Wed, 10 Apr 2019 11:38:47 +0200
8221355: Performance regression after JDK-8155635 backport into 8u
Reviewed-by: shade, roland
1 /*
2 * Copyright (c) 1999, 2018, 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 */
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"
45 class LibraryIntrinsic : public InlineCallGenerator {
46 // Extend the set of intrinsics known to the runtime:
47 public:
48 private:
49 bool _is_virtual;
50 bool _does_virtual_dispatch;
51 int8_t _predicates_count; // Intrinsic is predicated by several conditions
52 int8_t _last_predicate; // Last generated predicate
53 vmIntrinsics::ID _intrinsic_id;
55 public:
56 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
57 : InlineCallGenerator(m),
58 _is_virtual(is_virtual),
59 _does_virtual_dispatch(does_virtual_dispatch),
60 _predicates_count((int8_t)predicates_count),
61 _last_predicate((int8_t)-1),
62 _intrinsic_id(id)
63 {
64 }
65 virtual bool is_intrinsic() const { return true; }
66 virtual bool is_virtual() const { return _is_virtual; }
67 virtual bool is_predicated() const { return _predicates_count > 0; }
68 virtual int predicates_count() const { return _predicates_count; }
69 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
70 virtual JVMState* generate(JVMState* jvms);
71 virtual Node* generate_predicate(JVMState* jvms, int predicate);
72 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
73 };
76 // Local helper class for LibraryIntrinsic:
77 class LibraryCallKit : public GraphKit {
78 private:
79 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
80 Node* _result; // the result node, if any
81 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
83 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
85 public:
86 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
87 : GraphKit(jvms),
88 _intrinsic(intrinsic),
89 _result(NULL)
90 {
91 // Check if this is a root compile. In that case we don't have a caller.
92 if (!jvms->has_method()) {
93 _reexecute_sp = sp();
94 } else {
95 // Find out how many arguments the interpreter needs when deoptimizing
96 // and save the stack pointer value so it can used by uncommon_trap.
97 // We find the argument count by looking at the declared signature.
98 bool ignored_will_link;
99 ciSignature* declared_signature = NULL;
100 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
101 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
102 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
103 }
104 }
106 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
108 ciMethod* caller() const { return jvms()->method(); }
109 int bci() const { return jvms()->bci(); }
110 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
111 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
112 ciMethod* callee() const { return _intrinsic->method(); }
114 bool try_to_inline(int predicate);
115 Node* try_to_predicate(int predicate);
117 void push_result() {
118 // Push the result onto the stack.
119 if (!stopped() && result() != NULL) {
120 BasicType bt = result()->bottom_type()->basic_type();
121 push_node(bt, result());
122 }
123 }
125 private:
126 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
127 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
128 }
130 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
131 void set_result(RegionNode* region, PhiNode* value);
132 Node* result() { return _result; }
134 virtual int reexecute_sp() { return _reexecute_sp; }
136 // Helper functions to inline natives
137 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
138 Node* generate_slow_guard(Node* test, RegionNode* region);
139 Node* generate_fair_guard(Node* test, RegionNode* region);
140 Node* generate_negative_guard(Node* index, RegionNode* region,
141 // resulting CastII of index:
142 Node* *pos_index = NULL);
143 Node* generate_nonpositive_guard(Node* index, bool never_negative,
144 // resulting CastII of index:
145 Node* *pos_index = NULL);
146 Node* generate_limit_guard(Node* offset, Node* subseq_length,
147 Node* array_length,
148 RegionNode* region);
149 Node* generate_current_thread(Node* &tls_output);
150 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
151 bool disjoint_bases, const char* &name, bool dest_uninitialized);
152 Node* load_mirror_from_klass(Node* klass);
153 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
154 RegionNode* region, int null_path,
155 int offset);
156 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
157 RegionNode* region, int null_path) {
158 int offset = java_lang_Class::klass_offset_in_bytes();
159 return load_klass_from_mirror_common(mirror, never_see_null,
160 region, null_path,
161 offset);
162 }
163 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path) {
165 int offset = java_lang_Class::array_klass_offset_in_bytes();
166 return load_klass_from_mirror_common(mirror, never_see_null,
167 region, null_path,
168 offset);
169 }
170 Node* generate_access_flags_guard(Node* kls,
171 int modifier_mask, int modifier_bits,
172 RegionNode* region);
173 Node* generate_interface_guard(Node* kls, RegionNode* region);
174 Node* generate_array_guard(Node* kls, RegionNode* region) {
175 return generate_array_guard_common(kls, region, false, false);
176 }
177 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
178 return generate_array_guard_common(kls, region, false, true);
179 }
180 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, true, false);
182 }
183 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, true, true);
185 }
186 Node* generate_array_guard_common(Node* kls, RegionNode* region,
187 bool obj_array, bool not_array);
188 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
189 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
190 bool is_virtual = false, bool is_static = false);
191 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
192 return generate_method_call(method_id, false, true);
193 }
194 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
195 return generate_method_call(method_id, true, false);
196 }
197 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
199 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
200 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
201 bool inline_string_compareTo();
202 bool inline_string_indexOf();
203 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
204 bool inline_string_equals();
205 Node* round_double_node(Node* n);
206 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
207 bool inline_math_native(vmIntrinsics::ID id);
208 bool inline_trig(vmIntrinsics::ID id);
209 bool inline_math(vmIntrinsics::ID id);
210 template <typename OverflowOp>
211 bool inline_math_overflow(Node* arg1, Node* arg2);
212 void inline_math_mathExact(Node* math, Node* test);
213 bool inline_math_addExactI(bool is_increment);
214 bool inline_math_addExactL(bool is_increment);
215 bool inline_math_multiplyExactI();
216 bool inline_math_multiplyExactL();
217 bool inline_math_negateExactI();
218 bool inline_math_negateExactL();
219 bool inline_math_subtractExactI(bool is_decrement);
220 bool inline_math_subtractExactL(bool is_decrement);
221 bool inline_exp();
222 bool inline_pow();
223 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_min_max(vmIntrinsics::ID id);
225 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
226 // This returns Type::AnyPtr, RawPtr, or OopPtr.
227 int classify_unsafe_addr(Node* &base, Node* &offset);
228 Node* make_unsafe_address(Node* base, Node* offset);
229 // Helper for inline_unsafe_access.
230 // Generates the guards that check whether the result of
231 // Unsafe.getObject should be recorded in an SATB log buffer.
232 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
233 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
234 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
235 static bool klass_needs_init_guard(Node* kls);
236 bool inline_unsafe_allocate();
237 bool inline_unsafe_copyMemory();
238 bool inline_native_currentThread();
239 #ifdef TRACE_HAVE_INTRINSICS
240 bool inline_native_classID();
241 bool inline_native_threadID();
242 #endif
243 bool inline_native_time_funcs(address method, const char* funcName);
244 bool inline_native_isInterrupted();
245 bool inline_native_Class_query(vmIntrinsics::ID id);
246 bool inline_native_subtype_check();
248 bool inline_native_newArray();
249 bool inline_native_getLength();
250 bool inline_array_copyOf(bool is_copyOfRange);
251 bool inline_array_equals();
252 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
253 bool inline_native_clone(bool is_virtual);
254 bool inline_native_Reflection_getCallerClass();
255 // Helper function for inlining native object hash method
256 bool inline_native_hashcode(bool is_virtual, bool is_static);
257 bool inline_native_getClass();
259 // Helper functions for inlining arraycopy
260 bool inline_arraycopy();
261 void generate_arraycopy(const TypePtr* adr_type,
262 BasicType basic_elem_type,
263 Node* src, Node* src_offset,
264 Node* dest, Node* dest_offset,
265 Node* copy_length,
266 bool disjoint_bases = false,
267 bool length_never_negative = false,
268 RegionNode* slow_region = NULL);
269 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
270 RegionNode* slow_region);
271 void generate_clear_array(const TypePtr* adr_type,
272 Node* dest,
273 BasicType basic_elem_type,
274 Node* slice_off,
275 Node* slice_len,
276 Node* slice_end);
277 bool generate_block_arraycopy(const TypePtr* adr_type,
278 BasicType basic_elem_type,
279 AllocateNode* alloc,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* dest_size, bool dest_uninitialized);
283 void generate_slow_arraycopy(const TypePtr* adr_type,
284 Node* src, Node* src_offset,
285 Node* dest, Node* dest_offset,
286 Node* copy_length, bool dest_uninitialized);
287 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
288 Node* dest_elem_klass,
289 Node* src, Node* src_offset,
290 Node* dest, Node* dest_offset,
291 Node* copy_length, bool dest_uninitialized);
292 Node* generate_generic_arraycopy(const TypePtr* adr_type,
293 Node* src, Node* src_offset,
294 Node* dest, Node* dest_offset,
295 Node* copy_length, bool dest_uninitialized);
296 void generate_unchecked_arraycopy(const TypePtr* adr_type,
297 BasicType basic_elem_type,
298 bool disjoint_bases,
299 Node* src, Node* src_offset,
300 Node* dest, Node* dest_offset,
301 Node* copy_length, bool dest_uninitialized);
302 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
303 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
304 bool inline_unsafe_ordered_store(BasicType type);
305 bool inline_unsafe_fence(vmIntrinsics::ID id);
306 bool inline_fp_conversions(vmIntrinsics::ID id);
307 bool inline_number_methods(vmIntrinsics::ID id);
308 bool inline_reference_get();
309 bool inline_aescrypt_Block(vmIntrinsics::ID id);
310 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
311 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
312 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
313 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
314 bool inline_sha_implCompress(vmIntrinsics::ID id);
315 bool inline_digestBase_implCompressMB(int predicate);
316 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
317 bool long_state, address stubAddr, const char *stubName,
318 Node* src_start, Node* ofs, Node* limit);
319 Node* get_state_from_sha_object(Node *sha_object);
320 Node* get_state_from_sha5_object(Node *sha_object);
321 Node* inline_digestBase_implCompressMB_predicate(int predicate);
322 bool inline_encodeISOArray();
323 bool inline_updateCRC32();
324 bool inline_updateBytesCRC32();
325 bool inline_updateByteBufferCRC32();
326 bool inline_multiplyToLen();
327 bool inline_squareToLen();
328 bool inline_mulAdd();
329 bool inline_montgomeryMultiply();
330 bool inline_montgomerySquare();
332 bool inline_profileBoolean();
333 };
336 //---------------------------make_vm_intrinsic----------------------------
337 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
338 vmIntrinsics::ID id = m->intrinsic_id();
339 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
341 ccstr disable_intr = NULL;
343 if ((DisableIntrinsic[0] != '\0'
344 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
345 (method_has_option_value("DisableIntrinsic", disable_intr)
346 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
347 // disabled by a user request on the command line:
348 // example: -XX:DisableIntrinsic=_hashCode,_getClass
349 return NULL;
350 }
352 if (!m->is_loaded()) {
353 // do not attempt to inline unloaded methods
354 return NULL;
355 }
357 // Only a few intrinsics implement a virtual dispatch.
358 // They are expensive calls which are also frequently overridden.
359 if (is_virtual) {
360 switch (id) {
361 case vmIntrinsics::_hashCode:
362 case vmIntrinsics::_clone:
363 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
364 break;
365 default:
366 return NULL;
367 }
368 }
370 // -XX:-InlineNatives disables nearly all intrinsics:
371 if (!InlineNatives) {
372 switch (id) {
373 case vmIntrinsics::_indexOf:
374 case vmIntrinsics::_compareTo:
375 case vmIntrinsics::_equals:
376 case vmIntrinsics::_equalsC:
377 case vmIntrinsics::_getAndAddInt:
378 case vmIntrinsics::_getAndAddLong:
379 case vmIntrinsics::_getAndSetInt:
380 case vmIntrinsics::_getAndSetLong:
381 case vmIntrinsics::_getAndSetObject:
382 case vmIntrinsics::_loadFence:
383 case vmIntrinsics::_storeFence:
384 case vmIntrinsics::_fullFence:
385 break; // InlineNatives does not control String.compareTo
386 case vmIntrinsics::_Reference_get:
387 break; // InlineNatives does not control Reference.get
388 default:
389 return NULL;
390 }
391 }
393 int predicates = 0;
394 bool does_virtual_dispatch = false;
396 switch (id) {
397 case vmIntrinsics::_compareTo:
398 if (!SpecialStringCompareTo) return NULL;
399 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
400 break;
401 case vmIntrinsics::_indexOf:
402 if (!SpecialStringIndexOf) return NULL;
403 break;
404 case vmIntrinsics::_equals:
405 if (!SpecialStringEquals) return NULL;
406 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
407 break;
408 case vmIntrinsics::_equalsC:
409 if (!SpecialArraysEquals) return NULL;
410 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
411 break;
412 case vmIntrinsics::_arraycopy:
413 if (!InlineArrayCopy) return NULL;
414 break;
415 case vmIntrinsics::_copyMemory:
416 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
417 if (!InlineArrayCopy) return NULL;
418 break;
419 case vmIntrinsics::_hashCode:
420 if (!InlineObjectHash) return NULL;
421 does_virtual_dispatch = true;
422 break;
423 case vmIntrinsics::_clone:
424 does_virtual_dispatch = true;
425 case vmIntrinsics::_copyOf:
426 case vmIntrinsics::_copyOfRange:
427 if (!InlineObjectCopy) return NULL;
428 // These also use the arraycopy intrinsic mechanism:
429 if (!InlineArrayCopy) return NULL;
430 break;
431 case vmIntrinsics::_encodeISOArray:
432 if (!SpecialEncodeISOArray) return NULL;
433 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
434 break;
435 case vmIntrinsics::_checkIndex:
436 // We do not intrinsify this. The optimizer does fine with it.
437 return NULL;
439 case vmIntrinsics::_getCallerClass:
440 if (!UseNewReflection) return NULL;
441 if (!InlineReflectionGetCallerClass) return NULL;
442 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
443 break;
445 case vmIntrinsics::_bitCount_i:
446 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
447 break;
449 case vmIntrinsics::_bitCount_l:
450 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
451 break;
453 case vmIntrinsics::_numberOfLeadingZeros_i:
454 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
455 break;
457 case vmIntrinsics::_numberOfLeadingZeros_l:
458 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
459 break;
461 case vmIntrinsics::_numberOfTrailingZeros_i:
462 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
463 break;
465 case vmIntrinsics::_numberOfTrailingZeros_l:
466 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
467 break;
469 case vmIntrinsics::_reverseBytes_c:
470 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
471 break;
472 case vmIntrinsics::_reverseBytes_s:
473 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
474 break;
475 case vmIntrinsics::_reverseBytes_i:
476 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
477 break;
478 case vmIntrinsics::_reverseBytes_l:
479 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
480 break;
482 case vmIntrinsics::_Reference_get:
483 // Use the intrinsic version of Reference.get() so that the value in
484 // the referent field can be registered by the G1 pre-barrier code.
485 // Also add memory barrier to prevent commoning reads from this field
486 // across safepoint since GC can change it value.
487 break;
489 case vmIntrinsics::_compareAndSwapObject:
490 #ifdef _LP64
491 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
492 #endif
493 break;
495 case vmIntrinsics::_compareAndSwapLong:
496 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
497 break;
499 case vmIntrinsics::_getAndAddInt:
500 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
501 break;
503 case vmIntrinsics::_getAndAddLong:
504 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
505 break;
507 case vmIntrinsics::_getAndSetInt:
508 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
509 break;
511 case vmIntrinsics::_getAndSetLong:
512 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
513 break;
515 case vmIntrinsics::_getAndSetObject:
516 #ifdef _LP64
517 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
518 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
519 break;
520 #else
521 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
522 break;
523 #endif
525 case vmIntrinsics::_aescrypt_encryptBlock:
526 case vmIntrinsics::_aescrypt_decryptBlock:
527 if (!UseAESIntrinsics) return NULL;
528 break;
530 case vmIntrinsics::_multiplyToLen:
531 if (!UseMultiplyToLenIntrinsic) return NULL;
532 break;
534 case vmIntrinsics::_squareToLen:
535 if (!UseSquareToLenIntrinsic) return NULL;
536 break;
538 case vmIntrinsics::_mulAdd:
539 if (!UseMulAddIntrinsic) return NULL;
540 break;
542 case vmIntrinsics::_montgomeryMultiply:
543 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
544 break;
545 case vmIntrinsics::_montgomerySquare:
546 if (!UseMontgomerySquareIntrinsic) return NULL;
547 break;
549 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
550 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
551 if (!UseAESIntrinsics) return NULL;
552 // these two require the predicated logic
553 predicates = 1;
554 break;
556 case vmIntrinsics::_sha_implCompress:
557 if (!UseSHA1Intrinsics) return NULL;
558 break;
560 case vmIntrinsics::_sha2_implCompress:
561 if (!UseSHA256Intrinsics) return NULL;
562 break;
564 case vmIntrinsics::_sha5_implCompress:
565 if (!UseSHA512Intrinsics) return NULL;
566 break;
568 case vmIntrinsics::_digestBase_implCompressMB:
569 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
570 predicates = 3;
571 break;
573 case vmIntrinsics::_updateCRC32:
574 case vmIntrinsics::_updateBytesCRC32:
575 case vmIntrinsics::_updateByteBufferCRC32:
576 if (!UseCRC32Intrinsics) return NULL;
577 break;
579 case vmIntrinsics::_incrementExactI:
580 case vmIntrinsics::_addExactI:
581 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
582 break;
583 case vmIntrinsics::_incrementExactL:
584 case vmIntrinsics::_addExactL:
585 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
586 break;
587 case vmIntrinsics::_decrementExactI:
588 case vmIntrinsics::_subtractExactI:
589 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
590 break;
591 case vmIntrinsics::_decrementExactL:
592 case vmIntrinsics::_subtractExactL:
593 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
594 break;
595 case vmIntrinsics::_negateExactI:
596 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
597 break;
598 case vmIntrinsics::_negateExactL:
599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
600 break;
601 case vmIntrinsics::_multiplyExactI:
602 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
603 break;
604 case vmIntrinsics::_multiplyExactL:
605 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
606 break;
608 default:
609 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
610 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
611 break;
612 }
614 // -XX:-InlineClassNatives disables natives from the Class class.
615 // The flag applies to all reflective calls, notably Array.newArray
616 // (visible to Java programmers as Array.newInstance).
617 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
618 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
619 if (!InlineClassNatives) return NULL;
620 }
622 // -XX:-InlineThreadNatives disables natives from the Thread class.
623 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
624 if (!InlineThreadNatives) return NULL;
625 }
627 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
628 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
629 m->holder()->name() == ciSymbol::java_lang_Float() ||
630 m->holder()->name() == ciSymbol::java_lang_Double()) {
631 if (!InlineMathNatives) return NULL;
632 }
634 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
635 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
636 if (!InlineUnsafeOps) return NULL;
637 }
639 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
640 }
642 //----------------------register_library_intrinsics-----------------------
643 // Initialize this file's data structures, for each Compile instance.
644 void Compile::register_library_intrinsics() {
645 // Nothing to do here.
646 }
648 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
649 LibraryCallKit kit(jvms, this);
650 Compile* C = kit.C;
651 int nodes = C->unique();
652 #ifndef PRODUCT
653 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
654 char buf[1000];
655 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
656 tty->print_cr("Intrinsic %s", str);
657 }
658 #endif
659 ciMethod* callee = kit.callee();
660 const int bci = kit.bci();
662 // Try to inline the intrinsic.
663 if (kit.try_to_inline(_last_predicate)) {
664 if (C->print_intrinsics() || C->print_inlining()) {
665 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
666 }
667 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
668 if (C->log()) {
669 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
670 vmIntrinsics::name_at(intrinsic_id()),
671 (is_virtual() ? " virtual='1'" : ""),
672 C->unique() - nodes);
673 }
674 // Push the result from the inlined method onto the stack.
675 kit.push_result();
676 return kit.transfer_exceptions_into_jvms();
677 }
679 // The intrinsic bailed out
680 if (C->print_intrinsics() || C->print_inlining()) {
681 if (jvms->has_method()) {
682 // Not a root compile.
683 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
684 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
685 } else {
686 // Root compile
687 tty->print("Did not generate intrinsic %s%s at bci:%d in",
688 vmIntrinsics::name_at(intrinsic_id()),
689 (is_virtual() ? " (virtual)" : ""), bci);
690 }
691 }
692 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
693 return NULL;
694 }
696 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
697 LibraryCallKit kit(jvms, this);
698 Compile* C = kit.C;
699 int nodes = C->unique();
700 _last_predicate = predicate;
701 #ifndef PRODUCT
702 assert(is_predicated() && predicate < predicates_count(), "sanity");
703 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
704 char buf[1000];
705 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
706 tty->print_cr("Predicate for intrinsic %s", str);
707 }
708 #endif
709 ciMethod* callee = kit.callee();
710 const int bci = kit.bci();
712 Node* slow_ctl = kit.try_to_predicate(predicate);
713 if (!kit.failing()) {
714 if (C->print_intrinsics() || C->print_inlining()) {
715 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
716 }
717 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
718 if (C->log()) {
719 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
720 vmIntrinsics::name_at(intrinsic_id()),
721 (is_virtual() ? " virtual='1'" : ""),
722 C->unique() - nodes);
723 }
724 return slow_ctl; // Could be NULL if the check folds.
725 }
727 // The intrinsic bailed out
728 if (C->print_intrinsics() || C->print_inlining()) {
729 if (jvms->has_method()) {
730 // Not a root compile.
731 const char* msg = "failed to generate predicate for intrinsic";
732 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
733 } else {
734 // Root compile
735 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
736 vmIntrinsics::name_at(intrinsic_id()),
737 (is_virtual() ? " (virtual)" : ""), bci);
738 }
739 }
740 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
741 return NULL;
742 }
744 bool LibraryCallKit::try_to_inline(int predicate) {
745 // Handle symbolic names for otherwise undistinguished boolean switches:
746 const bool is_store = true;
747 const bool is_native_ptr = true;
748 const bool is_static = true;
749 const bool is_volatile = true;
751 if (!jvms()->has_method()) {
752 // Root JVMState has a null method.
753 assert(map()->memory()->Opcode() == Op_Parm, "");
754 // Insert the memory aliasing node
755 set_all_memory(reset_memory());
756 }
757 assert(merged_memory(), "");
760 switch (intrinsic_id()) {
761 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
762 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
763 case vmIntrinsics::_getClass: return inline_native_getClass();
765 case vmIntrinsics::_dsin:
766 case vmIntrinsics::_dcos:
767 case vmIntrinsics::_dtan:
768 case vmIntrinsics::_dabs:
769 case vmIntrinsics::_datan2:
770 case vmIntrinsics::_dsqrt:
771 case vmIntrinsics::_dexp:
772 case vmIntrinsics::_dlog:
773 case vmIntrinsics::_dlog10:
774 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
776 case vmIntrinsics::_min:
777 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
779 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
780 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
781 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
782 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
783 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
784 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
785 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
786 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
787 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
788 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
789 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
790 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
792 case vmIntrinsics::_arraycopy: return inline_arraycopy();
794 case vmIntrinsics::_compareTo: return inline_string_compareTo();
795 case vmIntrinsics::_indexOf: return inline_string_indexOf();
796 case vmIntrinsics::_equals: return inline_string_equals();
798 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
799 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
800 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
801 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
802 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
803 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
804 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
805 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
806 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
808 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
809 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
810 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
811 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
812 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
813 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
814 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
815 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
816 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
818 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
819 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
820 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
821 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
822 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
823 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
824 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
825 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
827 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
828 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
829 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
830 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
831 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
832 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
833 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
834 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
836 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
837 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
838 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
839 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
840 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
841 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
842 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
843 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
844 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
846 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
847 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
848 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
849 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
850 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
851 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
852 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
853 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
854 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
856 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
857 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
858 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
859 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
861 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
862 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
863 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
865 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
866 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
867 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
869 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
870 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
871 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
872 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
873 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
875 case vmIntrinsics::_loadFence:
876 case vmIntrinsics::_storeFence:
877 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
879 case vmIntrinsics::_currentThread: return inline_native_currentThread();
880 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
882 #ifdef TRACE_HAVE_INTRINSICS
883 case vmIntrinsics::_classID: return inline_native_classID();
884 case vmIntrinsics::_threadID: return inline_native_threadID();
885 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
886 #endif
887 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
888 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
889 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
890 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
891 case vmIntrinsics::_newArray: return inline_native_newArray();
892 case vmIntrinsics::_getLength: return inline_native_getLength();
893 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
894 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
895 case vmIntrinsics::_equalsC: return inline_array_equals();
896 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
898 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
900 case vmIntrinsics::_isInstance:
901 case vmIntrinsics::_getModifiers:
902 case vmIntrinsics::_isInterface:
903 case vmIntrinsics::_isArray:
904 case vmIntrinsics::_isPrimitive:
905 case vmIntrinsics::_getSuperclass:
906 case vmIntrinsics::_getComponentType:
907 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
909 case vmIntrinsics::_floatToRawIntBits:
910 case vmIntrinsics::_floatToIntBits:
911 case vmIntrinsics::_intBitsToFloat:
912 case vmIntrinsics::_doubleToRawLongBits:
913 case vmIntrinsics::_doubleToLongBits:
914 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
916 case vmIntrinsics::_numberOfLeadingZeros_i:
917 case vmIntrinsics::_numberOfLeadingZeros_l:
918 case vmIntrinsics::_numberOfTrailingZeros_i:
919 case vmIntrinsics::_numberOfTrailingZeros_l:
920 case vmIntrinsics::_bitCount_i:
921 case vmIntrinsics::_bitCount_l:
922 case vmIntrinsics::_reverseBytes_i:
923 case vmIntrinsics::_reverseBytes_l:
924 case vmIntrinsics::_reverseBytes_s:
925 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
927 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
929 case vmIntrinsics::_Reference_get: return inline_reference_get();
931 case vmIntrinsics::_aescrypt_encryptBlock:
932 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
934 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
935 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
936 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
938 case vmIntrinsics::_sha_implCompress:
939 case vmIntrinsics::_sha2_implCompress:
940 case vmIntrinsics::_sha5_implCompress:
941 return inline_sha_implCompress(intrinsic_id());
943 case vmIntrinsics::_digestBase_implCompressMB:
944 return inline_digestBase_implCompressMB(predicate);
946 case vmIntrinsics::_multiplyToLen:
947 return inline_multiplyToLen();
949 case vmIntrinsics::_squareToLen:
950 return inline_squareToLen();
952 case vmIntrinsics::_mulAdd:
953 return inline_mulAdd();
955 case vmIntrinsics::_montgomeryMultiply:
956 return inline_montgomeryMultiply();
957 case vmIntrinsics::_montgomerySquare:
958 return inline_montgomerySquare();
960 case vmIntrinsics::_encodeISOArray:
961 return inline_encodeISOArray();
963 case vmIntrinsics::_updateCRC32:
964 return inline_updateCRC32();
965 case vmIntrinsics::_updateBytesCRC32:
966 return inline_updateBytesCRC32();
967 case vmIntrinsics::_updateByteBufferCRC32:
968 return inline_updateByteBufferCRC32();
970 case vmIntrinsics::_profileBoolean:
971 return inline_profileBoolean();
973 default:
974 // If you get here, it may be that someone has added a new intrinsic
975 // to the list in vmSymbols.hpp without implementing it here.
976 #ifndef PRODUCT
977 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
978 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
979 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
980 }
981 #endif
982 return false;
983 }
984 }
986 Node* LibraryCallKit::try_to_predicate(int predicate) {
987 if (!jvms()->has_method()) {
988 // Root JVMState has a null method.
989 assert(map()->memory()->Opcode() == Op_Parm, "");
990 // Insert the memory aliasing node
991 set_all_memory(reset_memory());
992 }
993 assert(merged_memory(), "");
995 switch (intrinsic_id()) {
996 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
997 return inline_cipherBlockChaining_AESCrypt_predicate(false);
998 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
999 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1000 case vmIntrinsics::_digestBase_implCompressMB:
1001 return inline_digestBase_implCompressMB_predicate(predicate);
1003 default:
1004 // If you get here, it may be that someone has added a new intrinsic
1005 // to the list in vmSymbols.hpp without implementing it here.
1006 #ifndef PRODUCT
1007 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1008 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1009 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1010 }
1011 #endif
1012 Node* slow_ctl = control();
1013 set_control(top()); // No fast path instrinsic
1014 return slow_ctl;
1015 }
1016 }
1018 //------------------------------set_result-------------------------------
1019 // Helper function for finishing intrinsics.
1020 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1021 record_for_igvn(region);
1022 set_control(_gvn.transform(region));
1023 set_result( _gvn.transform(value));
1024 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1025 }
1027 //------------------------------generate_guard---------------------------
1028 // Helper function for generating guarded fast-slow graph structures.
1029 // The given 'test', if true, guards a slow path. If the test fails
1030 // then a fast path can be taken. (We generally hope it fails.)
1031 // In all cases, GraphKit::control() is updated to the fast path.
1032 // The returned value represents the control for the slow path.
1033 // The return value is never 'top'; it is either a valid control
1034 // or NULL if it is obvious that the slow path can never be taken.
1035 // Also, if region and the slow control are not NULL, the slow edge
1036 // is appended to the region.
1037 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1038 if (stopped()) {
1039 // Already short circuited.
1040 return NULL;
1041 }
1043 // Build an if node and its projections.
1044 // If test is true we take the slow path, which we assume is uncommon.
1045 if (_gvn.type(test) == TypeInt::ZERO) {
1046 // The slow branch is never taken. No need to build this guard.
1047 return NULL;
1048 }
1050 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1052 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1053 if (if_slow == top()) {
1054 // The slow branch is never taken. No need to build this guard.
1055 return NULL;
1056 }
1058 if (region != NULL)
1059 region->add_req(if_slow);
1061 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1062 set_control(if_fast);
1064 return if_slow;
1065 }
1067 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1068 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1069 }
1070 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1071 return generate_guard(test, region, PROB_FAIR);
1072 }
1074 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1075 Node* *pos_index) {
1076 if (stopped())
1077 return NULL; // already stopped
1078 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1079 return NULL; // index is already adequately typed
1080 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1081 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1082 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1083 if (is_neg != NULL && pos_index != NULL) {
1084 // Emulate effect of Parse::adjust_map_after_if.
1085 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1086 ccast->set_req(0, control());
1087 (*pos_index) = _gvn.transform(ccast);
1088 }
1089 return is_neg;
1090 }
1092 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1093 Node* *pos_index) {
1094 if (stopped())
1095 return NULL; // already stopped
1096 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1097 return NULL; // index is already adequately typed
1098 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1099 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1100 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1101 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1102 if (is_notp != NULL && pos_index != NULL) {
1103 // Emulate effect of Parse::adjust_map_after_if.
1104 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1105 ccast->set_req(0, control());
1106 (*pos_index) = _gvn.transform(ccast);
1107 }
1108 return is_notp;
1109 }
1111 // Make sure that 'position' is a valid limit index, in [0..length].
1112 // There are two equivalent plans for checking this:
1113 // A. (offset + copyLength) unsigned<= arrayLength
1114 // B. offset <= (arrayLength - copyLength)
1115 // We require that all of the values above, except for the sum and
1116 // difference, are already known to be non-negative.
1117 // Plan A is robust in the face of overflow, if offset and copyLength
1118 // are both hugely positive.
1119 //
1120 // Plan B is less direct and intuitive, but it does not overflow at
1121 // all, since the difference of two non-negatives is always
1122 // representable. Whenever Java methods must perform the equivalent
1123 // check they generally use Plan B instead of Plan A.
1124 // For the moment we use Plan A.
1125 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1126 Node* subseq_length,
1127 Node* array_length,
1128 RegionNode* region) {
1129 if (stopped())
1130 return NULL; // already stopped
1131 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1132 if (zero_offset && subseq_length->eqv_uncast(array_length))
1133 return NULL; // common case of whole-array copy
1134 Node* last = subseq_length;
1135 if (!zero_offset) // last += offset
1136 last = _gvn.transform(new (C) AddINode(last, offset));
1137 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1138 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1139 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1140 return is_over;
1141 }
1144 //--------------------------generate_current_thread--------------------
1145 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1146 ciKlass* thread_klass = env()->Thread_klass();
1147 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1148 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1149 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1150 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1151 tls_output = thread;
1152 return threadObj;
1153 }
1156 //------------------------------make_string_method_node------------------------
1157 // Helper method for String intrinsic functions. This version is called
1158 // with str1 and str2 pointing to String object nodes.
1159 //
1160 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1161 Node* no_ctrl = NULL;
1163 // Get start addr of string
1164 Node* str1_value = load_String_value(no_ctrl, str1);
1165 Node* str1_offset = load_String_offset(no_ctrl, str1);
1166 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1168 // Get length of string 1
1169 Node* str1_len = load_String_length(no_ctrl, str1);
1171 Node* str2_value = load_String_value(no_ctrl, str2);
1172 Node* str2_offset = load_String_offset(no_ctrl, str2);
1173 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1175 Node* str2_len = NULL;
1176 Node* result = NULL;
1178 switch (opcode) {
1179 case Op_StrIndexOf:
1180 // Get length of string 2
1181 str2_len = load_String_length(no_ctrl, str2);
1183 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1184 str1_start, str1_len, str2_start, str2_len);
1185 break;
1186 case Op_StrComp:
1187 // Get length of string 2
1188 str2_len = load_String_length(no_ctrl, str2);
1190 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1191 str1_start, str1_len, str2_start, str2_len);
1192 break;
1193 case Op_StrEquals:
1194 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1195 str1_start, str2_start, str1_len);
1196 break;
1197 default:
1198 ShouldNotReachHere();
1199 return NULL;
1200 }
1202 // All these intrinsics have checks.
1203 C->set_has_split_ifs(true); // Has chance for split-if optimization
1205 return _gvn.transform(result);
1206 }
1208 // Helper method for String intrinsic functions. This version is called
1209 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1210 // to Int nodes containing the lenghts of str1 and str2.
1211 //
1212 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1213 Node* result = NULL;
1214 switch (opcode) {
1215 case Op_StrIndexOf:
1216 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1217 str1_start, cnt1, str2_start, cnt2);
1218 break;
1219 case Op_StrComp:
1220 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1221 str1_start, cnt1, str2_start, cnt2);
1222 break;
1223 case Op_StrEquals:
1224 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1225 str1_start, str2_start, cnt1);
1226 break;
1227 default:
1228 ShouldNotReachHere();
1229 return NULL;
1230 }
1232 // All these intrinsics have checks.
1233 C->set_has_split_ifs(true); // Has chance for split-if optimization
1235 return _gvn.transform(result);
1236 }
1238 //------------------------------inline_string_compareTo------------------------
1239 // public int java.lang.String.compareTo(String anotherString);
1240 bool LibraryCallKit::inline_string_compareTo() {
1241 Node* receiver = null_check(argument(0));
1242 Node* arg = null_check(argument(1));
1243 if (stopped()) {
1244 return true;
1245 }
1246 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1247 return true;
1248 }
1250 //------------------------------inline_string_equals------------------------
1251 bool LibraryCallKit::inline_string_equals() {
1252 Node* receiver = null_check_receiver();
1253 // NOTE: Do not null check argument for String.equals() because spec
1254 // allows to specify NULL as argument.
1255 Node* argument = this->argument(1);
1256 if (stopped()) {
1257 return true;
1258 }
1260 // paths (plus control) merge
1261 RegionNode* region = new (C) RegionNode(5);
1262 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1264 // does source == target string?
1265 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1266 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1268 Node* if_eq = generate_slow_guard(bol, NULL);
1269 if (if_eq != NULL) {
1270 // receiver == argument
1271 phi->init_req(2, intcon(1));
1272 region->init_req(2, if_eq);
1273 }
1275 // get String klass for instanceOf
1276 ciInstanceKlass* klass = env()->String_klass();
1278 if (!stopped()) {
1279 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1280 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1281 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1283 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1284 //instanceOf == true, fallthrough
1286 if (inst_false != NULL) {
1287 phi->init_req(3, intcon(0));
1288 region->init_req(3, inst_false);
1289 }
1290 }
1292 if (!stopped()) {
1293 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1295 // Properly cast the argument to String
1296 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1297 // This path is taken only when argument's type is String:NotNull.
1298 argument = cast_not_null(argument, false);
1300 Node* no_ctrl = NULL;
1302 // Get start addr of receiver
1303 Node* receiver_val = load_String_value(no_ctrl, receiver);
1304 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1305 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1307 // Get length of receiver
1308 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1310 // Get start addr of argument
1311 Node* argument_val = load_String_value(no_ctrl, argument);
1312 Node* argument_offset = load_String_offset(no_ctrl, argument);
1313 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1315 // Get length of argument
1316 Node* argument_cnt = load_String_length(no_ctrl, argument);
1318 // Check for receiver count != argument count
1319 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1320 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1321 Node* if_ne = generate_slow_guard(bol, NULL);
1322 if (if_ne != NULL) {
1323 phi->init_req(4, intcon(0));
1324 region->init_req(4, if_ne);
1325 }
1327 // Check for count == 0 is done by assembler code for StrEquals.
1329 if (!stopped()) {
1330 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1331 phi->init_req(1, equals);
1332 region->init_req(1, control());
1333 }
1334 }
1336 // post merge
1337 set_control(_gvn.transform(region));
1338 record_for_igvn(region);
1340 set_result(_gvn.transform(phi));
1341 return true;
1342 }
1344 //------------------------------inline_array_equals----------------------------
1345 bool LibraryCallKit::inline_array_equals() {
1346 Node* arg1 = argument(0);
1347 Node* arg2 = argument(1);
1348 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1349 return true;
1350 }
1352 // Java version of String.indexOf(constant string)
1353 // class StringDecl {
1354 // StringDecl(char[] ca) {
1355 // offset = 0;
1356 // count = ca.length;
1357 // value = ca;
1358 // }
1359 // int offset;
1360 // int count;
1361 // char[] value;
1362 // }
1363 //
1364 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1365 // int targetOffset, int cache_i, int md2) {
1366 // int cache = cache_i;
1367 // int sourceOffset = string_object.offset;
1368 // int sourceCount = string_object.count;
1369 // int targetCount = target_object.length;
1370 //
1371 // int targetCountLess1 = targetCount - 1;
1372 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1373 //
1374 // char[] source = string_object.value;
1375 // char[] target = target_object;
1376 // int lastChar = target[targetCountLess1];
1377 //
1378 // outer_loop:
1379 // for (int i = sourceOffset; i < sourceEnd; ) {
1380 // int src = source[i + targetCountLess1];
1381 // if (src == lastChar) {
1382 // // With random strings and a 4-character alphabet,
1383 // // reverse matching at this point sets up 0.8% fewer
1384 // // frames, but (paradoxically) makes 0.3% more probes.
1385 // // Since those probes are nearer the lastChar probe,
1386 // // there is may be a net D$ win with reverse matching.
1387 // // But, reversing loop inhibits unroll of inner loop
1388 // // for unknown reason. So, does running outer loop from
1389 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1390 // for (int j = 0; j < targetCountLess1; j++) {
1391 // if (target[targetOffset + j] != source[i+j]) {
1392 // if ((cache & (1 << source[i+j])) == 0) {
1393 // if (md2 < j+1) {
1394 // i += j+1;
1395 // continue outer_loop;
1396 // }
1397 // }
1398 // i += md2;
1399 // continue outer_loop;
1400 // }
1401 // }
1402 // return i - sourceOffset;
1403 // }
1404 // if ((cache & (1 << src)) == 0) {
1405 // i += targetCountLess1;
1406 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1407 // i++;
1408 // }
1409 // return -1;
1410 // }
1412 //------------------------------string_indexOf------------------------
1413 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1414 jint cache_i, jint md2_i) {
1416 Node* no_ctrl = NULL;
1417 float likely = PROB_LIKELY(0.9);
1418 float unlikely = PROB_UNLIKELY(0.9);
1420 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1422 Node* source = load_String_value(no_ctrl, string_object);
1423 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1424 Node* sourceCount = load_String_length(no_ctrl, string_object);
1426 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1427 jint target_length = target_array->length();
1428 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1429 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1431 // String.value field is known to be @Stable.
1432 if (UseImplicitStableValues) {
1433 target = cast_array_to_stable(target, target_type);
1434 }
1436 IdealKit kit(this, false, true);
1437 #define __ kit.
1438 Node* zero = __ ConI(0);
1439 Node* one = __ ConI(1);
1440 Node* cache = __ ConI(cache_i);
1441 Node* md2 = __ ConI(md2_i);
1442 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1443 Node* targetCount = __ ConI(target_length);
1444 Node* targetCountLess1 = __ ConI(target_length - 1);
1445 Node* targetOffset = __ ConI(targetOffset_i);
1446 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1448 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1449 Node* outer_loop = __ make_label(2 /* goto */);
1450 Node* return_ = __ make_label(1);
1452 __ set(rtn,__ ConI(-1));
1453 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1454 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1455 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1456 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1457 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1458 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1459 Node* tpj = __ AddI(targetOffset, __ value(j));
1460 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1461 Node* ipj = __ AddI(__ value(i), __ value(j));
1462 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1463 __ if_then(targ, BoolTest::ne, src2); {
1464 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1465 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1466 __ increment(i, __ AddI(__ value(j), one));
1467 __ goto_(outer_loop);
1468 } __ end_if(); __ dead(j);
1469 }__ end_if(); __ dead(j);
1470 __ increment(i, md2);
1471 __ goto_(outer_loop);
1472 }__ end_if();
1473 __ increment(j, one);
1474 }__ end_loop(); __ dead(j);
1475 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1476 __ goto_(return_);
1477 }__ end_if();
1478 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1479 __ increment(i, targetCountLess1);
1480 }__ end_if();
1481 __ increment(i, one);
1482 __ bind(outer_loop);
1483 }__ end_loop(); __ dead(i);
1484 __ bind(return_);
1486 // Final sync IdealKit and GraphKit.
1487 final_sync(kit);
1488 Node* result = __ value(rtn);
1489 #undef __
1490 C->set_has_loops(true);
1491 return result;
1492 }
1494 //------------------------------inline_string_indexOf------------------------
1495 bool LibraryCallKit::inline_string_indexOf() {
1496 Node* receiver = argument(0);
1497 Node* arg = argument(1);
1499 Node* result;
1500 // Disable the use of pcmpestri until it can be guaranteed that
1501 // the load doesn't cross into the uncommited space.
1502 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1503 UseSSE42Intrinsics) {
1504 // Generate SSE4.2 version of indexOf
1505 // We currently only have match rules that use SSE4.2
1507 receiver = null_check(receiver);
1508 arg = null_check(arg);
1509 if (stopped()) {
1510 return true;
1511 }
1513 ciInstanceKlass* str_klass = env()->String_klass();
1514 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1516 // Make the merge point
1517 RegionNode* result_rgn = new (C) RegionNode(4);
1518 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1519 Node* no_ctrl = NULL;
1521 // Get start addr of source string
1522 Node* source = load_String_value(no_ctrl, receiver);
1523 Node* source_offset = load_String_offset(no_ctrl, receiver);
1524 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1526 // Get length of source string
1527 Node* source_cnt = load_String_length(no_ctrl, receiver);
1529 // Get start addr of substring
1530 Node* substr = load_String_value(no_ctrl, arg);
1531 Node* substr_offset = load_String_offset(no_ctrl, arg);
1532 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1534 // Get length of source string
1535 Node* substr_cnt = load_String_length(no_ctrl, arg);
1537 // Check for substr count > string count
1538 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1539 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1540 Node* if_gt = generate_slow_guard(bol, NULL);
1541 if (if_gt != NULL) {
1542 result_phi->init_req(2, intcon(-1));
1543 result_rgn->init_req(2, if_gt);
1544 }
1546 if (!stopped()) {
1547 // Check for substr count == 0
1548 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1549 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1550 Node* if_zero = generate_slow_guard(bol, NULL);
1551 if (if_zero != NULL) {
1552 result_phi->init_req(3, intcon(0));
1553 result_rgn->init_req(3, if_zero);
1554 }
1555 }
1557 if (!stopped()) {
1558 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1559 result_phi->init_req(1, result);
1560 result_rgn->init_req(1, control());
1561 }
1562 set_control(_gvn.transform(result_rgn));
1563 record_for_igvn(result_rgn);
1564 result = _gvn.transform(result_phi);
1566 } else { // Use LibraryCallKit::string_indexOf
1567 // don't intrinsify if argument isn't a constant string.
1568 if (!arg->is_Con()) {
1569 return false;
1570 }
1571 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1572 if (str_type == NULL) {
1573 return false;
1574 }
1575 ciInstanceKlass* klass = env()->String_klass();
1576 ciObject* str_const = str_type->const_oop();
1577 if (str_const == NULL || str_const->klass() != klass) {
1578 return false;
1579 }
1580 ciInstance* str = str_const->as_instance();
1581 assert(str != NULL, "must be instance");
1583 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1584 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1586 int o;
1587 int c;
1588 if (java_lang_String::has_offset_field()) {
1589 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1590 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1591 } else {
1592 o = 0;
1593 c = pat->length();
1594 }
1596 // constant strings have no offset and count == length which
1597 // simplifies the resulting code somewhat so lets optimize for that.
1598 if (o != 0 || c != pat->length()) {
1599 return false;
1600 }
1602 receiver = null_check(receiver, T_OBJECT);
1603 // NOTE: No null check on the argument is needed since it's a constant String oop.
1604 if (stopped()) {
1605 return true;
1606 }
1608 // The null string as a pattern always returns 0 (match at beginning of string)
1609 if (c == 0) {
1610 set_result(intcon(0));
1611 return true;
1612 }
1614 // Generate default indexOf
1615 jchar lastChar = pat->char_at(o + (c - 1));
1616 int cache = 0;
1617 int i;
1618 for (i = 0; i < c - 1; i++) {
1619 assert(i < pat->length(), "out of range");
1620 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1621 }
1623 int md2 = c;
1624 for (i = 0; i < c - 1; i++) {
1625 assert(i < pat->length(), "out of range");
1626 if (pat->char_at(o + i) == lastChar) {
1627 md2 = (c - 1) - i;
1628 }
1629 }
1631 result = string_indexOf(receiver, pat, o, cache, md2);
1632 }
1633 set_result(result);
1634 return true;
1635 }
1637 //--------------------------round_double_node--------------------------------
1638 // Round a double node if necessary.
1639 Node* LibraryCallKit::round_double_node(Node* n) {
1640 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1641 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1642 return n;
1643 }
1645 //------------------------------inline_math-----------------------------------
1646 // public static double Math.abs(double)
1647 // public static double Math.sqrt(double)
1648 // public static double Math.log(double)
1649 // public static double Math.log10(double)
1650 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1651 Node* arg = round_double_node(argument(0));
1652 Node* n = NULL;
1653 switch (id) {
1654 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1655 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1656 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1657 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1658 default: fatal_unexpected_iid(id); break;
1659 }
1660 set_result(_gvn.transform(n));
1661 return true;
1662 }
1664 //------------------------------inline_trig----------------------------------
1665 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1666 // argument reduction which will turn into a fast/slow diamond.
1667 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1668 Node* arg = round_double_node(argument(0));
1669 Node* n = NULL;
1671 switch (id) {
1672 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1673 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1674 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1675 default: fatal_unexpected_iid(id); break;
1676 }
1677 n = _gvn.transform(n);
1679 // Rounding required? Check for argument reduction!
1680 if (Matcher::strict_fp_requires_explicit_rounding) {
1681 static const double pi_4 = 0.7853981633974483;
1682 static const double neg_pi_4 = -0.7853981633974483;
1683 // pi/2 in 80-bit extended precision
1684 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1685 // -pi/2 in 80-bit extended precision
1686 // 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};
1687 // Cutoff value for using this argument reduction technique
1688 //static const double pi_2_minus_epsilon = 1.564660403643354;
1689 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1691 // Pseudocode for sin:
1692 // if (x <= Math.PI / 4.0) {
1693 // if (x >= -Math.PI / 4.0) return fsin(x);
1694 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1695 // } else {
1696 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1697 // }
1698 // return StrictMath.sin(x);
1700 // Pseudocode for cos:
1701 // if (x <= Math.PI / 4.0) {
1702 // if (x >= -Math.PI / 4.0) return fcos(x);
1703 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1704 // } else {
1705 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1706 // }
1707 // return StrictMath.cos(x);
1709 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1710 // requires a special machine instruction to load it. Instead we'll try
1711 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1712 // probably do the math inside the SIN encoding.
1714 // Make the merge point
1715 RegionNode* r = new (C) RegionNode(3);
1716 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1718 // Flatten arg so we need only 1 test
1719 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1720 // Node for PI/4 constant
1721 Node *pi4 = makecon(TypeD::make(pi_4));
1722 // Check PI/4 : abs(arg)
1723 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1724 // Check: If PI/4 < abs(arg) then go slow
1725 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1726 // Branch either way
1727 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1728 set_control(opt_iff(r,iff));
1730 // Set fast path result
1731 phi->init_req(2, n);
1733 // Slow path - non-blocking leaf call
1734 Node* call = NULL;
1735 switch (id) {
1736 case vmIntrinsics::_dsin:
1737 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1738 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1739 "Sin", NULL, arg, top());
1740 break;
1741 case vmIntrinsics::_dcos:
1742 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1743 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1744 "Cos", NULL, arg, top());
1745 break;
1746 case vmIntrinsics::_dtan:
1747 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1748 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1749 "Tan", NULL, arg, top());
1750 break;
1751 }
1752 assert(control()->in(0) == call, "");
1753 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1754 r->init_req(1, control());
1755 phi->init_req(1, slow_result);
1757 // Post-merge
1758 set_control(_gvn.transform(r));
1759 record_for_igvn(r);
1760 n = _gvn.transform(phi);
1762 C->set_has_split_ifs(true); // Has chance for split-if optimization
1763 }
1764 set_result(n);
1765 return true;
1766 }
1768 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1769 //-------------------
1770 //result=(result.isNaN())? funcAddr():result;
1771 // Check: If isNaN() by checking result!=result? then either trap
1772 // or go to runtime
1773 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1774 // Build the boolean node
1775 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1777 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1778 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1779 // The pow or exp intrinsic returned a NaN, which requires a call
1780 // to the runtime. Recompile with the runtime call.
1781 uncommon_trap(Deoptimization::Reason_intrinsic,
1782 Deoptimization::Action_make_not_entrant);
1783 }
1784 return result;
1785 } else {
1786 // If this inlining ever returned NaN in the past, we compile a call
1787 // to the runtime to properly handle corner cases
1789 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1790 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1791 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1793 if (!if_slow->is_top()) {
1794 RegionNode* result_region = new (C) RegionNode(3);
1795 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1797 result_region->init_req(1, if_fast);
1798 result_val->init_req(1, result);
1800 set_control(if_slow);
1802 const TypePtr* no_memory_effects = NULL;
1803 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1804 no_memory_effects,
1805 x, top(), y, y ? top() : NULL);
1806 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1807 #ifdef ASSERT
1808 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1809 assert(value_top == top(), "second value must be top");
1810 #endif
1812 result_region->init_req(2, control());
1813 result_val->init_req(2, value);
1814 set_control(_gvn.transform(result_region));
1815 return _gvn.transform(result_val);
1816 } else {
1817 return result;
1818 }
1819 }
1820 }
1822 //------------------------------inline_exp-------------------------------------
1823 // Inline exp instructions, if possible. The Intel hardware only misses
1824 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1825 bool LibraryCallKit::inline_exp() {
1826 Node* arg = round_double_node(argument(0));
1827 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1829 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1830 set_result(n);
1832 C->set_has_split_ifs(true); // Has chance for split-if optimization
1833 return true;
1834 }
1836 //------------------------------inline_pow-------------------------------------
1837 // Inline power instructions, if possible.
1838 bool LibraryCallKit::inline_pow() {
1839 // Pseudocode for pow
1840 // if (y == 2) {
1841 // return x * x;
1842 // } else {
1843 // if (x <= 0.0) {
1844 // long longy = (long)y;
1845 // if ((double)longy == y) { // if y is long
1846 // if (y + 1 == y) longy = 0; // huge number: even
1847 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1848 // } else {
1849 // result = NaN;
1850 // }
1851 // } else {
1852 // result = DPow(x,y);
1853 // }
1854 // if (result != result)? {
1855 // result = uncommon_trap() or runtime_call();
1856 // }
1857 // return result;
1858 // }
1860 Node* x = round_double_node(argument(0));
1861 Node* y = round_double_node(argument(2));
1863 Node* result = NULL;
1865 Node* const_two_node = makecon(TypeD::make(2.0));
1866 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1867 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1868 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1869 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1870 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1872 RegionNode* region_node = new (C) RegionNode(3);
1873 region_node->init_req(1, if_true);
1875 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1876 // special case for x^y where y == 2, we can convert it to x * x
1877 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1879 // set control to if_false since we will now process the false branch
1880 set_control(if_false);
1882 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1883 // Short form: skip the fancy tests and just check for NaN result.
1884 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1885 } else {
1886 // If this inlining ever returned NaN in the past, include all
1887 // checks + call to the runtime.
1889 // Set the merge point for If node with condition of (x <= 0.0)
1890 // There are four possible paths to region node and phi node
1891 RegionNode *r = new (C) RegionNode(4);
1892 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1894 // Build the first if node: if (x <= 0.0)
1895 // Node for 0 constant
1896 Node *zeronode = makecon(TypeD::ZERO);
1897 // Check x:0
1898 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1899 // Check: If (x<=0) then go complex path
1900 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1901 // Branch either way
1902 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1903 // Fast path taken; set region slot 3
1904 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1905 r->init_req(3,fast_taken); // Capture fast-control
1907 // Fast path not-taken, i.e. slow path
1908 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1910 // Set fast path result
1911 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1912 phi->init_req(3, fast_result);
1914 // Complex path
1915 // Build the second if node (if y is long)
1916 // Node for (long)y
1917 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1918 // Node for (double)((long) y)
1919 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1920 // Check (double)((long) y) : y
1921 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1922 // Check if (y isn't long) then go to slow path
1924 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1925 // Branch either way
1926 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1927 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1929 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1931 // Calculate DPow(abs(x), y)*(1 & (long)y)
1932 // Node for constant 1
1933 Node *conone = longcon(1);
1934 // 1& (long)y
1935 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1937 // A huge number is always even. Detect a huge number by checking
1938 // if y + 1 == y and set integer to be tested for parity to 0.
1939 // Required for corner case:
1940 // (long)9.223372036854776E18 = max_jlong
1941 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1942 // max_jlong is odd but 9.223372036854776E18 is even
1943 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1944 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1945 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1946 Node* correctedsign = NULL;
1947 if (ConditionalMoveLimit != 0) {
1948 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1949 } else {
1950 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1951 RegionNode *r = new (C) RegionNode(3);
1952 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1953 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1954 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1955 phi->init_req(1, signnode);
1956 phi->init_req(2, longcon(0));
1957 correctedsign = _gvn.transform(phi);
1958 ylong_path = _gvn.transform(r);
1959 record_for_igvn(r);
1960 }
1962 // zero node
1963 Node *conzero = longcon(0);
1964 // Check (1&(long)y)==0?
1965 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1966 // Check if (1&(long)y)!=0?, if so the result is negative
1967 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1968 // abs(x)
1969 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1970 // abs(x)^y
1971 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1972 // -abs(x)^y
1973 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1974 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1975 Node *signresult = NULL;
1976 if (ConditionalMoveLimit != 0) {
1977 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1978 } else {
1979 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1980 RegionNode *r = new (C) RegionNode(3);
1981 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1982 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1983 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1984 phi->init_req(1, absxpowy);
1985 phi->init_req(2, negabsxpowy);
1986 signresult = _gvn.transform(phi);
1987 ylong_path = _gvn.transform(r);
1988 record_for_igvn(r);
1989 }
1990 // Set complex path fast result
1991 r->init_req(2, ylong_path);
1992 phi->init_req(2, signresult);
1994 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1995 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1996 r->init_req(1,slow_path);
1997 phi->init_req(1,slow_result);
1999 // Post merge
2000 set_control(_gvn.transform(r));
2001 record_for_igvn(r);
2002 result = _gvn.transform(phi);
2003 }
2005 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2007 // control from finish_pow_exp is now input to the region node
2008 region_node->set_req(2, control());
2009 // the result from finish_pow_exp is now input to the phi node
2010 phi_node->init_req(2, result);
2011 set_control(_gvn.transform(region_node));
2012 record_for_igvn(region_node);
2013 set_result(_gvn.transform(phi_node));
2015 C->set_has_split_ifs(true); // Has chance for split-if optimization
2016 return true;
2017 }
2019 //------------------------------runtime_math-----------------------------
2020 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2021 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2022 "must be (DD)D or (D)D type");
2024 // Inputs
2025 Node* a = round_double_node(argument(0));
2026 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2028 const TypePtr* no_memory_effects = NULL;
2029 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2030 no_memory_effects,
2031 a, top(), b, b ? top() : NULL);
2032 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2033 #ifdef ASSERT
2034 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2035 assert(value_top == top(), "second value must be top");
2036 #endif
2038 set_result(value);
2039 return true;
2040 }
2042 //------------------------------inline_math_native-----------------------------
2043 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2044 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2045 switch (id) {
2046 // These intrinsics are not properly supported on all hardware
2047 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2048 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2049 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2050 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2051 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2052 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2054 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2055 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2056 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2057 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2059 // These intrinsics are supported on all hardware
2060 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2061 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2063 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2064 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2065 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2066 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2067 #undef FN_PTR
2069 // These intrinsics are not yet correctly implemented
2070 case vmIntrinsics::_datan2:
2071 return false;
2073 default:
2074 fatal_unexpected_iid(id);
2075 return false;
2076 }
2077 }
2079 static bool is_simple_name(Node* n) {
2080 return (n->req() == 1 // constant
2081 || (n->is_Type() && n->as_Type()->type()->singleton())
2082 || n->is_Proj() // parameter or return value
2083 || n->is_Phi() // local of some sort
2084 );
2085 }
2087 //----------------------------inline_min_max-----------------------------------
2088 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2089 set_result(generate_min_max(id, argument(0), argument(1)));
2090 return true;
2091 }
2093 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2094 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2095 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2096 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2097 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2099 {
2100 PreserveJVMState pjvms(this);
2101 PreserveReexecuteState preexecs(this);
2102 jvms()->set_should_reexecute(true);
2104 set_control(slow_path);
2105 set_i_o(i_o());
2107 uncommon_trap(Deoptimization::Reason_intrinsic,
2108 Deoptimization::Action_none);
2109 }
2111 set_control(fast_path);
2112 set_result(math);
2113 }
2115 template <typename OverflowOp>
2116 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2117 typedef typename OverflowOp::MathOp MathOp;
2119 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2120 Node* operation = _gvn.transform( mathOp );
2121 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2122 inline_math_mathExact(operation, ofcheck);
2123 return true;
2124 }
2126 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2127 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2128 }
2130 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2131 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2132 }
2134 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2135 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2136 }
2138 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2139 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2140 }
2142 bool LibraryCallKit::inline_math_negateExactI() {
2143 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2144 }
2146 bool LibraryCallKit::inline_math_negateExactL() {
2147 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2148 }
2150 bool LibraryCallKit::inline_math_multiplyExactI() {
2151 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2152 }
2154 bool LibraryCallKit::inline_math_multiplyExactL() {
2155 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2156 }
2158 Node*
2159 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2160 // These are the candidate return value:
2161 Node* xvalue = x0;
2162 Node* yvalue = y0;
2164 if (xvalue == yvalue) {
2165 return xvalue;
2166 }
2168 bool want_max = (id == vmIntrinsics::_max);
2170 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2171 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2172 if (txvalue == NULL || tyvalue == NULL) return top();
2173 // This is not really necessary, but it is consistent with a
2174 // hypothetical MaxINode::Value method:
2175 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2177 // %%% This folding logic should (ideally) be in a different place.
2178 // Some should be inside IfNode, and there to be a more reliable
2179 // transformation of ?: style patterns into cmoves. We also want
2180 // more powerful optimizations around cmove and min/max.
2182 // Try to find a dominating comparison of these guys.
2183 // It can simplify the index computation for Arrays.copyOf
2184 // and similar uses of System.arraycopy.
2185 // First, compute the normalized version of CmpI(x, y).
2186 int cmp_op = Op_CmpI;
2187 Node* xkey = xvalue;
2188 Node* ykey = yvalue;
2189 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2190 if (ideal_cmpxy->is_Cmp()) {
2191 // E.g., if we have CmpI(length - offset, count),
2192 // it might idealize to CmpI(length, count + offset)
2193 cmp_op = ideal_cmpxy->Opcode();
2194 xkey = ideal_cmpxy->in(1);
2195 ykey = ideal_cmpxy->in(2);
2196 }
2198 // Start by locating any relevant comparisons.
2199 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2200 Node* cmpxy = NULL;
2201 Node* cmpyx = NULL;
2202 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2203 Node* cmp = start_from->fast_out(k);
2204 if (cmp->outcnt() > 0 && // must have prior uses
2205 cmp->in(0) == NULL && // must be context-independent
2206 cmp->Opcode() == cmp_op) { // right kind of compare
2207 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2208 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2209 }
2210 }
2212 const int NCMPS = 2;
2213 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2214 int cmpn;
2215 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2216 if (cmps[cmpn] != NULL) break; // find a result
2217 }
2218 if (cmpn < NCMPS) {
2219 // Look for a dominating test that tells us the min and max.
2220 int depth = 0; // Limit search depth for speed
2221 Node* dom = control();
2222 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2223 if (++depth >= 100) break;
2224 Node* ifproj = dom;
2225 if (!ifproj->is_Proj()) continue;
2226 Node* iff = ifproj->in(0);
2227 if (!iff->is_If()) continue;
2228 Node* bol = iff->in(1);
2229 if (!bol->is_Bool()) continue;
2230 Node* cmp = bol->in(1);
2231 if (cmp == NULL) continue;
2232 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2233 if (cmps[cmpn] == cmp) break;
2234 if (cmpn == NCMPS) continue;
2235 BoolTest::mask btest = bol->as_Bool()->_test._test;
2236 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2237 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2238 // At this point, we know that 'x btest y' is true.
2239 switch (btest) {
2240 case BoolTest::eq:
2241 // They are proven equal, so we can collapse the min/max.
2242 // Either value is the answer. Choose the simpler.
2243 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2244 return yvalue;
2245 return xvalue;
2246 case BoolTest::lt: // x < y
2247 case BoolTest::le: // x <= y
2248 return (want_max ? yvalue : xvalue);
2249 case BoolTest::gt: // x > y
2250 case BoolTest::ge: // x >= y
2251 return (want_max ? xvalue : yvalue);
2252 }
2253 }
2254 }
2256 // We failed to find a dominating test.
2257 // Let's pick a test that might GVN with prior tests.
2258 Node* best_bol = NULL;
2259 BoolTest::mask best_btest = BoolTest::illegal;
2260 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2261 Node* cmp = cmps[cmpn];
2262 if (cmp == NULL) continue;
2263 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2264 Node* bol = cmp->fast_out(j);
2265 if (!bol->is_Bool()) continue;
2266 BoolTest::mask btest = bol->as_Bool()->_test._test;
2267 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2268 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2269 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2270 best_bol = bol->as_Bool();
2271 best_btest = btest;
2272 }
2273 }
2274 }
2276 Node* answer_if_true = NULL;
2277 Node* answer_if_false = NULL;
2278 switch (best_btest) {
2279 default:
2280 if (cmpxy == NULL)
2281 cmpxy = ideal_cmpxy;
2282 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2283 // and fall through:
2284 case BoolTest::lt: // x < y
2285 case BoolTest::le: // x <= y
2286 answer_if_true = (want_max ? yvalue : xvalue);
2287 answer_if_false = (want_max ? xvalue : yvalue);
2288 break;
2289 case BoolTest::gt: // x > y
2290 case BoolTest::ge: // x >= y
2291 answer_if_true = (want_max ? xvalue : yvalue);
2292 answer_if_false = (want_max ? yvalue : xvalue);
2293 break;
2294 }
2296 jint hi, lo;
2297 if (want_max) {
2298 // We can sharpen the minimum.
2299 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2300 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2301 } else {
2302 // We can sharpen the maximum.
2303 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2304 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2305 }
2307 // Use a flow-free graph structure, to avoid creating excess control edges
2308 // which could hinder other optimizations.
2309 // Since Math.min/max is often used with arraycopy, we want
2310 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2311 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2312 answer_if_false, answer_if_true,
2313 TypeInt::make(lo, hi, widen));
2315 return _gvn.transform(cmov);
2317 /*
2318 // This is not as desirable as it may seem, since Min and Max
2319 // nodes do not have a full set of optimizations.
2320 // And they would interfere, anyway, with 'if' optimizations
2321 // and with CMoveI canonical forms.
2322 switch (id) {
2323 case vmIntrinsics::_min:
2324 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2325 case vmIntrinsics::_max:
2326 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2327 default:
2328 ShouldNotReachHere();
2329 }
2330 */
2331 }
2333 inline int
2334 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2335 const TypePtr* base_type = TypePtr::NULL_PTR;
2336 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2337 if (base_type == NULL) {
2338 // Unknown type.
2339 return Type::AnyPtr;
2340 } else if (base_type == TypePtr::NULL_PTR) {
2341 // Since this is a NULL+long form, we have to switch to a rawptr.
2342 base = _gvn.transform(new (C) CastX2PNode(offset));
2343 offset = MakeConX(0);
2344 return Type::RawPtr;
2345 } else if (base_type->base() == Type::RawPtr) {
2346 return Type::RawPtr;
2347 } else if (base_type->isa_oopptr()) {
2348 // Base is never null => always a heap address.
2349 if (base_type->ptr() == TypePtr::NotNull) {
2350 return Type::OopPtr;
2351 }
2352 // Offset is small => always a heap address.
2353 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2354 if (offset_type != NULL &&
2355 base_type->offset() == 0 && // (should always be?)
2356 offset_type->_lo >= 0 &&
2357 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2358 return Type::OopPtr;
2359 }
2360 // Otherwise, it might either be oop+off or NULL+addr.
2361 return Type::AnyPtr;
2362 } else {
2363 // No information:
2364 return Type::AnyPtr;
2365 }
2366 }
2368 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2369 int kind = classify_unsafe_addr(base, offset);
2370 if (kind == Type::RawPtr) {
2371 return basic_plus_adr(top(), base, offset);
2372 } else {
2373 return basic_plus_adr(base, offset);
2374 }
2375 }
2377 //--------------------------inline_number_methods-----------------------------
2378 // inline int Integer.numberOfLeadingZeros(int)
2379 // inline int Long.numberOfLeadingZeros(long)
2380 //
2381 // inline int Integer.numberOfTrailingZeros(int)
2382 // inline int Long.numberOfTrailingZeros(long)
2383 //
2384 // inline int Integer.bitCount(int)
2385 // inline int Long.bitCount(long)
2386 //
2387 // inline char Character.reverseBytes(char)
2388 // inline short Short.reverseBytes(short)
2389 // inline int Integer.reverseBytes(int)
2390 // inline long Long.reverseBytes(long)
2391 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2392 Node* arg = argument(0);
2393 Node* n = NULL;
2394 switch (id) {
2395 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2396 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2397 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2398 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2399 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2400 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2401 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2402 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2403 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2404 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2405 default: fatal_unexpected_iid(id); break;
2406 }
2407 set_result(_gvn.transform(n));
2408 return true;
2409 }
2411 //----------------------------inline_unsafe_access----------------------------
2413 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2415 // Helper that guards and inserts a pre-barrier.
2416 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2417 Node* pre_val, bool need_mem_bar) {
2418 // We could be accessing the referent field of a reference object. If so, when G1
2419 // is enabled, we need to log the value in the referent field in an SATB buffer.
2420 // This routine performs some compile time filters and generates suitable
2421 // runtime filters that guard the pre-barrier code.
2422 // Also add memory barrier for non volatile load from the referent field
2423 // to prevent commoning of loads across safepoint.
2424 if (!UseG1GC && !need_mem_bar)
2425 return;
2427 // Some compile time checks.
2429 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2430 const TypeX* otype = offset->find_intptr_t_type();
2431 if (otype != NULL && otype->is_con() &&
2432 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2433 // Constant offset but not the reference_offset so just return
2434 return;
2435 }
2437 // We only need to generate the runtime guards for instances.
2438 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2439 if (btype != NULL) {
2440 if (btype->isa_aryptr()) {
2441 // Array type so nothing to do
2442 return;
2443 }
2445 const TypeInstPtr* itype = btype->isa_instptr();
2446 if (itype != NULL) {
2447 // Can the klass of base_oop be statically determined to be
2448 // _not_ a sub-class of Reference and _not_ Object?
2449 ciKlass* klass = itype->klass();
2450 if ( klass->is_loaded() &&
2451 !klass->is_subtype_of(env()->Reference_klass()) &&
2452 !env()->Object_klass()->is_subtype_of(klass)) {
2453 return;
2454 }
2455 }
2456 }
2458 // The compile time filters did not reject base_oop/offset so
2459 // we need to generate the following runtime filters
2460 //
2461 // if (offset == java_lang_ref_Reference::_reference_offset) {
2462 // if (instance_of(base, java.lang.ref.Reference)) {
2463 // pre_barrier(_, pre_val, ...);
2464 // }
2465 // }
2467 float likely = PROB_LIKELY( 0.999);
2468 float unlikely = PROB_UNLIKELY(0.999);
2470 IdealKit ideal(this);
2471 #define __ ideal.
2473 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2475 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2476 // Update graphKit memory and control from IdealKit.
2477 sync_kit(ideal);
2479 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2480 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2482 // Update IdealKit memory and control from graphKit.
2483 __ sync_kit(this);
2485 Node* one = __ ConI(1);
2486 // is_instof == 0 if base_oop == NULL
2487 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2489 // Update graphKit from IdeakKit.
2490 sync_kit(ideal);
2492 // Use the pre-barrier to record the value in the referent field
2493 pre_barrier(false /* do_load */,
2494 __ ctrl(),
2495 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2496 pre_val /* pre_val */,
2497 T_OBJECT);
2498 if (need_mem_bar) {
2499 // Add memory barrier to prevent commoning reads from this field
2500 // across safepoint since GC can change its value.
2501 insert_mem_bar(Op_MemBarCPUOrder);
2502 }
2503 // Update IdealKit from graphKit.
2504 __ sync_kit(this);
2506 } __ end_if(); // _ref_type != ref_none
2507 } __ end_if(); // offset == referent_offset
2509 // Final sync IdealKit and GraphKit.
2510 final_sync(ideal);
2511 #undef __
2512 }
2515 // Interpret Unsafe.fieldOffset cookies correctly:
2516 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2518 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2519 // Attempt to infer a sharper value type from the offset and base type.
2520 ciKlass* sharpened_klass = NULL;
2522 // See if it is an instance field, with an object type.
2523 if (alias_type->field() != NULL) {
2524 assert(!is_native_ptr, "native pointer op cannot use a java address");
2525 if (alias_type->field()->type()->is_klass()) {
2526 sharpened_klass = alias_type->field()->type()->as_klass();
2527 }
2528 }
2530 // See if it is a narrow oop array.
2531 if (adr_type->isa_aryptr()) {
2532 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2533 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2534 if (elem_type != NULL) {
2535 sharpened_klass = elem_type->klass();
2536 }
2537 }
2538 }
2540 // The sharpened class might be unloaded if there is no class loader
2541 // contraint in place.
2542 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2543 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2545 #ifndef PRODUCT
2546 if (C->print_intrinsics() || C->print_inlining()) {
2547 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2548 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2549 }
2550 #endif
2551 // Sharpen the value type.
2552 return tjp;
2553 }
2554 return NULL;
2555 }
2557 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2558 if (callee()->is_static()) return false; // caller must have the capability!
2559 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2561 #ifndef PRODUCT
2562 {
2563 ResourceMark rm;
2564 // Check the signatures.
2565 ciSignature* sig = callee()->signature();
2566 #ifdef ASSERT
2567 if (!is_store) {
2568 // Object getObject(Object base, int/long offset), etc.
2569 BasicType rtype = sig->return_type()->basic_type();
2570 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2571 rtype = T_ADDRESS; // it is really a C void*
2572 assert(rtype == type, "getter must return the expected value");
2573 if (!is_native_ptr) {
2574 assert(sig->count() == 2, "oop getter has 2 arguments");
2575 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2576 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2577 } else {
2578 assert(sig->count() == 1, "native getter has 1 argument");
2579 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2580 }
2581 } else {
2582 // void putObject(Object base, int/long offset, Object x), etc.
2583 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2584 if (!is_native_ptr) {
2585 assert(sig->count() == 3, "oop putter has 3 arguments");
2586 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2587 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2588 } else {
2589 assert(sig->count() == 2, "native putter has 2 arguments");
2590 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2591 }
2592 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2593 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2594 vtype = T_ADDRESS; // it is really a C void*
2595 assert(vtype == type, "putter must accept the expected value");
2596 }
2597 #endif // ASSERT
2598 }
2599 #endif //PRODUCT
2601 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2603 Node* receiver = argument(0); // type: oop
2605 // Build address expression. See the code in inline_unsafe_prefetch.
2606 Node* adr;
2607 Node* heap_base_oop = top();
2608 Node* offset = top();
2609 Node* val;
2611 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2612 Node* base = argument(1); // type: oop
2614 if (!is_native_ptr) {
2615 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2616 offset = argument(2); // type: long
2617 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2618 // to be plain byte offsets, which are also the same as those accepted
2619 // by oopDesc::field_base.
2620 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2621 "fieldOffset must be byte-scaled");
2622 // 32-bit machines ignore the high half!
2623 offset = ConvL2X(offset);
2624 adr = make_unsafe_address(base, offset);
2625 heap_base_oop = base;
2626 val = is_store ? argument(4) : NULL;
2627 } else {
2628 Node* ptr = argument(1); // type: long
2629 ptr = ConvL2X(ptr); // adjust Java long to machine word
2630 adr = make_unsafe_address(NULL, ptr);
2631 val = is_store ? argument(3) : NULL;
2632 }
2634 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) {
2635 return false; // off-heap oop accesses are not supported
2636 }
2638 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2640 // Try to categorize the address.
2641 Compile::AliasType* alias_type = C->alias_type(adr_type);
2642 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2644 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2645 alias_type->adr_type() == TypeAryPtr::RANGE) {
2646 return false; // not supported
2647 }
2649 bool mismatched = false;
2650 BasicType bt = alias_type->basic_type();
2651 if (bt != T_ILLEGAL) {
2652 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2653 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2654 // Alias type doesn't differentiate between byte[] and boolean[]).
2655 // Use address type to get the element type.
2656 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2657 }
2658 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2659 // accessing an array field with getObject is not a mismatch
2660 bt = T_OBJECT;
2661 }
2662 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2663 // Don't intrinsify mismatched object accesses
2664 return false;
2665 }
2666 mismatched = (bt != type);
2667 }
2669 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2671 // First guess at the value type.
2672 const Type *value_type = Type::get_const_basic_type(type);
2674 // We will need memory barriers unless we can determine a unique
2675 // alias category for this reference. (Note: If for some reason
2676 // the barriers get omitted and the unsafe reference begins to "pollute"
2677 // the alias analysis of the rest of the graph, either Compile::can_alias
2678 // or Compile::must_alias will throw a diagnostic assert.)
2679 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2681 // If we are reading the value of the referent field of a Reference
2682 // object (either by using Unsafe directly or through reflection)
2683 // then, if G1 is enabled, we need to record the referent in an
2684 // SATB log buffer using the pre-barrier mechanism.
2685 // Also we need to add memory barrier to prevent commoning reads
2686 // from this field across safepoint since GC can change its value.
2687 bool need_read_barrier = !is_native_ptr && !is_store &&
2688 offset != top() && heap_base_oop != top();
2690 if (!is_store && type == T_OBJECT) {
2691 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2692 if (tjp != NULL) {
2693 value_type = tjp;
2694 }
2695 }
2697 receiver = null_check(receiver);
2698 if (stopped()) {
2699 return true;
2700 }
2701 // Heap pointers get a null-check from the interpreter,
2702 // as a courtesy. However, this is not guaranteed by Unsafe,
2703 // and it is not possible to fully distinguish unintended nulls
2704 // from intended ones in this API.
2706 if (is_volatile) {
2707 // We need to emit leading and trailing CPU membars (see below) in
2708 // addition to memory membars when is_volatile. This is a little
2709 // too strong, but avoids the need to insert per-alias-type
2710 // volatile membars (for stores; compare Parse::do_put_xxx), which
2711 // we cannot do effectively here because we probably only have a
2712 // rough approximation of type.
2713 need_mem_bar = true;
2714 // For Stores, place a memory ordering barrier now.
2715 if (is_store) {
2716 insert_mem_bar(Op_MemBarRelease);
2717 } else {
2718 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2719 insert_mem_bar(Op_MemBarVolatile);
2720 }
2721 }
2722 }
2724 // Memory barrier to prevent normal and 'unsafe' accesses from
2725 // bypassing each other. Happens after null checks, so the
2726 // exception paths do not take memory state from the memory barrier,
2727 // so there's no problems making a strong assert about mixing users
2728 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2729 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2730 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2732 if (!is_store) {
2733 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2734 // To be valid, unsafe loads may depend on other conditions than
2735 // the one that guards them: pin the Load node
2736 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2737 // load value
2738 switch (type) {
2739 case T_BOOLEAN:
2740 case T_CHAR:
2741 case T_BYTE:
2742 case T_SHORT:
2743 case T_INT:
2744 case T_LONG:
2745 case T_FLOAT:
2746 case T_DOUBLE:
2747 break;
2748 case T_OBJECT:
2749 if (need_read_barrier) {
2750 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2751 }
2752 break;
2753 case T_ADDRESS:
2754 // Cast to an int type.
2755 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2756 p = ConvX2UL(p);
2757 break;
2758 default:
2759 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2760 break;
2761 }
2762 // The load node has the control of the preceding MemBarCPUOrder. All
2763 // following nodes will have the control of the MemBarCPUOrder inserted at
2764 // the end of this method. So, pushing the load onto the stack at a later
2765 // point is fine.
2766 set_result(p);
2767 } else {
2768 // place effect of store into memory
2769 switch (type) {
2770 case T_DOUBLE:
2771 val = dstore_rounding(val);
2772 break;
2773 case T_ADDRESS:
2774 // Repackage the long as a pointer.
2775 val = ConvL2X(val);
2776 val = _gvn.transform(new (C) CastX2PNode(val));
2777 break;
2778 }
2780 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2781 if (type == T_OBJECT ) {
2782 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2783 } else {
2784 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2785 }
2786 }
2788 if (is_volatile) {
2789 if (!is_store) {
2790 insert_mem_bar(Op_MemBarAcquire);
2791 } else {
2792 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2793 insert_mem_bar(Op_MemBarVolatile);
2794 }
2795 }
2796 }
2798 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2800 return true;
2801 }
2803 //----------------------------inline_unsafe_prefetch----------------------------
2805 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2806 #ifndef PRODUCT
2807 {
2808 ResourceMark rm;
2809 // Check the signatures.
2810 ciSignature* sig = callee()->signature();
2811 #ifdef ASSERT
2812 // Object getObject(Object base, int/long offset), etc.
2813 BasicType rtype = sig->return_type()->basic_type();
2814 if (!is_native_ptr) {
2815 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2816 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2817 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2818 } else {
2819 assert(sig->count() == 1, "native prefetch has 1 argument");
2820 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2821 }
2822 #endif // ASSERT
2823 }
2824 #endif // !PRODUCT
2826 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2828 const int idx = is_static ? 0 : 1;
2829 if (!is_static) {
2830 null_check_receiver();
2831 if (stopped()) {
2832 return true;
2833 }
2834 }
2836 // Build address expression. See the code in inline_unsafe_access.
2837 Node *adr;
2838 if (!is_native_ptr) {
2839 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2840 Node* base = argument(idx + 0); // type: oop
2841 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2842 Node* offset = argument(idx + 1); // type: long
2843 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2844 // to be plain byte offsets, which are also the same as those accepted
2845 // by oopDesc::field_base.
2846 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2847 "fieldOffset must be byte-scaled");
2848 // 32-bit machines ignore the high half!
2849 offset = ConvL2X(offset);
2850 adr = make_unsafe_address(base, offset);
2851 } else {
2852 Node* ptr = argument(idx + 0); // type: long
2853 ptr = ConvL2X(ptr); // adjust Java long to machine word
2854 adr = make_unsafe_address(NULL, ptr);
2855 }
2857 // Generate the read or write prefetch
2858 Node *prefetch;
2859 if (is_store) {
2860 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2861 } else {
2862 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2863 }
2864 prefetch->init_req(0, control());
2865 set_i_o(_gvn.transform(prefetch));
2867 return true;
2868 }
2870 //----------------------------inline_unsafe_load_store----------------------------
2871 // This method serves a couple of different customers (depending on LoadStoreKind):
2872 //
2873 // LS_cmpxchg:
2874 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2875 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2876 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2877 //
2878 // LS_xadd:
2879 // public int getAndAddInt( Object o, long offset, int delta)
2880 // public long getAndAddLong(Object o, long offset, long delta)
2881 //
2882 // LS_xchg:
2883 // int getAndSet(Object o, long offset, int newValue)
2884 // long getAndSet(Object o, long offset, long newValue)
2885 // Object getAndSet(Object o, long offset, Object newValue)
2886 //
2887 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2888 // This basic scheme here is the same as inline_unsafe_access, but
2889 // differs in enough details that combining them would make the code
2890 // overly confusing. (This is a true fact! I originally combined
2891 // them, but even I was confused by it!) As much code/comments as
2892 // possible are retained from inline_unsafe_access though to make
2893 // the correspondences clearer. - dl
2895 if (callee()->is_static()) return false; // caller must have the capability!
2897 #ifndef PRODUCT
2898 BasicType rtype;
2899 {
2900 ResourceMark rm;
2901 // Check the signatures.
2902 ciSignature* sig = callee()->signature();
2903 rtype = sig->return_type()->basic_type();
2904 if (kind == LS_xadd || kind == LS_xchg) {
2905 // Check the signatures.
2906 #ifdef ASSERT
2907 assert(rtype == type, "get and set must return the expected type");
2908 assert(sig->count() == 3, "get and set has 3 arguments");
2909 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2910 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2911 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2912 #endif // ASSERT
2913 } else if (kind == LS_cmpxchg) {
2914 // Check the signatures.
2915 #ifdef ASSERT
2916 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2917 assert(sig->count() == 4, "CAS has 4 arguments");
2918 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2919 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2920 #endif // ASSERT
2921 } else {
2922 ShouldNotReachHere();
2923 }
2924 }
2925 #endif //PRODUCT
2927 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2929 // Get arguments:
2930 Node* receiver = NULL;
2931 Node* base = NULL;
2932 Node* offset = NULL;
2933 Node* oldval = NULL;
2934 Node* newval = NULL;
2935 if (kind == LS_cmpxchg) {
2936 const bool two_slot_type = type2size[type] == 2;
2937 receiver = argument(0); // type: oop
2938 base = argument(1); // type: oop
2939 offset = argument(2); // type: long
2940 oldval = argument(4); // type: oop, int, or long
2941 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2942 } else if (kind == LS_xadd || kind == LS_xchg){
2943 receiver = argument(0); // type: oop
2944 base = argument(1); // type: oop
2945 offset = argument(2); // type: long
2946 oldval = NULL;
2947 newval = argument(4); // type: oop, int, or long
2948 }
2950 // Build field offset expression.
2951 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2952 // to be plain byte offsets, which are also the same as those accepted
2953 // by oopDesc::field_base.
2954 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2955 // 32-bit machines ignore the high half of long offsets
2956 offset = ConvL2X(offset);
2957 Node* adr = make_unsafe_address(base, offset);
2958 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2960 Compile::AliasType* alias_type = C->alias_type(adr_type);
2961 BasicType bt = alias_type->basic_type();
2962 if (bt != T_ILLEGAL &&
2963 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2964 // Don't intrinsify mismatched object accesses.
2965 return false;
2966 }
2968 // For CAS, unlike inline_unsafe_access, there seems no point in
2969 // trying to refine types. Just use the coarse types here.
2970 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2971 const Type *value_type = Type::get_const_basic_type(type);
2973 if (kind == LS_xchg && type == T_OBJECT) {
2974 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2975 if (tjp != NULL) {
2976 value_type = tjp;
2977 }
2978 }
2980 // Null check receiver.
2981 receiver = null_check(receiver);
2982 if (stopped()) {
2983 return true;
2984 }
2986 int alias_idx = C->get_alias_index(adr_type);
2988 // Memory-model-wise, a LoadStore acts like a little synchronized
2989 // block, so needs barriers on each side. These don't translate
2990 // into actual barriers on most machines, but we still need rest of
2991 // compiler to respect ordering.
2993 insert_mem_bar(Op_MemBarRelease);
2994 insert_mem_bar(Op_MemBarCPUOrder);
2996 // 4984716: MemBars must be inserted before this
2997 // memory node in order to avoid a false
2998 // dependency which will confuse the scheduler.
2999 Node *mem = memory(alias_idx);
3001 // For now, we handle only those cases that actually exist: ints,
3002 // longs, and Object. Adding others should be straightforward.
3003 Node* load_store = NULL;
3004 switch(type) {
3005 case T_INT:
3006 if (kind == LS_xadd) {
3007 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3008 } else if (kind == LS_xchg) {
3009 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3010 } else if (kind == LS_cmpxchg) {
3011 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3012 } else {
3013 ShouldNotReachHere();
3014 }
3015 break;
3016 case T_LONG:
3017 if (kind == LS_xadd) {
3018 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3019 } else if (kind == LS_xchg) {
3020 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3021 } else if (kind == LS_cmpxchg) {
3022 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3023 } else {
3024 ShouldNotReachHere();
3025 }
3026 break;
3027 case T_OBJECT:
3028 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3029 // could be delayed during Parse (for example, in adjust_map_after_if()).
3030 // Execute transformation here to avoid barrier generation in such case.
3031 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3032 newval = _gvn.makecon(TypePtr::NULL_PTR);
3034 // Reference stores need a store barrier.
3035 if (kind == LS_xchg) {
3036 // If pre-barrier must execute before the oop store, old value will require do_load here.
3037 if (!can_move_pre_barrier()) {
3038 pre_barrier(true /* do_load*/,
3039 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3040 NULL /* pre_val*/,
3041 T_OBJECT);
3042 } // Else move pre_barrier to use load_store value, see below.
3043 } else if (kind == LS_cmpxchg) {
3044 // Same as for newval above:
3045 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3046 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3047 }
3048 // The only known value which might get overwritten is oldval.
3049 pre_barrier(false /* do_load */,
3050 control(), NULL, NULL, max_juint, NULL, NULL,
3051 oldval /* pre_val */,
3052 T_OBJECT);
3053 } else {
3054 ShouldNotReachHere();
3055 }
3057 #ifdef _LP64
3058 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3059 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3060 if (kind == LS_xchg) {
3061 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3062 newval_enc, adr_type, value_type->make_narrowoop()));
3063 } else {
3064 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3065 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3066 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3067 newval_enc, oldval_enc));
3068 }
3069 } else
3070 #endif
3071 {
3072 if (kind == LS_xchg) {
3073 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3074 } else {
3075 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3076 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3077 }
3078 }
3079 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3080 break;
3081 default:
3082 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3083 break;
3084 }
3086 // SCMemProjNodes represent the memory state of a LoadStore. Their
3087 // main role is to prevent LoadStore nodes from being optimized away
3088 // when their results aren't used.
3089 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3090 set_memory(proj, alias_idx);
3092 if (type == T_OBJECT && kind == LS_xchg) {
3093 #ifdef _LP64
3094 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3095 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3096 }
3097 #endif
3098 if (can_move_pre_barrier()) {
3099 // Don't need to load pre_val. The old value is returned by load_store.
3100 // The pre_barrier can execute after the xchg as long as no safepoint
3101 // gets inserted between them.
3102 pre_barrier(false /* do_load */,
3103 control(), NULL, NULL, max_juint, NULL, NULL,
3104 load_store /* pre_val */,
3105 T_OBJECT);
3106 }
3107 }
3109 // Add the trailing membar surrounding the access
3110 insert_mem_bar(Op_MemBarCPUOrder);
3111 insert_mem_bar(Op_MemBarAcquire);
3113 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3114 set_result(load_store);
3115 return true;
3116 }
3118 //----------------------------inline_unsafe_ordered_store----------------------
3119 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3120 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3121 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3122 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3123 // This is another variant of inline_unsafe_access, differing in
3124 // that it always issues store-store ("release") barrier and ensures
3125 // store-atomicity (which only matters for "long").
3127 if (callee()->is_static()) return false; // caller must have the capability!
3129 #ifndef PRODUCT
3130 {
3131 ResourceMark rm;
3132 // Check the signatures.
3133 ciSignature* sig = callee()->signature();
3134 #ifdef ASSERT
3135 BasicType rtype = sig->return_type()->basic_type();
3136 assert(rtype == T_VOID, "must return void");
3137 assert(sig->count() == 3, "has 3 arguments");
3138 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3139 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3140 #endif // ASSERT
3141 }
3142 #endif //PRODUCT
3144 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3146 // Get arguments:
3147 Node* receiver = argument(0); // type: oop
3148 Node* base = argument(1); // type: oop
3149 Node* offset = argument(2); // type: long
3150 Node* val = argument(4); // type: oop, int, or long
3152 // Null check receiver.
3153 receiver = null_check(receiver);
3154 if (stopped()) {
3155 return true;
3156 }
3158 // Build field offset expression.
3159 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3160 // 32-bit machines ignore the high half of long offsets
3161 offset = ConvL2X(offset);
3162 Node* adr = make_unsafe_address(base, offset);
3163 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3164 const Type *value_type = Type::get_const_basic_type(type);
3165 Compile::AliasType* alias_type = C->alias_type(adr_type);
3167 insert_mem_bar(Op_MemBarRelease);
3168 insert_mem_bar(Op_MemBarCPUOrder);
3169 // Ensure that the store is atomic for longs:
3170 const bool require_atomic_access = true;
3171 Node* store;
3172 if (type == T_OBJECT) // reference stores need a store barrier.
3173 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3174 else {
3175 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3176 }
3177 insert_mem_bar(Op_MemBarCPUOrder);
3178 return true;
3179 }
3181 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3182 // Regardless of form, don't allow previous ld/st to move down,
3183 // then issue acquire, release, or volatile mem_bar.
3184 insert_mem_bar(Op_MemBarCPUOrder);
3185 switch(id) {
3186 case vmIntrinsics::_loadFence:
3187 insert_mem_bar(Op_LoadFence);
3188 return true;
3189 case vmIntrinsics::_storeFence:
3190 insert_mem_bar(Op_StoreFence);
3191 return true;
3192 case vmIntrinsics::_fullFence:
3193 insert_mem_bar(Op_MemBarVolatile);
3194 return true;
3195 default:
3196 fatal_unexpected_iid(id);
3197 return false;
3198 }
3199 }
3201 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3202 if (!kls->is_Con()) {
3203 return true;
3204 }
3205 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3206 if (klsptr == NULL) {
3207 return true;
3208 }
3209 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3210 // don't need a guard for a klass that is already initialized
3211 return !ik->is_initialized();
3212 }
3214 //----------------------------inline_unsafe_allocate---------------------------
3215 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3216 bool LibraryCallKit::inline_unsafe_allocate() {
3217 if (callee()->is_static()) return false; // caller must have the capability!
3219 null_check_receiver(); // null-check, then ignore
3220 Node* cls = null_check(argument(1));
3221 if (stopped()) return true;
3223 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3224 kls = null_check(kls);
3225 if (stopped()) return true; // argument was like int.class
3227 Node* test = NULL;
3228 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3229 // Note: The argument might still be an illegal value like
3230 // Serializable.class or Object[].class. The runtime will handle it.
3231 // But we must make an explicit check for initialization.
3232 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3233 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3234 // can generate code to load it as unsigned byte.
3235 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3236 Node* bits = intcon(InstanceKlass::fully_initialized);
3237 test = _gvn.transform(new (C) SubINode(inst, bits));
3238 // The 'test' is non-zero if we need to take a slow path.
3239 }
3241 Node* obj = new_instance(kls, test);
3242 set_result(obj);
3243 return true;
3244 }
3246 #ifdef TRACE_HAVE_INTRINSICS
3247 /*
3248 * oop -> myklass
3249 * myklass->trace_id |= USED
3250 * return myklass->trace_id & ~0x3
3251 */
3252 bool LibraryCallKit::inline_native_classID() {
3253 null_check_receiver(); // null-check, then ignore
3254 Node* cls = null_check(argument(1), T_OBJECT);
3255 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3256 kls = null_check(kls, T_OBJECT);
3257 ByteSize offset = TRACE_ID_OFFSET;
3258 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3259 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3260 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3261 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3262 Node* clsused = longcon(0x01l); // set the class bit
3263 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3265 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3266 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3267 set_result(andl);
3268 return true;
3269 }
3271 bool LibraryCallKit::inline_native_threadID() {
3272 Node* tls_ptr = NULL;
3273 Node* cur_thr = generate_current_thread(tls_ptr);
3274 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3275 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3276 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3278 Node* threadid = NULL;
3279 size_t thread_id_size = OSThread::thread_id_size();
3280 if (thread_id_size == (size_t) BytesPerLong) {
3281 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3282 } else if (thread_id_size == (size_t) BytesPerInt) {
3283 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3284 } else {
3285 ShouldNotReachHere();
3286 }
3287 set_result(threadid);
3288 return true;
3289 }
3290 #endif
3292 //------------------------inline_native_time_funcs--------------
3293 // inline code for System.currentTimeMillis() and System.nanoTime()
3294 // these have the same type and signature
3295 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3296 const TypeFunc* tf = OptoRuntime::void_long_Type();
3297 const TypePtr* no_memory_effects = NULL;
3298 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3299 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3300 #ifdef ASSERT
3301 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3302 assert(value_top == top(), "second value must be top");
3303 #endif
3304 set_result(value);
3305 return true;
3306 }
3308 //------------------------inline_native_currentThread------------------
3309 bool LibraryCallKit::inline_native_currentThread() {
3310 Node* junk = NULL;
3311 set_result(generate_current_thread(junk));
3312 return true;
3313 }
3315 //------------------------inline_native_isInterrupted------------------
3316 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3317 bool LibraryCallKit::inline_native_isInterrupted() {
3318 // Add a fast path to t.isInterrupted(clear_int):
3319 // (t == Thread.current() &&
3320 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3321 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3322 // So, in the common case that the interrupt bit is false,
3323 // we avoid making a call into the VM. Even if the interrupt bit
3324 // is true, if the clear_int argument is false, we avoid the VM call.
3325 // However, if the receiver is not currentThread, we must call the VM,
3326 // because there must be some locking done around the operation.
3328 // We only go to the fast case code if we pass two guards.
3329 // Paths which do not pass are accumulated in the slow_region.
3331 enum {
3332 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3333 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3334 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3335 PATH_LIMIT
3336 };
3338 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3339 // out of the function.
3340 insert_mem_bar(Op_MemBarCPUOrder);
3342 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3343 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3345 RegionNode* slow_region = new (C) RegionNode(1);
3346 record_for_igvn(slow_region);
3348 // (a) Receiving thread must be the current thread.
3349 Node* rec_thr = argument(0);
3350 Node* tls_ptr = NULL;
3351 Node* cur_thr = generate_current_thread(tls_ptr);
3352 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3353 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3355 generate_slow_guard(bol_thr, slow_region);
3357 // (b) Interrupt bit on TLS must be false.
3358 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3359 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3360 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3362 // Set the control input on the field _interrupted read to prevent it floating up.
3363 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3364 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3365 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3367 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3369 // First fast path: if (!TLS._interrupted) return false;
3370 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3371 result_rgn->init_req(no_int_result_path, false_bit);
3372 result_val->init_req(no_int_result_path, intcon(0));
3374 // drop through to next case
3375 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3377 #ifndef TARGET_OS_FAMILY_windows
3378 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3379 Node* clr_arg = argument(1);
3380 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3381 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3382 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3384 // Second fast path: ... else if (!clear_int) return true;
3385 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3386 result_rgn->init_req(no_clear_result_path, false_arg);
3387 result_val->init_req(no_clear_result_path, intcon(1));
3389 // drop through to next case
3390 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3391 #else
3392 // To return true on Windows you must read the _interrupted field
3393 // and check the the event state i.e. take the slow path.
3394 #endif // TARGET_OS_FAMILY_windows
3396 // (d) Otherwise, go to the slow path.
3397 slow_region->add_req(control());
3398 set_control( _gvn.transform(slow_region));
3400 if (stopped()) {
3401 // There is no slow path.
3402 result_rgn->init_req(slow_result_path, top());
3403 result_val->init_req(slow_result_path, top());
3404 } else {
3405 // non-virtual because it is a private non-static
3406 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3408 Node* slow_val = set_results_for_java_call(slow_call);
3409 // this->control() comes from set_results_for_java_call
3411 Node* fast_io = slow_call->in(TypeFunc::I_O);
3412 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3414 // These two phis are pre-filled with copies of of the fast IO and Memory
3415 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3416 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3418 result_rgn->init_req(slow_result_path, control());
3419 result_io ->init_req(slow_result_path, i_o());
3420 result_mem->init_req(slow_result_path, reset_memory());
3421 result_val->init_req(slow_result_path, slow_val);
3423 set_all_memory(_gvn.transform(result_mem));
3424 set_i_o( _gvn.transform(result_io));
3425 }
3427 C->set_has_split_ifs(true); // Has chance for split-if optimization
3428 set_result(result_rgn, result_val);
3429 return true;
3430 }
3432 //---------------------------load_mirror_from_klass----------------------------
3433 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3434 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3435 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3436 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3437 }
3439 //-----------------------load_klass_from_mirror_common-------------------------
3440 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3441 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3442 // and branch to the given path on the region.
3443 // If never_see_null, take an uncommon trap on null, so we can optimistically
3444 // compile for the non-null case.
3445 // If the region is NULL, force never_see_null = true.
3446 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3447 bool never_see_null,
3448 RegionNode* region,
3449 int null_path,
3450 int offset) {
3451 if (region == NULL) never_see_null = true;
3452 Node* p = basic_plus_adr(mirror, offset);
3453 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3454 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3455 Node* null_ctl = top();
3456 kls = null_check_oop(kls, &null_ctl, never_see_null);
3457 if (region != NULL) {
3458 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3459 region->init_req(null_path, null_ctl);
3460 } else {
3461 assert(null_ctl == top(), "no loose ends");
3462 }
3463 return kls;
3464 }
3466 //--------------------(inline_native_Class_query helpers)---------------------
3467 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3468 // Fall through if (mods & mask) == bits, take the guard otherwise.
3469 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3470 // Branch around if the given klass has the given modifier bit set.
3471 // Like generate_guard, adds a new path onto the region.
3472 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3473 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3474 Node* mask = intcon(modifier_mask);
3475 Node* bits = intcon(modifier_bits);
3476 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3477 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3478 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3479 return generate_fair_guard(bol, region);
3480 }
3481 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3482 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3483 }
3485 //-------------------------inline_native_Class_query-------------------
3486 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3487 const Type* return_type = TypeInt::BOOL;
3488 Node* prim_return_value = top(); // what happens if it's a primitive class?
3489 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3490 bool expect_prim = false; // most of these guys expect to work on refs
3492 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3494 Node* mirror = argument(0);
3495 Node* obj = top();
3497 switch (id) {
3498 case vmIntrinsics::_isInstance:
3499 // nothing is an instance of a primitive type
3500 prim_return_value = intcon(0);
3501 obj = argument(1);
3502 break;
3503 case vmIntrinsics::_getModifiers:
3504 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3505 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3506 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3507 break;
3508 case vmIntrinsics::_isInterface:
3509 prim_return_value = intcon(0);
3510 break;
3511 case vmIntrinsics::_isArray:
3512 prim_return_value = intcon(0);
3513 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3514 break;
3515 case vmIntrinsics::_isPrimitive:
3516 prim_return_value = intcon(1);
3517 expect_prim = true; // obviously
3518 break;
3519 case vmIntrinsics::_getSuperclass:
3520 prim_return_value = null();
3521 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3522 break;
3523 case vmIntrinsics::_getComponentType:
3524 prim_return_value = null();
3525 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3526 break;
3527 case vmIntrinsics::_getClassAccessFlags:
3528 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3529 return_type = TypeInt::INT; // not bool! 6297094
3530 break;
3531 default:
3532 fatal_unexpected_iid(id);
3533 break;
3534 }
3536 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3537 if (mirror_con == NULL) return false; // cannot happen?
3539 #ifndef PRODUCT
3540 if (C->print_intrinsics() || C->print_inlining()) {
3541 ciType* k = mirror_con->java_mirror_type();
3542 if (k) {
3543 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3544 k->print_name();
3545 tty->cr();
3546 }
3547 }
3548 #endif
3550 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3551 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3552 record_for_igvn(region);
3553 PhiNode* phi = new (C) PhiNode(region, return_type);
3555 // The mirror will never be null of Reflection.getClassAccessFlags, however
3556 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3557 // if it is. See bug 4774291.
3559 // For Reflection.getClassAccessFlags(), the null check occurs in
3560 // the wrong place; see inline_unsafe_access(), above, for a similar
3561 // situation.
3562 mirror = null_check(mirror);
3563 // If mirror or obj is dead, only null-path is taken.
3564 if (stopped()) return true;
3566 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3568 // Now load the mirror's klass metaobject, and null-check it.
3569 // Side-effects region with the control path if the klass is null.
3570 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3571 // If kls is null, we have a primitive mirror.
3572 phi->init_req(_prim_path, prim_return_value);
3573 if (stopped()) { set_result(region, phi); return true; }
3574 bool safe_for_replace = (region->in(_prim_path) == top());
3576 Node* p; // handy temp
3577 Node* null_ctl;
3579 // Now that we have the non-null klass, we can perform the real query.
3580 // For constant classes, the query will constant-fold in LoadNode::Value.
3581 Node* query_value = top();
3582 switch (id) {
3583 case vmIntrinsics::_isInstance:
3584 // nothing is an instance of a primitive type
3585 query_value = gen_instanceof(obj, kls, safe_for_replace);
3586 break;
3588 case vmIntrinsics::_getModifiers:
3589 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3590 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3591 break;
3593 case vmIntrinsics::_isInterface:
3594 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3595 if (generate_interface_guard(kls, region) != NULL)
3596 // A guard was added. If the guard is taken, it was an interface.
3597 phi->add_req(intcon(1));
3598 // If we fall through, it's a plain class.
3599 query_value = intcon(0);
3600 break;
3602 case vmIntrinsics::_isArray:
3603 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3604 if (generate_array_guard(kls, region) != NULL)
3605 // A guard was added. If the guard is taken, it was an array.
3606 phi->add_req(intcon(1));
3607 // If we fall through, it's a plain class.
3608 query_value = intcon(0);
3609 break;
3611 case vmIntrinsics::_isPrimitive:
3612 query_value = intcon(0); // "normal" path produces false
3613 break;
3615 case vmIntrinsics::_getSuperclass:
3616 // The rules here are somewhat unfortunate, but we can still do better
3617 // with random logic than with a JNI call.
3618 // Interfaces store null or Object as _super, but must report null.
3619 // Arrays store an intermediate super as _super, but must report Object.
3620 // Other types can report the actual _super.
3621 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3622 if (generate_interface_guard(kls, region) != NULL)
3623 // A guard was added. If the guard is taken, it was an interface.
3624 phi->add_req(null());
3625 if (generate_array_guard(kls, region) != NULL)
3626 // A guard was added. If the guard is taken, it was an array.
3627 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3628 // If we fall through, it's a plain class. Get its _super.
3629 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3630 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3631 null_ctl = top();
3632 kls = null_check_oop(kls, &null_ctl);
3633 if (null_ctl != top()) {
3634 // If the guard is taken, Object.superClass is null (both klass and mirror).
3635 region->add_req(null_ctl);
3636 phi ->add_req(null());
3637 }
3638 if (!stopped()) {
3639 query_value = load_mirror_from_klass(kls);
3640 }
3641 break;
3643 case vmIntrinsics::_getComponentType:
3644 if (generate_array_guard(kls, region) != NULL) {
3645 // Be sure to pin the oop load to the guard edge just created:
3646 Node* is_array_ctrl = region->in(region->req()-1);
3647 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3648 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3649 phi->add_req(cmo);
3650 }
3651 query_value = null(); // non-array case is null
3652 break;
3654 case vmIntrinsics::_getClassAccessFlags:
3655 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3656 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3657 break;
3659 default:
3660 fatal_unexpected_iid(id);
3661 break;
3662 }
3664 // Fall-through is the normal case of a query to a real class.
3665 phi->init_req(1, query_value);
3666 region->init_req(1, control());
3668 C->set_has_split_ifs(true); // Has chance for split-if optimization
3669 set_result(region, phi);
3670 return true;
3671 }
3673 //--------------------------inline_native_subtype_check------------------------
3674 // This intrinsic takes the JNI calls out of the heart of
3675 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3676 bool LibraryCallKit::inline_native_subtype_check() {
3677 // Pull both arguments off the stack.
3678 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3679 args[0] = argument(0);
3680 args[1] = argument(1);
3681 Node* klasses[2]; // corresponding Klasses: superk, subk
3682 klasses[0] = klasses[1] = top();
3684 enum {
3685 // A full decision tree on {superc is prim, subc is prim}:
3686 _prim_0_path = 1, // {P,N} => false
3687 // {P,P} & superc!=subc => false
3688 _prim_same_path, // {P,P} & superc==subc => true
3689 _prim_1_path, // {N,P} => false
3690 _ref_subtype_path, // {N,N} & subtype check wins => true
3691 _both_ref_path, // {N,N} & subtype check loses => false
3692 PATH_LIMIT
3693 };
3695 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3696 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3697 record_for_igvn(region);
3699 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3700 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3701 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3703 // First null-check both mirrors and load each mirror's klass metaobject.
3704 int which_arg;
3705 for (which_arg = 0; which_arg <= 1; which_arg++) {
3706 Node* arg = args[which_arg];
3707 arg = null_check(arg);
3708 if (stopped()) break;
3709 args[which_arg] = arg;
3711 Node* p = basic_plus_adr(arg, class_klass_offset);
3712 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3713 klasses[which_arg] = _gvn.transform(kls);
3714 }
3716 // Having loaded both klasses, test each for null.
3717 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3718 for (which_arg = 0; which_arg <= 1; which_arg++) {
3719 Node* kls = klasses[which_arg];
3720 Node* null_ctl = top();
3721 kls = null_check_oop(kls, &null_ctl, never_see_null);
3722 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3723 region->init_req(prim_path, null_ctl);
3724 if (stopped()) break;
3725 klasses[which_arg] = kls;
3726 }
3728 if (!stopped()) {
3729 // now we have two reference types, in klasses[0..1]
3730 Node* subk = klasses[1]; // the argument to isAssignableFrom
3731 Node* superk = klasses[0]; // the receiver
3732 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3733 // now we have a successful reference subtype check
3734 region->set_req(_ref_subtype_path, control());
3735 }
3737 // If both operands are primitive (both klasses null), then
3738 // we must return true when they are identical primitives.
3739 // It is convenient to test this after the first null klass check.
3740 set_control(region->in(_prim_0_path)); // go back to first null check
3741 if (!stopped()) {
3742 // Since superc is primitive, make a guard for the superc==subc case.
3743 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3744 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3745 generate_guard(bol_eq, region, PROB_FAIR);
3746 if (region->req() == PATH_LIMIT+1) {
3747 // A guard was added. If the added guard is taken, superc==subc.
3748 region->swap_edges(PATH_LIMIT, _prim_same_path);
3749 region->del_req(PATH_LIMIT);
3750 }
3751 region->set_req(_prim_0_path, control()); // Not equal after all.
3752 }
3754 // these are the only paths that produce 'true':
3755 phi->set_req(_prim_same_path, intcon(1));
3756 phi->set_req(_ref_subtype_path, intcon(1));
3758 // pull together the cases:
3759 assert(region->req() == PATH_LIMIT, "sane region");
3760 for (uint i = 1; i < region->req(); i++) {
3761 Node* ctl = region->in(i);
3762 if (ctl == NULL || ctl == top()) {
3763 region->set_req(i, top());
3764 phi ->set_req(i, top());
3765 } else if (phi->in(i) == NULL) {
3766 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3767 }
3768 }
3770 set_control(_gvn.transform(region));
3771 set_result(_gvn.transform(phi));
3772 return true;
3773 }
3775 //---------------------generate_array_guard_common------------------------
3776 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3777 bool obj_array, bool not_array) {
3778 // If obj_array/non_array==false/false:
3779 // Branch around if the given klass is in fact an array (either obj or prim).
3780 // If obj_array/non_array==false/true:
3781 // Branch around if the given klass is not an array klass of any kind.
3782 // If obj_array/non_array==true/true:
3783 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3784 // If obj_array/non_array==true/false:
3785 // Branch around if the kls is an oop array (Object[] or subtype)
3786 //
3787 // Like generate_guard, adds a new path onto the region.
3788 jint layout_con = 0;
3789 Node* layout_val = get_layout_helper(kls, layout_con);
3790 if (layout_val == NULL) {
3791 bool query = (obj_array
3792 ? Klass::layout_helper_is_objArray(layout_con)
3793 : Klass::layout_helper_is_array(layout_con));
3794 if (query == not_array) {
3795 return NULL; // never a branch
3796 } else { // always a branch
3797 Node* always_branch = control();
3798 if (region != NULL)
3799 region->add_req(always_branch);
3800 set_control(top());
3801 return always_branch;
3802 }
3803 }
3804 // Now test the correct condition.
3805 jint nval = (obj_array
3806 ? ((jint)Klass::_lh_array_tag_type_value
3807 << Klass::_lh_array_tag_shift)
3808 : Klass::_lh_neutral_value);
3809 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3810 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3811 // invert the test if we are looking for a non-array
3812 if (not_array) btest = BoolTest(btest).negate();
3813 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3814 return generate_fair_guard(bol, region);
3815 }
3818 //-----------------------inline_native_newArray--------------------------
3819 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3820 bool LibraryCallKit::inline_native_newArray() {
3821 Node* mirror = argument(0);
3822 Node* count_val = argument(1);
3824 mirror = null_check(mirror);
3825 // If mirror or obj is dead, only null-path is taken.
3826 if (stopped()) return true;
3828 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3829 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3830 PhiNode* result_val = new(C) PhiNode(result_reg,
3831 TypeInstPtr::NOTNULL);
3832 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3833 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3834 TypePtr::BOTTOM);
3836 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3837 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3838 result_reg, _slow_path);
3839 Node* normal_ctl = control();
3840 Node* no_array_ctl = result_reg->in(_slow_path);
3842 // Generate code for the slow case. We make a call to newArray().
3843 set_control(no_array_ctl);
3844 if (!stopped()) {
3845 // Either the input type is void.class, or else the
3846 // array klass has not yet been cached. Either the
3847 // ensuing call will throw an exception, or else it
3848 // will cache the array klass for next time.
3849 PreserveJVMState pjvms(this);
3850 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3851 Node* slow_result = set_results_for_java_call(slow_call);
3852 // this->control() comes from set_results_for_java_call
3853 result_reg->set_req(_slow_path, control());
3854 result_val->set_req(_slow_path, slow_result);
3855 result_io ->set_req(_slow_path, i_o());
3856 result_mem->set_req(_slow_path, reset_memory());
3857 }
3859 set_control(normal_ctl);
3860 if (!stopped()) {
3861 // Normal case: The array type has been cached in the java.lang.Class.
3862 // The following call works fine even if the array type is polymorphic.
3863 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3864 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3865 result_reg->init_req(_normal_path, control());
3866 result_val->init_req(_normal_path, obj);
3867 result_io ->init_req(_normal_path, i_o());
3868 result_mem->init_req(_normal_path, reset_memory());
3869 }
3871 // Return the combined state.
3872 set_i_o( _gvn.transform(result_io) );
3873 set_all_memory( _gvn.transform(result_mem));
3875 C->set_has_split_ifs(true); // Has chance for split-if optimization
3876 set_result(result_reg, result_val);
3877 return true;
3878 }
3880 //----------------------inline_native_getLength--------------------------
3881 // public static native int java.lang.reflect.Array.getLength(Object array);
3882 bool LibraryCallKit::inline_native_getLength() {
3883 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3885 Node* array = null_check(argument(0));
3886 // If array is dead, only null-path is taken.
3887 if (stopped()) return true;
3889 // Deoptimize if it is a non-array.
3890 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3892 if (non_array != NULL) {
3893 PreserveJVMState pjvms(this);
3894 set_control(non_array);
3895 uncommon_trap(Deoptimization::Reason_intrinsic,
3896 Deoptimization::Action_maybe_recompile);
3897 }
3899 // If control is dead, only non-array-path is taken.
3900 if (stopped()) return true;
3902 // The works fine even if the array type is polymorphic.
3903 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3904 Node* result = load_array_length(array);
3906 C->set_has_split_ifs(true); // Has chance for split-if optimization
3907 set_result(result);
3908 return true;
3909 }
3911 //------------------------inline_array_copyOf----------------------------
3912 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3913 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3914 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3915 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3917 // Get the arguments.
3918 Node* original = argument(0);
3919 Node* start = is_copyOfRange? argument(1): intcon(0);
3920 Node* end = is_copyOfRange? argument(2): argument(1);
3921 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3923 Node* newcopy = NULL;
3925 // Set the original stack and the reexecute bit for the interpreter to reexecute
3926 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3927 { PreserveReexecuteState preexecs(this);
3928 jvms()->set_should_reexecute(true);
3930 array_type_mirror = null_check(array_type_mirror);
3931 original = null_check(original);
3933 // Check if a null path was taken unconditionally.
3934 if (stopped()) return true;
3936 Node* orig_length = load_array_length(original);
3938 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3939 klass_node = null_check(klass_node);
3941 RegionNode* bailout = new (C) RegionNode(1);
3942 record_for_igvn(bailout);
3944 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3945 // Bail out if that is so.
3946 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3947 if (not_objArray != NULL) {
3948 // Improve the klass node's type from the new optimistic assumption:
3949 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3950 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3951 Node* cast = new (C) CastPPNode(klass_node, akls);
3952 cast->init_req(0, control());
3953 klass_node = _gvn.transform(cast);
3954 }
3956 // Bail out if either start or end is negative.
3957 generate_negative_guard(start, bailout, &start);
3958 generate_negative_guard(end, bailout, &end);
3960 Node* length = end;
3961 if (_gvn.type(start) != TypeInt::ZERO) {
3962 length = _gvn.transform(new (C) SubINode(end, start));
3963 }
3965 // Bail out if length is negative.
3966 // Without this the new_array would throw
3967 // NegativeArraySizeException but IllegalArgumentException is what
3968 // should be thrown
3969 generate_negative_guard(length, bailout, &length);
3971 if (bailout->req() > 1) {
3972 PreserveJVMState pjvms(this);
3973 set_control(_gvn.transform(bailout));
3974 uncommon_trap(Deoptimization::Reason_intrinsic,
3975 Deoptimization::Action_maybe_recompile);
3976 }
3978 if (!stopped()) {
3979 // How many elements will we copy from the original?
3980 // The answer is MinI(orig_length - start, length).
3981 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3982 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3984 newcopy = new_array(klass_node, length, 0); // no argments to push
3986 // Generate a direct call to the right arraycopy function(s).
3987 // We know the copy is disjoint but we might not know if the
3988 // oop stores need checking.
3989 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3990 // This will fail a store-check if x contains any non-nulls.
3991 bool disjoint_bases = true;
3992 // if start > orig_length then the length of the copy may be
3993 // negative.
3994 bool length_never_negative = !is_copyOfRange;
3995 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3996 original, start, newcopy, intcon(0), moved,
3997 disjoint_bases, length_never_negative);
3998 }
3999 } // original reexecute is set back here
4001 C->set_has_split_ifs(true); // Has chance for split-if optimization
4002 if (!stopped()) {
4003 set_result(newcopy);
4004 }
4005 return true;
4006 }
4009 //----------------------generate_virtual_guard---------------------------
4010 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4011 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4012 RegionNode* slow_region) {
4013 ciMethod* method = callee();
4014 int vtable_index = method->vtable_index();
4015 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4016 err_msg_res("bad index %d", vtable_index));
4017 // Get the Method* out of the appropriate vtable entry.
4018 int entry_offset = (InstanceKlass::vtable_start_offset() +
4019 vtable_index*vtableEntry::size()) * wordSize +
4020 vtableEntry::method_offset_in_bytes();
4021 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4022 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4024 // Compare the target method with the expected method (e.g., Object.hashCode).
4025 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4027 Node* native_call = makecon(native_call_addr);
4028 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4029 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4031 return generate_slow_guard(test_native, slow_region);
4032 }
4034 //-----------------------generate_method_call----------------------------
4035 // Use generate_method_call to make a slow-call to the real
4036 // method if the fast path fails. An alternative would be to
4037 // use a stub like OptoRuntime::slow_arraycopy_Java.
4038 // This only works for expanding the current library call,
4039 // not another intrinsic. (E.g., don't use this for making an
4040 // arraycopy call inside of the copyOf intrinsic.)
4041 CallJavaNode*
4042 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4043 // When compiling the intrinsic method itself, do not use this technique.
4044 guarantee(callee() != C->method(), "cannot make slow-call to self");
4046 ciMethod* method = callee();
4047 // ensure the JVMS we have will be correct for this call
4048 guarantee(method_id == method->intrinsic_id(), "must match");
4050 const TypeFunc* tf = TypeFunc::make(method);
4051 CallJavaNode* slow_call;
4052 if (is_static) {
4053 assert(!is_virtual, "");
4054 slow_call = new(C) CallStaticJavaNode(C, tf,
4055 SharedRuntime::get_resolve_static_call_stub(),
4056 method, bci());
4057 } else if (is_virtual) {
4058 null_check_receiver();
4059 int vtable_index = Method::invalid_vtable_index;
4060 if (UseInlineCaches) {
4061 // Suppress the vtable call
4062 } else {
4063 // hashCode and clone are not a miranda methods,
4064 // so the vtable index is fixed.
4065 // No need to use the linkResolver to get it.
4066 vtable_index = method->vtable_index();
4067 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4068 err_msg_res("bad index %d", vtable_index));
4069 }
4070 slow_call = new(C) CallDynamicJavaNode(tf,
4071 SharedRuntime::get_resolve_virtual_call_stub(),
4072 method, vtable_index, bci());
4073 } else { // neither virtual nor static: opt_virtual
4074 null_check_receiver();
4075 slow_call = new(C) CallStaticJavaNode(C, tf,
4076 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4077 method, bci());
4078 slow_call->set_optimized_virtual(true);
4079 }
4080 set_arguments_for_java_call(slow_call);
4081 set_edges_for_java_call(slow_call);
4082 return slow_call;
4083 }
4086 /**
4087 * Build special case code for calls to hashCode on an object. This call may
4088 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4089 * slightly different code.
4090 */
4091 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4092 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4093 assert(!(is_virtual && is_static), "either virtual, special, or static");
4095 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4097 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4098 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4099 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4100 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4101 Node* obj = NULL;
4102 if (!is_static) {
4103 // Check for hashing null object
4104 obj = null_check_receiver();
4105 if (stopped()) return true; // unconditionally null
4106 result_reg->init_req(_null_path, top());
4107 result_val->init_req(_null_path, top());
4108 } else {
4109 // Do a null check, and return zero if null.
4110 // System.identityHashCode(null) == 0
4111 obj = argument(0);
4112 Node* null_ctl = top();
4113 obj = null_check_oop(obj, &null_ctl);
4114 result_reg->init_req(_null_path, null_ctl);
4115 result_val->init_req(_null_path, _gvn.intcon(0));
4116 }
4118 // Unconditionally null? Then return right away.
4119 if (stopped()) {
4120 set_control( result_reg->in(_null_path));
4121 if (!stopped())
4122 set_result(result_val->in(_null_path));
4123 return true;
4124 }
4126 // We only go to the fast case code if we pass a number of guards. The
4127 // paths which do not pass are accumulated in the slow_region.
4128 RegionNode* slow_region = new (C) RegionNode(1);
4129 record_for_igvn(slow_region);
4131 // If this is a virtual call, we generate a funny guard. We pull out
4132 // the vtable entry corresponding to hashCode() from the target object.
4133 // If the target method which we are calling happens to be the native
4134 // Object hashCode() method, we pass the guard. We do not need this
4135 // guard for non-virtual calls -- the caller is known to be the native
4136 // Object hashCode().
4137 if (is_virtual) {
4138 // After null check, get the object's klass.
4139 Node* obj_klass = load_object_klass(obj);
4140 generate_virtual_guard(obj_klass, slow_region);
4141 }
4143 // Get the header out of the object, use LoadMarkNode when available
4144 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4145 // The control of the load must be NULL. Otherwise, the load can move before
4146 // the null check after castPP removal.
4147 Node* no_ctrl = NULL;
4148 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4150 // Test the header to see if it is unlocked.
4151 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4152 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4153 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4154 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4155 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4157 generate_slow_guard(test_unlocked, slow_region);
4159 // Get the hash value and check to see that it has been properly assigned.
4160 // We depend on hash_mask being at most 32 bits and avoid the use of
4161 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4162 // vm: see markOop.hpp.
4163 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4164 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4165 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4166 // This hack lets the hash bits live anywhere in the mark object now, as long
4167 // as the shift drops the relevant bits into the low 32 bits. Note that
4168 // Java spec says that HashCode is an int so there's no point in capturing
4169 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4170 hshifted_header = ConvX2I(hshifted_header);
4171 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4173 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4174 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4175 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4177 generate_slow_guard(test_assigned, slow_region);
4179 Node* init_mem = reset_memory();
4180 // fill in the rest of the null path:
4181 result_io ->init_req(_null_path, i_o());
4182 result_mem->init_req(_null_path, init_mem);
4184 result_val->init_req(_fast_path, hash_val);
4185 result_reg->init_req(_fast_path, control());
4186 result_io ->init_req(_fast_path, i_o());
4187 result_mem->init_req(_fast_path, init_mem);
4189 // Generate code for the slow case. We make a call to hashCode().
4190 set_control(_gvn.transform(slow_region));
4191 if (!stopped()) {
4192 // No need for PreserveJVMState, because we're using up the present state.
4193 set_all_memory(init_mem);
4194 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4195 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4196 Node* slow_result = set_results_for_java_call(slow_call);
4197 // this->control() comes from set_results_for_java_call
4198 result_reg->init_req(_slow_path, control());
4199 result_val->init_req(_slow_path, slow_result);
4200 result_io ->set_req(_slow_path, i_o());
4201 result_mem ->set_req(_slow_path, reset_memory());
4202 }
4204 // Return the combined state.
4205 set_i_o( _gvn.transform(result_io) );
4206 set_all_memory( _gvn.transform(result_mem));
4208 set_result(result_reg, result_val);
4209 return true;
4210 }
4212 //---------------------------inline_native_getClass----------------------------
4213 // public final native Class<?> java.lang.Object.getClass();
4214 //
4215 // Build special case code for calls to getClass on an object.
4216 bool LibraryCallKit::inline_native_getClass() {
4217 Node* obj = null_check_receiver();
4218 if (stopped()) return true;
4219 set_result(load_mirror_from_klass(load_object_klass(obj)));
4220 return true;
4221 }
4223 //-----------------inline_native_Reflection_getCallerClass---------------------
4224 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4225 //
4226 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4227 //
4228 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4229 // in that it must skip particular security frames and checks for
4230 // caller sensitive methods.
4231 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4232 #ifndef PRODUCT
4233 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4234 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4235 }
4236 #endif
4238 if (!jvms()->has_method()) {
4239 #ifndef PRODUCT
4240 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4241 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4242 }
4243 #endif
4244 return false;
4245 }
4247 // Walk back up the JVM state to find the caller at the required
4248 // depth.
4249 JVMState* caller_jvms = jvms();
4251 // Cf. JVM_GetCallerClass
4252 // NOTE: Start the loop at depth 1 because the current JVM state does
4253 // not include the Reflection.getCallerClass() frame.
4254 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4255 ciMethod* m = caller_jvms->method();
4256 switch (n) {
4257 case 0:
4258 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4259 break;
4260 case 1:
4261 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4262 if (!m->caller_sensitive()) {
4263 #ifndef PRODUCT
4264 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4265 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4266 }
4267 #endif
4268 return false; // bail-out; let JVM_GetCallerClass do the work
4269 }
4270 break;
4271 default:
4272 if (!m->is_ignored_by_security_stack_walk()) {
4273 // We have reached the desired frame; return the holder class.
4274 // Acquire method holder as java.lang.Class and push as constant.
4275 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4276 ciInstance* caller_mirror = caller_klass->java_mirror();
4277 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4279 #ifndef PRODUCT
4280 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4281 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());
4282 tty->print_cr(" JVM state at this point:");
4283 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4284 ciMethod* m = jvms()->of_depth(i)->method();
4285 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4286 }
4287 }
4288 #endif
4289 return true;
4290 }
4291 break;
4292 }
4293 }
4295 #ifndef PRODUCT
4296 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4297 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4298 tty->print_cr(" JVM state at this point:");
4299 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4300 ciMethod* m = jvms()->of_depth(i)->method();
4301 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4302 }
4303 }
4304 #endif
4306 return false; // bail-out; let JVM_GetCallerClass do the work
4307 }
4309 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4310 Node* arg = argument(0);
4311 Node* result = NULL;
4313 switch (id) {
4314 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4315 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4316 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4317 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4319 case vmIntrinsics::_doubleToLongBits: {
4320 // two paths (plus control) merge in a wood
4321 RegionNode *r = new (C) RegionNode(3);
4322 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4324 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4325 // Build the boolean node
4326 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4328 // Branch either way.
4329 // NaN case is less traveled, which makes all the difference.
4330 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4331 Node *opt_isnan = _gvn.transform(ifisnan);
4332 assert( opt_isnan->is_If(), "Expect an IfNode");
4333 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4334 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4336 set_control(iftrue);
4338 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4339 Node *slow_result = longcon(nan_bits); // return NaN
4340 phi->init_req(1, _gvn.transform( slow_result ));
4341 r->init_req(1, iftrue);
4343 // Else fall through
4344 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4345 set_control(iffalse);
4347 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4348 r->init_req(2, iffalse);
4350 // Post merge
4351 set_control(_gvn.transform(r));
4352 record_for_igvn(r);
4354 C->set_has_split_ifs(true); // Has chance for split-if optimization
4355 result = phi;
4356 assert(result->bottom_type()->isa_long(), "must be");
4357 break;
4358 }
4360 case vmIntrinsics::_floatToIntBits: {
4361 // two paths (plus control) merge in a wood
4362 RegionNode *r = new (C) RegionNode(3);
4363 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4365 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4366 // Build the boolean node
4367 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4369 // Branch either way.
4370 // NaN case is less traveled, which makes all the difference.
4371 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4372 Node *opt_isnan = _gvn.transform(ifisnan);
4373 assert( opt_isnan->is_If(), "Expect an IfNode");
4374 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4375 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4377 set_control(iftrue);
4379 static const jint nan_bits = 0x7fc00000;
4380 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4381 phi->init_req(1, _gvn.transform( slow_result ));
4382 r->init_req(1, iftrue);
4384 // Else fall through
4385 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4386 set_control(iffalse);
4388 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4389 r->init_req(2, iffalse);
4391 // Post merge
4392 set_control(_gvn.transform(r));
4393 record_for_igvn(r);
4395 C->set_has_split_ifs(true); // Has chance for split-if optimization
4396 result = phi;
4397 assert(result->bottom_type()->isa_int(), "must be");
4398 break;
4399 }
4401 default:
4402 fatal_unexpected_iid(id);
4403 break;
4404 }
4405 set_result(_gvn.transform(result));
4406 return true;
4407 }
4409 #ifdef _LP64
4410 #define XTOP ,top() /*additional argument*/
4411 #else //_LP64
4412 #define XTOP /*no additional argument*/
4413 #endif //_LP64
4415 //----------------------inline_unsafe_copyMemory-------------------------
4416 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4417 bool LibraryCallKit::inline_unsafe_copyMemory() {
4418 if (callee()->is_static()) return false; // caller must have the capability!
4419 null_check_receiver(); // null-check receiver
4420 if (stopped()) return true;
4422 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4424 Node* src_ptr = argument(1); // type: oop
4425 Node* src_off = ConvL2X(argument(2)); // type: long
4426 Node* dst_ptr = argument(4); // type: oop
4427 Node* dst_off = ConvL2X(argument(5)); // type: long
4428 Node* size = ConvL2X(argument(7)); // type: long
4430 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4431 "fieldOffset must be byte-scaled");
4433 Node* src = make_unsafe_address(src_ptr, src_off);
4434 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4436 // Conservatively insert a memory barrier on all memory slices.
4437 // Do not let writes of the copy source or destination float below the copy.
4438 insert_mem_bar(Op_MemBarCPUOrder);
4440 // Call it. Note that the length argument is not scaled.
4441 make_runtime_call(RC_LEAF|RC_NO_FP,
4442 OptoRuntime::fast_arraycopy_Type(),
4443 StubRoutines::unsafe_arraycopy(),
4444 "unsafe_arraycopy",
4445 TypeRawPtr::BOTTOM,
4446 src, dst, size XTOP);
4448 // Do not let reads of the copy destination float above the copy.
4449 insert_mem_bar(Op_MemBarCPUOrder);
4451 return true;
4452 }
4454 //------------------------clone_coping-----------------------------------
4455 // Helper function for inline_native_clone.
4456 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4457 assert(obj_size != NULL, "");
4458 Node* raw_obj = alloc_obj->in(1);
4459 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4461 AllocateNode* alloc = NULL;
4462 if (ReduceBulkZeroing) {
4463 // We will be completely responsible for initializing this object -
4464 // mark Initialize node as complete.
4465 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4466 // The object was just allocated - there should be no any stores!
4467 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4468 // Mark as complete_with_arraycopy so that on AllocateNode
4469 // expansion, we know this AllocateNode is initialized by an array
4470 // copy and a StoreStore barrier exists after the array copy.
4471 alloc->initialization()->set_complete_with_arraycopy();
4472 }
4474 // Copy the fastest available way.
4475 // TODO: generate fields copies for small objects instead.
4476 Node* src = obj;
4477 Node* dest = alloc_obj;
4478 Node* size = _gvn.transform(obj_size);
4480 // Exclude the header but include array length to copy by 8 bytes words.
4481 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4482 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4483 instanceOopDesc::base_offset_in_bytes();
4484 // base_off:
4485 // 8 - 32-bit VM
4486 // 12 - 64-bit VM, compressed klass
4487 // 16 - 64-bit VM, normal klass
4488 if (base_off % BytesPerLong != 0) {
4489 assert(UseCompressedClassPointers, "");
4490 if (is_array) {
4491 // Exclude length to copy by 8 bytes words.
4492 base_off += sizeof(int);
4493 } else {
4494 // Include klass to copy by 8 bytes words.
4495 base_off = instanceOopDesc::klass_offset_in_bytes();
4496 }
4497 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4498 }
4499 src = basic_plus_adr(src, base_off);
4500 dest = basic_plus_adr(dest, base_off);
4502 // Compute the length also, if needed:
4503 Node* countx = size;
4504 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4505 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4507 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4508 bool disjoint_bases = true;
4509 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4510 src, NULL, dest, NULL, countx,
4511 /*dest_uninitialized*/true);
4513 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4514 if (card_mark) {
4515 assert(!is_array, "");
4516 // Put in store barrier for any and all oops we are sticking
4517 // into this object. (We could avoid this if we could prove
4518 // that the object type contains no oop fields at all.)
4519 Node* no_particular_value = NULL;
4520 Node* no_particular_field = NULL;
4521 int raw_adr_idx = Compile::AliasIdxRaw;
4522 post_barrier(control(),
4523 memory(raw_adr_type),
4524 alloc_obj,
4525 no_particular_field,
4526 raw_adr_idx,
4527 no_particular_value,
4528 T_OBJECT,
4529 false);
4530 }
4532 // Do not let reads from the cloned object float above the arraycopy.
4533 if (alloc != NULL) {
4534 // Do not let stores that initialize this object be reordered with
4535 // a subsequent store that would make this object accessible by
4536 // other threads.
4537 // Record what AllocateNode this StoreStore protects so that
4538 // escape analysis can go from the MemBarStoreStoreNode to the
4539 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4540 // based on the escape status of the AllocateNode.
4541 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4542 } else {
4543 insert_mem_bar(Op_MemBarCPUOrder);
4544 }
4545 }
4547 //------------------------inline_native_clone----------------------------
4548 // protected native Object java.lang.Object.clone();
4549 //
4550 // Here are the simple edge cases:
4551 // null receiver => normal trap
4552 // virtual and clone was overridden => slow path to out-of-line clone
4553 // not cloneable or finalizer => slow path to out-of-line Object.clone
4554 //
4555 // The general case has two steps, allocation and copying.
4556 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4557 //
4558 // Copying also has two cases, oop arrays and everything else.
4559 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4560 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4561 //
4562 // These steps fold up nicely if and when the cloned object's klass
4563 // can be sharply typed as an object array, a type array, or an instance.
4564 //
4565 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4566 PhiNode* result_val;
4568 // Set the reexecute bit for the interpreter to reexecute
4569 // the bytecode that invokes Object.clone if deoptimization happens.
4570 { PreserveReexecuteState preexecs(this);
4571 jvms()->set_should_reexecute(true);
4573 Node* obj = null_check_receiver();
4574 if (stopped()) return true;
4576 Node* obj_klass = load_object_klass(obj);
4577 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4578 const TypeOopPtr* toop = ((tklass != NULL)
4579 ? tklass->as_instance_type()
4580 : TypeInstPtr::NOTNULL);
4582 // Conservatively insert a memory barrier on all memory slices.
4583 // Do not let writes into the original float below the clone.
4584 insert_mem_bar(Op_MemBarCPUOrder);
4586 // paths into result_reg:
4587 enum {
4588 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4589 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4590 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4591 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4592 PATH_LIMIT
4593 };
4594 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4595 result_val = new(C) PhiNode(result_reg,
4596 TypeInstPtr::NOTNULL);
4597 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4598 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4599 TypePtr::BOTTOM);
4600 record_for_igvn(result_reg);
4602 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4603 int raw_adr_idx = Compile::AliasIdxRaw;
4605 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4606 if (array_ctl != NULL) {
4607 // It's an array.
4608 PreserveJVMState pjvms(this);
4609 set_control(array_ctl);
4610 Node* obj_length = load_array_length(obj);
4611 Node* obj_size = NULL;
4612 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4614 if (!use_ReduceInitialCardMarks()) {
4615 // If it is an oop array, it requires very special treatment,
4616 // because card marking is required on each card of the array.
4617 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4618 if (is_obja != NULL) {
4619 PreserveJVMState pjvms2(this);
4620 set_control(is_obja);
4621 // Generate a direct call to the right arraycopy function(s).
4622 bool disjoint_bases = true;
4623 bool length_never_negative = true;
4624 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4625 obj, intcon(0), alloc_obj, intcon(0),
4626 obj_length,
4627 disjoint_bases, length_never_negative);
4628 result_reg->init_req(_objArray_path, control());
4629 result_val->init_req(_objArray_path, alloc_obj);
4630 result_i_o ->set_req(_objArray_path, i_o());
4631 result_mem ->set_req(_objArray_path, reset_memory());
4632 }
4633 }
4634 // Otherwise, there are no card marks to worry about.
4635 // (We can dispense with card marks if we know the allocation
4636 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4637 // causes the non-eden paths to take compensating steps to
4638 // simulate a fresh allocation, so that no further
4639 // card marks are required in compiled code to initialize
4640 // the object.)
4642 if (!stopped()) {
4643 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4645 // Present the results of the copy.
4646 result_reg->init_req(_array_path, control());
4647 result_val->init_req(_array_path, alloc_obj);
4648 result_i_o ->set_req(_array_path, i_o());
4649 result_mem ->set_req(_array_path, reset_memory());
4650 }
4651 }
4653 // We only go to the instance fast case code if we pass a number of guards.
4654 // The paths which do not pass are accumulated in the slow_region.
4655 RegionNode* slow_region = new (C) RegionNode(1);
4656 record_for_igvn(slow_region);
4657 if (!stopped()) {
4658 // It's an instance (we did array above). Make the slow-path tests.
4659 // If this is a virtual call, we generate a funny guard. We grab
4660 // the vtable entry corresponding to clone() from the target object.
4661 // If the target method which we are calling happens to be the
4662 // Object clone() method, we pass the guard. We do not need this
4663 // guard for non-virtual calls; the caller is known to be the native
4664 // Object clone().
4665 if (is_virtual) {
4666 generate_virtual_guard(obj_klass, slow_region);
4667 }
4669 // The object must be cloneable and must not have a finalizer.
4670 // Both of these conditions may be checked in a single test.
4671 // We could optimize the cloneable test further, but we don't care.
4672 generate_access_flags_guard(obj_klass,
4673 // Test both conditions:
4674 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4675 // Must be cloneable but not finalizer:
4676 JVM_ACC_IS_CLONEABLE,
4677 slow_region);
4678 }
4680 if (!stopped()) {
4681 // It's an instance, and it passed the slow-path tests.
4682 PreserveJVMState pjvms(this);
4683 Node* obj_size = NULL;
4684 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4685 // is reexecuted if deoptimization occurs and there could be problems when merging
4686 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4687 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4689 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4691 // Present the results of the slow call.
4692 result_reg->init_req(_instance_path, control());
4693 result_val->init_req(_instance_path, alloc_obj);
4694 result_i_o ->set_req(_instance_path, i_o());
4695 result_mem ->set_req(_instance_path, reset_memory());
4696 }
4698 // Generate code for the slow case. We make a call to clone().
4699 set_control(_gvn.transform(slow_region));
4700 if (!stopped()) {
4701 PreserveJVMState pjvms(this);
4702 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4703 Node* slow_result = set_results_for_java_call(slow_call);
4704 // this->control() comes from set_results_for_java_call
4705 result_reg->init_req(_slow_path, control());
4706 result_val->init_req(_slow_path, slow_result);
4707 result_i_o ->set_req(_slow_path, i_o());
4708 result_mem ->set_req(_slow_path, reset_memory());
4709 }
4711 // Return the combined state.
4712 set_control( _gvn.transform(result_reg));
4713 set_i_o( _gvn.transform(result_i_o));
4714 set_all_memory( _gvn.transform(result_mem));
4715 } // original reexecute is set back here
4717 set_result(_gvn.transform(result_val));
4718 return true;
4719 }
4721 //------------------------------basictype2arraycopy----------------------------
4722 address LibraryCallKit::basictype2arraycopy(BasicType t,
4723 Node* src_offset,
4724 Node* dest_offset,
4725 bool disjoint_bases,
4726 const char* &name,
4727 bool dest_uninitialized) {
4728 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4729 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4731 bool aligned = false;
4732 bool disjoint = disjoint_bases;
4734 // if the offsets are the same, we can treat the memory regions as
4735 // disjoint, because either the memory regions are in different arrays,
4736 // or they are identical (which we can treat as disjoint.) We can also
4737 // treat a copy with a destination index less that the source index
4738 // as disjoint since a low->high copy will work correctly in this case.
4739 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4740 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4741 // both indices are constants
4742 int s_offs = src_offset_inttype->get_con();
4743 int d_offs = dest_offset_inttype->get_con();
4744 int element_size = type2aelembytes(t);
4745 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4746 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4747 if (s_offs >= d_offs) disjoint = true;
4748 } else if (src_offset == dest_offset && src_offset != NULL) {
4749 // This can occur if the offsets are identical non-constants.
4750 disjoint = true;
4751 }
4753 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4754 }
4757 //------------------------------inline_arraycopy-----------------------
4758 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4759 // Object dest, int destPos,
4760 // int length);
4761 bool LibraryCallKit::inline_arraycopy() {
4762 // Get the arguments.
4763 Node* src = argument(0); // type: oop
4764 Node* src_offset = argument(1); // type: int
4765 Node* dest = argument(2); // type: oop
4766 Node* dest_offset = argument(3); // type: int
4767 Node* length = argument(4); // type: int
4769 // Compile time checks. If any of these checks cannot be verified at compile time,
4770 // we do not make a fast path for this call. Instead, we let the call remain as it
4771 // is. The checks we choose to mandate at compile time are:
4772 //
4773 // (1) src and dest are arrays.
4774 const Type* src_type = src->Value(&_gvn);
4775 const Type* dest_type = dest->Value(&_gvn);
4776 const TypeAryPtr* top_src = src_type->isa_aryptr();
4777 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4779 // Do we have the type of src?
4780 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4781 // Do we have the type of dest?
4782 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4783 // Is the type for src from speculation?
4784 bool src_spec = false;
4785 // Is the type for dest from speculation?
4786 bool dest_spec = false;
4788 if (!has_src || !has_dest) {
4789 // We don't have sufficient type information, let's see if
4790 // speculative types can help. We need to have types for both src
4791 // and dest so that it pays off.
4793 // Do we already have or could we have type information for src
4794 bool could_have_src = has_src;
4795 // Do we already have or could we have type information for dest
4796 bool could_have_dest = has_dest;
4798 ciKlass* src_k = NULL;
4799 if (!has_src) {
4800 src_k = src_type->speculative_type();
4801 if (src_k != NULL && src_k->is_array_klass()) {
4802 could_have_src = true;
4803 }
4804 }
4806 ciKlass* dest_k = NULL;
4807 if (!has_dest) {
4808 dest_k = dest_type->speculative_type();
4809 if (dest_k != NULL && dest_k->is_array_klass()) {
4810 could_have_dest = true;
4811 }
4812 }
4814 if (could_have_src && could_have_dest) {
4815 // This is going to pay off so emit the required guards
4816 if (!has_src) {
4817 src = maybe_cast_profiled_obj(src, src_k);
4818 src_type = _gvn.type(src);
4819 top_src = src_type->isa_aryptr();
4820 has_src = (top_src != NULL && top_src->klass() != NULL);
4821 src_spec = true;
4822 }
4823 if (!has_dest) {
4824 dest = maybe_cast_profiled_obj(dest, dest_k);
4825 dest_type = _gvn.type(dest);
4826 top_dest = dest_type->isa_aryptr();
4827 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4828 dest_spec = true;
4829 }
4830 }
4831 }
4833 if (!has_src || !has_dest) {
4834 // Conservatively insert a memory barrier on all memory slices.
4835 // Do not let writes into the source float below the arraycopy.
4836 insert_mem_bar(Op_MemBarCPUOrder);
4838 // Call StubRoutines::generic_arraycopy stub.
4839 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4840 src, src_offset, dest, dest_offset, length);
4842 // Do not let reads from the destination float above the arraycopy.
4843 // Since we cannot type the arrays, we don't know which slices
4844 // might be affected. We could restrict this barrier only to those
4845 // memory slices which pertain to array elements--but don't bother.
4846 if (!InsertMemBarAfterArraycopy)
4847 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4848 insert_mem_bar(Op_MemBarCPUOrder);
4849 return true;
4850 }
4852 // (2) src and dest arrays must have elements of the same BasicType
4853 // Figure out the size and type of the elements we will be copying.
4854 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4855 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4856 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4857 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4859 if (src_elem != dest_elem || dest_elem == T_VOID) {
4860 // The component types are not the same or are not recognized. Punt.
4861 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4862 generate_slow_arraycopy(TypePtr::BOTTOM,
4863 src, src_offset, dest, dest_offset, length,
4864 /*dest_uninitialized*/false);
4865 return true;
4866 }
4868 if (src_elem == T_OBJECT) {
4869 // If both arrays are object arrays then having the exact types
4870 // for both will remove the need for a subtype check at runtime
4871 // before the call and may make it possible to pick a faster copy
4872 // routine (without a subtype check on every element)
4873 // Do we have the exact type of src?
4874 bool could_have_src = src_spec;
4875 // Do we have the exact type of dest?
4876 bool could_have_dest = dest_spec;
4877 ciKlass* src_k = top_src->klass();
4878 ciKlass* dest_k = top_dest->klass();
4879 if (!src_spec) {
4880 src_k = src_type->speculative_type();
4881 if (src_k != NULL && src_k->is_array_klass()) {
4882 could_have_src = true;
4883 }
4884 }
4885 if (!dest_spec) {
4886 dest_k = dest_type->speculative_type();
4887 if (dest_k != NULL && dest_k->is_array_klass()) {
4888 could_have_dest = true;
4889 }
4890 }
4891 if (could_have_src && could_have_dest) {
4892 // If we can have both exact types, emit the missing guards
4893 if (could_have_src && !src_spec) {
4894 src = maybe_cast_profiled_obj(src, src_k);
4895 }
4896 if (could_have_dest && !dest_spec) {
4897 dest = maybe_cast_profiled_obj(dest, dest_k);
4898 }
4899 }
4900 }
4902 //---------------------------------------------------------------------------
4903 // We will make a fast path for this call to arraycopy.
4905 // We have the following tests left to perform:
4906 //
4907 // (3) src and dest must not be null.
4908 // (4) src_offset must not be negative.
4909 // (5) dest_offset must not be negative.
4910 // (6) length must not be negative.
4911 // (7) src_offset + length must not exceed length of src.
4912 // (8) dest_offset + length must not exceed length of dest.
4913 // (9) each element of an oop array must be assignable
4915 RegionNode* slow_region = new (C) RegionNode(1);
4916 record_for_igvn(slow_region);
4918 // (3) operands must not be null
4919 // We currently perform our null checks with the null_check routine.
4920 // This means that the null exceptions will be reported in the caller
4921 // rather than (correctly) reported inside of the native arraycopy call.
4922 // This should be corrected, given time. We do our null check with the
4923 // stack pointer restored.
4924 src = null_check(src, T_ARRAY);
4925 dest = null_check(dest, T_ARRAY);
4927 // (4) src_offset must not be negative.
4928 generate_negative_guard(src_offset, slow_region);
4930 // (5) dest_offset must not be negative.
4931 generate_negative_guard(dest_offset, slow_region);
4933 // (6) length must not be negative (moved to generate_arraycopy()).
4934 // generate_negative_guard(length, slow_region);
4936 // (7) src_offset + length must not exceed length of src.
4937 generate_limit_guard(src_offset, length,
4938 load_array_length(src),
4939 slow_region);
4941 // (8) dest_offset + length must not exceed length of dest.
4942 generate_limit_guard(dest_offset, length,
4943 load_array_length(dest),
4944 slow_region);
4946 // (9) each element of an oop array must be assignable
4947 // The generate_arraycopy subroutine checks this.
4949 // This is where the memory effects are placed:
4950 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4951 generate_arraycopy(adr_type, dest_elem,
4952 src, src_offset, dest, dest_offset, length,
4953 false, false, slow_region);
4955 return true;
4956 }
4958 //-----------------------------generate_arraycopy----------------------
4959 // Generate an optimized call to arraycopy.
4960 // Caller must guard against non-arrays.
4961 // Caller must determine a common array basic-type for both arrays.
4962 // Caller must validate offsets against array bounds.
4963 // The slow_region has already collected guard failure paths
4964 // (such as out of bounds length or non-conformable array types).
4965 // The generated code has this shape, in general:
4966 //
4967 // if (length == 0) return // via zero_path
4968 // slowval = -1
4969 // if (types unknown) {
4970 // slowval = call generic copy loop
4971 // if (slowval == 0) return // via checked_path
4972 // } else if (indexes in bounds) {
4973 // if ((is object array) && !(array type check)) {
4974 // slowval = call checked copy loop
4975 // if (slowval == 0) return // via checked_path
4976 // } else {
4977 // call bulk copy loop
4978 // return // via fast_path
4979 // }
4980 // }
4981 // // adjust params for remaining work:
4982 // if (slowval != -1) {
4983 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4984 // }
4985 // slow_region:
4986 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4987 // return // via slow_call_path
4988 //
4989 // This routine is used from several intrinsics: System.arraycopy,
4990 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4991 //
4992 void
4993 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4994 BasicType basic_elem_type,
4995 Node* src, Node* src_offset,
4996 Node* dest, Node* dest_offset,
4997 Node* copy_length,
4998 bool disjoint_bases,
4999 bool length_never_negative,
5000 RegionNode* slow_region) {
5002 if (slow_region == NULL) {
5003 slow_region = new(C) RegionNode(1);
5004 record_for_igvn(slow_region);
5005 }
5007 Node* original_dest = dest;
5008 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5009 bool dest_uninitialized = false;
5011 // See if this is the initialization of a newly-allocated array.
5012 // If so, we will take responsibility here for initializing it to zero.
5013 // (Note: Because tightly_coupled_allocation performs checks on the
5014 // out-edges of the dest, we need to avoid making derived pointers
5015 // from it until we have checked its uses.)
5016 if (ReduceBulkZeroing
5017 && !ZeroTLAB // pointless if already zeroed
5018 && basic_elem_type != T_CONFLICT // avoid corner case
5019 && !src->eqv_uncast(dest)
5020 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5021 != NULL)
5022 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5023 && alloc->maybe_set_complete(&_gvn)) {
5024 // "You break it, you buy it."
5025 InitializeNode* init = alloc->initialization();
5026 assert(init->is_complete(), "we just did this");
5027 init->set_complete_with_arraycopy();
5028 assert(dest->is_CheckCastPP(), "sanity");
5029 assert(dest->in(0)->in(0) == init, "dest pinned");
5030 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5031 // From this point on, every exit path is responsible for
5032 // initializing any non-copied parts of the object to zero.
5033 // Also, if this flag is set we make sure that arraycopy interacts properly
5034 // with G1, eliding pre-barriers. See CR 6627983.
5035 dest_uninitialized = true;
5036 } else {
5037 // No zeroing elimination here.
5038 alloc = NULL;
5039 //original_dest = dest;
5040 //dest_uninitialized = false;
5041 }
5043 // Results are placed here:
5044 enum { fast_path = 1, // normal void-returning assembly stub
5045 checked_path = 2, // special assembly stub with cleanup
5046 slow_call_path = 3, // something went wrong; call the VM
5047 zero_path = 4, // bypass when length of copy is zero
5048 bcopy_path = 5, // copy primitive array by 64-bit blocks
5049 PATH_LIMIT = 6
5050 };
5051 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5052 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5053 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5054 record_for_igvn(result_region);
5055 _gvn.set_type_bottom(result_i_o);
5056 _gvn.set_type_bottom(result_memory);
5057 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5059 // The slow_control path:
5060 Node* slow_control;
5061 Node* slow_i_o = i_o();
5062 Node* slow_mem = memory(adr_type);
5063 debug_only(slow_control = (Node*) badAddress);
5065 // Checked control path:
5066 Node* checked_control = top();
5067 Node* checked_mem = NULL;
5068 Node* checked_i_o = NULL;
5069 Node* checked_value = NULL;
5071 if (basic_elem_type == T_CONFLICT) {
5072 assert(!dest_uninitialized, "");
5073 Node* cv = generate_generic_arraycopy(adr_type,
5074 src, src_offset, dest, dest_offset,
5075 copy_length, dest_uninitialized);
5076 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5077 checked_control = control();
5078 checked_i_o = i_o();
5079 checked_mem = memory(adr_type);
5080 checked_value = cv;
5081 set_control(top()); // no fast path
5082 }
5084 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5085 if (not_pos != NULL) {
5086 PreserveJVMState pjvms(this);
5087 set_control(not_pos);
5089 // (6) length must not be negative.
5090 if (!length_never_negative) {
5091 generate_negative_guard(copy_length, slow_region);
5092 }
5094 // copy_length is 0.
5095 if (!stopped() && dest_uninitialized) {
5096 Node* dest_length = alloc->in(AllocateNode::ALength);
5097 if (copy_length->eqv_uncast(dest_length)
5098 || _gvn.find_int_con(dest_length, 1) <= 0) {
5099 // There is no zeroing to do. No need for a secondary raw memory barrier.
5100 } else {
5101 // Clear the whole thing since there are no source elements to copy.
5102 generate_clear_array(adr_type, dest, basic_elem_type,
5103 intcon(0), NULL,
5104 alloc->in(AllocateNode::AllocSize));
5105 // Use a secondary InitializeNode as raw memory barrier.
5106 // Currently it is needed only on this path since other
5107 // paths have stub or runtime calls as raw memory barriers.
5108 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5109 Compile::AliasIdxRaw,
5110 top())->as_Initialize();
5111 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5112 }
5113 }
5115 // Present the results of the fast call.
5116 result_region->init_req(zero_path, control());
5117 result_i_o ->init_req(zero_path, i_o());
5118 result_memory->init_req(zero_path, memory(adr_type));
5119 }
5121 if (!stopped() && dest_uninitialized) {
5122 // We have to initialize the *uncopied* part of the array to zero.
5123 // The copy destination is the slice dest[off..off+len]. The other slices
5124 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5125 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5126 Node* dest_length = alloc->in(AllocateNode::ALength);
5127 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5128 copy_length));
5130 // If there is a head section that needs zeroing, do it now.
5131 if (find_int_con(dest_offset, -1) != 0) {
5132 generate_clear_array(adr_type, dest, basic_elem_type,
5133 intcon(0), dest_offset,
5134 NULL);
5135 }
5137 // Next, perform a dynamic check on the tail length.
5138 // It is often zero, and we can win big if we prove this.
5139 // There are two wins: Avoid generating the ClearArray
5140 // with its attendant messy index arithmetic, and upgrade
5141 // the copy to a more hardware-friendly word size of 64 bits.
5142 Node* tail_ctl = NULL;
5143 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5144 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5145 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5146 tail_ctl = generate_slow_guard(bol_lt, NULL);
5147 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5148 }
5150 // At this point, let's assume there is no tail.
5151 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5152 // There is no tail. Try an upgrade to a 64-bit copy.
5153 bool didit = false;
5154 { PreserveJVMState pjvms(this);
5155 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5156 src, src_offset, dest, dest_offset,
5157 dest_size, dest_uninitialized);
5158 if (didit) {
5159 // Present the results of the block-copying fast call.
5160 result_region->init_req(bcopy_path, control());
5161 result_i_o ->init_req(bcopy_path, i_o());
5162 result_memory->init_req(bcopy_path, memory(adr_type));
5163 }
5164 }
5165 if (didit)
5166 set_control(top()); // no regular fast path
5167 }
5169 // Clear the tail, if any.
5170 if (tail_ctl != NULL) {
5171 Node* notail_ctl = stopped() ? NULL : control();
5172 set_control(tail_ctl);
5173 if (notail_ctl == NULL) {
5174 generate_clear_array(adr_type, dest, basic_elem_type,
5175 dest_tail, NULL,
5176 dest_size);
5177 } else {
5178 // Make a local merge.
5179 Node* done_ctl = new(C) RegionNode(3);
5180 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5181 done_ctl->init_req(1, notail_ctl);
5182 done_mem->init_req(1, memory(adr_type));
5183 generate_clear_array(adr_type, dest, basic_elem_type,
5184 dest_tail, NULL,
5185 dest_size);
5186 done_ctl->init_req(2, control());
5187 done_mem->init_req(2, memory(adr_type));
5188 set_control( _gvn.transform(done_ctl));
5189 set_memory( _gvn.transform(done_mem), adr_type );
5190 }
5191 }
5192 }
5194 BasicType copy_type = basic_elem_type;
5195 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5196 if (!stopped() && copy_type == T_OBJECT) {
5197 // If src and dest have compatible element types, we can copy bits.
5198 // Types S[] and D[] are compatible if D is a supertype of S.
5199 //
5200 // If they are not, we will use checked_oop_disjoint_arraycopy,
5201 // which performs a fast optimistic per-oop check, and backs off
5202 // further to JVM_ArrayCopy on the first per-oop check that fails.
5203 // (Actually, we don't move raw bits only; the GC requires card marks.)
5205 // Get the Klass* for both src and dest
5206 Node* src_klass = load_object_klass(src);
5207 Node* dest_klass = load_object_klass(dest);
5209 // Generate the subtype check.
5210 // This might fold up statically, or then again it might not.
5211 //
5212 // Non-static example: Copying List<String>.elements to a new String[].
5213 // The backing store for a List<String> is always an Object[],
5214 // but its elements are always type String, if the generic types
5215 // are correct at the source level.
5216 //
5217 // Test S[] against D[], not S against D, because (probably)
5218 // the secondary supertype cache is less busy for S[] than S.
5219 // This usually only matters when D is an interface.
5220 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5221 // Plug failing path into checked_oop_disjoint_arraycopy
5222 if (not_subtype_ctrl != top()) {
5223 PreserveJVMState pjvms(this);
5224 set_control(not_subtype_ctrl);
5225 // (At this point we can assume disjoint_bases, since types differ.)
5226 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5227 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5228 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5229 Node* dest_elem_klass = _gvn.transform(n1);
5230 Node* cv = generate_checkcast_arraycopy(adr_type,
5231 dest_elem_klass,
5232 src, src_offset, dest, dest_offset,
5233 ConvI2X(copy_length), dest_uninitialized);
5234 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5235 checked_control = control();
5236 checked_i_o = i_o();
5237 checked_mem = memory(adr_type);
5238 checked_value = cv;
5239 }
5240 // At this point we know we do not need type checks on oop stores.
5242 // Let's see if we need card marks:
5243 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5244 // If we do not need card marks, copy using the jint or jlong stub.
5245 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5246 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5247 "sizes agree");
5248 }
5249 }
5251 if (!stopped()) {
5252 // Generate the fast path, if possible.
5253 PreserveJVMState pjvms(this);
5254 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5255 src, src_offset, dest, dest_offset,
5256 ConvI2X(copy_length), dest_uninitialized);
5258 // Present the results of the fast call.
5259 result_region->init_req(fast_path, control());
5260 result_i_o ->init_req(fast_path, i_o());
5261 result_memory->init_req(fast_path, memory(adr_type));
5262 }
5264 // Here are all the slow paths up to this point, in one bundle:
5265 slow_control = top();
5266 if (slow_region != NULL)
5267 slow_control = _gvn.transform(slow_region);
5268 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5270 set_control(checked_control);
5271 if (!stopped()) {
5272 // Clean up after the checked call.
5273 // The returned value is either 0 or -1^K,
5274 // where K = number of partially transferred array elements.
5275 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5276 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5277 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5279 // If it is 0, we are done, so transfer to the end.
5280 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5281 result_region->init_req(checked_path, checks_done);
5282 result_i_o ->init_req(checked_path, checked_i_o);
5283 result_memory->init_req(checked_path, checked_mem);
5285 // If it is not zero, merge into the slow call.
5286 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5287 RegionNode* slow_reg2 = new(C) RegionNode(3);
5288 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5289 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5290 record_for_igvn(slow_reg2);
5291 slow_reg2 ->init_req(1, slow_control);
5292 slow_i_o2 ->init_req(1, slow_i_o);
5293 slow_mem2 ->init_req(1, slow_mem);
5294 slow_reg2 ->init_req(2, control());
5295 slow_i_o2 ->init_req(2, checked_i_o);
5296 slow_mem2 ->init_req(2, checked_mem);
5298 slow_control = _gvn.transform(slow_reg2);
5299 slow_i_o = _gvn.transform(slow_i_o2);
5300 slow_mem = _gvn.transform(slow_mem2);
5302 if (alloc != NULL) {
5303 // We'll restart from the very beginning, after zeroing the whole thing.
5304 // This can cause double writes, but that's OK since dest is brand new.
5305 // So we ignore the low 31 bits of the value returned from the stub.
5306 } else {
5307 // We must continue the copy exactly where it failed, or else
5308 // another thread might see the wrong number of writes to dest.
5309 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5310 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5311 slow_offset->init_req(1, intcon(0));
5312 slow_offset->init_req(2, checked_offset);
5313 slow_offset = _gvn.transform(slow_offset);
5315 // Adjust the arguments by the conditionally incoming offset.
5316 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5317 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5318 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5320 // Tweak the node variables to adjust the code produced below:
5321 src_offset = src_off_plus;
5322 dest_offset = dest_off_plus;
5323 copy_length = length_minus;
5324 }
5325 }
5327 set_control(slow_control);
5328 if (!stopped()) {
5329 // Generate the slow path, if needed.
5330 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5332 set_memory(slow_mem, adr_type);
5333 set_i_o(slow_i_o);
5335 if (dest_uninitialized) {
5336 generate_clear_array(adr_type, dest, basic_elem_type,
5337 intcon(0), NULL,
5338 alloc->in(AllocateNode::AllocSize));
5339 }
5341 generate_slow_arraycopy(adr_type,
5342 src, src_offset, dest, dest_offset,
5343 copy_length, /*dest_uninitialized*/false);
5345 result_region->init_req(slow_call_path, control());
5346 result_i_o ->init_req(slow_call_path, i_o());
5347 result_memory->init_req(slow_call_path, memory(adr_type));
5348 }
5350 // Remove unused edges.
5351 for (uint i = 1; i < result_region->req(); i++) {
5352 if (result_region->in(i) == NULL)
5353 result_region->init_req(i, top());
5354 }
5356 // Finished; return the combined state.
5357 set_control( _gvn.transform(result_region));
5358 set_i_o( _gvn.transform(result_i_o) );
5359 set_memory( _gvn.transform(result_memory), adr_type );
5361 // The memory edges above are precise in order to model effects around
5362 // array copies accurately to allow value numbering of field loads around
5363 // arraycopy. Such field loads, both before and after, are common in Java
5364 // collections and similar classes involving header/array data structures.
5365 //
5366 // But with low number of register or when some registers are used or killed
5367 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5368 // The next memory barrier is added to avoid it. If the arraycopy can be
5369 // optimized away (which it can, sometimes) then we can manually remove
5370 // the membar also.
5371 //
5372 // Do not let reads from the cloned object float above the arraycopy.
5373 if (alloc != NULL) {
5374 // Do not let stores that initialize this object be reordered with
5375 // a subsequent store that would make this object accessible by
5376 // other threads.
5377 // Record what AllocateNode this StoreStore protects so that
5378 // escape analysis can go from the MemBarStoreStoreNode to the
5379 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5380 // based on the escape status of the AllocateNode.
5381 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5382 } else if (InsertMemBarAfterArraycopy)
5383 insert_mem_bar(Op_MemBarCPUOrder);
5384 }
5387 // Helper function which determines if an arraycopy immediately follows
5388 // an allocation, with no intervening tests or other escapes for the object.
5389 AllocateArrayNode*
5390 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5391 RegionNode* slow_region) {
5392 if (stopped()) return NULL; // no fast path
5393 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5395 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5396 if (alloc == NULL) return NULL;
5398 Node* rawmem = memory(Compile::AliasIdxRaw);
5399 // Is the allocation's memory state untouched?
5400 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5401 // Bail out if there have been raw-memory effects since the allocation.
5402 // (Example: There might have been a call or safepoint.)
5403 return NULL;
5404 }
5405 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5406 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5407 return NULL;
5408 }
5410 // There must be no unexpected observers of this allocation.
5411 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5412 Node* obs = ptr->fast_out(i);
5413 if (obs != this->map()) {
5414 return NULL;
5415 }
5416 }
5418 // This arraycopy must unconditionally follow the allocation of the ptr.
5419 Node* alloc_ctl = ptr->in(0);
5420 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5422 Node* ctl = control();
5423 while (ctl != alloc_ctl) {
5424 // There may be guards which feed into the slow_region.
5425 // Any other control flow means that we might not get a chance
5426 // to finish initializing the allocated object.
5427 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5428 IfNode* iff = ctl->in(0)->as_If();
5429 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5430 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5431 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5432 ctl = iff->in(0); // This test feeds the known slow_region.
5433 continue;
5434 }
5435 // One more try: Various low-level checks bottom out in
5436 // uncommon traps. If the debug-info of the trap omits
5437 // any reference to the allocation, as we've already
5438 // observed, then there can be no objection to the trap.
5439 bool found_trap = false;
5440 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5441 Node* obs = not_ctl->fast_out(j);
5442 if (obs->in(0) == not_ctl && obs->is_Call() &&
5443 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5444 found_trap = true; break;
5445 }
5446 }
5447 if (found_trap) {
5448 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5449 continue;
5450 }
5451 }
5452 return NULL;
5453 }
5455 // If we get this far, we have an allocation which immediately
5456 // precedes the arraycopy, and we can take over zeroing the new object.
5457 // The arraycopy will finish the initialization, and provide
5458 // a new control state to which we will anchor the destination pointer.
5460 return alloc;
5461 }
5463 // Helper for initialization of arrays, creating a ClearArray.
5464 // It writes zero bits in [start..end), within the body of an array object.
5465 // The memory effects are all chained onto the 'adr_type' alias category.
5466 //
5467 // Since the object is otherwise uninitialized, we are free
5468 // to put a little "slop" around the edges of the cleared area,
5469 // as long as it does not go back into the array's header,
5470 // or beyond the array end within the heap.
5471 //
5472 // The lower edge can be rounded down to the nearest jint and the
5473 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5474 //
5475 // Arguments:
5476 // adr_type memory slice where writes are generated
5477 // dest oop of the destination array
5478 // basic_elem_type element type of the destination
5479 // slice_idx array index of first element to store
5480 // slice_len number of elements to store (or NULL)
5481 // dest_size total size in bytes of the array object
5482 //
5483 // Exactly one of slice_len or dest_size must be non-NULL.
5484 // If dest_size is non-NULL, zeroing extends to the end of the object.
5485 // If slice_len is non-NULL, the slice_idx value must be a constant.
5486 void
5487 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5488 Node* dest,
5489 BasicType basic_elem_type,
5490 Node* slice_idx,
5491 Node* slice_len,
5492 Node* dest_size) {
5493 // one or the other but not both of slice_len and dest_size:
5494 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5495 if (slice_len == NULL) slice_len = top();
5496 if (dest_size == NULL) dest_size = top();
5498 // operate on this memory slice:
5499 Node* mem = memory(adr_type); // memory slice to operate on
5501 // scaling and rounding of indexes:
5502 int scale = exact_log2(type2aelembytes(basic_elem_type));
5503 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5504 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5505 int bump_bit = (-1 << scale) & BytesPerInt;
5507 // determine constant starts and ends
5508 const intptr_t BIG_NEG = -128;
5509 assert(BIG_NEG + 2*abase < 0, "neg enough");
5510 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5511 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5512 if (slice_len_con == 0) {
5513 return; // nothing to do here
5514 }
5515 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5516 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5517 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5518 assert(end_con < 0, "not two cons");
5519 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5520 BytesPerLong);
5521 }
5523 if (start_con >= 0 && end_con >= 0) {
5524 // Constant start and end. Simple.
5525 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5526 start_con, end_con, &_gvn);
5527 } else if (start_con >= 0 && dest_size != top()) {
5528 // Constant start, pre-rounded end after the tail of the array.
5529 Node* end = dest_size;
5530 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5531 start_con, end, &_gvn);
5532 } else if (start_con >= 0 && slice_len != top()) {
5533 // Constant start, non-constant end. End needs rounding up.
5534 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5535 intptr_t end_base = abase + (slice_idx_con << scale);
5536 int end_round = (-1 << scale) & (BytesPerLong - 1);
5537 Node* end = ConvI2X(slice_len);
5538 if (scale != 0)
5539 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5540 end_base += end_round;
5541 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5542 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5543 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5544 start_con, end, &_gvn);
5545 } else if (start_con < 0 && dest_size != top()) {
5546 // Non-constant start, pre-rounded end after the tail of the array.
5547 // This is almost certainly a "round-to-end" operation.
5548 Node* start = slice_idx;
5549 start = ConvI2X(start);
5550 if (scale != 0)
5551 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5552 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5553 if ((bump_bit | clear_low) != 0) {
5554 int to_clear = (bump_bit | clear_low);
5555 // Align up mod 8, then store a jint zero unconditionally
5556 // just before the mod-8 boundary.
5557 if (((abase + bump_bit) & ~to_clear) - bump_bit
5558 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5559 bump_bit = 0;
5560 assert((abase & to_clear) == 0, "array base must be long-aligned");
5561 } else {
5562 // Bump 'start' up to (or past) the next jint boundary:
5563 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5564 assert((abase & clear_low) == 0, "array base must be int-aligned");
5565 }
5566 // Round bumped 'start' down to jlong boundary in body of array.
5567 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5568 if (bump_bit != 0) {
5569 // Store a zero to the immediately preceding jint:
5570 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5571 Node* p1 = basic_plus_adr(dest, x1);
5572 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5573 mem = _gvn.transform(mem);
5574 }
5575 }
5576 Node* end = dest_size; // pre-rounded
5577 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5578 start, end, &_gvn);
5579 } else {
5580 // Non-constant start, unrounded non-constant end.
5581 // (Nobody zeroes a random midsection of an array using this routine.)
5582 ShouldNotReachHere(); // fix caller
5583 }
5585 // Done.
5586 set_memory(mem, adr_type);
5587 }
5590 bool
5591 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5592 BasicType basic_elem_type,
5593 AllocateNode* alloc,
5594 Node* src, Node* src_offset,
5595 Node* dest, Node* dest_offset,
5596 Node* dest_size, bool dest_uninitialized) {
5597 // See if there is an advantage from block transfer.
5598 int scale = exact_log2(type2aelembytes(basic_elem_type));
5599 if (scale >= LogBytesPerLong)
5600 return false; // it is already a block transfer
5602 // Look at the alignment of the starting offsets.
5603 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5605 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5606 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5607 if (src_off_con < 0 || dest_off_con < 0)
5608 // At present, we can only understand constants.
5609 return false;
5611 intptr_t src_off = abase + (src_off_con << scale);
5612 intptr_t dest_off = abase + (dest_off_con << scale);
5614 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5615 // Non-aligned; too bad.
5616 // One more chance: Pick off an initial 32-bit word.
5617 // This is a common case, since abase can be odd mod 8.
5618 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5619 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5620 Node* sptr = basic_plus_adr(src, src_off);
5621 Node* dptr = basic_plus_adr(dest, dest_off);
5622 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5623 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5624 src_off += BytesPerInt;
5625 dest_off += BytesPerInt;
5626 } else {
5627 return false;
5628 }
5629 }
5630 assert(src_off % BytesPerLong == 0, "");
5631 assert(dest_off % BytesPerLong == 0, "");
5633 // Do this copy by giant steps.
5634 Node* sptr = basic_plus_adr(src, src_off);
5635 Node* dptr = basic_plus_adr(dest, dest_off);
5636 Node* countx = dest_size;
5637 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5638 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5640 bool disjoint_bases = true; // since alloc != NULL
5641 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5642 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5644 return true;
5645 }
5648 // Helper function; generates code for the slow case.
5649 // We make a call to a runtime method which emulates the native method,
5650 // but without the native wrapper overhead.
5651 void
5652 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5653 Node* src, Node* src_offset,
5654 Node* dest, Node* dest_offset,
5655 Node* copy_length, bool dest_uninitialized) {
5656 assert(!dest_uninitialized, "Invariant");
5657 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5658 OptoRuntime::slow_arraycopy_Type(),
5659 OptoRuntime::slow_arraycopy_Java(),
5660 "slow_arraycopy", adr_type,
5661 src, src_offset, dest, dest_offset,
5662 copy_length);
5664 // Handle exceptions thrown by this fellow:
5665 make_slow_call_ex(call, env()->Throwable_klass(), false);
5666 }
5668 // Helper function; generates code for cases requiring runtime checks.
5669 Node*
5670 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5671 Node* dest_elem_klass,
5672 Node* src, Node* src_offset,
5673 Node* dest, Node* dest_offset,
5674 Node* copy_length, bool dest_uninitialized) {
5675 if (stopped()) return NULL;
5677 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5678 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5679 return NULL;
5680 }
5682 // Pick out the parameters required to perform a store-check
5683 // for the target array. This is an optimistic check. It will
5684 // look in each non-null element's class, at the desired klass's
5685 // super_check_offset, for the desired klass.
5686 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5687 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5688 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5689 Node* check_offset = ConvI2X(_gvn.transform(n3));
5690 Node* check_value = dest_elem_klass;
5692 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5693 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5695 // (We know the arrays are never conjoint, because their types differ.)
5696 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5697 OptoRuntime::checkcast_arraycopy_Type(),
5698 copyfunc_addr, "checkcast_arraycopy", adr_type,
5699 // five arguments, of which two are
5700 // intptr_t (jlong in LP64)
5701 src_start, dest_start,
5702 copy_length XTOP,
5703 check_offset XTOP,
5704 check_value);
5706 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5707 }
5710 // Helper function; generates code for cases requiring runtime checks.
5711 Node*
5712 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5713 Node* src, Node* src_offset,
5714 Node* dest, Node* dest_offset,
5715 Node* copy_length, bool dest_uninitialized) {
5716 assert(!dest_uninitialized, "Invariant");
5717 if (stopped()) return NULL;
5718 address copyfunc_addr = StubRoutines::generic_arraycopy();
5719 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5720 return NULL;
5721 }
5723 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5724 OptoRuntime::generic_arraycopy_Type(),
5725 copyfunc_addr, "generic_arraycopy", adr_type,
5726 src, src_offset, dest, dest_offset, copy_length);
5728 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5729 }
5731 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5732 void
5733 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5734 BasicType basic_elem_type,
5735 bool disjoint_bases,
5736 Node* src, Node* src_offset,
5737 Node* dest, Node* dest_offset,
5738 Node* copy_length, bool dest_uninitialized) {
5739 if (stopped()) return; // nothing to do
5741 Node* src_start = src;
5742 Node* dest_start = dest;
5743 if (src_offset != NULL || dest_offset != NULL) {
5744 assert(src_offset != NULL && dest_offset != NULL, "");
5745 src_start = array_element_address(src, src_offset, basic_elem_type);
5746 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5747 }
5749 // Figure out which arraycopy runtime method to call.
5750 const char* copyfunc_name = "arraycopy";
5751 address copyfunc_addr =
5752 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5753 disjoint_bases, copyfunc_name, dest_uninitialized);
5755 // Call it. Note that the count_ix value is not scaled to a byte-size.
5756 make_runtime_call(RC_LEAF|RC_NO_FP,
5757 OptoRuntime::fast_arraycopy_Type(),
5758 copyfunc_addr, copyfunc_name, adr_type,
5759 src_start, dest_start, copy_length XTOP);
5760 }
5762 //-------------inline_encodeISOArray-----------------------------------
5763 // encode char[] to byte[] in ISO_8859_1
5764 bool LibraryCallKit::inline_encodeISOArray() {
5765 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5766 // no receiver since it is static method
5767 Node *src = argument(0);
5768 Node *src_offset = argument(1);
5769 Node *dst = argument(2);
5770 Node *dst_offset = argument(3);
5771 Node *length = argument(4);
5773 const Type* src_type = src->Value(&_gvn);
5774 const Type* dst_type = dst->Value(&_gvn);
5775 const TypeAryPtr* top_src = src_type->isa_aryptr();
5776 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5777 if (top_src == NULL || top_src->klass() == NULL ||
5778 top_dest == NULL || top_dest->klass() == NULL) {
5779 // failed array check
5780 return false;
5781 }
5783 // Figure out the size and type of the elements we will be copying.
5784 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5785 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5786 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5787 return false;
5788 }
5789 Node* src_start = array_element_address(src, src_offset, src_elem);
5790 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5791 // 'src_start' points to src array + scaled offset
5792 // 'dst_start' points to dst array + scaled offset
5794 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5795 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5796 enc = _gvn.transform(enc);
5797 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5798 set_memory(res_mem, mtype);
5799 set_result(enc);
5800 return true;
5801 }
5803 //-------------inline_multiplyToLen-----------------------------------
5804 bool LibraryCallKit::inline_multiplyToLen() {
5805 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5807 address stubAddr = StubRoutines::multiplyToLen();
5808 if (stubAddr == NULL) {
5809 return false; // Intrinsic's stub is not implemented on this platform
5810 }
5811 const char* stubName = "multiplyToLen";
5813 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5815 // no receiver because it is a static method
5816 Node* x = argument(0);
5817 Node* xlen = argument(1);
5818 Node* y = argument(2);
5819 Node* ylen = argument(3);
5820 Node* z = argument(4);
5822 const Type* x_type = x->Value(&_gvn);
5823 const Type* y_type = y->Value(&_gvn);
5824 const TypeAryPtr* top_x = x_type->isa_aryptr();
5825 const TypeAryPtr* top_y = y_type->isa_aryptr();
5826 if (top_x == NULL || top_x->klass() == NULL ||
5827 top_y == NULL || top_y->klass() == NULL) {
5828 // failed array check
5829 return false;
5830 }
5832 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5833 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5834 if (x_elem != T_INT || y_elem != T_INT) {
5835 return false;
5836 }
5838 // Set the original stack and the reexecute bit for the interpreter to reexecute
5839 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5840 // on the return from z array allocation in runtime.
5841 { PreserveReexecuteState preexecs(this);
5842 jvms()->set_should_reexecute(true);
5844 Node* x_start = array_element_address(x, intcon(0), x_elem);
5845 Node* y_start = array_element_address(y, intcon(0), y_elem);
5846 // 'x_start' points to x array + scaled xlen
5847 // 'y_start' points to y array + scaled ylen
5849 // Allocate the result array
5850 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5851 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5852 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5854 IdealKit ideal(this);
5856 #define __ ideal.
5857 Node* one = __ ConI(1);
5858 Node* zero = __ ConI(0);
5859 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5860 __ set(need_alloc, zero);
5861 __ set(z_alloc, z);
5862 __ if_then(z, BoolTest::eq, null()); {
5863 __ increment (need_alloc, one);
5864 } __ else_(); {
5865 // Update graphKit memory and control from IdealKit.
5866 sync_kit(ideal);
5867 Node* zlen_arg = load_array_length(z);
5868 // Update IdealKit memory and control from graphKit.
5869 __ sync_kit(this);
5870 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5871 __ increment (need_alloc, one);
5872 } __ end_if();
5873 } __ end_if();
5875 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5876 // Update graphKit memory and control from IdealKit.
5877 sync_kit(ideal);
5878 Node * narr = new_array(klass_node, zlen, 1);
5879 // Update IdealKit memory and control from graphKit.
5880 __ sync_kit(this);
5881 __ set(z_alloc, narr);
5882 } __ end_if();
5884 sync_kit(ideal);
5885 z = __ value(z_alloc);
5886 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5887 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5888 // Final sync IdealKit and GraphKit.
5889 final_sync(ideal);
5890 #undef __
5892 Node* z_start = array_element_address(z, intcon(0), T_INT);
5894 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5895 OptoRuntime::multiplyToLen_Type(),
5896 stubAddr, stubName, TypePtr::BOTTOM,
5897 x_start, xlen, y_start, ylen, z_start, zlen);
5898 } // original reexecute is set back here
5900 C->set_has_split_ifs(true); // Has chance for split-if optimization
5901 set_result(z);
5902 return true;
5903 }
5905 //-------------inline_squareToLen------------------------------------
5906 bool LibraryCallKit::inline_squareToLen() {
5907 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5909 address stubAddr = StubRoutines::squareToLen();
5910 if (stubAddr == NULL) {
5911 return false; // Intrinsic's stub is not implemented on this platform
5912 }
5913 const char* stubName = "squareToLen";
5915 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5917 Node* x = argument(0);
5918 Node* len = argument(1);
5919 Node* z = argument(2);
5920 Node* zlen = argument(3);
5922 const Type* x_type = x->Value(&_gvn);
5923 const Type* z_type = z->Value(&_gvn);
5924 const TypeAryPtr* top_x = x_type->isa_aryptr();
5925 const TypeAryPtr* top_z = z_type->isa_aryptr();
5926 if (top_x == NULL || top_x->klass() == NULL ||
5927 top_z == NULL || top_z->klass() == NULL) {
5928 // failed array check
5929 return false;
5930 }
5932 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5933 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5934 if (x_elem != T_INT || z_elem != T_INT) {
5935 return false;
5936 }
5939 Node* x_start = array_element_address(x, intcon(0), x_elem);
5940 Node* z_start = array_element_address(z, intcon(0), z_elem);
5942 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5943 OptoRuntime::squareToLen_Type(),
5944 stubAddr, stubName, TypePtr::BOTTOM,
5945 x_start, len, z_start, zlen);
5947 set_result(z);
5948 return true;
5949 }
5951 //-------------inline_mulAdd------------------------------------------
5952 bool LibraryCallKit::inline_mulAdd() {
5953 assert(UseMulAddIntrinsic, "not implementated on this platform");
5955 address stubAddr = StubRoutines::mulAdd();
5956 if (stubAddr == NULL) {
5957 return false; // Intrinsic's stub is not implemented on this platform
5958 }
5959 const char* stubName = "mulAdd";
5961 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5963 Node* out = argument(0);
5964 Node* in = argument(1);
5965 Node* offset = argument(2);
5966 Node* len = argument(3);
5967 Node* k = argument(4);
5969 const Type* out_type = out->Value(&_gvn);
5970 const Type* in_type = in->Value(&_gvn);
5971 const TypeAryPtr* top_out = out_type->isa_aryptr();
5972 const TypeAryPtr* top_in = in_type->isa_aryptr();
5973 if (top_out == NULL || top_out->klass() == NULL ||
5974 top_in == NULL || top_in->klass() == NULL) {
5975 // failed array check
5976 return false;
5977 }
5979 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5980 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5981 if (out_elem != T_INT || in_elem != T_INT) {
5982 return false;
5983 }
5985 Node* outlen = load_array_length(out);
5986 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
5987 Node* out_start = array_element_address(out, intcon(0), out_elem);
5988 Node* in_start = array_element_address(in, intcon(0), in_elem);
5990 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5991 OptoRuntime::mulAdd_Type(),
5992 stubAddr, stubName, TypePtr::BOTTOM,
5993 out_start,in_start, new_offset, len, k);
5994 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5995 set_result(result);
5996 return true;
5997 }
5999 //-------------inline_montgomeryMultiply-----------------------------------
6000 bool LibraryCallKit::inline_montgomeryMultiply() {
6001 address stubAddr = StubRoutines::montgomeryMultiply();
6002 if (stubAddr == NULL) {
6003 return false; // Intrinsic's stub is not implemented on this platform
6004 }
6006 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6007 const char* stubName = "montgomery_multiply";
6009 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6011 Node* a = argument(0);
6012 Node* b = argument(1);
6013 Node* n = argument(2);
6014 Node* len = argument(3);
6015 Node* inv = argument(4);
6016 Node* m = argument(6);
6018 const Type* a_type = a->Value(&_gvn);
6019 const TypeAryPtr* top_a = a_type->isa_aryptr();
6020 const Type* b_type = b->Value(&_gvn);
6021 const TypeAryPtr* top_b = b_type->isa_aryptr();
6022 const Type* n_type = a->Value(&_gvn);
6023 const TypeAryPtr* top_n = n_type->isa_aryptr();
6024 const Type* m_type = a->Value(&_gvn);
6025 const TypeAryPtr* top_m = m_type->isa_aryptr();
6026 if (top_a == NULL || top_a->klass() == NULL ||
6027 top_b == NULL || top_b->klass() == NULL ||
6028 top_n == NULL || top_n->klass() == NULL ||
6029 top_m == NULL || top_m->klass() == NULL) {
6030 // failed array check
6031 return false;
6032 }
6034 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6035 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6036 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6037 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6038 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6039 return false;
6040 }
6042 // Make the call
6043 {
6044 Node* a_start = array_element_address(a, intcon(0), a_elem);
6045 Node* b_start = array_element_address(b, intcon(0), b_elem);
6046 Node* n_start = array_element_address(n, intcon(0), n_elem);
6047 Node* m_start = array_element_address(m, intcon(0), m_elem);
6049 Node* call = NULL;
6050 if (CCallingConventionRequiresIntsAsLongs) {
6051 Node* len_I2L = ConvI2L(len);
6052 call = make_runtime_call(RC_LEAF,
6053 OptoRuntime::montgomeryMultiply_Type(),
6054 stubAddr, stubName, TypePtr::BOTTOM,
6055 a_start, b_start, n_start, len_I2L XTOP, inv,
6056 top(), m_start);
6057 } else {
6058 call = make_runtime_call(RC_LEAF,
6059 OptoRuntime::montgomeryMultiply_Type(),
6060 stubAddr, stubName, TypePtr::BOTTOM,
6061 a_start, b_start, n_start, len, inv, top(),
6062 m_start);
6063 }
6064 set_result(m);
6065 }
6067 return true;
6068 }
6070 bool LibraryCallKit::inline_montgomerySquare() {
6071 address stubAddr = StubRoutines::montgomerySquare();
6072 if (stubAddr == NULL) {
6073 return false; // Intrinsic's stub is not implemented on this platform
6074 }
6076 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6077 const char* stubName = "montgomery_square";
6079 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6081 Node* a = argument(0);
6082 Node* n = argument(1);
6083 Node* len = argument(2);
6084 Node* inv = argument(3);
6085 Node* m = argument(5);
6087 const Type* a_type = a->Value(&_gvn);
6088 const TypeAryPtr* top_a = a_type->isa_aryptr();
6089 const Type* n_type = a->Value(&_gvn);
6090 const TypeAryPtr* top_n = n_type->isa_aryptr();
6091 const Type* m_type = a->Value(&_gvn);
6092 const TypeAryPtr* top_m = m_type->isa_aryptr();
6093 if (top_a == NULL || top_a->klass() == NULL ||
6094 top_n == NULL || top_n->klass() == NULL ||
6095 top_m == NULL || top_m->klass() == NULL) {
6096 // failed array check
6097 return false;
6098 }
6100 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6101 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6102 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6103 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6104 return false;
6105 }
6107 // Make the call
6108 {
6109 Node* a_start = array_element_address(a, intcon(0), a_elem);
6110 Node* n_start = array_element_address(n, intcon(0), n_elem);
6111 Node* m_start = array_element_address(m, intcon(0), m_elem);
6113 Node* call = NULL;
6114 if (CCallingConventionRequiresIntsAsLongs) {
6115 Node* len_I2L = ConvI2L(len);
6116 call = make_runtime_call(RC_LEAF,
6117 OptoRuntime::montgomerySquare_Type(),
6118 stubAddr, stubName, TypePtr::BOTTOM,
6119 a_start, n_start, len_I2L XTOP, inv, top(),
6120 m_start);
6121 } else {
6122 call = make_runtime_call(RC_LEAF,
6123 OptoRuntime::montgomerySquare_Type(),
6124 stubAddr, stubName, TypePtr::BOTTOM,
6125 a_start, n_start, len, inv, top(),
6126 m_start);
6127 }
6129 set_result(m);
6130 }
6132 return true;
6133 }
6136 /**
6137 * Calculate CRC32 for byte.
6138 * int java.util.zip.CRC32.update(int crc, int b)
6139 */
6140 bool LibraryCallKit::inline_updateCRC32() {
6141 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6142 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6143 // no receiver since it is static method
6144 Node* crc = argument(0); // type: int
6145 Node* b = argument(1); // type: int
6147 /*
6148 * int c = ~ crc;
6149 * b = timesXtoThe32[(b ^ c) & 0xFF];
6150 * b = b ^ (c >>> 8);
6151 * crc = ~b;
6152 */
6154 Node* M1 = intcon(-1);
6155 crc = _gvn.transform(new (C) XorINode(crc, M1));
6156 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6157 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6159 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6160 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6161 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6162 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6164 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6165 result = _gvn.transform(new (C) XorINode(crc, result));
6166 result = _gvn.transform(new (C) XorINode(result, M1));
6167 set_result(result);
6168 return true;
6169 }
6171 /**
6172 * Calculate CRC32 for byte[] array.
6173 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6174 */
6175 bool LibraryCallKit::inline_updateBytesCRC32() {
6176 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6177 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6178 // no receiver since it is static method
6179 Node* crc = argument(0); // type: int
6180 Node* src = argument(1); // type: oop
6181 Node* offset = argument(2); // type: int
6182 Node* length = argument(3); // type: int
6184 const Type* src_type = src->Value(&_gvn);
6185 const TypeAryPtr* top_src = src_type->isa_aryptr();
6186 if (top_src == NULL || top_src->klass() == NULL) {
6187 // failed array check
6188 return false;
6189 }
6191 // Figure out the size and type of the elements we will be copying.
6192 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6193 if (src_elem != T_BYTE) {
6194 return false;
6195 }
6197 // 'src_start' points to src array + scaled offset
6198 Node* src_start = array_element_address(src, offset, src_elem);
6200 // We assume that range check is done by caller.
6201 // TODO: generate range check (offset+length < src.length) in debug VM.
6203 // Call the stub.
6204 address stubAddr = StubRoutines::updateBytesCRC32();
6205 const char *stubName = "updateBytesCRC32";
6206 Node* call;
6207 if (CCallingConventionRequiresIntsAsLongs) {
6208 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6209 stubAddr, stubName, TypePtr::BOTTOM,
6210 crc XTOP, src_start, length XTOP);
6211 } else {
6212 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6213 stubAddr, stubName, TypePtr::BOTTOM,
6214 crc, src_start, length);
6215 }
6216 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6217 set_result(result);
6218 return true;
6219 }
6221 /**
6222 * Calculate CRC32 for ByteBuffer.
6223 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6224 */
6225 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6226 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6227 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6228 // no receiver since it is static method
6229 Node* crc = argument(0); // type: int
6230 Node* src = argument(1); // type: long
6231 Node* offset = argument(3); // type: int
6232 Node* length = argument(4); // type: int
6234 src = ConvL2X(src); // adjust Java long to machine word
6235 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6236 offset = ConvI2X(offset);
6238 // 'src_start' points to src array + scaled offset
6239 Node* src_start = basic_plus_adr(top(), base, offset);
6241 // Call the stub.
6242 address stubAddr = StubRoutines::updateBytesCRC32();
6243 const char *stubName = "updateBytesCRC32";
6244 Node* call;
6245 if (CCallingConventionRequiresIntsAsLongs) {
6246 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6247 stubAddr, stubName, TypePtr::BOTTOM,
6248 crc XTOP, src_start, length XTOP);
6249 } else {
6250 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6251 stubAddr, stubName, TypePtr::BOTTOM,
6252 crc, src_start, length);
6253 }
6254 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6255 set_result(result);
6256 return true;
6257 }
6259 //----------------------------inline_reference_get----------------------------
6260 // public T java.lang.ref.Reference.get();
6261 bool LibraryCallKit::inline_reference_get() {
6262 const int referent_offset = java_lang_ref_Reference::referent_offset;
6263 guarantee(referent_offset > 0, "should have already been set");
6265 // Get the argument:
6266 Node* reference_obj = null_check_receiver();
6267 if (stopped()) return true;
6269 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6271 ciInstanceKlass* klass = env()->Object_klass();
6272 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6274 Node* no_ctrl = NULL;
6275 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6277 // Use the pre-barrier to record the value in the referent field
6278 pre_barrier(false /* do_load */,
6279 control(),
6280 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6281 result /* pre_val */,
6282 T_OBJECT);
6284 // Add memory barrier to prevent commoning reads from this field
6285 // across safepoint since GC can change its value.
6286 insert_mem_bar(Op_MemBarCPUOrder);
6288 set_result(result);
6289 return true;
6290 }
6293 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6294 bool is_exact=true, bool is_static=false) {
6296 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6297 assert(tinst != NULL, "obj is null");
6298 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6299 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6301 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6302 ciSymbol::make(fieldTypeString),
6303 is_static);
6304 if (field == NULL) return (Node *) NULL;
6305 assert (field != NULL, "undefined field");
6307 // Next code copied from Parse::do_get_xxx():
6309 // Compute address and memory type.
6310 int offset = field->offset_in_bytes();
6311 bool is_vol = field->is_volatile();
6312 ciType* field_klass = field->type();
6313 assert(field_klass->is_loaded(), "should be loaded");
6314 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6315 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6316 BasicType bt = field->layout_type();
6318 // Build the resultant type of the load
6319 const Type *type;
6320 if (bt == T_OBJECT) {
6321 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6322 } else {
6323 type = Type::get_const_basic_type(bt);
6324 }
6326 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6327 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6328 }
6329 // Build the load.
6330 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6331 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6332 // If reference is volatile, prevent following memory ops from
6333 // floating up past the volatile read. Also prevents commoning
6334 // another volatile read.
6335 if (is_vol) {
6336 // Memory barrier includes bogus read of value to force load BEFORE membar
6337 insert_mem_bar(Op_MemBarAcquire, loadedField);
6338 }
6339 return loadedField;
6340 }
6343 //------------------------------inline_aescrypt_Block-----------------------
6344 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6345 address stubAddr = NULL;
6346 const char *stubName;
6347 assert(UseAES, "need AES instruction support");
6349 switch(id) {
6350 case vmIntrinsics::_aescrypt_encryptBlock:
6351 stubAddr = StubRoutines::aescrypt_encryptBlock();
6352 stubName = "aescrypt_encryptBlock";
6353 break;
6354 case vmIntrinsics::_aescrypt_decryptBlock:
6355 stubAddr = StubRoutines::aescrypt_decryptBlock();
6356 stubName = "aescrypt_decryptBlock";
6357 break;
6358 }
6359 if (stubAddr == NULL) return false;
6361 Node* aescrypt_object = argument(0);
6362 Node* src = argument(1);
6363 Node* src_offset = argument(2);
6364 Node* dest = argument(3);
6365 Node* dest_offset = argument(4);
6367 // (1) src and dest are arrays.
6368 const Type* src_type = src->Value(&_gvn);
6369 const Type* dest_type = dest->Value(&_gvn);
6370 const TypeAryPtr* top_src = src_type->isa_aryptr();
6371 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6372 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6374 // for the quick and dirty code we will skip all the checks.
6375 // we are just trying to get the call to be generated.
6376 Node* src_start = src;
6377 Node* dest_start = dest;
6378 if (src_offset != NULL || dest_offset != NULL) {
6379 assert(src_offset != NULL && dest_offset != NULL, "");
6380 src_start = array_element_address(src, src_offset, T_BYTE);
6381 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6382 }
6384 // now need to get the start of its expanded key array
6385 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6386 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6387 if (k_start == NULL) return false;
6389 if (Matcher::pass_original_key_for_aes()) {
6390 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6391 // compatibility issues between Java key expansion and SPARC crypto instructions
6392 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6393 if (original_k_start == NULL) return false;
6395 // Call the stub.
6396 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6397 stubAddr, stubName, TypePtr::BOTTOM,
6398 src_start, dest_start, k_start, original_k_start);
6399 } else {
6400 // Call the stub.
6401 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6402 stubAddr, stubName, TypePtr::BOTTOM,
6403 src_start, dest_start, k_start);
6404 }
6406 return true;
6407 }
6409 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6410 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6411 address stubAddr = NULL;
6412 const char *stubName = NULL;
6414 assert(UseAES, "need AES instruction support");
6416 switch(id) {
6417 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6418 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6419 stubName = "cipherBlockChaining_encryptAESCrypt";
6420 break;
6421 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6422 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6423 stubName = "cipherBlockChaining_decryptAESCrypt";
6424 break;
6425 }
6426 if (stubAddr == NULL) return false;
6428 Node* cipherBlockChaining_object = argument(0);
6429 Node* src = argument(1);
6430 Node* src_offset = argument(2);
6431 Node* len = argument(3);
6432 Node* dest = argument(4);
6433 Node* dest_offset = argument(5);
6435 // (1) src and dest are arrays.
6436 const Type* src_type = src->Value(&_gvn);
6437 const Type* dest_type = dest->Value(&_gvn);
6438 const TypeAryPtr* top_src = src_type->isa_aryptr();
6439 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6440 assert (top_src != NULL && top_src->klass() != NULL
6441 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6443 // checks are the responsibility of the caller
6444 Node* src_start = src;
6445 Node* dest_start = dest;
6446 if (src_offset != NULL || dest_offset != NULL) {
6447 assert(src_offset != NULL && dest_offset != NULL, "");
6448 src_start = array_element_address(src, src_offset, T_BYTE);
6449 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6450 }
6452 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6453 // (because of the predicated logic executed earlier).
6454 // so we cast it here safely.
6455 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6457 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6458 if (embeddedCipherObj == NULL) return false;
6460 // cast it to what we know it will be at runtime
6461 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6462 assert(tinst != NULL, "CBC obj is null");
6463 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6464 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6465 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6467 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6468 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6469 const TypeOopPtr* xtype = aklass->as_instance_type();
6470 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6471 aescrypt_object = _gvn.transform(aescrypt_object);
6473 // we need to get the start of the aescrypt_object's expanded key array
6474 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6475 if (k_start == NULL) return false;
6477 // similarly, get the start address of the r vector
6478 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6479 if (objRvec == NULL) return false;
6480 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6482 Node* cbcCrypt;
6483 if (Matcher::pass_original_key_for_aes()) {
6484 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6485 // compatibility issues between Java key expansion and SPARC crypto instructions
6486 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6487 if (original_k_start == NULL) return false;
6489 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6490 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6491 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6492 stubAddr, stubName, TypePtr::BOTTOM,
6493 src_start, dest_start, k_start, r_start, len, original_k_start);
6494 } else {
6495 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6496 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6497 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6498 stubAddr, stubName, TypePtr::BOTTOM,
6499 src_start, dest_start, k_start, r_start, len);
6500 }
6502 // return cipher length (int)
6503 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6504 set_result(retvalue);
6505 return true;
6506 }
6508 //------------------------------get_key_start_from_aescrypt_object-----------------------
6509 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6510 #ifdef PPC64
6511 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6512 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6513 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6514 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6515 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6516 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6517 if (objSessionK == NULL) {
6518 return (Node *) NULL;
6519 }
6520 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6521 #else
6522 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6523 #endif // PPC64
6524 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6525 if (objAESCryptKey == NULL) return (Node *) NULL;
6527 // now have the array, need to get the start address of the K array
6528 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6529 return k_start;
6530 }
6532 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6533 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6534 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6535 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6536 if (objAESCryptKey == NULL) return (Node *) NULL;
6538 // now have the array, need to get the start address of the lastKey array
6539 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6540 return original_k_start;
6541 }
6543 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6544 // Return node representing slow path of predicate check.
6545 // the pseudo code we want to emulate with this predicate is:
6546 // for encryption:
6547 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6548 // for decryption:
6549 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6550 // note cipher==plain is more conservative than the original java code but that's OK
6551 //
6552 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6553 // The receiver was checked for NULL already.
6554 Node* objCBC = argument(0);
6556 // Load embeddedCipher field of CipherBlockChaining object.
6557 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6559 // get AESCrypt klass for instanceOf check
6560 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6561 // will have same classloader as CipherBlockChaining object
6562 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6563 assert(tinst != NULL, "CBCobj is null");
6564 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6566 // we want to do an instanceof comparison against the AESCrypt class
6567 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6568 if (!klass_AESCrypt->is_loaded()) {
6569 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6570 Node* ctrl = control();
6571 set_control(top()); // no regular fast path
6572 return ctrl;
6573 }
6574 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6576 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6577 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6578 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6580 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6582 // for encryption, we are done
6583 if (!decrypting)
6584 return instof_false; // even if it is NULL
6586 // for decryption, we need to add a further check to avoid
6587 // taking the intrinsic path when cipher and plain are the same
6588 // see the original java code for why.
6589 RegionNode* region = new(C) RegionNode(3);
6590 region->init_req(1, instof_false);
6591 Node* src = argument(1);
6592 Node* dest = argument(4);
6593 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6594 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6595 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6596 region->init_req(2, src_dest_conjoint);
6598 record_for_igvn(region);
6599 return _gvn.transform(region);
6600 }
6602 //------------------------------inline_sha_implCompress-----------------------
6603 //
6604 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6605 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6606 //
6607 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6608 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6609 //
6610 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6611 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6612 //
6613 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6614 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6616 Node* sha_obj = argument(0);
6617 Node* src = argument(1); // type oop
6618 Node* ofs = argument(2); // type int
6620 const Type* src_type = src->Value(&_gvn);
6621 const TypeAryPtr* top_src = src_type->isa_aryptr();
6622 if (top_src == NULL || top_src->klass() == NULL) {
6623 // failed array check
6624 return false;
6625 }
6626 // Figure out the size and type of the elements we will be copying.
6627 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6628 if (src_elem != T_BYTE) {
6629 return false;
6630 }
6631 // 'src_start' points to src array + offset
6632 Node* src_start = array_element_address(src, ofs, src_elem);
6633 Node* state = NULL;
6634 address stubAddr;
6635 const char *stubName;
6637 switch(id) {
6638 case vmIntrinsics::_sha_implCompress:
6639 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6640 state = get_state_from_sha_object(sha_obj);
6641 stubAddr = StubRoutines::sha1_implCompress();
6642 stubName = "sha1_implCompress";
6643 break;
6644 case vmIntrinsics::_sha2_implCompress:
6645 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6646 state = get_state_from_sha_object(sha_obj);
6647 stubAddr = StubRoutines::sha256_implCompress();
6648 stubName = "sha256_implCompress";
6649 break;
6650 case vmIntrinsics::_sha5_implCompress:
6651 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6652 state = get_state_from_sha5_object(sha_obj);
6653 stubAddr = StubRoutines::sha512_implCompress();
6654 stubName = "sha512_implCompress";
6655 break;
6656 default:
6657 fatal_unexpected_iid(id);
6658 return false;
6659 }
6660 if (state == NULL) return false;
6662 // Call the stub.
6663 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6664 stubAddr, stubName, TypePtr::BOTTOM,
6665 src_start, state);
6667 return true;
6668 }
6670 //------------------------------inline_digestBase_implCompressMB-----------------------
6671 //
6672 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6673 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6674 //
6675 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6676 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6677 "need SHA1/SHA256/SHA512 instruction support");
6678 assert((uint)predicate < 3, "sanity");
6679 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6681 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6682 Node* src = argument(1); // byte[] array
6683 Node* ofs = argument(2); // type int
6684 Node* limit = argument(3); // type int
6686 const Type* src_type = src->Value(&_gvn);
6687 const TypeAryPtr* top_src = src_type->isa_aryptr();
6688 if (top_src == NULL || top_src->klass() == NULL) {
6689 // failed array check
6690 return false;
6691 }
6692 // Figure out the size and type of the elements we will be copying.
6693 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6694 if (src_elem != T_BYTE) {
6695 return false;
6696 }
6697 // 'src_start' points to src array + offset
6698 Node* src_start = array_element_address(src, ofs, src_elem);
6700 const char* klass_SHA_name = NULL;
6701 const char* stub_name = NULL;
6702 address stub_addr = NULL;
6703 bool long_state = false;
6705 switch (predicate) {
6706 case 0:
6707 if (UseSHA1Intrinsics) {
6708 klass_SHA_name = "sun/security/provider/SHA";
6709 stub_name = "sha1_implCompressMB";
6710 stub_addr = StubRoutines::sha1_implCompressMB();
6711 }
6712 break;
6713 case 1:
6714 if (UseSHA256Intrinsics) {
6715 klass_SHA_name = "sun/security/provider/SHA2";
6716 stub_name = "sha256_implCompressMB";
6717 stub_addr = StubRoutines::sha256_implCompressMB();
6718 }
6719 break;
6720 case 2:
6721 if (UseSHA512Intrinsics) {
6722 klass_SHA_name = "sun/security/provider/SHA5";
6723 stub_name = "sha512_implCompressMB";
6724 stub_addr = StubRoutines::sha512_implCompressMB();
6725 long_state = true;
6726 }
6727 break;
6728 default:
6729 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6730 }
6731 if (klass_SHA_name != NULL) {
6732 // get DigestBase klass to lookup for SHA klass
6733 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6734 assert(tinst != NULL, "digestBase_obj is not instance???");
6735 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6737 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6738 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6739 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6740 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6741 }
6742 return false;
6743 }
6744 //------------------------------inline_sha_implCompressMB-----------------------
6745 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6746 bool long_state, address stubAddr, const char *stubName,
6747 Node* src_start, Node* ofs, Node* limit) {
6748 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6749 const TypeOopPtr* xtype = aklass->as_instance_type();
6750 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6751 sha_obj = _gvn.transform(sha_obj);
6753 Node* state;
6754 if (long_state) {
6755 state = get_state_from_sha5_object(sha_obj);
6756 } else {
6757 state = get_state_from_sha_object(sha_obj);
6758 }
6759 if (state == NULL) return false;
6761 // Call the stub.
6762 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6763 OptoRuntime::digestBase_implCompressMB_Type(),
6764 stubAddr, stubName, TypePtr::BOTTOM,
6765 src_start, state, ofs, limit);
6766 // return ofs (int)
6767 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6768 set_result(result);
6770 return true;
6771 }
6773 //------------------------------get_state_from_sha_object-----------------------
6774 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6775 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6776 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6777 if (sha_state == NULL) return (Node *) NULL;
6779 // now have the array, need to get the start address of the state array
6780 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6781 return state;
6782 }
6784 //------------------------------get_state_from_sha5_object-----------------------
6785 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6786 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6787 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6788 if (sha_state == NULL) return (Node *) NULL;
6790 // now have the array, need to get the start address of the state array
6791 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6792 return state;
6793 }
6795 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6796 // Return node representing slow path of predicate check.
6797 // the pseudo code we want to emulate with this predicate is:
6798 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6799 //
6800 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6801 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6802 "need SHA1/SHA256/SHA512 instruction support");
6803 assert((uint)predicate < 3, "sanity");
6805 // The receiver was checked for NULL already.
6806 Node* digestBaseObj = argument(0);
6808 // get DigestBase klass for instanceOf check
6809 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6810 assert(tinst != NULL, "digestBaseObj is null");
6811 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6813 const char* klass_SHA_name = NULL;
6814 switch (predicate) {
6815 case 0:
6816 if (UseSHA1Intrinsics) {
6817 // we want to do an instanceof comparison against the SHA class
6818 klass_SHA_name = "sun/security/provider/SHA";
6819 }
6820 break;
6821 case 1:
6822 if (UseSHA256Intrinsics) {
6823 // we want to do an instanceof comparison against the SHA2 class
6824 klass_SHA_name = "sun/security/provider/SHA2";
6825 }
6826 break;
6827 case 2:
6828 if (UseSHA512Intrinsics) {
6829 // we want to do an instanceof comparison against the SHA5 class
6830 klass_SHA_name = "sun/security/provider/SHA5";
6831 }
6832 break;
6833 default:
6834 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6835 }
6837 ciKlass* klass_SHA = NULL;
6838 if (klass_SHA_name != NULL) {
6839 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6840 }
6841 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6842 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6843 Node* ctrl = control();
6844 set_control(top()); // no intrinsic path
6845 return ctrl;
6846 }
6847 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6849 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6850 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6851 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6852 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6854 return instof_false; // even if it is NULL
6855 }
6857 bool LibraryCallKit::inline_profileBoolean() {
6858 Node* counts = argument(1);
6859 const TypeAryPtr* ary = NULL;
6860 ciArray* aobj = NULL;
6861 if (counts->is_Con()
6862 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6863 && (aobj = ary->const_oop()->as_array()) != NULL
6864 && (aobj->length() == 2)) {
6865 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6866 jint false_cnt = aobj->element_value(0).as_int();
6867 jint true_cnt = aobj->element_value(1).as_int();
6869 if (C->log() != NULL) {
6870 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6871 false_cnt, true_cnt);
6872 }
6874 if (false_cnt + true_cnt == 0) {
6875 // According to profile, never executed.
6876 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6877 Deoptimization::Action_reinterpret);
6878 return true;
6879 }
6881 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6882 // is a number of each value occurrences.
6883 Node* result = argument(0);
6884 if (false_cnt == 0 || true_cnt == 0) {
6885 // According to profile, one value has been never seen.
6886 int expected_val = (false_cnt == 0) ? 1 : 0;
6888 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6889 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6891 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6892 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6893 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6895 { // Slow path: uncommon trap for never seen value and then reexecute
6896 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6897 // the value has been seen at least once.
6898 PreserveJVMState pjvms(this);
6899 PreserveReexecuteState preexecs(this);
6900 jvms()->set_should_reexecute(true);
6902 set_control(slow_path);
6903 set_i_o(i_o());
6905 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6906 Deoptimization::Action_reinterpret);
6907 }
6908 // The guard for never seen value enables sharpening of the result and
6909 // returning a constant. It allows to eliminate branches on the same value
6910 // later on.
6911 set_control(fast_path);
6912 result = intcon(expected_val);
6913 }
6914 // Stop profiling.
6915 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6916 // By replacing method body with profile data (represented as ProfileBooleanNode
6917 // on IR level) we effectively disable profiling.
6918 // It enables full speed execution once optimized code is generated.
6919 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
6920 C->record_for_igvn(profile);
6921 set_result(profile);
6922 return true;
6923 } else {
6924 // Continue profiling.
6925 // Profile data isn't available at the moment. So, execute method's bytecode version.
6926 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6927 // is compiled and counters aren't available since corresponding MethodHandle
6928 // isn't a compile-time constant.
6929 return false;
6930 }
6931 }