char *stubName = NULL;
6439
6440 assert(UseAES, "need AES instruction support");
6441
6442 switch(id) {
6443 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6444 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6445 stubName = "cipherBlockChaining_encryptAESCrypt";
6446 break;
6447 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6448 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6449 stubName = "cipherBlockChaining_decryptAESCrypt";
6450 break;
6451 }
6452 if (stubAddr == NULL) return false;
6453
6454 Node* cipherBlockChaining_object = argument(0);
6455 Node* src = argument(1);
6456 Node* src_offset = argument(2);
6457 Node* len = argument(3);
6458 Node* dest = argument(4);
6459 Node* dest_offset = argument(5);
6460
6461 // (1) src and dest are arrays.
6462 const Type* src_type = src->Value(&_gvn);
6463 const Type* dest_type = dest->Value(&_gvn);
6464 const TypeAryPtr* top_src = src_type->isa_aryptr();
6465 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6466 assert (top_src != NULL && top_src->klass() != NULL
6467 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6468
6469 // checks are the responsibility of the caller
6470 Node* src_start = src;
6471 Node* dest_start = dest;
6472 if (src_offset != NULL || dest_offset != NULL) {
6473 assert(src_offset != NULL && dest_offset != NULL, "");
6474 src_start = array_element_address(src, src_offset, T_BYTE);
6475 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6476 }
6477
6478 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6479 // (because of the predicated logic executed earlier).
6480 // so we cast it here safely.
6481 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6482
6483 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6484 if (embeddedCipherObj == NULL) return false;
6485
6486 // cast it to what we know it will be at runtime
6487 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6488 assert(tinst != NULL, "CBC obj is null");
6489 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6490 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6491 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6492
6493 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6494 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6495 const TypeOopPtr* xtype = aklass->as_instance_type();
6496 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6497 aescrypt_object = _gvn.transform(aescrypt_object);
6498
6499 // we need to get the start of the aescrypt_object's expanded key array
6500 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6501 if (k_start == NULL) return false;
6502
6503 // similarly, get the start address of the r vector
6504 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6505 if (objRvec == NULL) return false;
6506 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6507
6508 Node* cbcCrypt;
6509 if (Matcher::pass_original_key_for_aes()) {
6510 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6511 // compatibility issues between Java key expansion and SPARC crypto instructions
6512 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6513 if (original_k_start == NULL) return false;
6514
6515 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6516 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6517 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6518 stubAddr, stubName, TypePtr::BOTTOM,
6519 src_start, dest_start, k_start, r_start, len, original_k_start);
6520 } else {
6521 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6522 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6523 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6524 stubAddr, stubName, TypePtr::BOTTOM,
6525 src_start, dest_start, k_start, r_start, len);
6526 }
6527
6528 // return cipher length (int)
6529 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6530 set_result(retvalue);
6531 return true;
6532 }
6533
6534 //------------------------------get_key_start_from_aescrypt_object-----------------------
6535 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6536 #ifdef PPC64
6537 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6538 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6539 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6540 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6541 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6542 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6543 if (objSessionK == NULL) {
6544 return (Node *) NULL;
6545 }
6546 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6547 #else
6548 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6549 #endif // PPC64
6550 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6551 if (objAESCryptKey == NULL) return (Node *) NULL;
6552
6553 // now have the array, need to get the start address of the K array
6554 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6555 return k_start;
6556 }
6557
6558 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6559 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6560 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6561 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6562 if (objAESCryptKey == NULL) return (Node *) NULL;
6563
6564 // now have the array, need to get the start address of the lastKey array
6565 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6566 return original_k_start;
6567 }
6568
6569 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6570 // Return node representing slow path of predicate check.
6571 // the pseudo code we want to emulate with this predicate is:
6572 // for encryption:
6573 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6574 // for decryption:
6575 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6576 // note cipher==plain is more conservative than the original java code but that's OK
6577 //
6578 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6579 // The receiver was checked for NULL already.
6580 Node* objCBC = argument(0);
6581
6582 // Load embeddedCipher field of CipherBlockChaining object.
6583 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6584
6585 // get AESCrypt klass for instanceOf check
6586 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6587 // will have same classloader as CipherBlockChaining object
6588 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6589 assert(tinst != NULL, "CBCobj is null");
6590 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6591
6592 // we want to do an instanceof comparison against the AESCrypt class
6593 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6594 if (!klass_AESCrypt->is_loaded()) {
6595 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6596 Node* ctrl = control();
6597 set_control(top()); // no regular fast path
6598 return ctrl;
6599 }
6600 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6601
6602 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6603 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6604 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6605
6606 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6607
6608 // for encryption, we are done
6609 if (!decrypting)
6610 return instof_false; // even if it is NULL
6611
6612 // for decryption, we need to add a further check to avoid
6613 // taking the intrinsic path when cipher and plain are the same
6614 // see the original java code for why.
6615 RegionNode* region = new(C) RegionNode(3);
6616 region->init_req(1, instof_false);
6617 Node* src = argument(1);
6618 Node* dest = argument(4);
6619 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6620 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6621 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6622 region->init_req(2, src_dest_conjoint);
6623
6624 record_for_igvn(region);
6625 return _gvn.transform(region);
6626 }
6627
6628 //------------------------------inline_sha_implCompress-----------------------
6629 //
6630 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6631 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6632 //
6633 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6634 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6635 //
6636 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6637 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6638 //
6639 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6640 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6641
6642 Node* sha_obj = argument(0);
6643 Node* src = argument(1); // type oop
6644 Node* ofs = argument(2); // type int
6645
6646 const Type* src_type = src->Value(&_gvn);
6647 const TypeAryPtr* top_src = src_type->isa_aryptr();
6648 if (top_src == NULL || top_src->klass() == NULL) {
6649 // failed array check
6650 return false;
6651 }
6652 // Figure out the size and type of the elements we will be copying.
6653 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6654 if (src_elem != T_BYTE) {
6655 return false;
6656 }
6657 // 'src_start' points to src array + offset
6658 Node* src_start = array_element_address(src, ofs, src_elem);
6659 Node* state = NULL;
6660 address stubAddr;
6661 const char *stubName;
6662
6663 switch(id) {
6664 case vmIntrinsics::_sha_implCompress:
6665 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6666 state = get_state_from_sha_object(sha_obj);
6667 stubAddr = StubRoutines::sha1_implCompress();
6668 stubName = "sha1_implCompress";
6669 break;
6670 case vmIntrinsics::_sha2_implCompress:
6671 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6672 state = get_state_from_sha_object(sha_obj);
6673 stubAddr = StubRoutines::sha256_implCompress();
6674 stubName = "sha256_implCompress";
6675 break;
6676 case vmIntrinsics::_sha5_implCompress:
6677 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6678 state = get_state_from_sha5_object(sha_obj);
6679 stubAddr = StubRoutines::sha512_implCompress();
6680 stubName = "sha512_implCompress";
6681 break;
6682 default:
6683 fatal_unexpected_iid(id);
6684 return false;
6685 }
6686 if (state == NULL) return false;
6687
6688 // Call the stub.
6689 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6690 stubAddr, stubName, TypePtr::BOTTOM,
6691 src_start, state);
6692
6693 return true;
6694 }
6695
6696 //------------------------------inline_digestBase_implCompressMB-----------------------
6697 //
6698 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6699 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6700 //
6701 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6702 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6703 "need SHA1/SHA256/SHA512 instruction support");
6704 assert((uint)predicate < 3, "sanity");
6705 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6706
6707 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6708 Node* src = argument(1); // byte[] array
6709 Node* ofs = argument(2); // type int
6710 Node* limit = argument(3); // type int
6711
6712 const Type* src_type = src->Value(&_gvn);
6713 const TypeAryPtr* top_src = src_type->isa_aryptr();
6714 if (top_src == NULL || top_src->klass() == NULL) {
6715 // failed array check
6716 return false;
6717 }
6718 // Figure out the size and type of the elements we will be copying.
6719 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6720 if (src_elem != T_BYTE) {
6721 return false;
6722 }
6723 // 'src_start' points to src array + offset
6724 Node* src_start = array_element_address(src, ofs, src_elem);
6725
6726 const char* klass_SHA_name = NULL;
6727 const char* stub_name = NULL;
6728 address stub_addr = NULL;
6729 bool long_state = false;
6730
6731 switch (predicate) {
6732 case 0:
6733 if (UseSHA1Intrinsics) {
6734 klass_SHA_name = "sun/security/provider/SHA";
6735 stub_name = "sha1_implCompressMB";
6736 stub_addr = StubRoutines::sha1_implCompressMB();
6737 }
6738 break;
6739 case 1:
6740 if (UseSHA256Intrinsics) {
6741 klass_SHA_name = "sun/security/provider/SHA2";
6742 stub_name = "sha256_implCompressMB";
6743 stub_addr = StubRoutines::sha256_implCompressMB();
6744 }
6745 break;
6746 case 2:
6747 if (UseSHA512Intrinsics) {
6748 klass_SHA_name = "sun/security/provider/SHA5";
6749 stub_name = "sha512_implCompressMB";
6750 stub_addr = StubRoutines::sha512_implCompressMB();
6751 long_state = true;
6752 }
6753 break;
6754 default:
6755 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6756 }
6757 if (klass_SHA_name != NULL) {
6758 // get DigestBase klass to lookup for SHA klass
6759 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6760 assert(tinst != NULL, "digestBase_obj is not instance???");
6761 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6762
6763 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6764 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6765 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6766 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6767 }
6768 return false;
6769 }
6770 //------------------------------inline_sha_implCompressMB-----------------------
6771 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6772 bool long_state, address stubAddr, const char *stubName,
6773 Node* src_start, Node* ofs, Node* limit) {
6774 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6775 const TypeOopPtr* xtype = aklass->as_instance_type();
6776 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6777 sha_obj = _gvn.transform(sha_obj);
6778
6779 Node* state;
6780 if (long_state) {
6781 state = get_state_from_sha5_object(sha_obj);
6782 } else {
6783 state = get_state_from_sha_object(sha_obj);
6784 }
6785 if (state == NULL) return false;
6786
6787 // Call the stub.
6788 Node *call;
6789 if (CCallingConventionRequiresIntsAsLongs) {
6790 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6791 OptoRuntime::digestBase_implCompressMB_Type(),
6792 stubAddr, stubName, TypePtr::BOTTOM,
6793 src_start, state, ofs XTOP, limit XTOP);
6794 } else {
6795 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6796 OptoRuntime::digestBase_implCompressMB_Type(),
6797 stubAddr, stubName, TypePtr::BOTTOM,
6798 src_start, state, ofs, limit);
6799 }
6800 // return ofs (int)
6801 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6802 set_result(result);
6803
6804 return true;
6805 }
6806
6807 //------------------------------get_state_from_sha_object-----------------------
6808 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6809 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6810 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6811 if (sha_state == NULL) return (Node *) NULL;
6812
6813 // now have the array, need to get the start address of the state array
6814 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6815 return state;
6816 }
6817
6818 //------------------------------get_state_from_sha5_object-----------------------
6819 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6820 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6821 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6822 if (sha_state == NULL) return (Node *) NULL;
6823
6824 // now have the array, need to get the start address of the state array
6825 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6826 return state;
6827 }
6828
6829 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6830 // Return node representing slow path of predicate check.
6831 // the pseudo code we want to emulate with this predicate is:
6832 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6833 //
6834 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6835 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6836 "need SHA1/SHA256/SHA512 instruction support");
6837 assert((uint)predicate < 3, "sanity");
6838
6839 // The receiver was checked for NULL already.
6840 Node* digestBaseObj = argument(0);
6841
6842 // get DigestBase klass for instanceOf check
6843 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6844 assert(tinst != NULL, "digestBaseObj is null");
6845 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6846
6847 const char* klass_SHA_name = NULL;
6848 switch (predicate) {
6849 case 0:
6850 if (UseSHA1Intrinsics) {
6851 // we want to do an instanceof comparison against the SHA class
6852 klass_SHA_name = "sun/security/provider/SHA";
6853 }
6854 break;
6855 case 1:
6856 if (UseSHA256Intrinsics) {
6857 // we want to do an instanceof comparison against the SHA2 class
6858 klass_SHA_name = "sun/security/provider/SHA2";
6859 }
6860 break;
6861 case 2:
6862 if (UseSHA512Intrinsics) {
6863 // we want to do an instanceof comparison against the SHA5 class
6864 klass_SHA_name = "sun/security/provider/SHA5";
6865 }
6866 break;
6867 default:
6868 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6869 }
6870
6871 ciKlass* klass_SHA = NULL;
6872 if (klass_SHA_name != NULL) {
6873 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6874 }
6875 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6876 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6877 Node* ctrl = control();
6878 set_control(top()); // no intrinsic path
6879 return ctrl;
6880 }
6881 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6882
6883 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6884 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6885 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6886 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6887
6888 return instof_false; // even if it is NULL
6889 }
6890
6891 bool LibraryCallKit::inline_profileBoolean() {
6892 Node* counts = argument(1);
6893 const TypeAryPtr* ary = NULL;
6894 ciArray* aobj = NULL;
6895 if (counts->is_Con()
6896 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6897 && (aobj = ary->const_oop()->as_array()) != NULL
6898 && (aobj->length() == 2)) {
6899 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6900 jint false_cnt = aobj->element_value(0).as_int();
6901 jint true_cnt = aobj->element_value(1).as_int();
6902
6903 if (C->log() != NULL) {
6904 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6905 false_cnt, true_cnt);
6906 }
6907
6908 if (false_cnt + true_cnt == 0) {
6909 // According to profile, never executed.
6910 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6911 Deoptimization::Action_reinterpret);
6912 return true;
6913 }
6914
6915 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6916 // is a number of each value occurrences.
6917 Node* result = argument(0);
6918 if (false_cnt == 0 || true_cnt == 0) {
6919 // According to profile, one value has been never seen.
6920 int expected_val = (false_cnt == 0) ? 1 : 0;
6921
6922 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6923 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6924
6925 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6926 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6927 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6928
6929 { // Slow path: uncommon trap for never seen value and then reexecute
6930 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6931 // the value has been seen at least once.
6932 PreserveJVMState pjvms(this);
6933 PreserveReexecuteState preexecs(this);
6934 jvms()->set_should_reexecute(true);
6935
6936 set_control(slow_path);
6937 set_i_o(i_o());
6938
6939 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6940 Deoptimization::Action_reinterpret);
6941 }
6942 // The guard for never seen value enables sharpening of the result and
6943 // returning a constant. It allows to eliminate branches on the same value
6944 // later on.
6945 set_control(fast_path);
6946 result = intcon(expected_val);
6947 }
6948 // Stop profiling.
6949 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6950 // By replacing method body with profile data (represented as ProfileBooleanNode
6951 // on IR level) we effectively disable profiling.
6952 // It enables full speed execution once optimized code is generated.
6953 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6954 C->record_for_igvn(profile);
6955 set_result(profile);
6956 return true;
6957 } else {
6958 // Continue profiling.
6959 // Profile data isn't available at the moment. So, execute method's bytecode version.
6960 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6961 // is compiled and counters aren't available since corresponding MethodHandle
6962 // isn't a compile-time constant.
6963 return false;
6964 }
6965 }