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 "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/connode.hpp"
35 #include "opto/idealKit.hpp"
36 #include "opto/mathexactnode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/parse.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "prims/nativeLookup.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "trace/traceMacros.hpp"
44 #include "utilities/macros.hpp"
45 #if INCLUDE_ALL_GCS
46 #include "gc_implementation/shenandoah/shenandoahBarrierSetC2.hpp"
47 #include "gc_implementation/shenandoah/shenandoahSupport.hpp"
48 #endif
49
50
51 class LibraryIntrinsic : public InlineCallGenerator {
52 // Extend the set of intrinsics known to the runtime:
53 public:
54 private:
55 bool _is_virtual;
56 bool _does_virtual_dispatch;
57 int8_t _predicates_count; // Intrinsic is predicated by several conditions
58 int8_t _last_predicate; // Last generated predicate
59 vmIntrinsics::ID _intrinsic_id;
60
61 public:
62 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
63 : InlineCallGenerator(m),
64 _is_virtual(is_virtual),
65 _does_virtual_dispatch(does_virtual_dispatch),
66 _predicates_count((int8_t)predicates_count),
67 _last_predicate((int8_t)-1),
68 _intrinsic_id(id)
69 {
70 }
71 virtual bool is_intrinsic() const { return true; }
72 virtual bool is_virtual() const { return _is_virtual; }
73 virtual bool is_predicated() const { return _predicates_count > 0; }
74 virtual int predicates_count() const { return _predicates_count; }
75 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
76 virtual JVMState* generate(JVMState* jvms);
77 virtual Node* generate_predicate(JVMState* jvms, int predicate);
78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
79 };
80
81
82 // Local helper class for LibraryIntrinsic:
83 class LibraryCallKit : public GraphKit {
84 private:
85 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
86 Node* _result; // the result node, if any
87 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
88
89 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
90
91 public:
92 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
93 : GraphKit(jvms),
94 _intrinsic(intrinsic),
95 _result(NULL)
96 {
97 // Check if this is a root compile. In that case we don't have a caller.
98 if (!jvms->has_method()) {
99 _reexecute_sp = sp();
100 } else {
101 // Find out how many arguments the interpreter needs when deoptimizing
102 // and save the stack pointer value so it can used by uncommon_trap.
103 // We find the argument count by looking at the declared signature.
104 bool ignored_will_link;
105 ciSignature* declared_signature = NULL;
106 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
107 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
108 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
109 }
110 }
111
112 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
113
114 ciMethod* caller() const { return jvms()->method(); }
115 int bci() const { return jvms()->bci(); }
116 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
117 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
118 ciMethod* callee() const { return _intrinsic->method(); }
119
120 bool try_to_inline(int predicate);
121 Node* try_to_predicate(int predicate);
122
123 void push_result() {
124 // Push the result onto the stack.
125 if (!stopped() && result() != NULL) {
126 BasicType bt = result()->bottom_type()->basic_type();
127 push_node(bt, result());
128 }
129 }
130
131 private:
132 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
134 }
135
136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
137 void set_result(RegionNode* region, PhiNode* value);
138 Node* result() { return _result; }
139
140 virtual int reexecute_sp() { return _reexecute_sp; }
141
142 // Helper functions to inline natives
143 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
144 Node* generate_slow_guard(Node* test, RegionNode* region);
145 Node* generate_fair_guard(Node* test, RegionNode* region);
146 Node* generate_negative_guard(Node* index, RegionNode* region,
147 // resulting CastII of index:
148 Node* *pos_index = NULL);
149 Node* generate_nonpositive_guard(Node* index, bool never_negative,
150 // resulting CastII of index:
151 Node* *pos_index = NULL);
152 Node* generate_limit_guard(Node* offset, Node* subseq_length,
153 Node* array_length,
154 RegionNode* region);
155 Node* generate_current_thread(Node* &tls_output);
156 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
157 bool disjoint_bases, const char* &name, bool dest_uninitialized);
158 Node* load_mirror_from_klass(Node* klass);
159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path,
161 int offset);
162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
163 RegionNode* region, int null_path) {
164 int offset = java_lang_Class::klass_offset_in_bytes();
165 return load_klass_from_mirror_common(mirror, never_see_null,
166 region, null_path,
167 offset);
168 }
169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
170 RegionNode* region, int null_path) {
171 int offset = java_lang_Class::array_klass_offset_in_bytes();
172 return load_klass_from_mirror_common(mirror, never_see_null,
173 region, null_path,
174 offset);
175 }
176 Node* generate_access_flags_guard(Node* kls,
177 int modifier_mask, int modifier_bits,
178 RegionNode* region);
179 Node* generate_interface_guard(Node* kls, RegionNode* region);
180 Node* generate_array_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, false, false);
182 }
183 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, false, true);
185 }
186 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, true, false);
188 }
189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, true, true);
191 }
192 Node* generate_array_guard_common(Node* kls, RegionNode* region,
193 bool obj_array, bool not_array);
194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
196 bool is_virtual = false, bool is_static = false);
197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
198 return generate_method_call(method_id, false, true);
199 }
200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
201 return generate_method_call(method_id, true, false);
202 }
203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
204
205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
206 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
207 bool inline_string_compareTo();
208 bool inline_string_indexOf();
209 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
210 bool inline_string_equals();
211 Node* round_double_node(Node* n);
212 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
213 bool inline_math_native(vmIntrinsics::ID id);
214 bool inline_trig(vmIntrinsics::ID id);
215 bool inline_math(vmIntrinsics::ID id);
216 template <typename OverflowOp>
217 bool inline_math_overflow(Node* arg1, Node* arg2);
218 void inline_math_mathExact(Node* math, Node* test);
219 bool inline_math_addExactI(bool is_increment);
220 bool inline_math_addExactL(bool is_increment);
221 bool inline_math_multiplyExactI();
222 bool inline_math_multiplyExactL();
223 bool inline_math_negateExactI();
224 bool inline_math_negateExactL();
225 bool inline_math_subtractExactI(bool is_decrement);
226 bool inline_math_subtractExactL(bool is_decrement);
227 bool inline_exp();
228 bool inline_pow();
229 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
230 bool inline_min_max(vmIntrinsics::ID id);
231 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
232 // This returns Type::AnyPtr, RawPtr, or OopPtr.
233 int classify_unsafe_addr(Node* &base, Node* &offset);
234 Node* make_unsafe_address(Node* base, Node* offset);
235 // Helper for inline_unsafe_access.
236 // Generates the guards that check whether the result of
237 // Unsafe.getObject should be recorded in an SATB log buffer.
238 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
239 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
240 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
241 static bool klass_needs_init_guard(Node* kls);
242 bool inline_unsafe_allocate();
243 bool inline_unsafe_copyMemory();
244 bool inline_native_currentThread();
245 #ifdef TRACE_HAVE_INTRINSICS
246 bool inline_native_classID();
247 bool inline_native_threadID();
248 #endif
249 bool inline_native_time_funcs(address method, const char* funcName);
250 bool inline_native_isInterrupted();
251 bool inline_native_Class_query(vmIntrinsics::ID id);
252 bool inline_native_subtype_check();
253
254 bool inline_native_newArray();
255 bool inline_native_getLength();
256 bool inline_array_copyOf(bool is_copyOfRange);
257 bool inline_array_equals();
258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
259 bool inline_native_clone(bool is_virtual);
260 bool inline_native_Reflection_getCallerClass();
261 // Helper function for inlining native object hash method
262 bool inline_native_hashcode(bool is_virtual, bool is_static);
263 bool inline_native_getClass();
264
265 // Helper functions for inlining arraycopy
266 bool inline_arraycopy();
267 void generate_arraycopy(const TypePtr* adr_type,
268 BasicType basic_elem_type,
269 Node* src, Node* src_offset,
270 Node* dest, Node* dest_offset,
271 Node* copy_length,
272 bool disjoint_bases = false,
273 bool length_never_negative = false,
274 RegionNode* slow_region = NULL);
275 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
276 RegionNode* slow_region);
277 void generate_clear_array(const TypePtr* adr_type,
278 Node* dest,
279 BasicType basic_elem_type,
280 Node* slice_off,
281 Node* slice_len,
282 Node* slice_end);
283 bool generate_block_arraycopy(const TypePtr* adr_type,
284 BasicType basic_elem_type,
285 AllocateNode* alloc,
286 Node* src, Node* src_offset,
287 Node* dest, Node* dest_offset,
288 Node* dest_size, bool dest_uninitialized);
289 void generate_slow_arraycopy(const TypePtr* adr_type,
290 Node* src, Node* src_offset,
291 Node* dest, Node* dest_offset,
292 Node* copy_length, bool dest_uninitialized);
293 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
294 Node* dest_elem_klass,
295 Node* src, Node* src_offset,
296 Node* dest, Node* dest_offset,
297 Node* copy_length, bool dest_uninitialized);
298 Node* generate_generic_arraycopy(const TypePtr* adr_type,
299 Node* src, Node* src_offset,
300 Node* dest, Node* dest_offset,
301 Node* copy_length, bool dest_uninitialized);
302 void generate_unchecked_arraycopy(const TypePtr* adr_type,
303 BasicType basic_elem_type,
304 bool disjoint_bases,
305 Node* src, Node* src_offset,
306 Node* dest, Node* dest_offset,
307 Node* copy_length, bool dest_uninitialized);
308 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
309 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
310 bool inline_unsafe_ordered_store(BasicType type);
311 bool inline_unsafe_fence(vmIntrinsics::ID id);
312 bool inline_fp_conversions(vmIntrinsics::ID id);
313 bool inline_number_methods(vmIntrinsics::ID id);
314 bool inline_reference_get();
315 bool inline_aescrypt_Block(vmIntrinsics::ID id);
316 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
317 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
318 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
319 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
320 bool inline_sha_implCompress(vmIntrinsics::ID id);
321 bool inline_digestBase_implCompressMB(int predicate);
322 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
323 bool long_state, address stubAddr, const char *stubName,
324 Node* src_start, Node* ofs, Node* limit);
325 Node* get_state_from_sha_object(Node *sha_object);
326 Node* get_state_from_sha5_object(Node *sha_object);
327 Node* inline_digestBase_implCompressMB_predicate(int predicate);
328 bool inline_encodeISOArray();
329 bool inline_updateCRC32();
330 bool inline_updateBytesCRC32();
331 bool inline_updateByteBufferCRC32();
332 bool inline_multiplyToLen();
333 bool inline_squareToLen();
334 bool inline_mulAdd();
335 bool inline_montgomeryMultiply();
336 bool inline_montgomerySquare();
337
338 bool inline_profileBoolean();
339 };
340
341
342 //---------------------------make_vm_intrinsic----------------------------
343 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
344 vmIntrinsics::ID id = m->intrinsic_id();
345 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
346
347 ccstr disable_intr = NULL;
348
349 if ((DisableIntrinsic[0] != '\0'
350 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
351 (method_has_option_value("DisableIntrinsic", disable_intr)
352 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
353 // disabled by a user request on the command line:
354 // example: -XX:DisableIntrinsic=_hashCode,_getClass
355 return NULL;
356 }
357
358 if (!m->is_loaded()) {
359 // do not attempt to inline unloaded methods
360 return NULL;
361 }
362
363 // Only a few intrinsics implement a virtual dispatch.
364 // They are expensive calls which are also frequently overridden.
365 if (is_virtual) {
366 switch (id) {
367 case vmIntrinsics::_hashCode:
368 case vmIntrinsics::_clone:
369 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
370 break;
371 default:
372 return NULL;
373 }
374 }
375
376 // -XX:-InlineNatives disables nearly all intrinsics:
377 if (!InlineNatives) {
378 switch (id) {
379 case vmIntrinsics::_indexOf:
380 case vmIntrinsics::_compareTo:
381 case vmIntrinsics::_equals:
382 case vmIntrinsics::_equalsC:
383 case vmIntrinsics::_getAndAddInt:
384 case vmIntrinsics::_getAndAddLong:
385 case vmIntrinsics::_getAndSetInt:
386 case vmIntrinsics::_getAndSetLong:
387 case vmIntrinsics::_getAndSetObject:
388 case vmIntrinsics::_loadFence:
389 case vmIntrinsics::_storeFence:
390 case vmIntrinsics::_fullFence:
391 break; // InlineNatives does not control String.compareTo
392 case vmIntrinsics::_Reference_get:
393 break; // InlineNatives does not control Reference.get
394 default:
395 return NULL;
396 }
397 }
398
399 int predicates = 0;
400 bool does_virtual_dispatch = false;
401
402 switch (id) {
403 case vmIntrinsics::_compareTo:
404 if (!SpecialStringCompareTo) return NULL;
405 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
406 break;
407 case vmIntrinsics::_indexOf:
408 if (!SpecialStringIndexOf) return NULL;
409 break;
410 case vmIntrinsics::_equals:
411 if (!SpecialStringEquals) return NULL;
412 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
413 break;
414 case vmIntrinsics::_equalsC:
415 if (!SpecialArraysEquals) return NULL;
416 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
417 break;
418 case vmIntrinsics::_arraycopy:
419 if (!InlineArrayCopy) return NULL;
420 break;
421 case vmIntrinsics::_copyMemory:
422 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
423 if (!InlineArrayCopy) return NULL;
424 break;
425 case vmIntrinsics::_hashCode:
426 if (!InlineObjectHash) return NULL;
427 does_virtual_dispatch = true;
428 break;
429 case vmIntrinsics::_clone:
430 does_virtual_dispatch = true;
431 case vmIntrinsics::_copyOf:
432 case vmIntrinsics::_copyOfRange:
433 if (!InlineObjectCopy) return NULL;
434 // These also use the arraycopy intrinsic mechanism:
435 if (!InlineArrayCopy) return NULL;
436 break;
437 case vmIntrinsics::_encodeISOArray:
438 if (!SpecialEncodeISOArray) return NULL;
439 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
440 break;
441 case vmIntrinsics::_checkIndex:
442 // We do not intrinsify this. The optimizer does fine with it.
443 return NULL;
444
445 case vmIntrinsics::_getCallerClass:
446 if (!UseNewReflection) return NULL;
447 if (!InlineReflectionGetCallerClass) return NULL;
448 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
449 break;
450
451 case vmIntrinsics::_bitCount_i:
452 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
453 break;
454
455 case vmIntrinsics::_bitCount_l:
456 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
457 break;
458
459 case vmIntrinsics::_numberOfLeadingZeros_i:
460 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
461 break;
462
463 case vmIntrinsics::_numberOfLeadingZeros_l:
464 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
465 break;
466
467 case vmIntrinsics::_numberOfTrailingZeros_i:
468 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
469 break;
470
471 case vmIntrinsics::_numberOfTrailingZeros_l:
472 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
473 break;
474
475 case vmIntrinsics::_reverseBytes_c:
476 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
477 break;
478 case vmIntrinsics::_reverseBytes_s:
479 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
480 break;
481 case vmIntrinsics::_reverseBytes_i:
482 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
483 break;
484 case vmIntrinsics::_reverseBytes_l:
485 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
486 break;
487
488 case vmIntrinsics::_Reference_get:
489 // Use the intrinsic version of Reference.get() so that the value in
490 // the referent field can be registered by the G1 pre-barrier code.
491 // Also add memory barrier to prevent commoning reads from this field
492 // across safepoint since GC can change it value.
493 break;
494
495 case vmIntrinsics::_compareAndSwapObject:
496 #ifdef _LP64
497 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
498 #endif
499 break;
500
501 case vmIntrinsics::_compareAndSwapLong:
502 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
503 break;
504
505 case vmIntrinsics::_getAndAddInt:
506 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
507 break;
508
509 case vmIntrinsics::_getAndAddLong:
510 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
511 break;
512
513 case vmIntrinsics::_getAndSetInt:
514 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
515 break;
516
517 case vmIntrinsics::_getAndSetLong:
518 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
519 break;
520
521 case vmIntrinsics::_getAndSetObject:
522 #ifdef _LP64
523 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
524 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
525 break;
526 #else
527 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
528 break;
529 #endif
530
531 case vmIntrinsics::_aescrypt_encryptBlock:
532 case vmIntrinsics::_aescrypt_decryptBlock:
533 if (!UseAESIntrinsics) return NULL;
534 break;
535
536 case vmIntrinsics::_multiplyToLen:
537 if (!UseMultiplyToLenIntrinsic) return NULL;
538 break;
539
540 case vmIntrinsics::_squareToLen:
541 if (!UseSquareToLenIntrinsic) return NULL;
542 break;
543
544 case vmIntrinsics::_mulAdd:
545 if (!UseMulAddIntrinsic) return NULL;
546 break;
547
548 case vmIntrinsics::_montgomeryMultiply:
549 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
550 break;
551 case vmIntrinsics::_montgomerySquare:
552 if (!UseMontgomerySquareIntrinsic) return NULL;
553 break;
554
555 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
556 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
557 if (!UseAESIntrinsics) return NULL;
558 // these two require the predicated logic
559 predicates = 1;
560 break;
561
562 case vmIntrinsics::_sha_implCompress:
563 if (!UseSHA1Intrinsics) return NULL;
564 break;
565
566 case vmIntrinsics::_sha2_implCompress:
567 if (!UseSHA256Intrinsics) return NULL;
568 break;
569
570 case vmIntrinsics::_sha5_implCompress:
571 if (!UseSHA512Intrinsics) return NULL;
572 break;
573
574 case vmIntrinsics::_digestBase_implCompressMB:
575 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
576 predicates = 3;
577 break;
578
579 case vmIntrinsics::_updateCRC32:
580 case vmIntrinsics::_updateBytesCRC32:
581 case vmIntrinsics::_updateByteBufferCRC32:
582 if (!UseCRC32Intrinsics) return NULL;
583 break;
584
585 case vmIntrinsics::_incrementExactI:
586 case vmIntrinsics::_addExactI:
587 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
588 break;
589 case vmIntrinsics::_incrementExactL:
590 case vmIntrinsics::_addExactL:
591 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
592 break;
593 case vmIntrinsics::_decrementExactI:
594 case vmIntrinsics::_subtractExactI:
595 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
596 break;
597 case vmIntrinsics::_decrementExactL:
598 case vmIntrinsics::_subtractExactL:
599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
600 break;
601 case vmIntrinsics::_negateExactI:
602 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
603 break;
604 case vmIntrinsics::_negateExactL:
605 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
606 break;
607 case vmIntrinsics::_multiplyExactI:
608 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
609 break;
610 case vmIntrinsics::_multiplyExactL:
611 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
612 break;
613
614 default:
615 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
616 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
617 break;
618 }
619
620 // -XX:-InlineClassNatives disables natives from the Class class.
621 // The flag applies to all reflective calls, notably Array.newArray
622 // (visible to Java programmers as Array.newInstance).
623 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
624 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
625 if (!InlineClassNatives) return NULL;
626 }
627
628 // -XX:-InlineThreadNatives disables natives from the Thread class.
629 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
630 if (!InlineThreadNatives) return NULL;
631 }
632
633 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
634 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
635 m->holder()->name() == ciSymbol::java_lang_Float() ||
636 m->holder()->name() == ciSymbol::java_lang_Double()) {
637 if (!InlineMathNatives) return NULL;
638 }
639
640 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
641 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
642 if (!InlineUnsafeOps) return NULL;
643 }
644
645 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
646 }
647
648 //----------------------register_library_intrinsics-----------------------
649 // Initialize this file's data structures, for each Compile instance.
650 void Compile::register_library_intrinsics() {
651 // Nothing to do here.
652 }
653
654 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
655 LibraryCallKit kit(jvms, this);
656 Compile* C = kit.C;
657 int nodes = C->unique();
658 #ifndef PRODUCT
659 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
660 char buf[1000];
661 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
662 tty->print_cr("Intrinsic %s", str);
663 }
664 #endif
665 ciMethod* callee = kit.callee();
666 const int bci = kit.bci();
667
668 // Try to inline the intrinsic.
669 if (kit.try_to_inline(_last_predicate)) {
670 if (C->print_intrinsics() || C->print_inlining()) {
671 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
672 }
673 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
674 if (C->log()) {
675 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
676 vmIntrinsics::name_at(intrinsic_id()),
677 (is_virtual() ? " virtual='1'" : ""),
678 C->unique() - nodes);
679 }
680 // Push the result from the inlined method onto the stack.
681 kit.push_result();
682 return kit.transfer_exceptions_into_jvms();
683 }
684
685 // The intrinsic bailed out
686 if (C->print_intrinsics() || C->print_inlining()) {
687 if (jvms->has_method()) {
688 // Not a root compile.
689 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
690 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
691 } else {
692 // Root compile
693 tty->print("Did not generate intrinsic %s%s at bci:%d in",
694 vmIntrinsics::name_at(intrinsic_id()),
695 (is_virtual() ? " (virtual)" : ""), bci);
696 }
697 }
698 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
699 return NULL;
700 }
701
702 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
703 LibraryCallKit kit(jvms, this);
704 Compile* C = kit.C;
705 int nodes = C->unique();
706 _last_predicate = predicate;
707 #ifndef PRODUCT
708 assert(is_predicated() && predicate < predicates_count(), "sanity");
709 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
710 char buf[1000];
711 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
712 tty->print_cr("Predicate for intrinsic %s", str);
713 }
714 #endif
715 ciMethod* callee = kit.callee();
716 const int bci = kit.bci();
717
718 Node* slow_ctl = kit.try_to_predicate(predicate);
719 if (!kit.failing()) {
720 if (C->print_intrinsics() || C->print_inlining()) {
721 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
722 }
723 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
724 if (C->log()) {
725 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
726 vmIntrinsics::name_at(intrinsic_id()),
727 (is_virtual() ? " virtual='1'" : ""),
728 C->unique() - nodes);
729 }
730 return slow_ctl; // Could be NULL if the check folds.
731 }
732
733 // The intrinsic bailed out
734 if (C->print_intrinsics() || C->print_inlining()) {
735 if (jvms->has_method()) {
736 // Not a root compile.
737 const char* msg = "failed to generate predicate for intrinsic";
738 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
739 } else {
740 // Root compile
741 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
742 vmIntrinsics::name_at(intrinsic_id()),
743 (is_virtual() ? " (virtual)" : ""), bci);
744 }
745 }
746 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
747 return NULL;
748 }
749
750 bool LibraryCallKit::try_to_inline(int predicate) {
751 // Handle symbolic names for otherwise undistinguished boolean switches:
752 const bool is_store = true;
753 const bool is_native_ptr = true;
754 const bool is_static = true;
755 const bool is_volatile = true;
756
757 if (!jvms()->has_method()) {
758 // Root JVMState has a null method.
759 assert(map()->memory()->Opcode() == Op_Parm, "");
760 // Insert the memory aliasing node
761 set_all_memory(reset_memory());
762 }
763 assert(merged_memory(), "");
764
765
766 switch (intrinsic_id()) {
767 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
768 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
769 case vmIntrinsics::_getClass: return inline_native_getClass();
770
771 case vmIntrinsics::_dsin:
772 case vmIntrinsics::_dcos:
773 case vmIntrinsics::_dtan:
774 case vmIntrinsics::_dabs:
775 case vmIntrinsics::_datan2:
776 case vmIntrinsics::_dsqrt:
777 case vmIntrinsics::_dexp:
778 case vmIntrinsics::_dlog:
779 case vmIntrinsics::_dlog10:
780 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
781
782 case vmIntrinsics::_min:
783 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
784
785 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
786 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
787 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
788 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
789 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
790 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
791 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
792 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
793 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
794 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
795 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
796 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
797
798 case vmIntrinsics::_arraycopy: return inline_arraycopy();
799
800 case vmIntrinsics::_compareTo: return inline_string_compareTo();
801 case vmIntrinsics::_indexOf: return inline_string_indexOf();
802 case vmIntrinsics::_equals: return inline_string_equals();
803
804 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
805 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
806 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
807 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
808 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
809 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
810 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
811 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
812 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
813
814 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
815 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
816 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
817 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
818 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
819 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
820 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
821 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
822 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
823
824 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
825 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
826 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
827 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
828 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
829 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
830 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
831 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
832
833 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
834 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
835 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
836 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
837 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
838 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
839 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
840 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
841
842 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
843 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
844 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
845 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
846 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
847 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
848 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
849 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
850 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
851
852 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
853 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
854 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
855 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
856 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
857 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
858 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
859 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
860 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
861
862 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
863 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
864 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
865 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
866
867 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
868 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
869 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
870
871 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
872 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
873 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
874
875 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
876 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
877 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
878 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
879 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
880
881 case vmIntrinsics::_loadFence:
882 case vmIntrinsics::_storeFence:
883 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
884
885 case vmIntrinsics::_currentThread: return inline_native_currentThread();
886 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
887
888 #ifdef TRACE_HAVE_INTRINSICS
889 case vmIntrinsics::_classID: return inline_native_classID();
890 case vmIntrinsics::_threadID: return inline_native_threadID();
891 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
892 #endif
893 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
894 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
895 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
896 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
897 case vmIntrinsics::_newArray: return inline_native_newArray();
898 case vmIntrinsics::_getLength: return inline_native_getLength();
899 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
900 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
901 case vmIntrinsics::_equalsC: return inline_array_equals();
902 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
903
904 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
905
906 case vmIntrinsics::_isInstance:
907 case vmIntrinsics::_getModifiers:
908 case vmIntrinsics::_isInterface:
909 case vmIntrinsics::_isArray:
910 case vmIntrinsics::_isPrimitive:
911 case vmIntrinsics::_getSuperclass:
912 case vmIntrinsics::_getComponentType:
913 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
914
915 case vmIntrinsics::_floatToRawIntBits:
916 case vmIntrinsics::_floatToIntBits:
917 case vmIntrinsics::_intBitsToFloat:
918 case vmIntrinsics::_doubleToRawLongBits:
919 case vmIntrinsics::_doubleToLongBits:
920 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
921
922 case vmIntrinsics::_numberOfLeadingZeros_i:
923 case vmIntrinsics::_numberOfLeadingZeros_l:
924 case vmIntrinsics::_numberOfTrailingZeros_i:
925 case vmIntrinsics::_numberOfTrailingZeros_l:
926 case vmIntrinsics::_bitCount_i:
927 case vmIntrinsics::_bitCount_l:
928 case vmIntrinsics::_reverseBytes_i:
929 case vmIntrinsics::_reverseBytes_l:
930 case vmIntrinsics::_reverseBytes_s:
931 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
932
933 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
934
935 case vmIntrinsics::_Reference_get: return inline_reference_get();
936
937 case vmIntrinsics::_aescrypt_encryptBlock:
938 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
939
940 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
941 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
942 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
943
944 case vmIntrinsics::_sha_implCompress:
945 case vmIntrinsics::_sha2_implCompress:
946 case vmIntrinsics::_sha5_implCompress:
947 return inline_sha_implCompress(intrinsic_id());
948
949 case vmIntrinsics::_digestBase_implCompressMB:
950 return inline_digestBase_implCompressMB(predicate);
951
952 case vmIntrinsics::_multiplyToLen:
953 return inline_multiplyToLen();
954
955 case vmIntrinsics::_squareToLen:
956 return inline_squareToLen();
957
958 case vmIntrinsics::_mulAdd:
959 return inline_mulAdd();
960
961 case vmIntrinsics::_montgomeryMultiply:
962 return inline_montgomeryMultiply();
963 case vmIntrinsics::_montgomerySquare:
964 return inline_montgomerySquare();
965
966 case vmIntrinsics::_encodeISOArray:
967 return inline_encodeISOArray();
968
969 case vmIntrinsics::_updateCRC32:
970 return inline_updateCRC32();
971 case vmIntrinsics::_updateBytesCRC32:
972 return inline_updateBytesCRC32();
973 case vmIntrinsics::_updateByteBufferCRC32:
974 return inline_updateByteBufferCRC32();
975
976 case vmIntrinsics::_profileBoolean:
977 return inline_profileBoolean();
978
979 default:
980 // If you get here, it may be that someone has added a new intrinsic
981 // to the list in vmSymbols.hpp without implementing it here.
982 #ifndef PRODUCT
983 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
984 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
985 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
986 }
987 #endif
988 return false;
989 }
990 }
991
992 Node* LibraryCallKit::try_to_predicate(int predicate) {
993 if (!jvms()->has_method()) {
994 // Root JVMState has a null method.
995 assert(map()->memory()->Opcode() == Op_Parm, "");
996 // Insert the memory aliasing node
997 set_all_memory(reset_memory());
998 }
999 assert(merged_memory(), "");
1000
1001 switch (intrinsic_id()) {
1002 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1003 return inline_cipherBlockChaining_AESCrypt_predicate(false);
1004 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1005 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1006 case vmIntrinsics::_digestBase_implCompressMB:
1007 return inline_digestBase_implCompressMB_predicate(predicate);
1008
1009 default:
1010 // If you get here, it may be that someone has added a new intrinsic
1011 // to the list in vmSymbols.hpp without implementing it here.
1012 #ifndef PRODUCT
1013 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1014 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1015 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1016 }
1017 #endif
1018 Node* slow_ctl = control();
1019 set_control(top()); // No fast path instrinsic
1020 return slow_ctl;
1021 }
1022 }
1023
1024 //------------------------------set_result-------------------------------
1025 // Helper function for finishing intrinsics.
1026 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1027 record_for_igvn(region);
1028 set_control(_gvn.transform(region));
1029 set_result( _gvn.transform(value));
1030 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1031 }
1032
1033 //------------------------------generate_guard---------------------------
1034 // Helper function for generating guarded fast-slow graph structures.
1035 // The given 'test', if true, guards a slow path. If the test fails
1036 // then a fast path can be taken. (We generally hope it fails.)
1037 // In all cases, GraphKit::control() is updated to the fast path.
1038 // The returned value represents the control for the slow path.
1039 // The return value is never 'top'; it is either a valid control
1040 // or NULL if it is obvious that the slow path can never be taken.
1041 // Also, if region and the slow control are not NULL, the slow edge
1042 // is appended to the region.
1043 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1044 if (stopped()) {
1045 // Already short circuited.
1046 return NULL;
1047 }
1048
1049 // Build an if node and its projections.
1050 // If test is true we take the slow path, which we assume is uncommon.
1051 if (_gvn.type(test) == TypeInt::ZERO) {
1052 // The slow branch is never taken. No need to build this guard.
1053 return NULL;
1054 }
1055
1056 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1057
1058 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1059 if (if_slow == top()) {
1060 // The slow branch is never taken. No need to build this guard.
1061 return NULL;
1062 }
1063
1064 if (region != NULL)
1065 region->add_req(if_slow);
1066
1067 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1068 set_control(if_fast);
1069
1070 return if_slow;
1071 }
1072
1073 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1074 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1075 }
1076 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1077 return generate_guard(test, region, PROB_FAIR);
1078 }
1079
1080 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1081 Node* *pos_index) {
1082 if (stopped())
1083 return NULL; // already stopped
1084 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1085 return NULL; // index is already adequately typed
1086 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1087 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1088 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1089 if (is_neg != NULL && pos_index != NULL) {
1090 // Emulate effect of Parse::adjust_map_after_if.
1091 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1092 ccast->set_req(0, control());
1093 (*pos_index) = _gvn.transform(ccast);
1094 }
1095 return is_neg;
1096 }
1097
1098 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1099 Node* *pos_index) {
1100 if (stopped())
1101 return NULL; // already stopped
1102 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1103 return NULL; // index is already adequately typed
1104 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1105 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1106 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1107 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1108 if (is_notp != NULL && pos_index != NULL) {
1109 // Emulate effect of Parse::adjust_map_after_if.
1110 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1111 ccast->set_req(0, control());
1112 (*pos_index) = _gvn.transform(ccast);
1113 }
1114 return is_notp;
1115 }
1116
1117 // Make sure that 'position' is a valid limit index, in [0..length].
1118 // There are two equivalent plans for checking this:
1119 // A. (offset + copyLength) unsigned<= arrayLength
1120 // B. offset <= (arrayLength - copyLength)
1121 // We require that all of the values above, except for the sum and
1122 // difference, are already known to be non-negative.
1123 // Plan A is robust in the face of overflow, if offset and copyLength
1124 // are both hugely positive.
1125 //
1126 // Plan B is less direct and intuitive, but it does not overflow at
1127 // all, since the difference of two non-negatives is always
1128 // representable. Whenever Java methods must perform the equivalent
1129 // check they generally use Plan B instead of Plan A.
1130 // For the moment we use Plan A.
1131 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1132 Node* subseq_length,
1133 Node* array_length,
1134 RegionNode* region) {
1135 if (stopped())
1136 return NULL; // already stopped
1137 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1138 if (zero_offset && subseq_length->eqv_uncast(array_length))
1139 return NULL; // common case of whole-array copy
1140 Node* last = subseq_length;
1141 if (!zero_offset) // last += offset
1142 last = _gvn.transform(new (C) AddINode(last, offset));
1143 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1144 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1145 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1146 return is_over;
1147 }
1148
1149
1150 //--------------------------generate_current_thread--------------------
1151 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1152 ciKlass* thread_klass = env()->Thread_klass();
1153 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1154 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1155 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1156 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1157 tls_output = thread;
1158 return threadObj;
1159 }
1160
1161
1162 //------------------------------make_string_method_node------------------------
1163 // Helper method for String intrinsic functions. This version is called
1164 // with str1 and str2 pointing to String object nodes.
1165 //
1166 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1167 Node* no_ctrl = NULL;
1168
1169 // Get start addr of string
1170 Node* str1_value = load_String_value(no_ctrl, str1);
1171 Node* str1_offset = load_String_offset(no_ctrl, str1);
1172 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1173
1174 // Get length of string 1
1175 Node* str1_len = load_String_length(no_ctrl, str1);
1176
1177 Node* str2_value = load_String_value(no_ctrl, str2);
1178 Node* str2_offset = load_String_offset(no_ctrl, str2);
1179 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1180
1181 Node* str2_len = NULL;
1182 Node* result = NULL;
1183
1184 switch (opcode) {
1185 case Op_StrIndexOf:
1186 // Get length of string 2
1187 str2_len = load_String_length(no_ctrl, str2);
1188
1189 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1190 str1_start, str1_len, str2_start, str2_len);
1191 break;
1192 case Op_StrComp:
1193 // Get length of string 2
1194 str2_len = load_String_length(no_ctrl, str2);
1195
1196 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1197 str1_start, str1_len, str2_start, str2_len);
1198 break;
1199 case Op_StrEquals:
1200 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1201 str1_start, str2_start, str1_len);
1202 break;
1203 default:
1204 ShouldNotReachHere();
1205 return NULL;
1206 }
1207
1208 // All these intrinsics have checks.
1209 C->set_has_split_ifs(true); // Has chance for split-if optimization
1210
1211 return _gvn.transform(result);
1212 }
1213
1214 // Helper method for String intrinsic functions. This version is called
1215 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1216 // to Int nodes containing the lenghts of str1 and str2.
1217 //
1218 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1219 Node* result = NULL;
1220 switch (opcode) {
1221 case Op_StrIndexOf:
1222 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1223 str1_start, cnt1, str2_start, cnt2);
1224 break;
1225 case Op_StrComp:
1226 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1227 str1_start, cnt1, str2_start, cnt2);
1228 break;
1229 case Op_StrEquals:
1230 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1231 str1_start, str2_start, cnt1);
1232 break;
1233 default:
1234 ShouldNotReachHere();
1235 return NULL;
1236 }
1237
1238 // All these intrinsics have checks.
1239 C->set_has_split_ifs(true); // Has chance for split-if optimization
1240
1241 return _gvn.transform(result);
1242 }
1243
1244 //------------------------------inline_string_compareTo------------------------
1245 // public int java.lang.String.compareTo(String anotherString);
1246 bool LibraryCallKit::inline_string_compareTo() {
1247 Node* receiver = null_check(argument(0));
1248 Node* arg = null_check(argument(1));
1249 if (stopped()) {
1250 return true;
1251 }
1252 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1253 return true;
1254 }
1255
1256 //------------------------------inline_string_equals------------------------
1257 bool LibraryCallKit::inline_string_equals() {
1258 Node* receiver = null_check_receiver();
1259 // NOTE: Do not null check argument for String.equals() because spec
1260 // allows to specify NULL as argument.
1261 Node* argument = this->argument(1);
1262 if (stopped()) {
1263 return true;
1264 }
1265
1266 // paths (plus control) merge
1267 RegionNode* region = new (C) RegionNode(5);
1268 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1269
1270 // does source == target string?
1271 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1272 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1273
1274 Node* if_eq = generate_slow_guard(bol, NULL);
1275 if (if_eq != NULL) {
1276 // receiver == argument
1277 phi->init_req(2, intcon(1));
1278 region->init_req(2, if_eq);
1279 }
1280
1281 // get String klass for instanceOf
1282 ciInstanceKlass* klass = env()->String_klass();
1283
1284 if (!stopped()) {
1285 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1286 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1287 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1288
1289 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1290 //instanceOf == true, fallthrough
1291
1292 if (inst_false != NULL) {
1293 phi->init_req(3, intcon(0));
1294 region->init_req(3, inst_false);
1295 }
1296 }
1297
1298 if (!stopped()) {
1299 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1300
1301 // Properly cast the argument to String
1302 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1303 // This path is taken only when argument's type is String:NotNull.
1304 argument = cast_not_null(argument, false);
1305
1306 Node* no_ctrl = NULL;
1307
1308 // Get start addr of receiver
1309 Node* receiver_val = load_String_value(no_ctrl, receiver);
1310 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1311 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1312
1313 // Get length of receiver
1314 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1315
1316 // Get start addr of argument
1317 Node* argument_val = load_String_value(no_ctrl, argument);
1318 Node* argument_offset = load_String_offset(no_ctrl, argument);
1319 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1320
1321 // Get length of argument
1322 Node* argument_cnt = load_String_length(no_ctrl, argument);
1323
1324 // Check for receiver count != argument count
1325 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1326 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1327 Node* if_ne = generate_slow_guard(bol, NULL);
1328 if (if_ne != NULL) {
1329 phi->init_req(4, intcon(0));
1330 region->init_req(4, if_ne);
1331 }
1332
1333 // Check for count == 0 is done by assembler code for StrEquals.
1334
1335 if (!stopped()) {
1336 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1337 phi->init_req(1, equals);
1338 region->init_req(1, control());
1339 }
1340 }
1341
1342 // post merge
1343 set_control(_gvn.transform(region));
1344 record_for_igvn(region);
1345
1346 set_result(_gvn.transform(phi));
1347 return true;
1348 }
1349
1350 //------------------------------inline_array_equals----------------------------
1351 bool LibraryCallKit::inline_array_equals() {
1352 Node* arg1 = argument(0);
1353 Node* arg2 = argument(1);
1354 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1355 return true;
1356 }
1357
1358 // Java version of String.indexOf(constant string)
1359 // class StringDecl {
1360 // StringDecl(char[] ca) {
1361 // offset = 0;
1362 // count = ca.length;
1363 // value = ca;
1364 // }
1365 // int offset;
1366 // int count;
1367 // char[] value;
1368 // }
1369 //
1370 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1371 // int targetOffset, int cache_i, int md2) {
1372 // int cache = cache_i;
1373 // int sourceOffset = string_object.offset;
1374 // int sourceCount = string_object.count;
1375 // int targetCount = target_object.length;
1376 //
1377 // int targetCountLess1 = targetCount - 1;
1378 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1379 //
1380 // char[] source = string_object.value;
1381 // char[] target = target_object;
1382 // int lastChar = target[targetCountLess1];
1383 //
1384 // outer_loop:
1385 // for (int i = sourceOffset; i < sourceEnd; ) {
1386 // int src = source[i + targetCountLess1];
1387 // if (src == lastChar) {
1388 // // With random strings and a 4-character alphabet,
1389 // // reverse matching at this point sets up 0.8% fewer
1390 // // frames, but (paradoxically) makes 0.3% more probes.
1391 // // Since those probes are nearer the lastChar probe,
1392 // // there is may be a net D$ win with reverse matching.
1393 // // But, reversing loop inhibits unroll of inner loop
1394 // // for unknown reason. So, does running outer loop from
1395 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1396 // for (int j = 0; j < targetCountLess1; j++) {
1397 // if (target[targetOffset + j] != source[i+j]) {
1398 // if ((cache & (1 << source[i+j])) == 0) {
1399 // if (md2 < j+1) {
1400 // i += j+1;
1401 // continue outer_loop;
1402 // }
1403 // }
1404 // i += md2;
1405 // continue outer_loop;
1406 // }
1407 // }
1408 // return i - sourceOffset;
1409 // }
1410 // if ((cache & (1 << src)) == 0) {
1411 // i += targetCountLess1;
1412 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1413 // i++;
1414 // }
1415 // return -1;
1416 // }
1417
1418 //------------------------------string_indexOf------------------------
1419 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1420 jint cache_i, jint md2_i) {
1421
1422 Node* no_ctrl = NULL;
1423 float likely = PROB_LIKELY(0.9);
1424 float unlikely = PROB_UNLIKELY(0.9);
1425
1426 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1427
1428 Node* source = load_String_value(no_ctrl, string_object);
1429 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1430 Node* sourceCount = load_String_length(no_ctrl, string_object);
1431
1432 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1433 jint target_length = target_array->length();
1434 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1435 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1436
1437 // String.value field is known to be @Stable.
1438 if (UseImplicitStableValues) {
1439 target = cast_array_to_stable(target, target_type);
1440 }
1441
1442 IdealKit kit(this, false, true);
1443 #define __ kit.
1444 Node* zero = __ ConI(0);
1445 Node* one = __ ConI(1);
1446 Node* cache = __ ConI(cache_i);
1447 Node* md2 = __ ConI(md2_i);
1448 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1449 Node* targetCount = __ ConI(target_length);
1450 Node* targetCountLess1 = __ ConI(target_length - 1);
1451 Node* targetOffset = __ ConI(targetOffset_i);
1452 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1453
1454 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1455 Node* outer_loop = __ make_label(2 /* goto */);
1456 Node* return_ = __ make_label(1);
1457
1458 __ set(rtn,__ ConI(-1));
1459 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1460 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1461 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1462 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1463 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1464 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1465 Node* tpj = __ AddI(targetOffset, __ value(j));
1466 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1467 Node* ipj = __ AddI(__ value(i), __ value(j));
1468 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1469 __ if_then(targ, BoolTest::ne, src2); {
1470 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1471 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1472 __ increment(i, __ AddI(__ value(j), one));
1473 __ goto_(outer_loop);
1474 } __ end_if(); __ dead(j);
1475 }__ end_if(); __ dead(j);
1476 __ increment(i, md2);
1477 __ goto_(outer_loop);
1478 }__ end_if();
1479 __ increment(j, one);
1480 }__ end_loop(); __ dead(j);
1481 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1482 __ goto_(return_);
1483 }__ end_if();
1484 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1485 __ increment(i, targetCountLess1);
1486 }__ end_if();
1487 __ increment(i, one);
1488 __ bind(outer_loop);
1489 }__ end_loop(); __ dead(i);
1490 __ bind(return_);
1491
1492 // Final sync IdealKit and GraphKit.
1493 final_sync(kit);
1494 Node* result = __ value(rtn);
1495 #undef __
1496 C->set_has_loops(true);
1497 return result;
1498 }
1499
1500 //------------------------------inline_string_indexOf------------------------
1501 bool LibraryCallKit::inline_string_indexOf() {
1502 Node* receiver = argument(0);
1503 Node* arg = argument(1);
1504
1505 Node* result;
1506 // Disable the use of pcmpestri until it can be guaranteed that
1507 // the load doesn't cross into the uncommited space.
1508 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1509 UseSSE42Intrinsics) {
1510 // Generate SSE4.2 version of indexOf
1511 // We currently only have match rules that use SSE4.2
1512
1513 receiver = null_check(receiver);
1514 arg = null_check(arg);
1515 if (stopped()) {
1516 return true;
1517 }
1518
1519 ciInstanceKlass* str_klass = env()->String_klass();
1520 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1521
1522 // Make the merge point
1523 RegionNode* result_rgn = new (C) RegionNode(4);
1524 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1525 Node* no_ctrl = NULL;
1526
1527 // Get start addr of source string
1528 Node* source = load_String_value(no_ctrl, receiver);
1529 Node* source_offset = load_String_offset(no_ctrl, receiver);
1530 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1531
1532 // Get length of source string
1533 Node* source_cnt = load_String_length(no_ctrl, receiver);
1534
1535 // Get start addr of substring
1536 Node* substr = load_String_value(no_ctrl, arg);
1537 Node* substr_offset = load_String_offset(no_ctrl, arg);
1538 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1539
1540 // Get length of source string
1541 Node* substr_cnt = load_String_length(no_ctrl, arg);
1542
1543 // Check for substr count > string count
1544 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1545 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1546 Node* if_gt = generate_slow_guard(bol, NULL);
1547 if (if_gt != NULL) {
1548 result_phi->init_req(2, intcon(-1));
1549 result_rgn->init_req(2, if_gt);
1550 }
1551
1552 if (!stopped()) {
1553 // Check for substr count == 0
1554 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1555 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1556 Node* if_zero = generate_slow_guard(bol, NULL);
1557 if (if_zero != NULL) {
1558 result_phi->init_req(3, intcon(0));
1559 result_rgn->init_req(3, if_zero);
1560 }
1561 }
1562
1563 if (!stopped()) {
1564 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1565 result_phi->init_req(1, result);
1566 result_rgn->init_req(1, control());
1567 }
1568 set_control(_gvn.transform(result_rgn));
1569 record_for_igvn(result_rgn);
1570 result = _gvn.transform(result_phi);
1571
1572 } else { // Use LibraryCallKit::string_indexOf
1573 // don't intrinsify if argument isn't a constant string.
1574 if (!arg->is_Con()) {
1575 return false;
1576 }
1577 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1578 if (str_type == NULL) {
1579 return false;
1580 }
1581 ciInstanceKlass* klass = env()->String_klass();
1582 ciObject* str_const = str_type->const_oop();
1583 if (str_const == NULL || str_const->klass() != klass) {
1584 return false;
1585 }
1586 ciInstance* str = str_const->as_instance();
1587 assert(str != NULL, "must be instance");
1588
1589 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1590 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1591
1592 int o;
1593 int c;
1594 if (java_lang_String::has_offset_field()) {
1595 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1596 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1597 } else {
1598 o = 0;
1599 c = pat->length();
1600 }
1601
1602 // constant strings have no offset and count == length which
1603 // simplifies the resulting code somewhat so lets optimize for that.
1604 if (o != 0 || c != pat->length()) {
1605 return false;
1606 }
1607
1608 receiver = null_check(receiver, T_OBJECT);
1609 // NOTE: No null check on the argument is needed since it's a constant String oop.
1610 if (stopped()) {
1611 return true;
1612 }
1613
1614 // The null string as a pattern always returns 0 (match at beginning of string)
1615 if (c == 0) {
1616 set_result(intcon(0));
1617 return true;
1618 }
1619
1620 // Generate default indexOf
1621 jchar lastChar = pat->char_at(o + (c - 1));
1622 int cache = 0;
1623 int i;
1624 for (i = 0; i < c - 1; i++) {
1625 assert(i < pat->length(), "out of range");
1626 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1627 }
1628
1629 int md2 = c;
1630 for (i = 0; i < c - 1; i++) {
1631 assert(i < pat->length(), "out of range");
1632 if (pat->char_at(o + i) == lastChar) {
1633 md2 = (c - 1) - i;
1634 }
1635 }
1636
1637 result = string_indexOf(receiver, pat, o, cache, md2);
1638 }
1639 set_result(result);
1640 return true;
1641 }
1642
1643 //--------------------------round_double_node--------------------------------
1644 // Round a double node if necessary.
1645 Node* LibraryCallKit::round_double_node(Node* n) {
1646 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1647 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1648 return n;
1649 }
1650
1651 //------------------------------inline_math-----------------------------------
1652 // public static double Math.abs(double)
1653 // public static double Math.sqrt(double)
1654 // public static double Math.log(double)
1655 // public static double Math.log10(double)
1656 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1657 Node* arg = round_double_node(argument(0));
1658 Node* n = NULL;
1659 switch (id) {
1660 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1661 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1662 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1663 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1664 default: fatal_unexpected_iid(id); break;
1665 }
1666 set_result(_gvn.transform(n));
1667 return true;
1668 }
1669
1670 //------------------------------inline_trig----------------------------------
1671 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1672 // argument reduction which will turn into a fast/slow diamond.
1673 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1674 Node* arg = round_double_node(argument(0));
1675 Node* n = NULL;
1676
1677 switch (id) {
1678 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1679 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1680 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1681 default: fatal_unexpected_iid(id); break;
1682 }
1683 n = _gvn.transform(n);
1684
1685 // Rounding required? Check for argument reduction!
1686 if (Matcher::strict_fp_requires_explicit_rounding) {
1687 static const double pi_4 = 0.7853981633974483;
1688 static const double neg_pi_4 = -0.7853981633974483;
1689 // pi/2 in 80-bit extended precision
1690 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1691 // -pi/2 in 80-bit extended precision
1692 // 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};
1693 // Cutoff value for using this argument reduction technique
1694 //static const double pi_2_minus_epsilon = 1.564660403643354;
1695 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1696
1697 // Pseudocode for sin:
1698 // if (x <= Math.PI / 4.0) {
1699 // if (x >= -Math.PI / 4.0) return fsin(x);
1700 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1701 // } else {
1702 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1703 // }
1704 // return StrictMath.sin(x);
1705
1706 // Pseudocode for cos:
1707 // if (x <= Math.PI / 4.0) {
1708 // if (x >= -Math.PI / 4.0) return fcos(x);
1709 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1710 // } else {
1711 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1712 // }
1713 // return StrictMath.cos(x);
1714
1715 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1716 // requires a special machine instruction to load it. Instead we'll try
1717 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1718 // probably do the math inside the SIN encoding.
1719
1720 // Make the merge point
1721 RegionNode* r = new (C) RegionNode(3);
1722 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1723
1724 // Flatten arg so we need only 1 test
1725 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1726 // Node for PI/4 constant
1727 Node *pi4 = makecon(TypeD::make(pi_4));
1728 // Check PI/4 : abs(arg)
1729 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1730 // Check: If PI/4 < abs(arg) then go slow
1731 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1732 // Branch either way
1733 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1734 set_control(opt_iff(r,iff));
1735
1736 // Set fast path result
1737 phi->init_req(2, n);
1738
1739 // Slow path - non-blocking leaf call
1740 Node* call = NULL;
1741 switch (id) {
1742 case vmIntrinsics::_dsin:
1743 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1744 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1745 "Sin", NULL, arg, top());
1746 break;
1747 case vmIntrinsics::_dcos:
1748 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1749 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1750 "Cos", NULL, arg, top());
1751 break;
1752 case vmIntrinsics::_dtan:
1753 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1754 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1755 "Tan", NULL, arg, top());
1756 break;
1757 }
1758 assert(control()->in(0) == call, "");
1759 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1760 r->init_req(1, control());
1761 phi->init_req(1, slow_result);
1762
1763 // Post-merge
1764 set_control(_gvn.transform(r));
1765 record_for_igvn(r);
1766 n = _gvn.transform(phi);
1767
1768 C->set_has_split_ifs(true); // Has chance for split-if optimization
1769 }
1770 set_result(n);
1771 return true;
1772 }
1773
1774 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1775 //-------------------
1776 //result=(result.isNaN())? funcAddr():result;
1777 // Check: If isNaN() by checking result!=result? then either trap
1778 // or go to runtime
1779 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1780 // Build the boolean node
1781 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1782
1783 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1784 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1785 // The pow or exp intrinsic returned a NaN, which requires a call
1786 // to the runtime. Recompile with the runtime call.
1787 uncommon_trap(Deoptimization::Reason_intrinsic,
1788 Deoptimization::Action_make_not_entrant);
1789 }
1790 return result;
1791 } else {
1792 // If this inlining ever returned NaN in the past, we compile a call
1793 // to the runtime to properly handle corner cases
1794
1795 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1796 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1797 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1798
1799 if (!if_slow->is_top()) {
1800 RegionNode* result_region = new (C) RegionNode(3);
1801 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1802
1803 result_region->init_req(1, if_fast);
1804 result_val->init_req(1, result);
1805
1806 set_control(if_slow);
1807
1808 const TypePtr* no_memory_effects = NULL;
1809 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1810 no_memory_effects,
1811 x, top(), y, y ? top() : NULL);
1812 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1813 #ifdef ASSERT
1814 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1815 assert(value_top == top(), "second value must be top");
1816 #endif
1817
1818 result_region->init_req(2, control());
1819 result_val->init_req(2, value);
1820 set_control(_gvn.transform(result_region));
1821 return _gvn.transform(result_val);
1822 } else {
1823 return result;
1824 }
1825 }
1826 }
1827
1828 //------------------------------inline_exp-------------------------------------
1829 // Inline exp instructions, if possible. The Intel hardware only misses
1830 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1831 bool LibraryCallKit::inline_exp() {
1832 Node* arg = round_double_node(argument(0));
1833 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1834
1835 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1836 set_result(n);
1837
1838 C->set_has_split_ifs(true); // Has chance for split-if optimization
1839 return true;
1840 }
1841
1842 //------------------------------inline_pow-------------------------------------
1843 // Inline power instructions, if possible.
1844 bool LibraryCallKit::inline_pow() {
1845 // Pseudocode for pow
1846 // if (y == 2) {
1847 // return x * x;
1848 // } else {
1849 // if (x <= 0.0) {
1850 // long longy = (long)y;
1851 // if ((double)longy == y) { // if y is long
1852 // if (y + 1 == y) longy = 0; // huge number: even
1853 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1854 // } else {
1855 // result = NaN;
1856 // }
1857 // } else {
1858 // result = DPow(x,y);
1859 // }
1860 // if (result != result)? {
1861 // result = uncommon_trap() or runtime_call();
1862 // }
1863 // return result;
1864 // }
1865
1866 Node* x = round_double_node(argument(0));
1867 Node* y = round_double_node(argument(2));
1868
1869 Node* result = NULL;
1870
1871 Node* const_two_node = makecon(TypeD::make(2.0));
1872 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1873 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1874 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1875 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1876 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1877
1878 RegionNode* region_node = new (C) RegionNode(3);
1879 region_node->init_req(1, if_true);
1880
1881 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1882 // special case for x^y where y == 2, we can convert it to x * x
1883 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1884
1885 // set control to if_false since we will now process the false branch
1886 set_control(if_false);
1887
1888 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1889 // Short form: skip the fancy tests and just check for NaN result.
1890 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1891 } else {
1892 // If this inlining ever returned NaN in the past, include all
1893 // checks + call to the runtime.
1894
1895 // Set the merge point for If node with condition of (x <= 0.0)
1896 // There are four possible paths to region node and phi node
1897 RegionNode *r = new (C) RegionNode(4);
1898 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1899
1900 // Build the first if node: if (x <= 0.0)
1901 // Node for 0 constant
1902 Node *zeronode = makecon(TypeD::ZERO);
1903 // Check x:0
1904 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1905 // Check: If (x<=0) then go complex path
1906 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1907 // Branch either way
1908 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1909 // Fast path taken; set region slot 3
1910 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1911 r->init_req(3,fast_taken); // Capture fast-control
1912
1913 // Fast path not-taken, i.e. slow path
1914 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1915
1916 // Set fast path result
1917 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1918 phi->init_req(3, fast_result);
1919
1920 // Complex path
1921 // Build the second if node (if y is long)
1922 // Node for (long)y
1923 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1924 // Node for (double)((long) y)
1925 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1926 // Check (double)((long) y) : y
1927 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1928 // Check if (y isn't long) then go to slow path
1929
1930 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1931 // Branch either way
1932 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1933 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1934
1935 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1936
1937 // Calculate DPow(abs(x), y)*(1 & (long)y)
1938 // Node for constant 1
1939 Node *conone = longcon(1);
1940 // 1& (long)y
1941 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1942
1943 // A huge number is always even. Detect a huge number by checking
1944 // if y + 1 == y and set integer to be tested for parity to 0.
1945 // Required for corner case:
1946 // (long)9.223372036854776E18 = max_jlong
1947 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1948 // max_jlong is odd but 9.223372036854776E18 is even
1949 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1950 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1951 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1952 Node* correctedsign = NULL;
1953 if (ConditionalMoveLimit != 0) {
1954 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1955 } else {
1956 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1957 RegionNode *r = new (C) RegionNode(3);
1958 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1959 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1960 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1961 phi->init_req(1, signnode);
1962 phi->init_req(2, longcon(0));
1963 correctedsign = _gvn.transform(phi);
1964 ylong_path = _gvn.transform(r);
1965 record_for_igvn(r);
1966 }
1967
1968 // zero node
1969 Node *conzero = longcon(0);
1970 // Check (1&(long)y)==0?
1971 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1972 // Check if (1&(long)y)!=0?, if so the result is negative
1973 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1974 // abs(x)
1975 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1976 // abs(x)^y
1977 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1978 // -abs(x)^y
1979 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1980 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1981 Node *signresult = NULL;
1982 if (ConditionalMoveLimit != 0) {
1983 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1984 } else {
1985 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1986 RegionNode *r = new (C) RegionNode(3);
1987 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1988 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1989 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1990 phi->init_req(1, absxpowy);
1991 phi->init_req(2, negabsxpowy);
1992 signresult = _gvn.transform(phi);
1993 ylong_path = _gvn.transform(r);
1994 record_for_igvn(r);
1995 }
1996 // Set complex path fast result
1997 r->init_req(2, ylong_path);
1998 phi->init_req(2, signresult);
1999
2000 static const jlong nan_bits = CONST64(0x7ff8000000000000);
2001 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
2002 r->init_req(1,slow_path);
2003 phi->init_req(1,slow_result);
2004
2005 // Post merge
2006 set_control(_gvn.transform(r));
2007 record_for_igvn(r);
2008 result = _gvn.transform(phi);
2009 }
2010
2011 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2012
2013 // control from finish_pow_exp is now input to the region node
2014 region_node->set_req(2, control());
2015 // the result from finish_pow_exp is now input to the phi node
2016 phi_node->init_req(2, result);
2017 set_control(_gvn.transform(region_node));
2018 record_for_igvn(region_node);
2019 set_result(_gvn.transform(phi_node));
2020
2021 C->set_has_split_ifs(true); // Has chance for split-if optimization
2022 return true;
2023 }
2024
2025 //------------------------------runtime_math-----------------------------
2026 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2027 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2028 "must be (DD)D or (D)D type");
2029
2030 // Inputs
2031 Node* a = round_double_node(argument(0));
2032 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2033
2034 const TypePtr* no_memory_effects = NULL;
2035 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2036 no_memory_effects,
2037 a, top(), b, b ? top() : NULL);
2038 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2039 #ifdef ASSERT
2040 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2041 assert(value_top == top(), "second value must be top");
2042 #endif
2043
2044 set_result(value);
2045 return true;
2046 }
2047
2048 //------------------------------inline_math_native-----------------------------
2049 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2050 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2051 switch (id) {
2052 // These intrinsics are not properly supported on all hardware
2053 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2054 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2055 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2056 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2057 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2058 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2059
2060 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2061 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2062 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2063 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2064
2065 // These intrinsics are supported on all hardware
2066 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2067 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2068
2069 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2070 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2071 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2072 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2073 #undef FN_PTR
2074
2075 // These intrinsics are not yet correctly implemented
2076 case vmIntrinsics::_datan2:
2077 return false;
2078
2079 default:
2080 fatal_unexpected_iid(id);
2081 return false;
2082 }
2083 }
2084
2085 static bool is_simple_name(Node* n) {
2086 return (n->req() == 1 // constant
2087 || (n->is_Type() && n->as_Type()->type()->singleton())
2088 || n->is_Proj() // parameter or return value
2089 || n->is_Phi() // local of some sort
2090 );
2091 }
2092
2093 //----------------------------inline_min_max-----------------------------------
2094 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2095 set_result(generate_min_max(id, argument(0), argument(1)));
2096 return true;
2097 }
2098
2099 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2100 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2101 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2102 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2103 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2104
2105 {
2106 PreserveJVMState pjvms(this);
2107 PreserveReexecuteState preexecs(this);
2108 jvms()->set_should_reexecute(true);
2109
2110 set_control(slow_path);
2111 set_i_o(i_o());
2112
2113 uncommon_trap(Deoptimization::Reason_intrinsic,
2114 Deoptimization::Action_none);
2115 }
2116
2117 set_control(fast_path);
2118 set_result(math);
2119 }
2120
2121 template <typename OverflowOp>
2122 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2123 typedef typename OverflowOp::MathOp MathOp;
2124
2125 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2126 Node* operation = _gvn.transform( mathOp );
2127 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2128 inline_math_mathExact(operation, ofcheck);
2129 return true;
2130 }
2131
2132 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2133 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2134 }
2135
2136 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2137 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2138 }
2139
2140 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2141 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2142 }
2143
2144 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2145 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2146 }
2147
2148 bool LibraryCallKit::inline_math_negateExactI() {
2149 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2150 }
2151
2152 bool LibraryCallKit::inline_math_negateExactL() {
2153 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2154 }
2155
2156 bool LibraryCallKit::inline_math_multiplyExactI() {
2157 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2158 }
2159
2160 bool LibraryCallKit::inline_math_multiplyExactL() {
2161 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2162 }
2163
2164 Node*
2165 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2166 // These are the candidate return value:
2167 Node* xvalue = x0;
2168 Node* yvalue = y0;
2169
2170 if (xvalue == yvalue) {
2171 return xvalue;
2172 }
2173
2174 bool want_max = (id == vmIntrinsics::_max);
2175
2176 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2177 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2178 if (txvalue == NULL || tyvalue == NULL) return top();
2179 // This is not really necessary, but it is consistent with a
2180 // hypothetical MaxINode::Value method:
2181 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2182
2183 // %%% This folding logic should (ideally) be in a different place.
2184 // Some should be inside IfNode, and there to be a more reliable
2185 // transformation of ?: style patterns into cmoves. We also want
2186 // more powerful optimizations around cmove and min/max.
2187
2188 // Try to find a dominating comparison of these guys.
2189 // It can simplify the index computation for Arrays.copyOf
2190 // and similar uses of System.arraycopy.
2191 // First, compute the normalized version of CmpI(x, y).
2192 int cmp_op = Op_CmpI;
2193 Node* xkey = xvalue;
2194 Node* ykey = yvalue;
2195 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2196 if (ideal_cmpxy->is_Cmp()) {
2197 // E.g., if we have CmpI(length - offset, count),
2198 // it might idealize to CmpI(length, count + offset)
2199 cmp_op = ideal_cmpxy->Opcode();
2200 xkey = ideal_cmpxy->in(1);
2201 ykey = ideal_cmpxy->in(2);
2202 }
2203
2204 // Start by locating any relevant comparisons.
2205 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2206 Node* cmpxy = NULL;
2207 Node* cmpyx = NULL;
2208 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2209 Node* cmp = start_from->fast_out(k);
2210 if (cmp->outcnt() > 0 && // must have prior uses
2211 cmp->in(0) == NULL && // must be context-independent
2212 cmp->Opcode() == cmp_op) { // right kind of compare
2213 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2214 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2215 }
2216 }
2217
2218 const int NCMPS = 2;
2219 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2220 int cmpn;
2221 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2222 if (cmps[cmpn] != NULL) break; // find a result
2223 }
2224 if (cmpn < NCMPS) {
2225 // Look for a dominating test that tells us the min and max.
2226 int depth = 0; // Limit search depth for speed
2227 Node* dom = control();
2228 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2229 if (++depth >= 100) break;
2230 Node* ifproj = dom;
2231 if (!ifproj->is_Proj()) continue;
2232 Node* iff = ifproj->in(0);
2233 if (!iff->is_If()) continue;
2234 Node* bol = iff->in(1);
2235 if (!bol->is_Bool()) continue;
2236 Node* cmp = bol->in(1);
2237 if (cmp == NULL) continue;
2238 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2239 if (cmps[cmpn] == cmp) break;
2240 if (cmpn == NCMPS) continue;
2241 BoolTest::mask btest = bol->as_Bool()->_test._test;
2242 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2243 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2244 // At this point, we know that 'x btest y' is true.
2245 switch (btest) {
2246 case BoolTest::eq:
2247 // They are proven equal, so we can collapse the min/max.
2248 // Either value is the answer. Choose the simpler.
2249 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2250 return yvalue;
2251 return xvalue;
2252 case BoolTest::lt: // x < y
2253 case BoolTest::le: // x <= y
2254 return (want_max ? yvalue : xvalue);
2255 case BoolTest::gt: // x > y
2256 case BoolTest::ge: // x >= y
2257 return (want_max ? xvalue : yvalue);
2258 }
2259 }
2260 }
2261
2262 // We failed to find a dominating test.
2263 // Let's pick a test that might GVN with prior tests.
2264 Node* best_bol = NULL;
2265 BoolTest::mask best_btest = BoolTest::illegal;
2266 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2267 Node* cmp = cmps[cmpn];
2268 if (cmp == NULL) continue;
2269 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2270 Node* bol = cmp->fast_out(j);
2271 if (!bol->is_Bool()) continue;
2272 BoolTest::mask btest = bol->as_Bool()->_test._test;
2273 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2274 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2275 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2276 best_bol = bol->as_Bool();
2277 best_btest = btest;
2278 }
2279 }
2280 }
2281
2282 Node* answer_if_true = NULL;
2283 Node* answer_if_false = NULL;
2284 switch (best_btest) {
2285 default:
2286 if (cmpxy == NULL)
2287 cmpxy = ideal_cmpxy;
2288 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2289 // and fall through:
2290 case BoolTest::lt: // x < y
2291 case BoolTest::le: // x <= y
2292 answer_if_true = (want_max ? yvalue : xvalue);
2293 answer_if_false = (want_max ? xvalue : yvalue);
2294 break;
2295 case BoolTest::gt: // x > y
2296 case BoolTest::ge: // x >= y
2297 answer_if_true = (want_max ? xvalue : yvalue);
2298 answer_if_false = (want_max ? yvalue : xvalue);
2299 break;
2300 }
2301
2302 jint hi, lo;
2303 if (want_max) {
2304 // We can sharpen the minimum.
2305 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2306 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2307 } else {
2308 // We can sharpen the maximum.
2309 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2310 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2311 }
2312
2313 // Use a flow-free graph structure, to avoid creating excess control edges
2314 // which could hinder other optimizations.
2315 // Since Math.min/max is often used with arraycopy, we want
2316 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2317 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2318 answer_if_false, answer_if_true,
2319 TypeInt::make(lo, hi, widen));
2320
2321 return _gvn.transform(cmov);
2322
2323 /*
2324 // This is not as desirable as it may seem, since Min and Max
2325 // nodes do not have a full set of optimizations.
2326 // And they would interfere, anyway, with 'if' optimizations
2327 // and with CMoveI canonical forms.
2328 switch (id) {
2329 case vmIntrinsics::_min:
2330 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2331 case vmIntrinsics::_max:
2332 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2333 default:
2334 ShouldNotReachHere();
2335 }
2336 */
2337 }
2338
2339 inline int
2340 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2341 const TypePtr* base_type = TypePtr::NULL_PTR;
2342 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2343 if (base_type == NULL) {
2344 // Unknown type.
2345 return Type::AnyPtr;
2346 } else if (base_type == TypePtr::NULL_PTR) {
2347 // Since this is a NULL+long form, we have to switch to a rawptr.
2348 base = _gvn.transform(new (C) CastX2PNode(offset));
2349 offset = MakeConX(0);
2350 return Type::RawPtr;
2351 } else if (base_type->base() == Type::RawPtr) {
2352 return Type::RawPtr;
2353 } else if (base_type->isa_oopptr()) {
2354 // Base is never null => always a heap address.
2355 if (base_type->ptr() == TypePtr::NotNull) {
2356 return Type::OopPtr;
2357 }
2358 // Offset is small => always a heap address.
2359 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2360 if (offset_type != NULL &&
2361 base_type->offset() == 0 && // (should always be?)
2362 offset_type->_lo >= 0 &&
2363 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2364 return Type::OopPtr;
2365 }
2366 // Otherwise, it might either be oop+off or NULL+addr.
2367 return Type::AnyPtr;
2368 } else {
2369 // No information:
2370 return Type::AnyPtr;
2371 }
2372 }
2373
2374 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2375 int kind = classify_unsafe_addr(base, offset);
2376 if (kind == Type::RawPtr) {
2377 return basic_plus_adr(top(), base, offset);
2378 } else {
2379 return basic_plus_adr(base, offset);
2380 }
2381 }
2382
2383 //--------------------------inline_number_methods-----------------------------
2384 // inline int Integer.numberOfLeadingZeros(int)
2385 // inline int Long.numberOfLeadingZeros(long)
2386 //
2387 // inline int Integer.numberOfTrailingZeros(int)
2388 // inline int Long.numberOfTrailingZeros(long)
2389 //
2390 // inline int Integer.bitCount(int)
2391 // inline int Long.bitCount(long)
2392 //
2393 // inline char Character.reverseBytes(char)
2394 // inline short Short.reverseBytes(short)
2395 // inline int Integer.reverseBytes(int)
2396 // inline long Long.reverseBytes(long)
2397 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2398 Node* arg = argument(0);
2399 Node* n = NULL;
2400 switch (id) {
2401 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2402 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2403 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2404 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2405 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2406 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2407 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2408 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2409 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2410 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2411 default: fatal_unexpected_iid(id); break;
2412 }
2413 set_result(_gvn.transform(n));
2414 return true;
2415 }
2416
2417 //----------------------------inline_unsafe_access----------------------------
2418
2419 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2420
2421 // Helper that guards and inserts a pre-barrier.
2422 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2423 Node* pre_val, bool need_mem_bar) {
2424 // We could be accessing the referent field of a reference object. If so, when G1
2425 // is enabled, we need to log the value in the referent field in an SATB buffer.
2426 // This routine performs some compile time filters and generates suitable
2427 // runtime filters that guard the pre-barrier code.
2428 // Also add memory barrier for non volatile load from the referent field
2429 // to prevent commoning of loads across safepoint.
2430 if (!(UseG1GC || UseShenandoahGC) && !need_mem_bar)
2431 return;
2432
2433 // Some compile time checks.
2434
2435 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2436 const TypeX* otype = offset->find_intptr_t_type();
2437 if (otype != NULL && otype->is_con() &&
2438 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2439 // Constant offset but not the reference_offset so just return
2440 return;
2441 }
2442
2443 // We only need to generate the runtime guards for instances.
2444 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2445 if (btype != NULL) {
2446 if (btype->isa_aryptr()) {
2447 // Array type so nothing to do
2448 return;
2449 }
2450
2451 const TypeInstPtr* itype = btype->isa_instptr();
2452 if (itype != NULL) {
2453 // Can the klass of base_oop be statically determined to be
2454 // _not_ a sub-class of Reference and _not_ Object?
2455 ciKlass* klass = itype->klass();
2456 if ( klass->is_loaded() &&
2457 !klass->is_subtype_of(env()->Reference_klass()) &&
2458 !env()->Object_klass()->is_subtype_of(klass)) {
2459 return;
2460 }
2461 }
2462 }
2463
2464 // The compile time filters did not reject base_oop/offset so
2465 // we need to generate the following runtime filters
2466 //
2467 // if (offset == java_lang_ref_Reference::_reference_offset) {
2468 // if (instance_of(base, java.lang.ref.Reference)) {
2469 // pre_barrier(_, pre_val, ...);
2470 // }
2471 // }
2472
2473 float likely = PROB_LIKELY( 0.999);
2474 float unlikely = PROB_UNLIKELY(0.999);
2475
2476 IdealKit ideal(this);
2477 #define __ ideal.
2478
2479 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2480
2481 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2482 // Update graphKit memory and control from IdealKit.
2483 sync_kit(ideal);
2484
2485 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2486 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2487
2488 // Update IdealKit memory and control from graphKit.
2489 __ sync_kit(this);
2490
2491 Node* one = __ ConI(1);
2492 // is_instof == 0 if base_oop == NULL
2493 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2494
2495 // Update graphKit from IdeakKit.
2496 sync_kit(ideal);
2497
2498 // Use the pre-barrier to record the value in the referent field
2499 pre_barrier(false /* do_load */,
2500 __ ctrl(),
2501 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2502 pre_val /* pre_val */,
2503 T_OBJECT);
2504 if (need_mem_bar) {
2505 // Add memory barrier to prevent commoning reads from this field
2506 // across safepoint since GC can change its value.
2507 insert_mem_bar(Op_MemBarCPUOrder);
2508 }
2509 // Update IdealKit from graphKit.
2510 __ sync_kit(this);
2511
2512 } __ end_if(); // _ref_type != ref_none
2513 } __ end_if(); // offset == referent_offset
2514
2515 // Final sync IdealKit and GraphKit.
2516 final_sync(ideal);
2517 #undef __
2518 }
2519
2520
2521 // Interpret Unsafe.fieldOffset cookies correctly:
2522 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2523
2524 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2525 // Attempt to infer a sharper value type from the offset and base type.
2526 ciKlass* sharpened_klass = NULL;
2527
2528 // See if it is an instance field, with an object type.
2529 if (alias_type->field() != NULL) {
2530 assert(!is_native_ptr, "native pointer op cannot use a java address");
2531 if (alias_type->field()->type()->is_klass()) {
2532 sharpened_klass = alias_type->field()->type()->as_klass();
2533 }
2534 }
2535
2536 // See if it is a narrow oop array.
2537 if (adr_type->isa_aryptr()) {
2538 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2539 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2540 if (elem_type != NULL) {
2541 sharpened_klass = elem_type->klass();
2542 }
2543 }
2544 }
2545
2546 // The sharpened class might be unloaded if there is no class loader
2547 // contraint in place.
2548 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2549 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2550
2551 #ifndef PRODUCT
2552 if (C->print_intrinsics() || C->print_inlining()) {
2553 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2554 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2555 }
2556 #endif
2557 // Sharpen the value type.
2558 return tjp;
2559 }
2560 return NULL;
2561 }
2562
2563 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2564 if (callee()->is_static()) return false; // caller must have the capability!
2565 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2566
2567 #ifndef PRODUCT
2568 {
2569 ResourceMark rm;
2570 // Check the signatures.
2571 ciSignature* sig = callee()->signature();
2572 #ifdef ASSERT
2573 if (!is_store) {
2574 // Object getObject(Object base, int/long offset), etc.
2575 BasicType rtype = sig->return_type()->basic_type();
2576 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2577 rtype = T_ADDRESS; // it is really a C void*
2578 assert(rtype == type, "getter must return the expected value");
2579 if (!is_native_ptr) {
2580 assert(sig->count() == 2, "oop getter has 2 arguments");
2581 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2582 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2583 } else {
2584 assert(sig->count() == 1, "native getter has 1 argument");
2585 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2586 }
2587 } else {
2588 // void putObject(Object base, int/long offset, Object x), etc.
2589 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2590 if (!is_native_ptr) {
2591 assert(sig->count() == 3, "oop putter has 3 arguments");
2592 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2593 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2594 } else {
2595 assert(sig->count() == 2, "native putter has 2 arguments");
2596 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2597 }
2598 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2599 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2600 vtype = T_ADDRESS; // it is really a C void*
2601 assert(vtype == type, "putter must accept the expected value");
2602 }
2603 #endif // ASSERT
2604 }
2605 #endif //PRODUCT
2606
2607 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2608
2609 Node* receiver = argument(0); // type: oop
2610
2611 // Build address expression. See the code in inline_unsafe_prefetch.
2612 Node* adr;
2613 Node* heap_base_oop = top();
2614 Node* offset = top();
2615 Node* val;
2616
2617 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2618 Node* base = argument(1); // type: oop
2619
2620 if (!is_native_ptr) {
2621 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2622 offset = argument(2); // type: long
2623 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2624 // to be plain byte offsets, which are also the same as those accepted
2625 // by oopDesc::field_base.
2626 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2627 "fieldOffset must be byte-scaled");
2628 // 32-bit machines ignore the high half!
2629 offset = ConvL2X(offset);
2630 adr = make_unsafe_address(base, offset);
2631 heap_base_oop = base;
2632 val = is_store ? argument(4) : NULL;
2633 } else {
2634 Node* ptr = argument(1); // type: long
2635 ptr = ConvL2X(ptr); // adjust Java long to machine word
2636 adr = make_unsafe_address(NULL, ptr);
2637 val = is_store ? argument(3) : NULL;
2638 }
2639
2640 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) {
2641 return false; // off-heap oop accesses are not supported
2642 }
2643
2644 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2645
2646 // Try to categorize the address.
2647 Compile::AliasType* alias_type = C->alias_type(adr_type);
2648 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2649
2650 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2651 alias_type->adr_type() == TypeAryPtr::RANGE) {
2652 return false; // not supported
2653 }
2654
2655 bool mismatched = false;
2656 BasicType bt = alias_type->basic_type();
2657 if (bt != T_ILLEGAL) {
2658 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2659 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2660 // Alias type doesn't differentiate between byte[] and boolean[]).
2661 // Use address type to get the element type.
2662 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2663 }
2664 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2665 // accessing an array field with getObject is not a mismatch
2666 bt = T_OBJECT;
2667 }
2668 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2669 // Don't intrinsify mismatched object accesses
2670 return false;
2671 }
2672 mismatched = (bt != type);
2673 }
2674
2675 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2676
2677 // First guess at the value type.
2678 const Type *value_type = Type::get_const_basic_type(type);
2679
2680 // We will need memory barriers unless we can determine a unique
2681 // alias category for this reference. (Note: If for some reason
2682 // the barriers get omitted and the unsafe reference begins to "pollute"
2683 // the alias analysis of the rest of the graph, either Compile::can_alias
2684 // or Compile::must_alias will throw a diagnostic assert.)
2685 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2686
2687 #if INCLUDE_ALL_GCS
2688 // Work around JDK-8220714 bug. This is done for Shenandoah only, until
2689 // the shared code fix is upstreamed and properly tested there.
2690 if (UseShenandoahGC) {
2691 need_mem_bar |= is_native_ptr;
2692 }
2693 #endif
2694
2695 // If we are reading the value of the referent field of a Reference
2696 // object (either by using Unsafe directly or through reflection)
2697 // then, if G1 is enabled, we need to record the referent in an
2698 // SATB log buffer using the pre-barrier mechanism.
2699 // Also we need to add memory barrier to prevent commoning reads
2700 // from this field across safepoint since GC can change its value.
2701 bool need_read_barrier = !is_native_ptr && !is_store &&
2702 offset != top() && heap_base_oop != top();
2703
2704 if (!is_store && type == T_OBJECT) {
2705 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2706 if (tjp != NULL) {
2707 value_type = tjp;
2708 }
2709 }
2710
2711 receiver = null_check(receiver);
2712 if (stopped()) {
2713 return true;
2714 }
2715 // Heap pointers get a null-check from the interpreter,
2716 // as a courtesy. However, this is not guaranteed by Unsafe,
2717 // and it is not possible to fully distinguish unintended nulls
2718 // from intended ones in this API.
2719
2720 Node* load = NULL;
2721 Node* store = NULL;
2722 Node* leading_membar = NULL;
2723 if (is_volatile) {
2724 // We need to emit leading and trailing CPU membars (see below) in
2725 // addition to memory membars when is_volatile. This is a little
2726 // too strong, but avoids the need to insert per-alias-type
2727 // volatile membars (for stores; compare Parse::do_put_xxx), which
2728 // we cannot do effectively here because we probably only have a
2729 // rough approximation of type.
2730 need_mem_bar = true;
2731 // For Stores, place a memory ordering barrier now.
2732 if (is_store) {
2733 leading_membar = insert_mem_bar(Op_MemBarRelease);
2734 } else {
2735 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2736 leading_membar = insert_mem_bar(Op_MemBarVolatile);
2737 }
2738 }
2739 }
2740
2741 // Memory barrier to prevent normal and 'unsafe' accesses from
2742 // bypassing each other. Happens after null checks, so the
2743 // exception paths do not take memory state from the memory barrier,
2744 // so there's no problems making a strong assert about mixing users
2745 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2746 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2747 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2748
2749 if (!is_store) {
2750 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2751 // To be valid, unsafe loads may depend on other conditions than
2752 // the one that guards them: pin the Load node
2753 load = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2754 #if INCLUDE_ALL_GCS
2755 if (UseShenandoahGC && (type == T_OBJECT || type == T_ARRAY)) {
2756 load = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load);
2757 }
2758 #endif
2759 // load value
2760 switch (type) {
2761 case T_BOOLEAN:
2762 case T_CHAR:
2763 case T_BYTE:
2764 case T_SHORT:
2765 case T_INT:
2766 case T_LONG:
2767 case T_FLOAT:
2768 case T_DOUBLE:
2769 break;
2770 case T_OBJECT:
2771 if (need_read_barrier) {
2772 insert_pre_barrier(heap_base_oop, offset, load, !(is_volatile || need_mem_bar));
2773 }
2774 break;
2775 case T_ADDRESS:
2776 // Cast to an int type.
2777 load = _gvn.transform(new (C) CastP2XNode(NULL, load));
2778 load = ConvX2UL(load);
2779 break;
2780 default:
2781 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2782 break;
2783 }
2784 // The load node has the control of the preceding MemBarCPUOrder. All
2785 // following nodes will have the control of the MemBarCPUOrder inserted at
2786 // the end of this method. So, pushing the load onto the stack at a later
2787 // point is fine.
2788 set_result(load);
2789 } else {
2790 // place effect of store into memory
2791 switch (type) {
2792 case T_DOUBLE:
2793 val = dstore_rounding(val);
2794 break;
2795 case T_ADDRESS:
2796 // Repackage the long as a pointer.
2797 val = ConvL2X(val);
2798 val = _gvn.transform(new (C) CastX2PNode(val));
2799 break;
2800 }
2801
2802 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2803 if (type == T_OBJECT ) {
2804 store = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2805 } else {
2806 store = store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2807 }
2808 }
2809
2810 if (is_volatile) {
2811 if (!is_store) {
2812 Node* mb = insert_mem_bar(Op_MemBarAcquire, load);
2813 mb->as_MemBar()->set_trailing_load();
2814 } else {
2815 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2816 Node* mb = insert_mem_bar(Op_MemBarVolatile, store);
2817 MemBarNode::set_store_pair(leading_membar->as_MemBar(), mb->as_MemBar());
2818 }
2819 }
2820 }
2821
2822 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2823
2824 return true;
2825 }
2826
2827 //----------------------------inline_unsafe_prefetch----------------------------
2828
2829 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2830 #ifndef PRODUCT
2831 {
2832 ResourceMark rm;
2833 // Check the signatures.
2834 ciSignature* sig = callee()->signature();
2835 #ifdef ASSERT
2836 // Object getObject(Object base, int/long offset), etc.
2837 BasicType rtype = sig->return_type()->basic_type();
2838 if (!is_native_ptr) {
2839 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2840 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2841 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2842 } else {
2843 assert(sig->count() == 1, "native prefetch has 1 argument");
2844 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2845 }
2846 #endif // ASSERT
2847 }
2848 #endif // !PRODUCT
2849
2850 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2851
2852 const int idx = is_static ? 0 : 1;
2853 if (!is_static) {
2854 null_check_receiver();
2855 if (stopped()) {
2856 return true;
2857 }
2858 }
2859
2860 // Build address expression. See the code in inline_unsafe_access.
2861 Node *adr;
2862 if (!is_native_ptr) {
2863 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2864 Node* base = argument(idx + 0); // type: oop
2865 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2866 Node* offset = argument(idx + 1); // type: long
2867 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2868 // to be plain byte offsets, which are also the same as those accepted
2869 // by oopDesc::field_base.
2870 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2871 "fieldOffset must be byte-scaled");
2872 // 32-bit machines ignore the high half!
2873 offset = ConvL2X(offset);
2874 adr = make_unsafe_address(base, offset);
2875 } else {
2876 Node* ptr = argument(idx + 0); // type: long
2877 ptr = ConvL2X(ptr); // adjust Java long to machine word
2878 adr = make_unsafe_address(NULL, ptr);
2879 }
2880
2881 // Generate the read or write prefetch
2882 Node *prefetch;
2883 if (is_store) {
2884 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2885 } else {
2886 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2887 }
2888 prefetch->init_req(0, control());
2889 set_i_o(_gvn.transform(prefetch));
2890
2891 return true;
2892 }
2893
2894 //----------------------------inline_unsafe_load_store----------------------------
2895 // This method serves a couple of different customers (depending on LoadStoreKind):
2896 //
2897 // LS_cmpxchg:
2898 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2899 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2900 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2901 //
2902 // LS_xadd:
2903 // public int getAndAddInt( Object o, long offset, int delta)
2904 // public long getAndAddLong(Object o, long offset, long delta)
2905 //
2906 // LS_xchg:
2907 // int getAndSet(Object o, long offset, int newValue)
2908 // long getAndSet(Object o, long offset, long newValue)
2909 // Object getAndSet(Object o, long offset, Object newValue)
2910 //
2911 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2912 // This basic scheme here is the same as inline_unsafe_access, but
2913 // differs in enough details that combining them would make the code
2914 // overly confusing. (This is a true fact! I originally combined
2915 // them, but even I was confused by it!) As much code/comments as
2916 // possible are retained from inline_unsafe_access though to make
2917 // the correspondences clearer. - dl
2918
2919 if (callee()->is_static()) return false; // caller must have the capability!
2920
2921 #ifndef PRODUCT
2922 BasicType rtype;
2923 {
2924 ResourceMark rm;
2925 // Check the signatures.
2926 ciSignature* sig = callee()->signature();
2927 rtype = sig->return_type()->basic_type();
2928 if (kind == LS_xadd || kind == LS_xchg) {
2929 // Check the signatures.
2930 #ifdef ASSERT
2931 assert(rtype == type, "get and set must return the expected type");
2932 assert(sig->count() == 3, "get and set has 3 arguments");
2933 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2934 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2935 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2936 #endif // ASSERT
2937 } else if (kind == LS_cmpxchg) {
2938 // Check the signatures.
2939 #ifdef ASSERT
2940 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2941 assert(sig->count() == 4, "CAS has 4 arguments");
2942 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2943 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2944 #endif // ASSERT
2945 } else {
2946 ShouldNotReachHere();
2947 }
2948 }
2949 #endif //PRODUCT
2950
2951 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2952
2953 // Get arguments:
2954 Node* receiver = NULL;
2955 Node* base = NULL;
2956 Node* offset = NULL;
2957 Node* oldval = NULL;
2958 Node* newval = NULL;
2959 if (kind == LS_cmpxchg) {
2960 const bool two_slot_type = type2size[type] == 2;
2961 receiver = argument(0); // type: oop
2962 base = argument(1); // type: oop
2963 offset = argument(2); // type: long
2964 oldval = argument(4); // type: oop, int, or long
2965 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2966 } else if (kind == LS_xadd || kind == LS_xchg){
2967 receiver = argument(0); // type: oop
2968 base = argument(1); // type: oop
2969 offset = argument(2); // type: long
2970 oldval = NULL;
2971 newval = argument(4); // type: oop, int, or long
2972 }
2973
2974 // Build field offset expression.
2975 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2976 // to be plain byte offsets, which are also the same as those accepted
2977 // by oopDesc::field_base.
2978 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2979 // 32-bit machines ignore the high half of long offsets
2980 offset = ConvL2X(offset);
2981 Node* adr = make_unsafe_address(base, offset);
2982 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2983
2984 Compile::AliasType* alias_type = C->alias_type(adr_type);
2985 BasicType bt = alias_type->basic_type();
2986 if (bt != T_ILLEGAL &&
2987 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2988 // Don't intrinsify mismatched object accesses.
2989 return false;
2990 }
2991
2992 // For CAS, unlike inline_unsafe_access, there seems no point in
2993 // trying to refine types. Just use the coarse types here.
2994 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2995 const Type *value_type = Type::get_const_basic_type(type);
2996
2997 if (kind == LS_xchg && type == T_OBJECT) {
2998 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2999 if (tjp != NULL) {
3000 value_type = tjp;
3001 }
3002 }
3003
3004 // Null check receiver.
3005 receiver = null_check(receiver);
3006 if (stopped()) {
3007 return true;
3008 }
3009
3010 int alias_idx = C->get_alias_index(adr_type);
3011
3012 // Memory-model-wise, a LoadStore acts like a little synchronized
3013 // block, so needs barriers on each side. These don't translate
3014 // into actual barriers on most machines, but we still need rest of
3015 // compiler to respect ordering.
3016
3017 Node* leading_membar = insert_mem_bar(Op_MemBarRelease);
3018 insert_mem_bar(Op_MemBarCPUOrder);
3019
3020 // 4984716: MemBars must be inserted before this
3021 // memory node in order to avoid a false
3022 // dependency which will confuse the scheduler.
3023 Node *mem = memory(alias_idx);
3024
3025 // For now, we handle only those cases that actually exist: ints,
3026 // longs, and Object. Adding others should be straightforward.
3027 Node* load_store = NULL;
3028 switch(type) {
3029 case T_INT:
3030 if (kind == LS_xadd) {
3031 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3032 } else if (kind == LS_xchg) {
3033 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3034 } else if (kind == LS_cmpxchg) {
3035 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3036 } else {
3037 ShouldNotReachHere();
3038 }
3039 break;
3040 case T_LONG:
3041 if (kind == LS_xadd) {
3042 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3043 } else if (kind == LS_xchg) {
3044 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3045 } else if (kind == LS_cmpxchg) {
3046 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3047 } else {
3048 ShouldNotReachHere();
3049 }
3050 break;
3051 case T_OBJECT:
3052 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3053 // could be delayed during Parse (for example, in adjust_map_after_if()).
3054 // Execute transformation here to avoid barrier generation in such case.
3055 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3056 newval = _gvn.makecon(TypePtr::NULL_PTR);
3057
3058 // Reference stores need a store barrier.
3059 if (kind == LS_xchg) {
3060 // If pre-barrier must execute before the oop store, old value will require do_load here.
3061 if (!can_move_pre_barrier()) {
3062 pre_barrier(true /* do_load*/,
3063 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3064 NULL /* pre_val*/,
3065 T_OBJECT);
3066 } // Else move pre_barrier to use load_store value, see below.
3067 } else if (kind == LS_cmpxchg) {
3068 // Same as for newval above:
3069 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3070 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3071 }
3072 // The only known value which might get overwritten is oldval.
3073 pre_barrier(false /* do_load */,
3074 control(), NULL, NULL, max_juint, NULL, NULL,
3075 oldval /* pre_val */,
3076 T_OBJECT);
3077 } else {
3078 ShouldNotReachHere();
3079 }
3080
3081 #ifdef _LP64
3082 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3083 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3084 if (kind == LS_xchg) {
3085 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3086 newval_enc, adr_type, value_type->make_narrowoop()));
3087 } else {
3088 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3089 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3090 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3091 newval_enc, oldval_enc));
3092 }
3093 } else
3094 #endif
3095 {
3096 if (kind == LS_xchg) {
3097 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3098 } else {
3099 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3100 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3101 }
3102 }
3103 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3104 break;
3105 default:
3106 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3107 break;
3108 }
3109
3110 // SCMemProjNodes represent the memory state of a LoadStore. Their
3111 // main role is to prevent LoadStore nodes from being optimized away
3112 // when their results aren't used.
3113 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3114 set_memory(proj, alias_idx);
3115
3116 Node* access = load_store;
3117
3118 if (type == T_OBJECT && kind == LS_xchg) {
3119 #ifdef _LP64
3120 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3121 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3122 }
3123 #endif
3124 #if INCLUDE_ALL_GCS
3125 if (UseShenandoahGC) {
3126 load_store = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load_store);
3127 }
3128 #endif
3129 if (can_move_pre_barrier()) {
3130 // Don't need to load pre_val. The old value is returned by load_store.
3131 // The pre_barrier can execute after the xchg as long as no safepoint
3132 // gets inserted between them.
3133 pre_barrier(false /* do_load */,
3134 control(), NULL, NULL, max_juint, NULL, NULL,
3135 load_store /* pre_val */,
3136 T_OBJECT);
3137 }
3138 }
3139
3140 // Add the trailing membar surrounding the access
3141 insert_mem_bar(Op_MemBarCPUOrder);
3142 Node* mb = insert_mem_bar(Op_MemBarAcquire, access);
3143 MemBarNode::set_load_store_pair(leading_membar->as_MemBar(), mb->as_MemBar());
3144
3145 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3146 set_result(load_store);
3147 return true;
3148 }
3149
3150 //----------------------------inline_unsafe_ordered_store----------------------
3151 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3152 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3153 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3154 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3155 // This is another variant of inline_unsafe_access, differing in
3156 // that it always issues store-store ("release") barrier and ensures
3157 // store-atomicity (which only matters for "long").
3158
3159 if (callee()->is_static()) return false; // caller must have the capability!
3160
3161 #ifndef PRODUCT
3162 {
3163 ResourceMark rm;
3164 // Check the signatures.
3165 ciSignature* sig = callee()->signature();
3166 #ifdef ASSERT
3167 BasicType rtype = sig->return_type()->basic_type();
3168 assert(rtype == T_VOID, "must return void");
3169 assert(sig->count() == 3, "has 3 arguments");
3170 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3171 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3172 #endif // ASSERT
3173 }
3174 #endif //PRODUCT
3175
3176 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3177
3178 // Get arguments:
3179 Node* receiver = argument(0); // type: oop
3180 Node* base = argument(1); // type: oop
3181 Node* offset = argument(2); // type: long
3182 Node* val = argument(4); // type: oop, int, or long
3183
3184 // Null check receiver.
3185 receiver = null_check(receiver);
3186 if (stopped()) {
3187 return true;
3188 }
3189
3190 // Build field offset expression.
3191 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3192 // 32-bit machines ignore the high half of long offsets
3193 offset = ConvL2X(offset);
3194 Node* adr = make_unsafe_address(base, offset);
3195 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3196 const Type *value_type = Type::get_const_basic_type(type);
3197 Compile::AliasType* alias_type = C->alias_type(adr_type);
3198
3199 insert_mem_bar(Op_MemBarRelease);
3200 insert_mem_bar(Op_MemBarCPUOrder);
3201 // Ensure that the store is atomic for longs:
3202 const bool require_atomic_access = true;
3203 Node* store;
3204 if (type == T_OBJECT) // reference stores need a store barrier.
3205 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3206 else {
3207 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3208 }
3209 insert_mem_bar(Op_MemBarCPUOrder);
3210 return true;
3211 }
3212
3213 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3214 // Regardless of form, don't allow previous ld/st to move down,
3215 // then issue acquire, release, or volatile mem_bar.
3216 insert_mem_bar(Op_MemBarCPUOrder);
3217 switch(id) {
3218 case vmIntrinsics::_loadFence:
3219 insert_mem_bar(Op_LoadFence);
3220 return true;
3221 case vmIntrinsics::_storeFence:
3222 insert_mem_bar(Op_StoreFence);
3223 return true;
3224 case vmIntrinsics::_fullFence:
3225 insert_mem_bar(Op_MemBarVolatile);
3226 return true;
3227 default:
3228 fatal_unexpected_iid(id);
3229 return false;
3230 }
3231 }
3232
3233 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3234 if (!kls->is_Con()) {
3235 return true;
3236 }
3237 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3238 if (klsptr == NULL) {
3239 return true;
3240 }
3241 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3242 // don't need a guard for a klass that is already initialized
3243 return !ik->is_initialized();
3244 }
3245
3246 //----------------------------inline_unsafe_allocate---------------------------
3247 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3248 bool LibraryCallKit::inline_unsafe_allocate() {
3249 if (callee()->is_static()) return false; // caller must have the capability!
3250
3251 null_check_receiver(); // null-check, then ignore
3252 Node* cls = null_check(argument(1));
3253 if (stopped()) return true;
3254
3255 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3256 kls = null_check(kls);
3257 if (stopped()) return true; // argument was like int.class
3258
3259 Node* test = NULL;
3260 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3261 // Note: The argument might still be an illegal value like
3262 // Serializable.class or Object[].class. The runtime will handle it.
3263 // But we must make an explicit check for initialization.
3264 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3265 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3266 // can generate code to load it as unsigned byte.
3267 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3268 Node* bits = intcon(InstanceKlass::fully_initialized);
3269 test = _gvn.transform(new (C) SubINode(inst, bits));
3270 // The 'test' is non-zero if we need to take a slow path.
3271 }
3272
3273 Node* obj = new_instance(kls, test);
3274 set_result(obj);
3275 return true;
3276 }
3277
3278 #ifdef TRACE_HAVE_INTRINSICS
3279 /*
3280 * oop -> myklass
3281 * myklass->trace_id |= USED
3282 * return myklass->trace_id & ~0x3
3283 */
3284 bool LibraryCallKit::inline_native_classID() {
3285 null_check_receiver(); // null-check, then ignore
3286 Node* cls = null_check(argument(1), T_OBJECT);
3287 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3288 kls = null_check(kls, T_OBJECT);
3289 ByteSize offset = TRACE_ID_OFFSET;
3290 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3291 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3292 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3293 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3294 Node* clsused = longcon(0x01l); // set the class bit
3295 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3296
3297 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3298 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3299 set_result(andl);
3300 return true;
3301 }
3302
3303 bool LibraryCallKit::inline_native_threadID() {
3304 Node* tls_ptr = NULL;
3305 Node* cur_thr = generate_current_thread(tls_ptr);
3306 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3307 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3308 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3309
3310 Node* threadid = NULL;
3311 size_t thread_id_size = OSThread::thread_id_size();
3312 if (thread_id_size == (size_t) BytesPerLong) {
3313 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3314 } else if (thread_id_size == (size_t) BytesPerInt) {
3315 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3316 } else {
3317 ShouldNotReachHere();
3318 }
3319 set_result(threadid);
3320 return true;
3321 }
3322 #endif
3323
3324 //------------------------inline_native_time_funcs--------------
3325 // inline code for System.currentTimeMillis() and System.nanoTime()
3326 // these have the same type and signature
3327 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3328 const TypeFunc* tf = OptoRuntime::void_long_Type();
3329 const TypePtr* no_memory_effects = NULL;
3330 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3331 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3332 #ifdef ASSERT
3333 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3334 assert(value_top == top(), "second value must be top");
3335 #endif
3336 set_result(value);
3337 return true;
3338 }
3339
3340 //------------------------inline_native_currentThread------------------
3341 bool LibraryCallKit::inline_native_currentThread() {
3342 Node* junk = NULL;
3343 set_result(generate_current_thread(junk));
3344 return true;
3345 }
3346
3347 //------------------------inline_native_isInterrupted------------------
3348 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3349 bool LibraryCallKit::inline_native_isInterrupted() {
3350 // Add a fast path to t.isInterrupted(clear_int):
3351 // (t == Thread.current() &&
3352 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3353 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3354 // So, in the common case that the interrupt bit is false,
3355 // we avoid making a call into the VM. Even if the interrupt bit
3356 // is true, if the clear_int argument is false, we avoid the VM call.
3357 // However, if the receiver is not currentThread, we must call the VM,
3358 // because there must be some locking done around the operation.
3359
3360 // We only go to the fast case code if we pass two guards.
3361 // Paths which do not pass are accumulated in the slow_region.
3362
3363 enum {
3364 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3365 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3366 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3367 PATH_LIMIT
3368 };
3369
3370 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3371 // out of the function.
3372 insert_mem_bar(Op_MemBarCPUOrder);
3373
3374 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3375 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3376
3377 RegionNode* slow_region = new (C) RegionNode(1);
3378 record_for_igvn(slow_region);
3379
3380 // (a) Receiving thread must be the current thread.
3381 Node* rec_thr = argument(0);
3382 Node* tls_ptr = NULL;
3383 Node* cur_thr = generate_current_thread(tls_ptr);
3384 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3385 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3386
3387 generate_slow_guard(bol_thr, slow_region);
3388
3389 // (b) Interrupt bit on TLS must be false.
3390 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3391 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3392 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3393
3394 // Set the control input on the field _interrupted read to prevent it floating up.
3395 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3396 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3397 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3398
3399 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3400
3401 // First fast path: if (!TLS._interrupted) return false;
3402 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3403 result_rgn->init_req(no_int_result_path, false_bit);
3404 result_val->init_req(no_int_result_path, intcon(0));
3405
3406 // drop through to next case
3407 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3408
3409 #ifndef TARGET_OS_FAMILY_windows
3410 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3411 Node* clr_arg = argument(1);
3412 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3413 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3414 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3415
3416 // Second fast path: ... else if (!clear_int) return true;
3417 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3418 result_rgn->init_req(no_clear_result_path, false_arg);
3419 result_val->init_req(no_clear_result_path, intcon(1));
3420
3421 // drop through to next case
3422 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3423 #else
3424 // To return true on Windows you must read the _interrupted field
3425 // and check the the event state i.e. take the slow path.
3426 #endif // TARGET_OS_FAMILY_windows
3427
3428 // (d) Otherwise, go to the slow path.
3429 slow_region->add_req(control());
3430 set_control( _gvn.transform(slow_region));
3431
3432 if (stopped()) {
3433 // There is no slow path.
3434 result_rgn->init_req(slow_result_path, top());
3435 result_val->init_req(slow_result_path, top());
3436 } else {
3437 // non-virtual because it is a private non-static
3438 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3439
3440 Node* slow_val = set_results_for_java_call(slow_call);
3441 // this->control() comes from set_results_for_java_call
3442
3443 Node* fast_io = slow_call->in(TypeFunc::I_O);
3444 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3445
3446 // These two phis are pre-filled with copies of of the fast IO and Memory
3447 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3448 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3449
3450 result_rgn->init_req(slow_result_path, control());
3451 result_io ->init_req(slow_result_path, i_o());
3452 result_mem->init_req(slow_result_path, reset_memory());
3453 result_val->init_req(slow_result_path, slow_val);
3454
3455 set_all_memory(_gvn.transform(result_mem));
3456 set_i_o( _gvn.transform(result_io));
3457 }
3458
3459 C->set_has_split_ifs(true); // Has chance for split-if optimization
3460 set_result(result_rgn, result_val);
3461 return true;
3462 }
3463
3464 //---------------------------load_mirror_from_klass----------------------------
3465 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3466 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3467 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3468 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3469 }
3470
3471 //-----------------------load_klass_from_mirror_common-------------------------
3472 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3473 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3474 // and branch to the given path on the region.
3475 // If never_see_null, take an uncommon trap on null, so we can optimistically
3476 // compile for the non-null case.
3477 // If the region is NULL, force never_see_null = true.
3478 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3479 bool never_see_null,
3480 RegionNode* region,
3481 int null_path,
3482 int offset) {
3483 if (region == NULL) never_see_null = true;
3484 Node* p = basic_plus_adr(mirror, offset);
3485 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3486 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3487 Node* null_ctl = top();
3488 kls = null_check_oop(kls, &null_ctl, never_see_null);
3489 if (region != NULL) {
3490 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3491 region->init_req(null_path, null_ctl);
3492 } else {
3493 assert(null_ctl == top(), "no loose ends");
3494 }
3495 return kls;
3496 }
3497
3498 //--------------------(inline_native_Class_query helpers)---------------------
3499 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3500 // Fall through if (mods & mask) == bits, take the guard otherwise.
3501 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3502 // Branch around if the given klass has the given modifier bit set.
3503 // Like generate_guard, adds a new path onto the region.
3504 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3505 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3506 Node* mask = intcon(modifier_mask);
3507 Node* bits = intcon(modifier_bits);
3508 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3509 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3510 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3511 return generate_fair_guard(bol, region);
3512 }
3513 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3514 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3515 }
3516
3517 //-------------------------inline_native_Class_query-------------------
3518 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3519 const Type* return_type = TypeInt::BOOL;
3520 Node* prim_return_value = top(); // what happens if it's a primitive class?
3521 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3522 bool expect_prim = false; // most of these guys expect to work on refs
3523
3524 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3525
3526 Node* mirror = argument(0);
3527 Node* obj = top();
3528
3529 switch (id) {
3530 case vmIntrinsics::_isInstance:
3531 // nothing is an instance of a primitive type
3532 prim_return_value = intcon(0);
3533 obj = argument(1);
3534 break;
3535 case vmIntrinsics::_getModifiers:
3536 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3537 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3538 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3539 break;
3540 case vmIntrinsics::_isInterface:
3541 prim_return_value = intcon(0);
3542 break;
3543 case vmIntrinsics::_isArray:
3544 prim_return_value = intcon(0);
3545 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3546 break;
3547 case vmIntrinsics::_isPrimitive:
3548 prim_return_value = intcon(1);
3549 expect_prim = true; // obviously
3550 break;
3551 case vmIntrinsics::_getSuperclass:
3552 prim_return_value = null();
3553 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3554 break;
3555 case vmIntrinsics::_getComponentType:
3556 prim_return_value = null();
3557 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3558 break;
3559 case vmIntrinsics::_getClassAccessFlags:
3560 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3561 return_type = TypeInt::INT; // not bool! 6297094
3562 break;
3563 default:
3564 fatal_unexpected_iid(id);
3565 break;
3566 }
3567
3568 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3569 if (mirror_con == NULL) return false; // cannot happen?
3570
3571 #ifndef PRODUCT
3572 if (C->print_intrinsics() || C->print_inlining()) {
3573 ciType* k = mirror_con->java_mirror_type();
3574 if (k) {
3575 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3576 k->print_name();
3577 tty->cr();
3578 }
3579 }
3580 #endif
3581
3582 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3583 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3584 record_for_igvn(region);
3585 PhiNode* phi = new (C) PhiNode(region, return_type);
3586
3587 // The mirror will never be null of Reflection.getClassAccessFlags, however
3588 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3589 // if it is. See bug 4774291.
3590
3591 // For Reflection.getClassAccessFlags(), the null check occurs in
3592 // the wrong place; see inline_unsafe_access(), above, for a similar
3593 // situation.
3594 mirror = null_check(mirror);
3595 // If mirror or obj is dead, only null-path is taken.
3596 if (stopped()) return true;
3597
3598 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3599
3600 // Now load the mirror's klass metaobject, and null-check it.
3601 // Side-effects region with the control path if the klass is null.
3602 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3603 // If kls is null, we have a primitive mirror.
3604 phi->init_req(_prim_path, prim_return_value);
3605 if (stopped()) { set_result(region, phi); return true; }
3606 bool safe_for_replace = (region->in(_prim_path) == top());
3607
3608 Node* p; // handy temp
3609 Node* null_ctl;
3610
3611 // Now that we have the non-null klass, we can perform the real query.
3612 // For constant classes, the query will constant-fold in LoadNode::Value.
3613 Node* query_value = top();
3614 switch (id) {
3615 case vmIntrinsics::_isInstance:
3616 // nothing is an instance of a primitive type
3617 query_value = gen_instanceof(obj, kls, safe_for_replace);
3618 break;
3619
3620 case vmIntrinsics::_getModifiers:
3621 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3622 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3623 break;
3624
3625 case vmIntrinsics::_isInterface:
3626 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3627 if (generate_interface_guard(kls, region) != NULL)
3628 // A guard was added. If the guard is taken, it was an interface.
3629 phi->add_req(intcon(1));
3630 // If we fall through, it's a plain class.
3631 query_value = intcon(0);
3632 break;
3633
3634 case vmIntrinsics::_isArray:
3635 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3636 if (generate_array_guard(kls, region) != NULL)
3637 // A guard was added. If the guard is taken, it was an array.
3638 phi->add_req(intcon(1));
3639 // If we fall through, it's a plain class.
3640 query_value = intcon(0);
3641 break;
3642
3643 case vmIntrinsics::_isPrimitive:
3644 query_value = intcon(0); // "normal" path produces false
3645 break;
3646
3647 case vmIntrinsics::_getSuperclass:
3648 // The rules here are somewhat unfortunate, but we can still do better
3649 // with random logic than with a JNI call.
3650 // Interfaces store null or Object as _super, but must report null.
3651 // Arrays store an intermediate super as _super, but must report Object.
3652 // Other types can report the actual _super.
3653 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3654 if (generate_interface_guard(kls, region) != NULL)
3655 // A guard was added. If the guard is taken, it was an interface.
3656 phi->add_req(null());
3657 if (generate_array_guard(kls, region) != NULL)
3658 // A guard was added. If the guard is taken, it was an array.
3659 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3660 // If we fall through, it's a plain class. Get its _super.
3661 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3662 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3663 null_ctl = top();
3664 kls = null_check_oop(kls, &null_ctl);
3665 if (null_ctl != top()) {
3666 // If the guard is taken, Object.superClass is null (both klass and mirror).
3667 region->add_req(null_ctl);
3668 phi ->add_req(null());
3669 }
3670 if (!stopped()) {
3671 query_value = load_mirror_from_klass(kls);
3672 }
3673 break;
3674
3675 case vmIntrinsics::_getComponentType:
3676 if (generate_array_guard(kls, region) != NULL) {
3677 // Be sure to pin the oop load to the guard edge just created:
3678 Node* is_array_ctrl = region->in(region->req()-1);
3679 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3680 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3681 phi->add_req(cmo);
3682 }
3683 query_value = null(); // non-array case is null
3684 break;
3685
3686 case vmIntrinsics::_getClassAccessFlags:
3687 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3688 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3689 break;
3690
3691 default:
3692 fatal_unexpected_iid(id);
3693 break;
3694 }
3695
3696 // Fall-through is the normal case of a query to a real class.
3697 phi->init_req(1, query_value);
3698 region->init_req(1, control());
3699
3700 C->set_has_split_ifs(true); // Has chance for split-if optimization
3701 set_result(region, phi);
3702 return true;
3703 }
3704
3705 //--------------------------inline_native_subtype_check------------------------
3706 // This intrinsic takes the JNI calls out of the heart of
3707 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3708 bool LibraryCallKit::inline_native_subtype_check() {
3709 // Pull both arguments off the stack.
3710 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3711 args[0] = argument(0);
3712 args[1] = argument(1);
3713 Node* klasses[2]; // corresponding Klasses: superk, subk
3714 klasses[0] = klasses[1] = top();
3715
3716 enum {
3717 // A full decision tree on {superc is prim, subc is prim}:
3718 _prim_0_path = 1, // {P,N} => false
3719 // {P,P} & superc!=subc => false
3720 _prim_same_path, // {P,P} & superc==subc => true
3721 _prim_1_path, // {N,P} => false
3722 _ref_subtype_path, // {N,N} & subtype check wins => true
3723 _both_ref_path, // {N,N} & subtype check loses => false
3724 PATH_LIMIT
3725 };
3726
3727 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3728 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3729 record_for_igvn(region);
3730
3731 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3732 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3733 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3734
3735 // First null-check both mirrors and load each mirror's klass metaobject.
3736 int which_arg;
3737 for (which_arg = 0; which_arg <= 1; which_arg++) {
3738 Node* arg = args[which_arg];
3739 arg = null_check(arg);
3740 if (stopped()) break;
3741 args[which_arg] = arg;
3742
3743 Node* p = basic_plus_adr(arg, class_klass_offset);
3744 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3745 klasses[which_arg] = _gvn.transform(kls);
3746 }
3747
3748 // Having loaded both klasses, test each for null.
3749 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3750 for (which_arg = 0; which_arg <= 1; which_arg++) {
3751 Node* kls = klasses[which_arg];
3752 Node* null_ctl = top();
3753 kls = null_check_oop(kls, &null_ctl, never_see_null);
3754 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3755 region->init_req(prim_path, null_ctl);
3756 if (stopped()) break;
3757 klasses[which_arg] = kls;
3758 }
3759
3760 if (!stopped()) {
3761 // now we have two reference types, in klasses[0..1]
3762 Node* subk = klasses[1]; // the argument to isAssignableFrom
3763 Node* superk = klasses[0]; // the receiver
3764 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3765 // now we have a successful reference subtype check
3766 region->set_req(_ref_subtype_path, control());
3767 }
3768
3769 // If both operands are primitive (both klasses null), then
3770 // we must return true when they are identical primitives.
3771 // It is convenient to test this after the first null klass check.
3772 set_control(region->in(_prim_0_path)); // go back to first null check
3773 if (!stopped()) {
3774 // Since superc is primitive, make a guard for the superc==subc case.
3775 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3776 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3777 generate_guard(bol_eq, region, PROB_FAIR);
3778 if (region->req() == PATH_LIMIT+1) {
3779 // A guard was added. If the added guard is taken, superc==subc.
3780 region->swap_edges(PATH_LIMIT, _prim_same_path);
3781 region->del_req(PATH_LIMIT);
3782 }
3783 region->set_req(_prim_0_path, control()); // Not equal after all.
3784 }
3785
3786 // these are the only paths that produce 'true':
3787 phi->set_req(_prim_same_path, intcon(1));
3788 phi->set_req(_ref_subtype_path, intcon(1));
3789
3790 // pull together the cases:
3791 assert(region->req() == PATH_LIMIT, "sane region");
3792 for (uint i = 1; i < region->req(); i++) {
3793 Node* ctl = region->in(i);
3794 if (ctl == NULL || ctl == top()) {
3795 region->set_req(i, top());
3796 phi ->set_req(i, top());
3797 } else if (phi->in(i) == NULL) {
3798 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3799 }
3800 }
3801
3802 set_control(_gvn.transform(region));
3803 set_result(_gvn.transform(phi));
3804 return true;
3805 }
3806
3807 //---------------------generate_array_guard_common------------------------
3808 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3809 bool obj_array, bool not_array) {
3810 // If obj_array/non_array==false/false:
3811 // Branch around if the given klass is in fact an array (either obj or prim).
3812 // If obj_array/non_array==false/true:
3813 // Branch around if the given klass is not an array klass of any kind.
3814 // If obj_array/non_array==true/true:
3815 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3816 // If obj_array/non_array==true/false:
3817 // Branch around if the kls is an oop array (Object[] or subtype)
3818 //
3819 // Like generate_guard, adds a new path onto the region.
3820 jint layout_con = 0;
3821 Node* layout_val = get_layout_helper(kls, layout_con);
3822 if (layout_val == NULL) {
3823 bool query = (obj_array
3824 ? Klass::layout_helper_is_objArray(layout_con)
3825 : Klass::layout_helper_is_array(layout_con));
3826 if (query == not_array) {
3827 return NULL; // never a branch
3828 } else { // always a branch
3829 Node* always_branch = control();
3830 if (region != NULL)
3831 region->add_req(always_branch);
3832 set_control(top());
3833 return always_branch;
3834 }
3835 }
3836 // Now test the correct condition.
3837 jint nval = (obj_array
3838 ? (jint)(Klass::_lh_array_tag_type_value
3839 << Klass::_lh_array_tag_shift)
3840 : Klass::_lh_neutral_value);
3841 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3842 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3843 // invert the test if we are looking for a non-array
3844 if (not_array) btest = BoolTest(btest).negate();
3845 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3846 return generate_fair_guard(bol, region);
3847 }
3848
3849
3850 //-----------------------inline_native_newArray--------------------------
3851 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3852 bool LibraryCallKit::inline_native_newArray() {
3853 Node* mirror = argument(0);
3854 Node* count_val = argument(1);
3855
3856 mirror = null_check(mirror);
3857 // If mirror or obj is dead, only null-path is taken.
3858 if (stopped()) return true;
3859
3860 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3861 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3862 PhiNode* result_val = new(C) PhiNode(result_reg,
3863 TypeInstPtr::NOTNULL);
3864 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3865 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3866 TypePtr::BOTTOM);
3867
3868 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3869 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3870 result_reg, _slow_path);
3871 Node* normal_ctl = control();
3872 Node* no_array_ctl = result_reg->in(_slow_path);
3873
3874 // Generate code for the slow case. We make a call to newArray().
3875 set_control(no_array_ctl);
3876 if (!stopped()) {
3877 // Either the input type is void.class, or else the
3878 // array klass has not yet been cached. Either the
3879 // ensuing call will throw an exception, or else it
3880 // will cache the array klass for next time.
3881 PreserveJVMState pjvms(this);
3882 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3883 Node* slow_result = set_results_for_java_call(slow_call);
3884 // this->control() comes from set_results_for_java_call
3885 result_reg->set_req(_slow_path, control());
3886 result_val->set_req(_slow_path, slow_result);
3887 result_io ->set_req(_slow_path, i_o());
3888 result_mem->set_req(_slow_path, reset_memory());
3889 }
3890
3891 set_control(normal_ctl);
3892 if (!stopped()) {
3893 // Normal case: The array type has been cached in the java.lang.Class.
3894 // The following call works fine even if the array type is polymorphic.
3895 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3896 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3897 result_reg->init_req(_normal_path, control());
3898 result_val->init_req(_normal_path, obj);
3899 result_io ->init_req(_normal_path, i_o());
3900 result_mem->init_req(_normal_path, reset_memory());
3901 }
3902
3903 // Return the combined state.
3904 set_i_o( _gvn.transform(result_io) );
3905 set_all_memory( _gvn.transform(result_mem));
3906
3907 C->set_has_split_ifs(true); // Has chance for split-if optimization
3908 set_result(result_reg, result_val);
3909 return true;
3910 }
3911
3912 //----------------------inline_native_getLength--------------------------
3913 // public static native int java.lang.reflect.Array.getLength(Object array);
3914 bool LibraryCallKit::inline_native_getLength() {
3915 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3916
3917 Node* array = null_check(argument(0));
3918 // If array is dead, only null-path is taken.
3919 if (stopped()) return true;
3920
3921 // Deoptimize if it is a non-array.
3922 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3923
3924 if (non_array != NULL) {
3925 PreserveJVMState pjvms(this);
3926 set_control(non_array);
3927 uncommon_trap(Deoptimization::Reason_intrinsic,
3928 Deoptimization::Action_maybe_recompile);
3929 }
3930
3931 // If control is dead, only non-array-path is taken.
3932 if (stopped()) return true;
3933
3934 // The works fine even if the array type is polymorphic.
3935 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3936 Node* result = load_array_length(array);
3937
3938 C->set_has_split_ifs(true); // Has chance for split-if optimization
3939 set_result(result);
3940 return true;
3941 }
3942
3943 //------------------------inline_array_copyOf----------------------------
3944 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3945 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3946 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3947 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3948
3949 // Get the arguments.
3950 Node* original = argument(0);
3951 Node* start = is_copyOfRange? argument(1): intcon(0);
3952 Node* end = is_copyOfRange? argument(2): argument(1);
3953 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3954
3955 Node* newcopy = NULL;
3956
3957 // Set the original stack and the reexecute bit for the interpreter to reexecute
3958 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3959 { PreserveReexecuteState preexecs(this);
3960 jvms()->set_should_reexecute(true);
3961
3962 array_type_mirror = null_check(array_type_mirror);
3963 original = null_check(original);
3964
3965 // Check if a null path was taken unconditionally.
3966 if (stopped()) return true;
3967
3968 Node* orig_length = load_array_length(original);
3969
3970 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3971 klass_node = null_check(klass_node);
3972
3973 RegionNode* bailout = new (C) RegionNode(1);
3974 record_for_igvn(bailout);
3975
3976 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3977 // Bail out if that is so.
3978 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3979 if (not_objArray != NULL) {
3980 // Improve the klass node's type from the new optimistic assumption:
3981 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3982 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3983 Node* cast = new (C) CastPPNode(klass_node, akls);
3984 cast->init_req(0, control());
3985 klass_node = _gvn.transform(cast);
3986 }
3987
3988 // Bail out if either start or end is negative.
3989 generate_negative_guard(start, bailout, &start);
3990 generate_negative_guard(end, bailout, &end);
3991
3992 Node* length = end;
3993 if (_gvn.type(start) != TypeInt::ZERO) {
3994 length = _gvn.transform(new (C) SubINode(end, start));
3995 }
3996
3997 // Bail out if length is negative.
3998 // Without this the new_array would throw
3999 // NegativeArraySizeException but IllegalArgumentException is what
4000 // should be thrown
4001 generate_negative_guard(length, bailout, &length);
4002
4003 if (bailout->req() > 1) {
4004 PreserveJVMState pjvms(this);
4005 set_control(_gvn.transform(bailout));
4006 uncommon_trap(Deoptimization::Reason_intrinsic,
4007 Deoptimization::Action_maybe_recompile);
4008 }
4009
4010 if (!stopped()) {
4011 // How many elements will we copy from the original?
4012 // The answer is MinI(orig_length - start, length).
4013 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
4014 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4015
4016 newcopy = new_array(klass_node, length, 0); // no argments to push
4017
4018 // Generate a direct call to the right arraycopy function(s).
4019 // We know the copy is disjoint but we might not know if the
4020 // oop stores need checking.
4021 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4022 // This will fail a store-check if x contains any non-nulls.
4023 bool disjoint_bases = true;
4024 // if start > orig_length then the length of the copy may be
4025 // negative.
4026 bool length_never_negative = !is_copyOfRange;
4027 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4028 original, start, newcopy, intcon(0), moved,
4029 disjoint_bases, length_never_negative);
4030 }
4031 } // original reexecute is set back here
4032
4033 C->set_has_split_ifs(true); // Has chance for split-if optimization
4034 if (!stopped()) {
4035 set_result(newcopy);
4036 }
4037 return true;
4038 }
4039
4040
4041 //----------------------generate_virtual_guard---------------------------
4042 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4043 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4044 RegionNode* slow_region) {
4045 ciMethod* method = callee();
4046 int vtable_index = method->vtable_index();
4047 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4048 err_msg_res("bad index %d", vtable_index));
4049 // Get the Method* out of the appropriate vtable entry.
4050 int entry_offset = (InstanceKlass::vtable_start_offset() +
4051 vtable_index*vtableEntry::size()) * wordSize +
4052 vtableEntry::method_offset_in_bytes();
4053 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4054 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4055
4056 // Compare the target method with the expected method (e.g., Object.hashCode).
4057 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4058
4059 Node* native_call = makecon(native_call_addr);
4060 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4061 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4062
4063 return generate_slow_guard(test_native, slow_region);
4064 }
4065
4066 //-----------------------generate_method_call----------------------------
4067 // Use generate_method_call to make a slow-call to the real
4068 // method if the fast path fails. An alternative would be to
4069 // use a stub like OptoRuntime::slow_arraycopy_Java.
4070 // This only works for expanding the current library call,
4071 // not another intrinsic. (E.g., don't use this for making an
4072 // arraycopy call inside of the copyOf intrinsic.)
4073 CallJavaNode*
4074 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4075 // When compiling the intrinsic method itself, do not use this technique.
4076 guarantee(callee() != C->method(), "cannot make slow-call to self");
4077
4078 ciMethod* method = callee();
4079 // ensure the JVMS we have will be correct for this call
4080 guarantee(method_id == method->intrinsic_id(), "must match");
4081
4082 const TypeFunc* tf = TypeFunc::make(method);
4083 CallJavaNode* slow_call;
4084 if (is_static) {
4085 assert(!is_virtual, "");
4086 slow_call = new(C) CallStaticJavaNode(C, tf,
4087 SharedRuntime::get_resolve_static_call_stub(),
4088 method, bci());
4089 } else if (is_virtual) {
4090 null_check_receiver();
4091 int vtable_index = Method::invalid_vtable_index;
4092 if (UseInlineCaches) {
4093 // Suppress the vtable call
4094 } else {
4095 // hashCode and clone are not a miranda methods,
4096 // so the vtable index is fixed.
4097 // No need to use the linkResolver to get it.
4098 vtable_index = method->vtable_index();
4099 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4100 err_msg_res("bad index %d", vtable_index));
4101 }
4102 slow_call = new(C) CallDynamicJavaNode(tf,
4103 SharedRuntime::get_resolve_virtual_call_stub(),
4104 method, vtable_index, bci());
4105 } else { // neither virtual nor static: opt_virtual
4106 null_check_receiver();
4107 slow_call = new(C) CallStaticJavaNode(C, tf,
4108 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4109 method, bci());
4110 slow_call->set_optimized_virtual(true);
4111 }
4112 set_arguments_for_java_call(slow_call);
4113 set_edges_for_java_call(slow_call);
4114 return slow_call;
4115 }
4116
4117
4118 /**
4119 * Build special case code for calls to hashCode on an object. This call may
4120 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4121 * slightly different code.
4122 */
4123 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4124 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4125 assert(!(is_virtual && is_static), "either virtual, special, or static");
4126
4127 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4128
4129 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4130 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4131 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4132 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4133 Node* obj = NULL;
4134 if (!is_static) {
4135 // Check for hashing null object
4136 obj = null_check_receiver();
4137 if (stopped()) return true; // unconditionally null
4138 result_reg->init_req(_null_path, top());
4139 result_val->init_req(_null_path, top());
4140 } else {
4141 // Do a null check, and return zero if null.
4142 // System.identityHashCode(null) == 0
4143 obj = argument(0);
4144 Node* null_ctl = top();
4145 obj = null_check_oop(obj, &null_ctl);
4146 result_reg->init_req(_null_path, null_ctl);
4147 result_val->init_req(_null_path, _gvn.intcon(0));
4148 }
4149
4150 // Unconditionally null? Then return right away.
4151 if (stopped()) {
4152 set_control( result_reg->in(_null_path));
4153 if (!stopped())
4154 set_result(result_val->in(_null_path));
4155 return true;
4156 }
4157
4158 // We only go to the fast case code if we pass a number of guards. The
4159 // paths which do not pass are accumulated in the slow_region.
4160 RegionNode* slow_region = new (C) RegionNode(1);
4161 record_for_igvn(slow_region);
4162
4163 // If this is a virtual call, we generate a funny guard. We pull out
4164 // the vtable entry corresponding to hashCode() from the target object.
4165 // If the target method which we are calling happens to be the native
4166 // Object hashCode() method, we pass the guard. We do not need this
4167 // guard for non-virtual calls -- the caller is known to be the native
4168 // Object hashCode().
4169 if (is_virtual) {
4170 // After null check, get the object's klass.
4171 Node* obj_klass = load_object_klass(obj);
4172 generate_virtual_guard(obj_klass, slow_region);
4173 }
4174
4175 // Get the header out of the object, use LoadMarkNode when available
4176 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4177 // The control of the load must be NULL. Otherwise, the load can move before
4178 // the null check after castPP removal.
4179 Node* no_ctrl = NULL;
4180 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4181
4182 // Test the header to see if it is unlocked.
4183 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4184 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4185 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4186 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4187 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4188
4189 generate_slow_guard(test_unlocked, slow_region);
4190
4191 // Get the hash value and check to see that it has been properly assigned.
4192 // We depend on hash_mask being at most 32 bits and avoid the use of
4193 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4194 // vm: see markOop.hpp.
4195 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4196 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4197 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4198 // This hack lets the hash bits live anywhere in the mark object now, as long
4199 // as the shift drops the relevant bits into the low 32 bits. Note that
4200 // Java spec says that HashCode is an int so there's no point in capturing
4201 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4202 hshifted_header = ConvX2I(hshifted_header);
4203 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4204
4205 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4206 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4207 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4208
4209 generate_slow_guard(test_assigned, slow_region);
4210
4211 Node* init_mem = reset_memory();
4212 // fill in the rest of the null path:
4213 result_io ->init_req(_null_path, i_o());
4214 result_mem->init_req(_null_path, init_mem);
4215
4216 result_val->init_req(_fast_path, hash_val);
4217 result_reg->init_req(_fast_path, control());
4218 result_io ->init_req(_fast_path, i_o());
4219 result_mem->init_req(_fast_path, init_mem);
4220
4221 // Generate code for the slow case. We make a call to hashCode().
4222 set_control(_gvn.transform(slow_region));
4223 if (!stopped()) {
4224 // No need for PreserveJVMState, because we're using up the present state.
4225 set_all_memory(init_mem);
4226 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4227 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4228 Node* slow_result = set_results_for_java_call(slow_call);
4229 // this->control() comes from set_results_for_java_call
4230 result_reg->init_req(_slow_path, control());
4231 result_val->init_req(_slow_path, slow_result);
4232 result_io ->set_req(_slow_path, i_o());
4233 result_mem ->set_req(_slow_path, reset_memory());
4234 }
4235
4236 // Return the combined state.
4237 set_i_o( _gvn.transform(result_io) );
4238 set_all_memory( _gvn.transform(result_mem));
4239
4240 set_result(result_reg, result_val);
4241 return true;
4242 }
4243
4244 //---------------------------inline_native_getClass----------------------------
4245 // public final native Class<?> java.lang.Object.getClass();
4246 //
4247 // Build special case code for calls to getClass on an object.
4248 bool LibraryCallKit::inline_native_getClass() {
4249 Node* obj = null_check_receiver();
4250 if (stopped()) return true;
4251 set_result(load_mirror_from_klass(load_object_klass(obj)));
4252 return true;
4253 }
4254
4255 //-----------------inline_native_Reflection_getCallerClass---------------------
4256 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4257 //
4258 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4259 //
4260 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4261 // in that it must skip particular security frames and checks for
4262 // caller sensitive methods.
4263 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4264 #ifndef PRODUCT
4265 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4266 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4267 }
4268 #endif
4269
4270 if (!jvms()->has_method()) {
4271 #ifndef PRODUCT
4272 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4273 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4274 }
4275 #endif
4276 return false;
4277 }
4278
4279 // Walk back up the JVM state to find the caller at the required
4280 // depth.
4281 JVMState* caller_jvms = jvms();
4282
4283 // Cf. JVM_GetCallerClass
4284 // NOTE: Start the loop at depth 1 because the current JVM state does
4285 // not include the Reflection.getCallerClass() frame.
4286 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4287 ciMethod* m = caller_jvms->method();
4288 switch (n) {
4289 case 0:
4290 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4291 break;
4292 case 1:
4293 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4294 if (!m->caller_sensitive()) {
4295 #ifndef PRODUCT
4296 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4297 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4298 }
4299 #endif
4300 return false; // bail-out; let JVM_GetCallerClass do the work
4301 }
4302 break;
4303 default:
4304 if (!m->is_ignored_by_security_stack_walk()) {
4305 // We have reached the desired frame; return the holder class.
4306 // Acquire method holder as java.lang.Class and push as constant.
4307 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4308 ciInstance* caller_mirror = caller_klass->java_mirror();
4309 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4310
4311 #ifndef PRODUCT
4312 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4313 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());
4314 tty->print_cr(" JVM state at this point:");
4315 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4316 ciMethod* m = jvms()->of_depth(i)->method();
4317 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4318 }
4319 }
4320 #endif
4321 return true;
4322 }
4323 break;
4324 }
4325 }
4326
4327 #ifndef PRODUCT
4328 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4329 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4330 tty->print_cr(" JVM state at this point:");
4331 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4332 ciMethod* m = jvms()->of_depth(i)->method();
4333 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4334 }
4335 }
4336 #endif
4337
4338 return false; // bail-out; let JVM_GetCallerClass do the work
4339 }
4340
4341 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4342 Node* arg = argument(0);
4343 Node* result = NULL;
4344
4345 switch (id) {
4346 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4347 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4348 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4349 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4350
4351 case vmIntrinsics::_doubleToLongBits: {
4352 // two paths (plus control) merge in a wood
4353 RegionNode *r = new (C) RegionNode(3);
4354 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4355
4356 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4357 // Build the boolean node
4358 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4359
4360 // Branch either way.
4361 // NaN case is less traveled, which makes all the difference.
4362 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4363 Node *opt_isnan = _gvn.transform(ifisnan);
4364 assert( opt_isnan->is_If(), "Expect an IfNode");
4365 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4366 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4367
4368 set_control(iftrue);
4369
4370 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4371 Node *slow_result = longcon(nan_bits); // return NaN
4372 phi->init_req(1, _gvn.transform( slow_result ));
4373 r->init_req(1, iftrue);
4374
4375 // Else fall through
4376 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4377 set_control(iffalse);
4378
4379 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4380 r->init_req(2, iffalse);
4381
4382 // Post merge
4383 set_control(_gvn.transform(r));
4384 record_for_igvn(r);
4385
4386 C->set_has_split_ifs(true); // Has chance for split-if optimization
4387 result = phi;
4388 assert(result->bottom_type()->isa_long(), "must be");
4389 break;
4390 }
4391
4392 case vmIntrinsics::_floatToIntBits: {
4393 // two paths (plus control) merge in a wood
4394 RegionNode *r = new (C) RegionNode(3);
4395 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4396
4397 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4398 // Build the boolean node
4399 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4400
4401 // Branch either way.
4402 // NaN case is less traveled, which makes all the difference.
4403 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4404 Node *opt_isnan = _gvn.transform(ifisnan);
4405 assert( opt_isnan->is_If(), "Expect an IfNode");
4406 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4407 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4408
4409 set_control(iftrue);
4410
4411 static const jint nan_bits = 0x7fc00000;
4412 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4413 phi->init_req(1, _gvn.transform( slow_result ));
4414 r->init_req(1, iftrue);
4415
4416 // Else fall through
4417 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4418 set_control(iffalse);
4419
4420 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4421 r->init_req(2, iffalse);
4422
4423 // Post merge
4424 set_control(_gvn.transform(r));
4425 record_for_igvn(r);
4426
4427 C->set_has_split_ifs(true); // Has chance for split-if optimization
4428 result = phi;
4429 assert(result->bottom_type()->isa_int(), "must be");
4430 break;
4431 }
4432
4433 default:
4434 fatal_unexpected_iid(id);
4435 break;
4436 }
4437 set_result(_gvn.transform(result));
4438 return true;
4439 }
4440
4441 #ifdef _LP64
4442 #define XTOP ,top() /*additional argument*/
4443 #else //_LP64
4444 #define XTOP /*no additional argument*/
4445 #endif //_LP64
4446
4447 //----------------------inline_unsafe_copyMemory-------------------------
4448 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4449 bool LibraryCallKit::inline_unsafe_copyMemory() {
4450 if (callee()->is_static()) return false; // caller must have the capability!
4451 null_check_receiver(); // null-check receiver
4452 if (stopped()) return true;
4453
4454 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4455
4456 Node* src_ptr = argument(1); // type: oop
4457 Node* src_off = ConvL2X(argument(2)); // type: long
4458 Node* dst_ptr = argument(4); // type: oop
4459 Node* dst_off = ConvL2X(argument(5)); // type: long
4460 Node* size = ConvL2X(argument(7)); // type: long
4461
4462 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4463 "fieldOffset must be byte-scaled");
4464
4465 Node* src = make_unsafe_address(src_ptr, src_off);
4466 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4467
4468 // Conservatively insert a memory barrier on all memory slices.
4469 // Do not let writes of the copy source or destination float below the copy.
4470 insert_mem_bar(Op_MemBarCPUOrder);
4471
4472 // Call it. Note that the length argument is not scaled.
4473 make_runtime_call(RC_LEAF|RC_NO_FP,
4474 OptoRuntime::fast_arraycopy_Type(),
4475 StubRoutines::unsafe_arraycopy(),
4476 "unsafe_arraycopy",
4477 TypeRawPtr::BOTTOM,
4478 src, dst, size XTOP);
4479
4480 // Do not let reads of the copy destination float above the copy.
4481 insert_mem_bar(Op_MemBarCPUOrder);
4482
4483 return true;
4484 }
4485
4486 //------------------------clone_coping-----------------------------------
4487 // Helper function for inline_native_clone.
4488 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4489 assert(obj_size != NULL, "");
4490 Node* raw_obj = alloc_obj->in(1);
4491 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4492
4493 AllocateNode* alloc = NULL;
4494 if (ReduceBulkZeroing) {
4495 // We will be completely responsible for initializing this object -
4496 // mark Initialize node as complete.
4497 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4498 // The object was just allocated - there should be no any stores!
4499 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4500 // Mark as complete_with_arraycopy so that on AllocateNode
4501 // expansion, we know this AllocateNode is initialized by an array
4502 // copy and a StoreStore barrier exists after the array copy.
4503 alloc->initialization()->set_complete_with_arraycopy();
4504 }
4505
4506 // Copy the fastest available way.
4507 // TODO: generate fields copies for small objects instead.
4508 Node* src = obj;
4509 Node* dest = alloc_obj;
4510 Node* size = _gvn.transform(obj_size);
4511
4512 // Exclude the header but include array length to copy by 8 bytes words.
4513 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4514 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4515 instanceOopDesc::base_offset_in_bytes();
4516 // base_off:
4517 // 8 - 32-bit VM
4518 // 12 - 64-bit VM, compressed klass
4519 // 16 - 64-bit VM, normal klass
4520 if (base_off % BytesPerLong != 0) {
4521 assert(UseCompressedClassPointers, "");
4522 if (is_array) {
4523 // Exclude length to copy by 8 bytes words.
4524 base_off += sizeof(int);
4525 } else {
4526 // Include klass to copy by 8 bytes words.
4527 base_off = instanceOopDesc::klass_offset_in_bytes();
4528 }
4529 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4530 }
4531 src = basic_plus_adr(src, base_off);
4532 dest = basic_plus_adr(dest, base_off);
4533
4534 // Compute the length also, if needed:
4535 Node* countx = size;
4536 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4537 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4538
4539 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4540 bool disjoint_bases = true;
4541 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4542 src, NULL, dest, NULL, countx,
4543 /*dest_uninitialized*/true);
4544
4545 if (UseShenandoahGC && ShenandoahCloneBarrier) {
4546 // Make sure that references in the cloned object are updated for Shenandoah.
4547 make_runtime_call(RC_LEAF|RC_NO_FP,
4548 OptoRuntime::shenandoah_clone_barrier_Type(),
4549 CAST_FROM_FN_PTR(address, SharedRuntime::shenandoah_clone_barrier),
4550 "shenandoah_clone_barrier", TypePtr::BOTTOM,
4551 alloc_obj);
4552 }
4553
4554 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4555 if (card_mark) {
4556 assert(!is_array, "");
4557 // Put in store barrier for any and all oops we are sticking
4558 // into this object. (We could avoid this if we could prove
4559 // that the object type contains no oop fields at all.)
4560 Node* no_particular_value = NULL;
4561 Node* no_particular_field = NULL;
4562 int raw_adr_idx = Compile::AliasIdxRaw;
4563 post_barrier(control(),
4564 memory(raw_adr_type),
4565 alloc_obj,
4566 no_particular_field,
4567 raw_adr_idx,
4568 no_particular_value,
4569 T_OBJECT,
4570 false);
4571 }
4572
4573 // Do not let reads from the cloned object float above the arraycopy.
4574 if (alloc != NULL) {
4575 // Do not let stores that initialize this object be reordered with
4576 // a subsequent store that would make this object accessible by
4577 // other threads.
4578 // Record what AllocateNode this StoreStore protects so that
4579 // escape analysis can go from the MemBarStoreStoreNode to the
4580 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4581 // based on the escape status of the AllocateNode.
4582 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4583 } else {
4584 insert_mem_bar(Op_MemBarCPUOrder);
4585 }
4586 }
4587
4588 //------------------------inline_native_clone----------------------------
4589 // protected native Object java.lang.Object.clone();
4590 //
4591 // Here are the simple edge cases:
4592 // null receiver => normal trap
4593 // virtual and clone was overridden => slow path to out-of-line clone
4594 // not cloneable or finalizer => slow path to out-of-line Object.clone
4595 //
4596 // The general case has two steps, allocation and copying.
4597 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4598 //
4599 // Copying also has two cases, oop arrays and everything else.
4600 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4601 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4602 //
4603 // These steps fold up nicely if and when the cloned object's klass
4604 // can be sharply typed as an object array, a type array, or an instance.
4605 //
4606 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4607 PhiNode* result_val;
4608
4609 // Set the reexecute bit for the interpreter to reexecute
4610 // the bytecode that invokes Object.clone if deoptimization happens.
4611 { PreserveReexecuteState preexecs(this);
4612 jvms()->set_should_reexecute(true);
4613
4614 Node* obj = null_check_receiver();
4615 if (stopped()) return true;
4616
4617 Node* obj_klass = load_object_klass(obj);
4618 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4619 const TypeOopPtr* toop = ((tklass != NULL)
4620 ? tklass->as_instance_type()
4621 : TypeInstPtr::NOTNULL);
4622
4623 // Conservatively insert a memory barrier on all memory slices.
4624 // Do not let writes into the original float below the clone.
4625 insert_mem_bar(Op_MemBarCPUOrder);
4626
4627 // paths into result_reg:
4628 enum {
4629 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4630 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4631 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4632 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4633 PATH_LIMIT
4634 };
4635 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4636 result_val = new(C) PhiNode(result_reg,
4637 TypeInstPtr::NOTNULL);
4638 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4639 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4640 TypePtr::BOTTOM);
4641 record_for_igvn(result_reg);
4642
4643 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4644 int raw_adr_idx = Compile::AliasIdxRaw;
4645
4646 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4647 if (array_ctl != NULL) {
4648 // It's an array.
4649 PreserveJVMState pjvms(this);
4650 set_control(array_ctl);
4651 Node* obj_length = load_array_length(obj);
4652 Node* obj_size = NULL;
4653 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4654
4655 if (!use_ReduceInitialCardMarks()) {
4656 // If it is an oop array, it requires very special treatment,
4657 // because card marking is required on each card of the array.
4658 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4659 if (is_obja != NULL) {
4660 PreserveJVMState pjvms2(this);
4661 set_control(is_obja);
4662 // Generate a direct call to the right arraycopy function(s).
4663 bool disjoint_bases = true;
4664 bool length_never_negative = true;
4665 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4666 obj, intcon(0), alloc_obj, intcon(0),
4667 obj_length,
4668 disjoint_bases, length_never_negative);
4669 result_reg->init_req(_objArray_path, control());
4670 result_val->init_req(_objArray_path, alloc_obj);
4671 result_i_o ->set_req(_objArray_path, i_o());
4672 result_mem ->set_req(_objArray_path, reset_memory());
4673 }
4674 }
4675 // Otherwise, there are no card marks to worry about.
4676 // (We can dispense with card marks if we know the allocation
4677 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4678 // causes the non-eden paths to take compensating steps to
4679 // simulate a fresh allocation, so that no further
4680 // card marks are required in compiled code to initialize
4681 // the object.)
4682
4683 if (!stopped()) {
4684 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4685
4686 // Present the results of the copy.
4687 result_reg->init_req(_array_path, control());
4688 result_val->init_req(_array_path, alloc_obj);
4689 result_i_o ->set_req(_array_path, i_o());
4690 result_mem ->set_req(_array_path, reset_memory());
4691 }
4692 }
4693
4694 // We only go to the instance fast case code if we pass a number of guards.
4695 // The paths which do not pass are accumulated in the slow_region.
4696 RegionNode* slow_region = new (C) RegionNode(1);
4697 record_for_igvn(slow_region);
4698 if (!stopped()) {
4699 // It's an instance (we did array above). Make the slow-path tests.
4700 // If this is a virtual call, we generate a funny guard. We grab
4701 // the vtable entry corresponding to clone() from the target object.
4702 // If the target method which we are calling happens to be the
4703 // Object clone() method, we pass the guard. We do not need this
4704 // guard for non-virtual calls; the caller is known to be the native
4705 // Object clone().
4706 if (is_virtual) {
4707 generate_virtual_guard(obj_klass, slow_region);
4708 }
4709
4710 // The object must be cloneable and must not have a finalizer.
4711 // Both of these conditions may be checked in a single test.
4712 // We could optimize the cloneable test further, but we don't care.
4713 generate_access_flags_guard(obj_klass,
4714 // Test both conditions:
4715 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4716 // Must be cloneable but not finalizer:
4717 JVM_ACC_IS_CLONEABLE,
4718 slow_region);
4719 }
4720
4721 if (!stopped()) {
4722 // It's an instance, and it passed the slow-path tests.
4723 PreserveJVMState pjvms(this);
4724 Node* obj_size = NULL;
4725 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4726 // is reexecuted if deoptimization occurs and there could be problems when merging
4727 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4728 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4729
4730 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4731
4732 // Present the results of the slow call.
4733 result_reg->init_req(_instance_path, control());
4734 result_val->init_req(_instance_path, alloc_obj);
4735 result_i_o ->set_req(_instance_path, i_o());
4736 result_mem ->set_req(_instance_path, reset_memory());
4737 }
4738
4739 // Generate code for the slow case. We make a call to clone().
4740 set_control(_gvn.transform(slow_region));
4741 if (!stopped()) {
4742 PreserveJVMState pjvms(this);
4743 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4744 Node* slow_result = set_results_for_java_call(slow_call);
4745 // this->control() comes from set_results_for_java_call
4746 result_reg->init_req(_slow_path, control());
4747 result_val->init_req(_slow_path, slow_result);
4748 result_i_o ->set_req(_slow_path, i_o());
4749 result_mem ->set_req(_slow_path, reset_memory());
4750 }
4751
4752 // Return the combined state.
4753 set_control( _gvn.transform(result_reg));
4754 set_i_o( _gvn.transform(result_i_o));
4755 set_all_memory( _gvn.transform(result_mem));
4756 } // original reexecute is set back here
4757
4758 set_result(_gvn.transform(result_val));
4759 return true;
4760 }
4761
4762 //------------------------------basictype2arraycopy----------------------------
4763 address LibraryCallKit::basictype2arraycopy(BasicType t,
4764 Node* src_offset,
4765 Node* dest_offset,
4766 bool disjoint_bases,
4767 const char* &name,
4768 bool dest_uninitialized) {
4769 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4770 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4771
4772 bool aligned = false;
4773 bool disjoint = disjoint_bases;
4774
4775 // if the offsets are the same, we can treat the memory regions as
4776 // disjoint, because either the memory regions are in different arrays,
4777 // or they are identical (which we can treat as disjoint.) We can also
4778 // treat a copy with a destination index less that the source index
4779 // as disjoint since a low->high copy will work correctly in this case.
4780 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4781 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4782 // both indices are constants
4783 int s_offs = src_offset_inttype->get_con();
4784 int d_offs = dest_offset_inttype->get_con();
4785 int element_size = type2aelembytes(t);
4786 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4787 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4788 if (s_offs >= d_offs) disjoint = true;
4789 } else if (src_offset == dest_offset && src_offset != NULL) {
4790 // This can occur if the offsets are identical non-constants.
4791 disjoint = true;
4792 }
4793
4794 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4795 }
4796
4797
4798 //------------------------------inline_arraycopy-----------------------
4799 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4800 // Object dest, int destPos,
4801 // int length);
4802 bool LibraryCallKit::inline_arraycopy() {
4803 // Get the arguments.
4804 Node* src = argument(0); // type: oop
4805 Node* src_offset = argument(1); // type: int
4806 Node* dest = argument(2); // type: oop
4807 Node* dest_offset = argument(3); // type: int
4808 Node* length = argument(4); // type: int
4809
4810 // Compile time checks. If any of these checks cannot be verified at compile time,
4811 // we do not make a fast path for this call. Instead, we let the call remain as it
4812 // is. The checks we choose to mandate at compile time are:
4813 //
4814 // (1) src and dest are arrays.
4815 const Type* src_type = src->Value(&_gvn);
4816 const Type* dest_type = dest->Value(&_gvn);
4817 const TypeAryPtr* top_src = src_type->isa_aryptr();
4818 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4819
4820 // Do we have the type of src?
4821 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4822 // Do we have the type of dest?
4823 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4824 // Is the type for src from speculation?
4825 bool src_spec = false;
4826 // Is the type for dest from speculation?
4827 bool dest_spec = false;
4828
4829 if (!has_src || !has_dest) {
4830 // We don't have sufficient type information, let's see if
4831 // speculative types can help. We need to have types for both src
4832 // and dest so that it pays off.
4833
4834 // Do we already have or could we have type information for src
4835 bool could_have_src = has_src;
4836 // Do we already have or could we have type information for dest
4837 bool could_have_dest = has_dest;
4838
4839 ciKlass* src_k = NULL;
4840 if (!has_src) {
4841 src_k = src_type->speculative_type();
4842 if (src_k != NULL && src_k->is_array_klass()) {
4843 could_have_src = true;
4844 }
4845 }
4846
4847 ciKlass* dest_k = NULL;
4848 if (!has_dest) {
4849 dest_k = dest_type->speculative_type();
4850 if (dest_k != NULL && dest_k->is_array_klass()) {
4851 could_have_dest = true;
4852 }
4853 }
4854
4855 if (could_have_src && could_have_dest) {
4856 // This is going to pay off so emit the required guards
4857 if (!has_src) {
4858 src = maybe_cast_profiled_obj(src, src_k);
4859 src_type = _gvn.type(src);
4860 top_src = src_type->isa_aryptr();
4861 has_src = (top_src != NULL && top_src->klass() != NULL);
4862 src_spec = true;
4863 }
4864 if (!has_dest) {
4865 dest = maybe_cast_profiled_obj(dest, dest_k);
4866 dest_type = _gvn.type(dest);
4867 top_dest = dest_type->isa_aryptr();
4868 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4869 dest_spec = true;
4870 }
4871 }
4872 }
4873
4874 if (!has_src || !has_dest) {
4875 // Conservatively insert a memory barrier on all memory slices.
4876 // Do not let writes into the source float below the arraycopy.
4877 insert_mem_bar(Op_MemBarCPUOrder);
4878
4879 // Call StubRoutines::generic_arraycopy stub.
4880 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4881 src, src_offset, dest, dest_offset, length);
4882
4883 // Do not let reads from the destination float above the arraycopy.
4884 // Since we cannot type the arrays, we don't know which slices
4885 // might be affected. We could restrict this barrier only to those
4886 // memory slices which pertain to array elements--but don't bother.
4887 if (!InsertMemBarAfterArraycopy)
4888 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4889 insert_mem_bar(Op_MemBarCPUOrder);
4890 return true;
4891 }
4892
4893 // (2) src and dest arrays must have elements of the same BasicType
4894 // Figure out the size and type of the elements we will be copying.
4895 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4896 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4897 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4898 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4899
4900 if (src_elem != dest_elem || dest_elem == T_VOID) {
4901 // The component types are not the same or are not recognized. Punt.
4902 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4903 generate_slow_arraycopy(TypePtr::BOTTOM,
4904 src, src_offset, dest, dest_offset, length,
4905 /*dest_uninitialized*/false);
4906 return true;
4907 }
4908
4909 if (src_elem == T_OBJECT) {
4910 // If both arrays are object arrays then having the exact types
4911 // for both will remove the need for a subtype check at runtime
4912 // before the call and may make it possible to pick a faster copy
4913 // routine (without a subtype check on every element)
4914 // Do we have the exact type of src?
4915 bool could_have_src = src_spec;
4916 // Do we have the exact type of dest?
4917 bool could_have_dest = dest_spec;
4918 ciKlass* src_k = top_src->klass();
4919 ciKlass* dest_k = top_dest->klass();
4920 if (!src_spec) {
4921 src_k = src_type->speculative_type();
4922 if (src_k != NULL && src_k->is_array_klass()) {
4923 could_have_src = true;
4924 }
4925 }
4926 if (!dest_spec) {
4927 dest_k = dest_type->speculative_type();
4928 if (dest_k != NULL && dest_k->is_array_klass()) {
4929 could_have_dest = true;
4930 }
4931 }
4932 if (could_have_src && could_have_dest) {
4933 // If we can have both exact types, emit the missing guards
4934 if (could_have_src && !src_spec) {
4935 src = maybe_cast_profiled_obj(src, src_k);
4936 }
4937 if (could_have_dest && !dest_spec) {
4938 dest = maybe_cast_profiled_obj(dest, dest_k);
4939 }
4940 }
4941 }
4942
4943 //---------------------------------------------------------------------------
4944 // We will make a fast path for this call to arraycopy.
4945
4946 // We have the following tests left to perform:
4947 //
4948 // (3) src and dest must not be null.
4949 // (4) src_offset must not be negative.
4950 // (5) dest_offset must not be negative.
4951 // (6) length must not be negative.
4952 // (7) src_offset + length must not exceed length of src.
4953 // (8) dest_offset + length must not exceed length of dest.
4954 // (9) each element of an oop array must be assignable
4955
4956 RegionNode* slow_region = new (C) RegionNode(1);
4957 record_for_igvn(slow_region);
4958
4959 // (3) operands must not be null
4960 // We currently perform our null checks with the null_check routine.
4961 // This means that the null exceptions will be reported in the caller
4962 // rather than (correctly) reported inside of the native arraycopy call.
4963 // This should be corrected, given time. We do our null check with the
4964 // stack pointer restored.
4965 src = null_check(src, T_ARRAY);
4966 dest = null_check(dest, T_ARRAY);
4967
4968 // (4) src_offset must not be negative.
4969 generate_negative_guard(src_offset, slow_region);
4970
4971 // (5) dest_offset must not be negative.
4972 generate_negative_guard(dest_offset, slow_region);
4973
4974 // (6) length must not be negative (moved to generate_arraycopy()).
4975 // generate_negative_guard(length, slow_region);
4976
4977 // (7) src_offset + length must not exceed length of src.
4978 generate_limit_guard(src_offset, length,
4979 load_array_length(src),
4980 slow_region);
4981
4982 // (8) dest_offset + length must not exceed length of dest.
4983 generate_limit_guard(dest_offset, length,
4984 load_array_length(dest),
4985 slow_region);
4986
4987 // (9) each element of an oop array must be assignable
4988 // The generate_arraycopy subroutine checks this.
4989
4990 // This is where the memory effects are placed:
4991 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4992 generate_arraycopy(adr_type, dest_elem,
4993 src, src_offset, dest, dest_offset, length,
4994 false, false, slow_region);
4995
4996 return true;
4997 }
4998
4999 //-----------------------------generate_arraycopy----------------------
5000 // Generate an optimized call to arraycopy.
5001 // Caller must guard against non-arrays.
5002 // Caller must determine a common array basic-type for both arrays.
5003 // Caller must validate offsets against array bounds.
5004 // The slow_region has already collected guard failure paths
5005 // (such as out of bounds length or non-conformable array types).
5006 // The generated code has this shape, in general:
5007 //
5008 // if (length == 0) return // via zero_path
5009 // slowval = -1
5010 // if (types unknown) {
5011 // slowval = call generic copy loop
5012 // if (slowval == 0) return // via checked_path
5013 // } else if (indexes in bounds) {
5014 // if ((is object array) && !(array type check)) {
5015 // slowval = call checked copy loop
5016 // if (slowval == 0) return // via checked_path
5017 // } else {
5018 // call bulk copy loop
5019 // return // via fast_path
5020 // }
5021 // }
5022 // // adjust params for remaining work:
5023 // if (slowval != -1) {
5024 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
5025 // }
5026 // slow_region:
5027 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
5028 // return // via slow_call_path
5029 //
5030 // This routine is used from several intrinsics: System.arraycopy,
5031 // Object.clone (the array subcase), and Arrays.copyOf[Range].
5032 //
5033 void
5034 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
5035 BasicType basic_elem_type,
5036 Node* src, Node* src_offset,
5037 Node* dest, Node* dest_offset,
5038 Node* copy_length,
5039 bool disjoint_bases,
5040 bool length_never_negative,
5041 RegionNode* slow_region) {
5042
5043 if (slow_region == NULL) {
5044 slow_region = new(C) RegionNode(1);
5045 record_for_igvn(slow_region);
5046 }
5047
5048 Node* original_dest = dest;
5049 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5050 bool dest_uninitialized = false;
5051
5052 // See if this is the initialization of a newly-allocated array.
5053 // If so, we will take responsibility here for initializing it to zero.
5054 // (Note: Because tightly_coupled_allocation performs checks on the
5055 // out-edges of the dest, we need to avoid making derived pointers
5056 // from it until we have checked its uses.)
5057 if (ReduceBulkZeroing
5058 && !ZeroTLAB // pointless if already zeroed
5059 && basic_elem_type != T_CONFLICT // avoid corner case
5060 && !src->eqv_uncast(dest)
5061 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5062 != NULL)
5063 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5064 && alloc->maybe_set_complete(&_gvn)) {
5065 // "You break it, you buy it."
5066 InitializeNode* init = alloc->initialization();
5067 assert(init->is_complete(), "we just did this");
5068 init->set_complete_with_arraycopy();
5069 assert(dest->is_CheckCastPP(), "sanity");
5070 assert(dest->in(0)->in(0) == init, "dest pinned");
5071 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5072 // From this point on, every exit path is responsible for
5073 // initializing any non-copied parts of the object to zero.
5074 // Also, if this flag is set we make sure that arraycopy interacts properly
5075 // with G1, eliding pre-barriers. See CR 6627983.
5076 dest_uninitialized = true;
5077 } else {
5078 // No zeroing elimination here.
5079 alloc = NULL;
5080 //original_dest = dest;
5081 //dest_uninitialized = false;
5082 }
5083
5084 // Results are placed here:
5085 enum { fast_path = 1, // normal void-returning assembly stub
5086 checked_path = 2, // special assembly stub with cleanup
5087 slow_call_path = 3, // something went wrong; call the VM
5088 zero_path = 4, // bypass when length of copy is zero
5089 bcopy_path = 5, // copy primitive array by 64-bit blocks
5090 PATH_LIMIT = 6
5091 };
5092 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5093 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5094 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5095 record_for_igvn(result_region);
5096 _gvn.set_type_bottom(result_i_o);
5097 _gvn.set_type_bottom(result_memory);
5098 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5099
5100 // The slow_control path:
5101 Node* slow_control;
5102 Node* slow_i_o = i_o();
5103 Node* slow_mem = memory(adr_type);
5104 debug_only(slow_control = (Node*) badAddress);
5105
5106 // Checked control path:
5107 Node* checked_control = top();
5108 Node* checked_mem = NULL;
5109 Node* checked_i_o = NULL;
5110 Node* checked_value = NULL;
5111
5112 if (basic_elem_type == T_CONFLICT) {
5113 assert(!dest_uninitialized, "");
5114 Node* cv = generate_generic_arraycopy(adr_type,
5115 src, src_offset, dest, dest_offset,
5116 copy_length, dest_uninitialized);
5117 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5118 checked_control = control();
5119 checked_i_o = i_o();
5120 checked_mem = memory(adr_type);
5121 checked_value = cv;
5122 set_control(top()); // no fast path
5123 }
5124
5125 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5126 if (not_pos != NULL) {
5127 PreserveJVMState pjvms(this);
5128 set_control(not_pos);
5129
5130 // (6) length must not be negative.
5131 if (!length_never_negative) {
5132 generate_negative_guard(copy_length, slow_region);
5133 }
5134
5135 // copy_length is 0.
5136 if (!stopped() && dest_uninitialized) {
5137 Node* dest_length = alloc->in(AllocateNode::ALength);
5138 if (copy_length->eqv_uncast(dest_length)
5139 || _gvn.find_int_con(dest_length, 1) <= 0) {
5140 // There is no zeroing to do. No need for a secondary raw memory barrier.
5141 } else {
5142 // Clear the whole thing since there are no source elements to copy.
5143 generate_clear_array(adr_type, dest, basic_elem_type,
5144 intcon(0), NULL,
5145 alloc->in(AllocateNode::AllocSize));
5146 // Use a secondary InitializeNode as raw memory barrier.
5147 // Currently it is needed only on this path since other
5148 // paths have stub or runtime calls as raw memory barriers.
5149 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5150 Compile::AliasIdxRaw,
5151 top())->as_Initialize();
5152 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5153 }
5154 }
5155
5156 // Present the results of the fast call.
5157 result_region->init_req(zero_path, control());
5158 result_i_o ->init_req(zero_path, i_o());
5159 result_memory->init_req(zero_path, memory(adr_type));
5160 }
5161
5162 if (!stopped() && dest_uninitialized) {
5163 // We have to initialize the *uncopied* part of the array to zero.
5164 // The copy destination is the slice dest[off..off+len]. The other slices
5165 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5166 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5167 Node* dest_length = alloc->in(AllocateNode::ALength);
5168 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5169 copy_length));
5170
5171 // If there is a head section that needs zeroing, do it now.
5172 if (find_int_con(dest_offset, -1) != 0) {
5173 generate_clear_array(adr_type, dest, basic_elem_type,
5174 intcon(0), dest_offset,
5175 NULL);
5176 }
5177
5178 // Next, perform a dynamic check on the tail length.
5179 // It is often zero, and we can win big if we prove this.
5180 // There are two wins: Avoid generating the ClearArray
5181 // with its attendant messy index arithmetic, and upgrade
5182 // the copy to a more hardware-friendly word size of 64 bits.
5183 Node* tail_ctl = NULL;
5184 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5185 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5186 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5187 tail_ctl = generate_slow_guard(bol_lt, NULL);
5188 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5189 }
5190
5191 // At this point, let's assume there is no tail.
5192 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5193 // There is no tail. Try an upgrade to a 64-bit copy.
5194 bool didit = false;
5195 { PreserveJVMState pjvms(this);
5196 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5197 src, src_offset, dest, dest_offset,
5198 dest_size, dest_uninitialized);
5199 if (didit) {
5200 // Present the results of the block-copying fast call.
5201 result_region->init_req(bcopy_path, control());
5202 result_i_o ->init_req(bcopy_path, i_o());
5203 result_memory->init_req(bcopy_path, memory(adr_type));
5204 }
5205 }
5206 if (didit)
5207 set_control(top()); // no regular fast path
5208 }
5209
5210 // Clear the tail, if any.
5211 if (tail_ctl != NULL) {
5212 Node* notail_ctl = stopped() ? NULL : control();
5213 set_control(tail_ctl);
5214 if (notail_ctl == NULL) {
5215 generate_clear_array(adr_type, dest, basic_elem_type,
5216 dest_tail, NULL,
5217 dest_size);
5218 } else {
5219 // Make a local merge.
5220 Node* done_ctl = new(C) RegionNode(3);
5221 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5222 done_ctl->init_req(1, notail_ctl);
5223 done_mem->init_req(1, memory(adr_type));
5224 generate_clear_array(adr_type, dest, basic_elem_type,
5225 dest_tail, NULL,
5226 dest_size);
5227 done_ctl->init_req(2, control());
5228 done_mem->init_req(2, memory(adr_type));
5229 set_control( _gvn.transform(done_ctl));
5230 set_memory( _gvn.transform(done_mem), adr_type );
5231 }
5232 }
5233 }
5234
5235 BasicType copy_type = basic_elem_type;
5236 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5237 if (!stopped() && copy_type == T_OBJECT) {
5238 // If src and dest have compatible element types, we can copy bits.
5239 // Types S[] and D[] are compatible if D is a supertype of S.
5240 //
5241 // If they are not, we will use checked_oop_disjoint_arraycopy,
5242 // which performs a fast optimistic per-oop check, and backs off
5243 // further to JVM_ArrayCopy on the first per-oop check that fails.
5244 // (Actually, we don't move raw bits only; the GC requires card marks.)
5245
5246 // Get the Klass* for both src and dest
5247 Node* src_klass = load_object_klass(src);
5248 Node* dest_klass = load_object_klass(dest);
5249
5250 // Generate the subtype check.
5251 // This might fold up statically, or then again it might not.
5252 //
5253 // Non-static example: Copying List<String>.elements to a new String[].
5254 // The backing store for a List<String> is always an Object[],
5255 // but its elements are always type String, if the generic types
5256 // are correct at the source level.
5257 //
5258 // Test S[] against D[], not S against D, because (probably)
5259 // the secondary supertype cache is less busy for S[] than S.
5260 // This usually only matters when D is an interface.
5261 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5262 // Plug failing path into checked_oop_disjoint_arraycopy
5263 if (not_subtype_ctrl != top()) {
5264 PreserveJVMState pjvms(this);
5265 set_control(not_subtype_ctrl);
5266 // (At this point we can assume disjoint_bases, since types differ.)
5267 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5268 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5269 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5270 Node* dest_elem_klass = _gvn.transform(n1);
5271 Node* cv = generate_checkcast_arraycopy(adr_type,
5272 dest_elem_klass,
5273 src, src_offset, dest, dest_offset,
5274 ConvI2X(copy_length), dest_uninitialized);
5275 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5276 checked_control = control();
5277 checked_i_o = i_o();
5278 checked_mem = memory(adr_type);
5279 checked_value = cv;
5280 }
5281 // At this point we know we do not need type checks on oop stores.
5282
5283 // Let's see if we need card marks:
5284 if (alloc != NULL && use_ReduceInitialCardMarks() && ! UseShenandoahGC) {
5285 // If we do not need card marks, copy using the jint or jlong stub.
5286 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5287 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5288 "sizes agree");
5289 }
5290 }
5291
5292 if (!stopped()) {
5293 // Generate the fast path, if possible.
5294 PreserveJVMState pjvms(this);
5295 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5296 src, src_offset, dest, dest_offset,
5297 ConvI2X(copy_length), dest_uninitialized);
5298
5299 // Present the results of the fast call.
5300 result_region->init_req(fast_path, control());
5301 result_i_o ->init_req(fast_path, i_o());
5302 result_memory->init_req(fast_path, memory(adr_type));
5303 }
5304
5305 // Here are all the slow paths up to this point, in one bundle:
5306 slow_control = top();
5307 if (slow_region != NULL)
5308 slow_control = _gvn.transform(slow_region);
5309 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5310
5311 set_control(checked_control);
5312 if (!stopped()) {
5313 // Clean up after the checked call.
5314 // The returned value is either 0 or -1^K,
5315 // where K = number of partially transferred array elements.
5316 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5317 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5318 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5319
5320 // If it is 0, we are done, so transfer to the end.
5321 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5322 result_region->init_req(checked_path, checks_done);
5323 result_i_o ->init_req(checked_path, checked_i_o);
5324 result_memory->init_req(checked_path, checked_mem);
5325
5326 // If it is not zero, merge into the slow call.
5327 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5328 RegionNode* slow_reg2 = new(C) RegionNode(3);
5329 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5330 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5331 record_for_igvn(slow_reg2);
5332 slow_reg2 ->init_req(1, slow_control);
5333 slow_i_o2 ->init_req(1, slow_i_o);
5334 slow_mem2 ->init_req(1, slow_mem);
5335 slow_reg2 ->init_req(2, control());
5336 slow_i_o2 ->init_req(2, checked_i_o);
5337 slow_mem2 ->init_req(2, checked_mem);
5338
5339 slow_control = _gvn.transform(slow_reg2);
5340 slow_i_o = _gvn.transform(slow_i_o2);
5341 slow_mem = _gvn.transform(slow_mem2);
5342
5343 if (alloc != NULL) {
5344 // We'll restart from the very beginning, after zeroing the whole thing.
5345 // This can cause double writes, but that's OK since dest is brand new.
5346 // So we ignore the low 31 bits of the value returned from the stub.
5347 } else {
5348 // We must continue the copy exactly where it failed, or else
5349 // another thread might see the wrong number of writes to dest.
5350 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5351 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5352 slow_offset->init_req(1, intcon(0));
5353 slow_offset->init_req(2, checked_offset);
5354 slow_offset = _gvn.transform(slow_offset);
5355
5356 // Adjust the arguments by the conditionally incoming offset.
5357 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5358 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5359 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5360
5361 // Tweak the node variables to adjust the code produced below:
5362 src_offset = src_off_plus;
5363 dest_offset = dest_off_plus;
5364 copy_length = length_minus;
5365 }
5366 }
5367
5368 set_control(slow_control);
5369 if (!stopped()) {
5370 // Generate the slow path, if needed.
5371 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5372
5373 set_memory(slow_mem, adr_type);
5374 set_i_o(slow_i_o);
5375
5376 if (dest_uninitialized) {
5377 generate_clear_array(adr_type, dest, basic_elem_type,
5378 intcon(0), NULL,
5379 alloc->in(AllocateNode::AllocSize));
5380 }
5381
5382 generate_slow_arraycopy(adr_type,
5383 src, src_offset, dest, dest_offset,
5384 copy_length, /*dest_uninitialized*/false);
5385
5386 result_region->init_req(slow_call_path, control());
5387 result_i_o ->init_req(slow_call_path, i_o());
5388 result_memory->init_req(slow_call_path, memory(adr_type));
5389 }
5390
5391 // Remove unused edges.
5392 for (uint i = 1; i < result_region->req(); i++) {
5393 if (result_region->in(i) == NULL)
5394 result_region->init_req(i, top());
5395 }
5396
5397 // Finished; return the combined state.
5398 set_control( _gvn.transform(result_region));
5399 set_i_o( _gvn.transform(result_i_o) );
5400 set_memory( _gvn.transform(result_memory), adr_type );
5401
5402 // The memory edges above are precise in order to model effects around
5403 // array copies accurately to allow value numbering of field loads around
5404 // arraycopy. Such field loads, both before and after, are common in Java
5405 // collections and similar classes involving header/array data structures.
5406 //
5407 // But with low number of register or when some registers are used or killed
5408 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5409 // The next memory barrier is added to avoid it. If the arraycopy can be
5410 // optimized away (which it can, sometimes) then we can manually remove
5411 // the membar also.
5412 //
5413 // Do not let reads from the cloned object float above the arraycopy.
5414 if (alloc != NULL) {
5415 // Do not let stores that initialize this object be reordered with
5416 // a subsequent store that would make this object accessible by
5417 // other threads.
5418 // Record what AllocateNode this StoreStore protects so that
5419 // escape analysis can go from the MemBarStoreStoreNode to the
5420 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5421 // based on the escape status of the AllocateNode.
5422 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5423 } else if (InsertMemBarAfterArraycopy)
5424 insert_mem_bar(Op_MemBarCPUOrder);
5425 }
5426
5427
5428 // Helper function which determines if an arraycopy immediately follows
5429 // an allocation, with no intervening tests or other escapes for the object.
5430 AllocateArrayNode*
5431 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5432 RegionNode* slow_region) {
5433 if (stopped()) return NULL; // no fast path
5434 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5435
5436 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5437 if (alloc == NULL) return NULL;
5438
5439 Node* rawmem = memory(Compile::AliasIdxRaw);
5440 // Is the allocation's memory state untouched?
5441 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5442 // Bail out if there have been raw-memory effects since the allocation.
5443 // (Example: There might have been a call or safepoint.)
5444 return NULL;
5445 }
5446 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5447 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5448 return NULL;
5449 }
5450
5451 // There must be no unexpected observers of this allocation.
5452 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5453 Node* obs = ptr->fast_out(i);
5454 if (obs != this->map()) {
5455 return NULL;
5456 }
5457 }
5458
5459 // This arraycopy must unconditionally follow the allocation of the ptr.
5460 Node* alloc_ctl = ptr->in(0);
5461 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5462
5463 Node* ctl = control();
5464 while (ctl != alloc_ctl) {
5465 // There may be guards which feed into the slow_region.
5466 // Any other control flow means that we might not get a chance
5467 // to finish initializing the allocated object.
5468 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5469 IfNode* iff = ctl->in(0)->as_If();
5470 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5471 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5472 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5473 ctl = iff->in(0); // This test feeds the known slow_region.
5474 continue;
5475 }
5476 // One more try: Various low-level checks bottom out in
5477 // uncommon traps. If the debug-info of the trap omits
5478 // any reference to the allocation, as we've already
5479 // observed, then there can be no objection to the trap.
5480 bool found_trap = false;
5481 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5482 Node* obs = not_ctl->fast_out(j);
5483 if (obs->in(0) == not_ctl && obs->is_Call() &&
5484 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5485 found_trap = true; break;
5486 }
5487 }
5488 if (found_trap) {
5489 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5490 continue;
5491 }
5492 }
5493 return NULL;
5494 }
5495
5496 // If we get this far, we have an allocation which immediately
5497 // precedes the arraycopy, and we can take over zeroing the new object.
5498 // The arraycopy will finish the initialization, and provide
5499 // a new control state to which we will anchor the destination pointer.
5500
5501 return alloc;
5502 }
5503
5504 // Helper for initialization of arrays, creating a ClearArray.
5505 // It writes zero bits in [start..end), within the body of an array object.
5506 // The memory effects are all chained onto the 'adr_type' alias category.
5507 //
5508 // Since the object is otherwise uninitialized, we are free
5509 // to put a little "slop" around the edges of the cleared area,
5510 // as long as it does not go back into the array's header,
5511 // or beyond the array end within the heap.
5512 //
5513 // The lower edge can be rounded down to the nearest jint and the
5514 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5515 //
5516 // Arguments:
5517 // adr_type memory slice where writes are generated
5518 // dest oop of the destination array
5519 // basic_elem_type element type of the destination
5520 // slice_idx array index of first element to store
5521 // slice_len number of elements to store (or NULL)
5522 // dest_size total size in bytes of the array object
5523 //
5524 // Exactly one of slice_len or dest_size must be non-NULL.
5525 // If dest_size is non-NULL, zeroing extends to the end of the object.
5526 // If slice_len is non-NULL, the slice_idx value must be a constant.
5527 void
5528 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5529 Node* dest,
5530 BasicType basic_elem_type,
5531 Node* slice_idx,
5532 Node* slice_len,
5533 Node* dest_size) {
5534 // one or the other but not both of slice_len and dest_size:
5535 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5536 if (slice_len == NULL) slice_len = top();
5537 if (dest_size == NULL) dest_size = top();
5538
5539 // operate on this memory slice:
5540 Node* mem = memory(adr_type); // memory slice to operate on
5541
5542 // scaling and rounding of indexes:
5543 int scale = exact_log2(type2aelembytes(basic_elem_type));
5544 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5545 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5546 int bump_bit = (-1 << scale) & BytesPerInt;
5547
5548 // determine constant starts and ends
5549 const intptr_t BIG_NEG = -128;
5550 assert(BIG_NEG + 2*abase < 0, "neg enough");
5551 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5552 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5553 if (slice_len_con == 0) {
5554 return; // nothing to do here
5555 }
5556 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5557 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5558 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5559 assert(end_con < 0, "not two cons");
5560 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5561 BytesPerLong);
5562 }
5563
5564 if (start_con >= 0 && end_con >= 0) {
5565 // Constant start and end. Simple.
5566 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5567 start_con, end_con, &_gvn);
5568 } else if (start_con >= 0 && dest_size != top()) {
5569 // Constant start, pre-rounded end after the tail of the array.
5570 Node* end = dest_size;
5571 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5572 start_con, end, &_gvn);
5573 } else if (start_con >= 0 && slice_len != top()) {
5574 // Constant start, non-constant end. End needs rounding up.
5575 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5576 intptr_t end_base = abase + (slice_idx_con << scale);
5577 int end_round = (-1 << scale) & (BytesPerLong - 1);
5578 Node* end = ConvI2X(slice_len);
5579 if (scale != 0)
5580 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5581 end_base += end_round;
5582 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5583 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5584 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5585 start_con, end, &_gvn);
5586 } else if (start_con < 0 && dest_size != top()) {
5587 // Non-constant start, pre-rounded end after the tail of the array.
5588 // This is almost certainly a "round-to-end" operation.
5589 Node* start = slice_idx;
5590 start = ConvI2X(start);
5591 if (scale != 0)
5592 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5593 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5594 if ((bump_bit | clear_low) != 0) {
5595 int to_clear = (bump_bit | clear_low);
5596 // Align up mod 8, then store a jint zero unconditionally
5597 // just before the mod-8 boundary.
5598 if (((abase + bump_bit) & ~to_clear) - bump_bit
5599 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5600 bump_bit = 0;
5601 assert((abase & to_clear) == 0, "array base must be long-aligned");
5602 } else {
5603 // Bump 'start' up to (or past) the next jint boundary:
5604 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5605 assert((abase & clear_low) == 0, "array base must be int-aligned");
5606 }
5607 // Round bumped 'start' down to jlong boundary in body of array.
5608 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5609 if (bump_bit != 0) {
5610 // Store a zero to the immediately preceding jint:
5611 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5612 Node* p1 = basic_plus_adr(dest, x1);
5613 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5614 mem = _gvn.transform(mem);
5615 }
5616 }
5617 Node* end = dest_size; // pre-rounded
5618 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5619 start, end, &_gvn);
5620 } else {
5621 // Non-constant start, unrounded non-constant end.
5622 // (Nobody zeroes a random midsection of an array using this routine.)
5623 ShouldNotReachHere(); // fix caller
5624 }
5625
5626 // Done.
5627 set_memory(mem, adr_type);
5628 }
5629
5630
5631 bool
5632 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5633 BasicType basic_elem_type,
5634 AllocateNode* alloc,
5635 Node* src, Node* src_offset,
5636 Node* dest, Node* dest_offset,
5637 Node* dest_size, bool dest_uninitialized) {
5638 // See if there is an advantage from block transfer.
5639 int scale = exact_log2(type2aelembytes(basic_elem_type));
5640 if (scale >= LogBytesPerLong)
5641 return false; // it is already a block transfer
5642
5643 // Look at the alignment of the starting offsets.
5644 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5645
5646 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5647 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5648 if (src_off_con < 0 || dest_off_con < 0)
5649 // At present, we can only understand constants.
5650 return false;
5651
5652 intptr_t src_off = abase + (src_off_con << scale);
5653 intptr_t dest_off = abase + (dest_off_con << scale);
5654
5655 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5656 // Non-aligned; too bad.
5657 // One more chance: Pick off an initial 32-bit word.
5658 // This is a common case, since abase can be odd mod 8.
5659 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5660 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5661 Node* sptr = basic_plus_adr(src, src_off);
5662 Node* dptr = basic_plus_adr(dest, dest_off);
5663 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5664 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5665 src_off += BytesPerInt;
5666 dest_off += BytesPerInt;
5667 } else {
5668 return false;
5669 }
5670 }
5671 assert(src_off % BytesPerLong == 0, "");
5672 assert(dest_off % BytesPerLong == 0, "");
5673
5674 // Do this copy by giant steps.
5675 Node* sptr = basic_plus_adr(src, src_off);
5676 Node* dptr = basic_plus_adr(dest, dest_off);
5677 Node* countx = dest_size;
5678 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5679 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5680
5681 bool disjoint_bases = true; // since alloc != NULL
5682 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5683 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5684
5685 return true;
5686 }
5687
5688
5689 // Helper function; generates code for the slow case.
5690 // We make a call to a runtime method which emulates the native method,
5691 // but without the native wrapper overhead.
5692 void
5693 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5694 Node* src, Node* src_offset,
5695 Node* dest, Node* dest_offset,
5696 Node* copy_length, bool dest_uninitialized) {
5697 assert(!dest_uninitialized, "Invariant");
5698 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5699 OptoRuntime::slow_arraycopy_Type(),
5700 OptoRuntime::slow_arraycopy_Java(),
5701 "slow_arraycopy", adr_type,
5702 src, src_offset, dest, dest_offset,
5703 copy_length);
5704
5705 // Handle exceptions thrown by this fellow:
5706 make_slow_call_ex(call, env()->Throwable_klass(), false);
5707 }
5708
5709 // Helper function; generates code for cases requiring runtime checks.
5710 Node*
5711 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5712 Node* dest_elem_klass,
5713 Node* src, Node* src_offset,
5714 Node* dest, Node* dest_offset,
5715 Node* copy_length, bool dest_uninitialized) {
5716 if (stopped()) return NULL;
5717
5718 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5719 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5720 return NULL;
5721 }
5722
5723 // Pick out the parameters required to perform a store-check
5724 // for the target array. This is an optimistic check. It will
5725 // look in each non-null element's class, at the desired klass's
5726 // super_check_offset, for the desired klass.
5727 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5728 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5729 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5730 Node* check_offset = ConvI2X(_gvn.transform(n3));
5731 Node* check_value = dest_elem_klass;
5732
5733 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5734 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5735
5736 // (We know the arrays are never conjoint, because their types differ.)
5737 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5738 OptoRuntime::checkcast_arraycopy_Type(),
5739 copyfunc_addr, "checkcast_arraycopy", adr_type,
5740 // five arguments, of which two are
5741 // intptr_t (jlong in LP64)
5742 src_start, dest_start,
5743 copy_length XTOP,
5744 check_offset XTOP,
5745 check_value);
5746
5747 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5748 }
5749
5750
5751 // Helper function; generates code for cases requiring runtime checks.
5752 Node*
5753 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5754 Node* src, Node* src_offset,
5755 Node* dest, Node* dest_offset,
5756 Node* copy_length, bool dest_uninitialized) {
5757 assert(!dest_uninitialized, "Invariant");
5758 if (stopped()) return NULL;
5759 address copyfunc_addr = StubRoutines::generic_arraycopy();
5760 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5761 return NULL;
5762 }
5763
5764 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5765 OptoRuntime::generic_arraycopy_Type(),
5766 copyfunc_addr, "generic_arraycopy", adr_type,
5767 src, src_offset, dest, dest_offset, copy_length);
5768
5769 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5770 }
5771
5772 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5773 void
5774 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5775 BasicType basic_elem_type,
5776 bool disjoint_bases,
5777 Node* src, Node* src_offset,
5778 Node* dest, Node* dest_offset,
5779 Node* copy_length, bool dest_uninitialized) {
5780 if (stopped()) return; // nothing to do
5781
5782 Node* src_start = src;
5783 Node* dest_start = dest;
5784 if (src_offset != NULL || dest_offset != NULL) {
5785 assert(src_offset != NULL && dest_offset != NULL, "");
5786 src_start = array_element_address(src, src_offset, basic_elem_type);
5787 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5788 }
5789
5790 // Figure out which arraycopy runtime method to call.
5791 const char* copyfunc_name = "arraycopy";
5792 address copyfunc_addr =
5793 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5794 disjoint_bases, copyfunc_name, dest_uninitialized);
5795
5796 // Call it. Note that the count_ix value is not scaled to a byte-size.
5797 make_runtime_call(RC_LEAF|RC_NO_FP,
5798 OptoRuntime::fast_arraycopy_Type(),
5799 copyfunc_addr, copyfunc_name, adr_type,
5800 src_start, dest_start, copy_length XTOP);
5801 }
5802
5803 //-------------inline_encodeISOArray-----------------------------------
5804 // encode char[] to byte[] in ISO_8859_1
5805 bool LibraryCallKit::inline_encodeISOArray() {
5806 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5807 // no receiver since it is static method
5808 Node *src = argument(0);
5809 Node *src_offset = argument(1);
5810 Node *dst = argument(2);
5811 Node *dst_offset = argument(3);
5812 Node *length = argument(4);
5813
5814 const Type* src_type = src->Value(&_gvn);
5815 const Type* dst_type = dst->Value(&_gvn);
5816 const TypeAryPtr* top_src = src_type->isa_aryptr();
5817 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5818 if (top_src == NULL || top_src->klass() == NULL ||
5819 top_dest == NULL || top_dest->klass() == NULL) {
5820 // failed array check
5821 return false;
5822 }
5823
5824 // Figure out the size and type of the elements we will be copying.
5825 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5826 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5827 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5828 return false;
5829 }
5830 Node* src_start = array_element_address(src, src_offset, src_elem);
5831 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5832 // 'src_start' points to src array + scaled offset
5833 // 'dst_start' points to dst array + scaled offset
5834
5835 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5836 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5837 enc = _gvn.transform(enc);
5838 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5839 set_memory(res_mem, mtype);
5840 set_result(enc);
5841 return true;
5842 }
5843
5844 //-------------inline_multiplyToLen-----------------------------------
5845 bool LibraryCallKit::inline_multiplyToLen() {
5846 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5847
5848 address stubAddr = StubRoutines::multiplyToLen();
5849 if (stubAddr == NULL) {
5850 return false; // Intrinsic's stub is not implemented on this platform
5851 }
5852 const char* stubName = "multiplyToLen";
5853
5854 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5855
5856 // no receiver because it is a static method
5857 Node* x = argument(0);
5858 Node* xlen = argument(1);
5859 Node* y = argument(2);
5860 Node* ylen = argument(3);
5861 Node* z = argument(4);
5862
5863 const Type* x_type = x->Value(&_gvn);
5864 const Type* y_type = y->Value(&_gvn);
5865 const TypeAryPtr* top_x = x_type->isa_aryptr();
5866 const TypeAryPtr* top_y = y_type->isa_aryptr();
5867 if (top_x == NULL || top_x->klass() == NULL ||
5868 top_y == NULL || top_y->klass() == NULL) {
5869 // failed array check
5870 return false;
5871 }
5872
5873 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5874 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5875 if (x_elem != T_INT || y_elem != T_INT) {
5876 return false;
5877 }
5878
5879 // Set the original stack and the reexecute bit for the interpreter to reexecute
5880 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5881 // on the return from z array allocation in runtime.
5882 { PreserveReexecuteState preexecs(this);
5883 jvms()->set_should_reexecute(true);
5884
5885 Node* x_start = array_element_address(x, intcon(0), x_elem);
5886 Node* y_start = array_element_address(y, intcon(0), y_elem);
5887 // 'x_start' points to x array + scaled xlen
5888 // 'y_start' points to y array + scaled ylen
5889
5890 // Allocate the result array
5891 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5892 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5893 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5894
5895 IdealKit ideal(this);
5896
5897 #define __ ideal.
5898 Node* one = __ ConI(1);
5899 Node* zero = __ ConI(0);
5900 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5901 __ set(need_alloc, zero);
5902 __ set(z_alloc, z);
5903 __ if_then(z, BoolTest::eq, null()); {
5904 __ increment (need_alloc, one);
5905 } __ else_(); {
5906 // Update graphKit memory and control from IdealKit.
5907 sync_kit(ideal);
5908 Node* zlen_arg = load_array_length(z);
5909 // Update IdealKit memory and control from graphKit.
5910 __ sync_kit(this);
5911 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5912 __ increment (need_alloc, one);
5913 } __ end_if();
5914 } __ end_if();
5915
5916 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5917 // Update graphKit memory and control from IdealKit.
5918 sync_kit(ideal);
5919 Node * narr = new_array(klass_node, zlen, 1);
5920 // Update IdealKit memory and control from graphKit.
5921 __ sync_kit(this);
5922 __ set(z_alloc, narr);
5923 } __ end_if();
5924
5925 sync_kit(ideal);
5926 z = __ value(z_alloc);
5927 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5928 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5929 // Final sync IdealKit and GraphKit.
5930 final_sync(ideal);
5931 #undef __
5932
5933 Node* z_start = array_element_address(z, intcon(0), T_INT);
5934
5935 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5936 OptoRuntime::multiplyToLen_Type(),
5937 stubAddr, stubName, TypePtr::BOTTOM,
5938 x_start, xlen, y_start, ylen, z_start, zlen);
5939 } // original reexecute is set back here
5940
5941 C->set_has_split_ifs(true); // Has chance for split-if optimization
5942 set_result(z);
5943 return true;
5944 }
5945
5946 //-------------inline_squareToLen------------------------------------
5947 bool LibraryCallKit::inline_squareToLen() {
5948 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5949
5950 address stubAddr = StubRoutines::squareToLen();
5951 if (stubAddr == NULL) {
5952 return false; // Intrinsic's stub is not implemented on this platform
5953 }
5954 const char* stubName = "squareToLen";
5955
5956 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5957
5958 Node* x = argument(0);
5959 Node* len = argument(1);
5960 Node* z = argument(2);
5961 Node* zlen = argument(3);
5962
5963 const Type* x_type = x->Value(&_gvn);
5964 const Type* z_type = z->Value(&_gvn);
5965 const TypeAryPtr* top_x = x_type->isa_aryptr();
5966 const TypeAryPtr* top_z = z_type->isa_aryptr();
5967 if (top_x == NULL || top_x->klass() == NULL ||
5968 top_z == NULL || top_z->klass() == NULL) {
5969 // failed array check
5970 return false;
5971 }
5972
5973 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5974 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5975 if (x_elem != T_INT || z_elem != T_INT) {
5976 return false;
5977 }
5978
5979
5980 Node* x_start = array_element_address(x, intcon(0), x_elem);
5981 Node* z_start = array_element_address(z, intcon(0), z_elem);
5982
5983 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5984 OptoRuntime::squareToLen_Type(),
5985 stubAddr, stubName, TypePtr::BOTTOM,
5986 x_start, len, z_start, zlen);
5987
5988 set_result(z);
5989 return true;
5990 }
5991
5992 //-------------inline_mulAdd------------------------------------------
5993 bool LibraryCallKit::inline_mulAdd() {
5994 assert(UseMulAddIntrinsic, "not implementated on this platform");
5995
5996 address stubAddr = StubRoutines::mulAdd();
5997 if (stubAddr == NULL) {
5998 return false; // Intrinsic's stub is not implemented on this platform
5999 }
6000 const char* stubName = "mulAdd";
6001
6002 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
6003
6004 Node* out = argument(0);
6005 Node* in = argument(1);
6006 Node* offset = argument(2);
6007 Node* len = argument(3);
6008 Node* k = argument(4);
6009
6010 const Type* out_type = out->Value(&_gvn);
6011 const Type* in_type = in->Value(&_gvn);
6012 const TypeAryPtr* top_out = out_type->isa_aryptr();
6013 const TypeAryPtr* top_in = in_type->isa_aryptr();
6014 if (top_out == NULL || top_out->klass() == NULL ||
6015 top_in == NULL || top_in->klass() == NULL) {
6016 // failed array check
6017 return false;
6018 }
6019
6020 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6021 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6022 if (out_elem != T_INT || in_elem != T_INT) {
6023 return false;
6024 }
6025
6026 Node* outlen = load_array_length(out);
6027 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
6028 Node* out_start = array_element_address(out, intcon(0), out_elem);
6029 Node* in_start = array_element_address(in, intcon(0), in_elem);
6030
6031 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6032 OptoRuntime::mulAdd_Type(),
6033 stubAddr, stubName, TypePtr::BOTTOM,
6034 out_start,in_start, new_offset, len, k);
6035 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6036 set_result(result);
6037 return true;
6038 }
6039
6040 //-------------inline_montgomeryMultiply-----------------------------------
6041 bool LibraryCallKit::inline_montgomeryMultiply() {
6042 address stubAddr = StubRoutines::montgomeryMultiply();
6043 if (stubAddr == NULL) {
6044 return false; // Intrinsic's stub is not implemented on this platform
6045 }
6046
6047 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6048 const char* stubName = "montgomery_multiply";
6049
6050 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6051
6052 Node* a = argument(0);
6053 Node* b = argument(1);
6054 Node* n = argument(2);
6055 Node* len = argument(3);
6056 Node* inv = argument(4);
6057 Node* m = argument(6);
6058
6059 const Type* a_type = a->Value(&_gvn);
6060 const TypeAryPtr* top_a = a_type->isa_aryptr();
6061 const Type* b_type = b->Value(&_gvn);
6062 const TypeAryPtr* top_b = b_type->isa_aryptr();
6063 const Type* n_type = a->Value(&_gvn);
6064 const TypeAryPtr* top_n = n_type->isa_aryptr();
6065 const Type* m_type = a->Value(&_gvn);
6066 const TypeAryPtr* top_m = m_type->isa_aryptr();
6067 if (top_a == NULL || top_a->klass() == NULL ||
6068 top_b == NULL || top_b->klass() == NULL ||
6069 top_n == NULL || top_n->klass() == NULL ||
6070 top_m == NULL || top_m->klass() == NULL) {
6071 // failed array check
6072 return false;
6073 }
6074
6075 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6076 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6077 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6078 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6079 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6080 return false;
6081 }
6082
6083 // Make the call
6084 {
6085 Node* a_start = array_element_address(a, intcon(0), a_elem);
6086 Node* b_start = array_element_address(b, intcon(0), b_elem);
6087 Node* n_start = array_element_address(n, intcon(0), n_elem);
6088 Node* m_start = array_element_address(m, intcon(0), m_elem);
6089
6090 Node* call = NULL;
6091 if (CCallingConventionRequiresIntsAsLongs) {
6092 Node* len_I2L = ConvI2L(len);
6093 call = make_runtime_call(RC_LEAF,
6094 OptoRuntime::montgomeryMultiply_Type(),
6095 stubAddr, stubName, TypePtr::BOTTOM,
6096 a_start, b_start, n_start, len_I2L XTOP, inv,
6097 top(), m_start);
6098 } else {
6099 call = make_runtime_call(RC_LEAF,
6100 OptoRuntime::montgomeryMultiply_Type(),
6101 stubAddr, stubName, TypePtr::BOTTOM,
6102 a_start, b_start, n_start, len, inv, top(),
6103 m_start);
6104 }
6105 set_result(m);
6106 }
6107
6108 return true;
6109 }
6110
6111 bool LibraryCallKit::inline_montgomerySquare() {
6112 address stubAddr = StubRoutines::montgomerySquare();
6113 if (stubAddr == NULL) {
6114 return false; // Intrinsic's stub is not implemented on this platform
6115 }
6116
6117 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6118 const char* stubName = "montgomery_square";
6119
6120 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6121
6122 Node* a = argument(0);
6123 Node* n = argument(1);
6124 Node* len = argument(2);
6125 Node* inv = argument(3);
6126 Node* m = argument(5);
6127
6128 const Type* a_type = a->Value(&_gvn);
6129 const TypeAryPtr* top_a = a_type->isa_aryptr();
6130 const Type* n_type = a->Value(&_gvn);
6131 const TypeAryPtr* top_n = n_type->isa_aryptr();
6132 const Type* m_type = a->Value(&_gvn);
6133 const TypeAryPtr* top_m = m_type->isa_aryptr();
6134 if (top_a == NULL || top_a->klass() == NULL ||
6135 top_n == NULL || top_n->klass() == NULL ||
6136 top_m == NULL || top_m->klass() == NULL) {
6137 // failed array check
6138 return false;
6139 }
6140
6141 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6142 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6143 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6144 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6145 return false;
6146 }
6147
6148 // Make the call
6149 {
6150 Node* a_start = array_element_address(a, intcon(0), a_elem);
6151 Node* n_start = array_element_address(n, intcon(0), n_elem);
6152 Node* m_start = array_element_address(m, intcon(0), m_elem);
6153
6154 Node* call = NULL;
6155 if (CCallingConventionRequiresIntsAsLongs) {
6156 Node* len_I2L = ConvI2L(len);
6157 call = make_runtime_call(RC_LEAF,
6158 OptoRuntime::montgomerySquare_Type(),
6159 stubAddr, stubName, TypePtr::BOTTOM,
6160 a_start, n_start, len_I2L XTOP, inv, top(),
6161 m_start);
6162 } else {
6163 call = make_runtime_call(RC_LEAF,
6164 OptoRuntime::montgomerySquare_Type(),
6165 stubAddr, stubName, TypePtr::BOTTOM,
6166 a_start, n_start, len, inv, top(),
6167 m_start);
6168 }
6169
6170 set_result(m);
6171 }
6172
6173 return true;
6174 }
6175
6176
6177 /**
6178 * Calculate CRC32 for byte.
6179 * int java.util.zip.CRC32.update(int crc, int b)
6180 */
6181 bool LibraryCallKit::inline_updateCRC32() {
6182 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6183 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6184 // no receiver since it is static method
6185 Node* crc = argument(0); // type: int
6186 Node* b = argument(1); // type: int
6187
6188 /*
6189 * int c = ~ crc;
6190 * b = timesXtoThe32[(b ^ c) & 0xFF];
6191 * b = b ^ (c >>> 8);
6192 * crc = ~b;
6193 */
6194
6195 Node* M1 = intcon(-1);
6196 crc = _gvn.transform(new (C) XorINode(crc, M1));
6197 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6198 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6199
6200 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6201 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6202 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6203 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6204
6205 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6206 result = _gvn.transform(new (C) XorINode(crc, result));
6207 result = _gvn.transform(new (C) XorINode(result, M1));
6208 set_result(result);
6209 return true;
6210 }
6211
6212 /**
6213 * Calculate CRC32 for byte[] array.
6214 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6215 */
6216 bool LibraryCallKit::inline_updateBytesCRC32() {
6217 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6218 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6219 // no receiver since it is static method
6220 Node* crc = argument(0); // type: int
6221 Node* src = argument(1); // type: oop
6222 Node* offset = argument(2); // type: int
6223 Node* length = argument(3); // type: int
6224
6225 const Type* src_type = src->Value(&_gvn);
6226 const TypeAryPtr* top_src = src_type->isa_aryptr();
6227 if (top_src == NULL || top_src->klass() == NULL) {
6228 // failed array check
6229 return false;
6230 }
6231
6232 // Figure out the size and type of the elements we will be copying.
6233 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6234 if (src_elem != T_BYTE) {
6235 return false;
6236 }
6237
6238 // 'src_start' points to src array + scaled offset
6239 Node* src_start = array_element_address(src, offset, src_elem);
6240
6241 // We assume that range check is done by caller.
6242 // TODO: generate range check (offset+length < src.length) in debug VM.
6243
6244 // Call the stub.
6245 address stubAddr = StubRoutines::updateBytesCRC32();
6246 const char *stubName = "updateBytesCRC32";
6247 Node* call;
6248 if (CCallingConventionRequiresIntsAsLongs) {
6249 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6250 stubAddr, stubName, TypePtr::BOTTOM,
6251 crc XTOP, src_start, length XTOP);
6252 } else {
6253 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6254 stubAddr, stubName, TypePtr::BOTTOM,
6255 crc, src_start, length);
6256 }
6257 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6258 set_result(result);
6259 return true;
6260 }
6261
6262 /**
6263 * Calculate CRC32 for ByteBuffer.
6264 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6265 */
6266 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6267 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6268 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6269 // no receiver since it is static method
6270 Node* crc = argument(0); // type: int
6271 Node* src = argument(1); // type: long
6272 Node* offset = argument(3); // type: int
6273 Node* length = argument(4); // type: int
6274
6275 src = ConvL2X(src); // adjust Java long to machine word
6276 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6277 offset = ConvI2X(offset);
6278
6279 // 'src_start' points to src array + scaled offset
6280 Node* src_start = basic_plus_adr(top(), base, offset);
6281
6282 // Call the stub.
6283 address stubAddr = StubRoutines::updateBytesCRC32();
6284 const char *stubName = "updateBytesCRC32";
6285 Node* call;
6286 if (CCallingConventionRequiresIntsAsLongs) {
6287 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6288 stubAddr, stubName, TypePtr::BOTTOM,
6289 crc XTOP, src_start, length XTOP);
6290 } else {
6291 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6292 stubAddr, stubName, TypePtr::BOTTOM,
6293 crc, src_start, length);
6294 }
6295 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6296 set_result(result);
6297 return true;
6298 }
6299
6300 //----------------------------inline_reference_get----------------------------
6301 // public T java.lang.ref.Reference.get();
6302 bool LibraryCallKit::inline_reference_get() {
6303 const int referent_offset = java_lang_ref_Reference::referent_offset;
6304 guarantee(referent_offset > 0, "should have already been set");
6305
6306 // Get the argument:
6307 Node* reference_obj = null_check_receiver();
6308 if (stopped()) return true;
6309
6310 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6311
6312 ciInstanceKlass* klass = env()->Object_klass();
6313 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6314
6315 Node* no_ctrl = NULL;
6316 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6317
6318 #if INCLUDE_ALL_GCS
6319 if (UseShenandoahGC) {
6320 result = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, result);
6321 }
6322 #endif
6323
6324 // Use the pre-barrier to record the value in the referent field
6325 pre_barrier(false /* do_load */,
6326 control(),
6327 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6328 result /* pre_val */,
6329 T_OBJECT);
6330
6331 // Add memory barrier to prevent commoning reads from this field
6332 // across safepoint since GC can change its value.
6333 insert_mem_bar(Op_MemBarCPUOrder);
6334
6335 set_result(result);
6336 return true;
6337 }
6338
6339
6340 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6341 bool is_exact=true, bool is_static=false) {
6342
6343 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6344 assert(tinst != NULL, "obj is null");
6345 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6346 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6347
6348 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6349 ciSymbol::make(fieldTypeString),
6350 is_static);
6351 if (field == NULL) return (Node *) NULL;
6352 assert (field != NULL, "undefined field");
6353
6354 // Next code copied from Parse::do_get_xxx():
6355
6356 // Compute address and memory type.
6357 int offset = field->offset_in_bytes();
6358 bool is_vol = field->is_volatile();
6359 ciType* field_klass = field->type();
6360 assert(field_klass->is_loaded(), "should be loaded");
6361 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6362 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6363 BasicType bt = field->layout_type();
6364
6365 // Build the resultant type of the load
6366 const Type *type;
6367 if (bt == T_OBJECT) {
6368 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6369 } else {
6370 type = Type::get_const_basic_type(bt);
6371 }
6372
6373 Node* leading_membar = NULL;
6374 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6375 leading_membar = insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6376 }
6377 // Build the load.
6378 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6379 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6380 #if INCLUDE_ALL_GCS
6381 if (UseShenandoahGC && (bt == T_OBJECT || bt == T_ARRAY)) {
6382 loadedField = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, loadedField);
6383 }
6384 #endif
6385
6386 // If reference is volatile, prevent following memory ops from
6387 // floating up past the volatile read. Also prevents commoning
6388 // another volatile read.
6389 if (is_vol) {
6390 // Memory barrier includes bogus read of value to force load BEFORE membar
6391 Node* mb = insert_mem_bar(Op_MemBarAcquire, loadedField);
6392 mb->as_MemBar()->set_trailing_load();
6393 }
6394 return loadedField;
6395 }
6396
6397
6398 //------------------------------inline_aescrypt_Block-----------------------
6399 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6400 address stubAddr = NULL;
6401 const char *stubName;
6402 assert(UseAES, "need AES instruction support");
6403
6404 switch(id) {
6405 case vmIntrinsics::_aescrypt_encryptBlock:
6406 stubAddr = StubRoutines::aescrypt_encryptBlock();
6407 stubName = "aescrypt_encryptBlock";
6408 break;
6409 case vmIntrinsics::_aescrypt_decryptBlock:
6410 stubAddr = StubRoutines::aescrypt_decryptBlock();
6411 stubName = "aescrypt_decryptBlock";
6412 break;
6413 }
6414 if (stubAddr == NULL) return false;
6415
6416 Node* aescrypt_object = argument(0);
6417 Node* src = argument(1);
6418 Node* src_offset = argument(2);
6419 Node* dest = argument(3);
6420 Node* dest_offset = argument(4);
6421
6422 // (1) src and dest are arrays.
6423 const Type* src_type = src->Value(&_gvn);
6424 const Type* dest_type = dest->Value(&_gvn);
6425 const TypeAryPtr* top_src = src_type->isa_aryptr();
6426 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6427 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6428
6429 // for the quick and dirty code we will skip all the checks.
6430 // we are just trying to get the call to be generated.
6431 Node* src_start = src;
6432 Node* dest_start = dest;
6433 if (src_offset != NULL || dest_offset != NULL) {
6434 assert(src_offset != NULL && dest_offset != NULL, "");
6435 src_start = array_element_address(src, src_offset, T_BYTE);
6436 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6437 }
6438
6439 // now need to get the start of its expanded key array
6440 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6441 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6442 if (k_start == NULL) return false;
6443
6444 if (Matcher::pass_original_key_for_aes()) {
6445 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6446 // compatibility issues between Java key expansion and SPARC crypto instructions
6447 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6448 if (original_k_start == NULL) return false;
6449
6450 // Call the stub.
6451 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6452 stubAddr, stubName, TypePtr::BOTTOM,
6453 src_start, dest_start, k_start, original_k_start);
6454 } else {
6455 // Call the stub.
6456 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6457 stubAddr, stubName, TypePtr::BOTTOM,
6458 src_start, dest_start, k_start);
6459 }
6460
6461 return true;
6462 }
6463
6464 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6465 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6466 address stubAddr = NULL;
6467 const char *stubName = NULL;
6468
6469 assert(UseAES, "need AES instruction support");
6470
6471 switch(id) {
6472 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6473 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6474 stubName = "cipherBlockChaining_encryptAESCrypt";
6475 break;
6476 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6477 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6478 stubName = "cipherBlockChaining_decryptAESCrypt";
6479 break;
6480 }
6481 if (stubAddr == NULL) return false;
6482
6483 Node* cipherBlockChaining_object = argument(0);
6484 Node* src = argument(1);
6485 Node* src_offset = argument(2);
6486 Node* len = argument(3);
6487 Node* dest = argument(4);
6488 Node* dest_offset = argument(5);
6489
6490 // (1) src and dest are arrays.
6491 const Type* src_type = src->Value(&_gvn);
6492 const Type* dest_type = dest->Value(&_gvn);
6493 const TypeAryPtr* top_src = src_type->isa_aryptr();
6494 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6495 assert (top_src != NULL && top_src->klass() != NULL
6496 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6497
6498 // checks are the responsibility of the caller
6499 Node* src_start = src;
6500 Node* dest_start = dest;
6501 if (src_offset != NULL || dest_offset != NULL) {
6502 assert(src_offset != NULL && dest_offset != NULL, "");
6503 src_start = array_element_address(src, src_offset, T_BYTE);
6504 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6505 }
6506
6507 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6508 // (because of the predicated logic executed earlier).
6509 // so we cast it here safely.
6510 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6511
6512 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6513 if (embeddedCipherObj == NULL) return false;
6514
6515 // cast it to what we know it will be at runtime
6516 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6517 assert(tinst != NULL, "CBC obj is null");
6518 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6519 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6520 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6521
6522 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6523 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6524 const TypeOopPtr* xtype = aklass->as_instance_type();
6525 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6526 aescrypt_object = _gvn.transform(aescrypt_object);
6527
6528 // we need to get the start of the aescrypt_object's expanded key array
6529 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6530 if (k_start == NULL) return false;
6531
6532 // similarly, get the start address of the r vector
6533 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6534 if (objRvec == NULL) return false;
6535 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6536
6537 Node* cbcCrypt;
6538 if (Matcher::pass_original_key_for_aes()) {
6539 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6540 // compatibility issues between Java key expansion and SPARC crypto instructions
6541 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6542 if (original_k_start == NULL) return false;
6543
6544 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6545 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6546 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6547 stubAddr, stubName, TypePtr::BOTTOM,
6548 src_start, dest_start, k_start, r_start, len, original_k_start);
6549 } else {
6550 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6551 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6552 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6553 stubAddr, stubName, TypePtr::BOTTOM,
6554 src_start, dest_start, k_start, r_start, len);
6555 }
6556
6557 // return cipher length (int)
6558 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6559 set_result(retvalue);
6560 return true;
6561 }
6562
6563 //------------------------------get_key_start_from_aescrypt_object-----------------------
6564 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6565 #ifdef PPC64
6566 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6567 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6568 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6569 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6570 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6571 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6572 if (objSessionK == NULL) {
6573 return (Node *) NULL;
6574 }
6575 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6576 #else
6577 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6578 #endif // PPC64
6579 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6580 if (objAESCryptKey == NULL) return (Node *) NULL;
6581
6582 // now have the array, need to get the start address of the K array
6583 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6584 return k_start;
6585 }
6586
6587 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6588 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6589 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6590 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6591 if (objAESCryptKey == NULL) return (Node *) NULL;
6592
6593 // now have the array, need to get the start address of the lastKey array
6594 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6595 return original_k_start;
6596 }
6597
6598 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6599 // Return node representing slow path of predicate check.
6600 // the pseudo code we want to emulate with this predicate is:
6601 // for encryption:
6602 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6603 // for decryption:
6604 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6605 // note cipher==plain is more conservative than the original java code but that's OK
6606 //
6607 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6608 // The receiver was checked for NULL already.
6609 Node* objCBC = argument(0);
6610
6611 // Load embeddedCipher field of CipherBlockChaining object.
6612 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6613
6614 // get AESCrypt klass for instanceOf check
6615 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6616 // will have same classloader as CipherBlockChaining object
6617 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6618 assert(tinst != NULL, "CBCobj is null");
6619 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6620
6621 // we want to do an instanceof comparison against the AESCrypt class
6622 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6623 if (!klass_AESCrypt->is_loaded()) {
6624 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6625 Node* ctrl = control();
6626 set_control(top()); // no regular fast path
6627 return ctrl;
6628 }
6629 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6630
6631 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6632 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6633 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6634
6635 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6636
6637 // for encryption, we are done
6638 if (!decrypting)
6639 return instof_false; // even if it is NULL
6640
6641 // for decryption, we need to add a further check to avoid
6642 // taking the intrinsic path when cipher and plain are the same
6643 // see the original java code for why.
6644 RegionNode* region = new(C) RegionNode(3);
6645 region->init_req(1, instof_false);
6646 Node* src = argument(1);
6647 Node* dest = argument(4);
6648 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6649 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6650 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6651 region->init_req(2, src_dest_conjoint);
6652
6653 record_for_igvn(region);
6654 return _gvn.transform(region);
6655 }
6656
6657 //------------------------------inline_sha_implCompress-----------------------
6658 //
6659 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6660 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6661 //
6662 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6663 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6664 //
6665 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6666 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6667 //
6668 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6669 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6670
6671 Node* sha_obj = argument(0);
6672 Node* src = argument(1); // type oop
6673 Node* ofs = argument(2); // type int
6674
6675 const Type* src_type = src->Value(&_gvn);
6676 const TypeAryPtr* top_src = src_type->isa_aryptr();
6677 if (top_src == NULL || top_src->klass() == NULL) {
6678 // failed array check
6679 return false;
6680 }
6681 // Figure out the size and type of the elements we will be copying.
6682 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6683 if (src_elem != T_BYTE) {
6684 return false;
6685 }
6686 // 'src_start' points to src array + offset
6687 Node* src_start = array_element_address(src, ofs, src_elem);
6688 Node* state = NULL;
6689 address stubAddr;
6690 const char *stubName;
6691
6692 switch(id) {
6693 case vmIntrinsics::_sha_implCompress:
6694 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6695 state = get_state_from_sha_object(sha_obj);
6696 stubAddr = StubRoutines::sha1_implCompress();
6697 stubName = "sha1_implCompress";
6698 break;
6699 case vmIntrinsics::_sha2_implCompress:
6700 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6701 state = get_state_from_sha_object(sha_obj);
6702 stubAddr = StubRoutines::sha256_implCompress();
6703 stubName = "sha256_implCompress";
6704 break;
6705 case vmIntrinsics::_sha5_implCompress:
6706 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6707 state = get_state_from_sha5_object(sha_obj);
6708 stubAddr = StubRoutines::sha512_implCompress();
6709 stubName = "sha512_implCompress";
6710 break;
6711 default:
6712 fatal_unexpected_iid(id);
6713 return false;
6714 }
6715 if (state == NULL) return false;
6716
6717 // Call the stub.
6718 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6719 stubAddr, stubName, TypePtr::BOTTOM,
6720 src_start, state);
6721
6722 return true;
6723 }
6724
6725 //------------------------------inline_digestBase_implCompressMB-----------------------
6726 //
6727 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6728 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6729 //
6730 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6731 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6732 "need SHA1/SHA256/SHA512 instruction support");
6733 assert((uint)predicate < 3, "sanity");
6734 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6735
6736 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6737 Node* src = argument(1); // byte[] array
6738 Node* ofs = argument(2); // type int
6739 Node* limit = argument(3); // type int
6740
6741 const Type* src_type = src->Value(&_gvn);
6742 const TypeAryPtr* top_src = src_type->isa_aryptr();
6743 if (top_src == NULL || top_src->klass() == NULL) {
6744 // failed array check
6745 return false;
6746 }
6747 // Figure out the size and type of the elements we will be copying.
6748 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6749 if (src_elem != T_BYTE) {
6750 return false;
6751 }
6752 // 'src_start' points to src array + offset
6753 Node* src_start = array_element_address(src, ofs, src_elem);
6754
6755 const char* klass_SHA_name = NULL;
6756 const char* stub_name = NULL;
6757 address stub_addr = NULL;
6758 bool long_state = false;
6759
6760 switch (predicate) {
6761 case 0:
6762 if (UseSHA1Intrinsics) {
6763 klass_SHA_name = "sun/security/provider/SHA";
6764 stub_name = "sha1_implCompressMB";
6765 stub_addr = StubRoutines::sha1_implCompressMB();
6766 }
6767 break;
6768 case 1:
6769 if (UseSHA256Intrinsics) {
6770 klass_SHA_name = "sun/security/provider/SHA2";
6771 stub_name = "sha256_implCompressMB";
6772 stub_addr = StubRoutines::sha256_implCompressMB();
6773 }
6774 break;
6775 case 2:
6776 if (UseSHA512Intrinsics) {
6777 klass_SHA_name = "sun/security/provider/SHA5";
6778 stub_name = "sha512_implCompressMB";
6779 stub_addr = StubRoutines::sha512_implCompressMB();
6780 long_state = true;
6781 }
6782 break;
6783 default:
6784 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6785 }
6786 if (klass_SHA_name != NULL) {
6787 // get DigestBase klass to lookup for SHA klass
6788 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6789 assert(tinst != NULL, "digestBase_obj is not instance???");
6790 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6791
6792 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6793 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6794 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6795 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6796 }
6797 return false;
6798 }
6799 //------------------------------inline_sha_implCompressMB-----------------------
6800 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6801 bool long_state, address stubAddr, const char *stubName,
6802 Node* src_start, Node* ofs, Node* limit) {
6803 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6804 const TypeOopPtr* xtype = aklass->as_instance_type();
6805 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6806 sha_obj = _gvn.transform(sha_obj);
6807
6808 Node* state;
6809 if (long_state) {
6810 state = get_state_from_sha5_object(sha_obj);
6811 } else {
6812 state = get_state_from_sha_object(sha_obj);
6813 }
6814 if (state == NULL) return false;
6815
6816 // Call the stub.
6817 Node *call;
6818 if (CCallingConventionRequiresIntsAsLongs) {
6819 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6820 OptoRuntime::digestBase_implCompressMB_Type(),
6821 stubAddr, stubName, TypePtr::BOTTOM,
6822 src_start, state, ofs XTOP, limit XTOP);
6823 } else {
6824 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6825 OptoRuntime::digestBase_implCompressMB_Type(),
6826 stubAddr, stubName, TypePtr::BOTTOM,
6827 src_start, state, ofs, limit);
6828 }
6829 // return ofs (int)
6830 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6831 set_result(result);
6832
6833 return true;
6834 }
6835
6836 //------------------------------get_state_from_sha_object-----------------------
6837 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6838 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6839 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6840 if (sha_state == NULL) return (Node *) NULL;
6841
6842 // now have the array, need to get the start address of the state array
6843 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6844 return state;
6845 }
6846
6847 //------------------------------get_state_from_sha5_object-----------------------
6848 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6849 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6850 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6851 if (sha_state == NULL) return (Node *) NULL;
6852
6853 // now have the array, need to get the start address of the state array
6854 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6855 return state;
6856 }
6857
6858 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6859 // Return node representing slow path of predicate check.
6860 // the pseudo code we want to emulate with this predicate is:
6861 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6862 //
6863 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6864 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6865 "need SHA1/SHA256/SHA512 instruction support");
6866 assert((uint)predicate < 3, "sanity");
6867
6868 // The receiver was checked for NULL already.
6869 Node* digestBaseObj = argument(0);
6870
6871 // get DigestBase klass for instanceOf check
6872 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6873 assert(tinst != NULL, "digestBaseObj is null");
6874 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6875
6876 const char* klass_SHA_name = NULL;
6877 switch (predicate) {
6878 case 0:
6879 if (UseSHA1Intrinsics) {
6880 // we want to do an instanceof comparison against the SHA class
6881 klass_SHA_name = "sun/security/provider/SHA";
6882 }
6883 break;
6884 case 1:
6885 if (UseSHA256Intrinsics) {
6886 // we want to do an instanceof comparison against the SHA2 class
6887 klass_SHA_name = "sun/security/provider/SHA2";
6888 }
6889 break;
6890 case 2:
6891 if (UseSHA512Intrinsics) {
6892 // we want to do an instanceof comparison against the SHA5 class
6893 klass_SHA_name = "sun/security/provider/SHA5";
6894 }
6895 break;
6896 default:
6897 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6898 }
6899
6900 ciKlass* klass_SHA = NULL;
6901 if (klass_SHA_name != NULL) {
6902 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6903 }
6904 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6905 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6906 Node* ctrl = control();
6907 set_control(top()); // no intrinsic path
6908 return ctrl;
6909 }
6910 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6911
6912 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6913 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6914 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6915 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6916
6917 return instof_false; // even if it is NULL
6918 }
6919
6920 bool LibraryCallKit::inline_profileBoolean() {
6921 Node* counts = argument(1);
6922 const TypeAryPtr* ary = NULL;
6923 ciArray* aobj = NULL;
6924 if (counts->is_Con()
6925 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6926 && (aobj = ary->const_oop()->as_array()) != NULL
6927 && (aobj->length() == 2)) {
6928 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6929 jint false_cnt = aobj->element_value(0).as_int();
6930 jint true_cnt = aobj->element_value(1).as_int();
6931
6932 if (C->log() != NULL) {
6933 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6934 false_cnt, true_cnt);
6935 }
6936
6937 if (false_cnt + true_cnt == 0) {
6938 // According to profile, never executed.
6939 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6940 Deoptimization::Action_reinterpret);
6941 return true;
6942 }
6943
6944 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6945 // is a number of each value occurrences.
6946 Node* result = argument(0);
6947 if (false_cnt == 0 || true_cnt == 0) {
6948 // According to profile, one value has been never seen.
6949 int expected_val = (false_cnt == 0) ? 1 : 0;
6950
6951 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6952 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6953
6954 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6955 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6956 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6957
6958 { // Slow path: uncommon trap for never seen value and then reexecute
6959 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6960 // the value has been seen at least once.
6961 PreserveJVMState pjvms(this);
6962 PreserveReexecuteState preexecs(this);
6963 jvms()->set_should_reexecute(true);
6964
6965 set_control(slow_path);
6966 set_i_o(i_o());
6967
6968 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6969 Deoptimization::Action_reinterpret);
6970 }
6971 // The guard for never seen value enables sharpening of the result and
6972 // returning a constant. It allows to eliminate branches on the same value
6973 // later on.
6974 set_control(fast_path);
6975 result = intcon(expected_val);
6976 }
6977 // Stop profiling.
6978 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6979 // By replacing method body with profile data (represented as ProfileBooleanNode
6980 // on IR level) we effectively disable profiling.
6981 // It enables full speed execution once optimized code is generated.
6982 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6983 C->record_for_igvn(profile);
6984 set_result(profile);
6985 return true;
6986 } else {
6987 // Continue profiling.
6988 // Profile data isn't available at the moment. So, execute method's bytecode version.
6989 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6990 // is compiled and counters aren't available since corresponding MethodHandle
6991 // isn't a compile-time constant.
6992 return false;
6993 }
6994 }