1 /*
2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "jfr/support/jfrIntrinsics.hpp"
31 #include "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callGenerator.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/mathexactnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/parse.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "prims/nativeLookup.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "utilities/macros.hpp"
45
46 class LibraryIntrinsic : public InlineCallGenerator {
47 // Extend the set of intrinsics known to the runtime:
48 public:
49 private:
50 bool _is_virtual;
51 bool _does_virtual_dispatch;
52 int8_t _predicates_count; // Intrinsic is predicated by several conditions
53 int8_t _last_predicate; // Last generated predicate
54 vmIntrinsics::ID _intrinsic_id;
55
56 public:
57 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
58 : InlineCallGenerator(m),
59 _is_virtual(is_virtual),
60 _does_virtual_dispatch(does_virtual_dispatch),
61 _predicates_count((int8_t)predicates_count),
62 _last_predicate((int8_t)-1),
63 _intrinsic_id(id)
64 {
65 }
66 virtual bool is_intrinsic() const { return true; }
67 virtual bool is_virtual() const { return _is_virtual; }
68 virtual bool is_predicated() const { return _predicates_count > 0; }
69 virtual int predicates_count() const { return _predicates_count; }
70 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
71 virtual JVMState* generate(JVMState* jvms);
72 virtual Node* generate_predicate(JVMState* jvms, int predicate);
73 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
74 };
75
76
77 // Local helper class for LibraryIntrinsic:
78 class LibraryCallKit : public GraphKit {
79 private:
80 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
81 Node* _result; // the result node, if any
82 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
83
84 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
85
86 public:
87 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
88 : GraphKit(jvms),
89 _intrinsic(intrinsic),
90 _result(NULL)
91 {
92 // Check if this is a root compile. In that case we don't have a caller.
93 if (!jvms->has_method()) {
94 _reexecute_sp = sp();
95 } else {
96 // Find out how many arguments the interpreter needs when deoptimizing
97 // and save the stack pointer value so it can used by uncommon_trap.
98 // We find the argument count by looking at the declared signature.
99 bool ignored_will_link;
100 ciSignature* declared_signature = NULL;
101 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
102 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
103 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
104 }
105 }
106
107 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
108
109 ciMethod* caller() const { return jvms()->method(); }
110 int bci() const { return jvms()->bci(); }
111 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
112 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
113 ciMethod* callee() const { return _intrinsic->method(); }
114
115 bool try_to_inline(int predicate);
116 Node* try_to_predicate(int predicate);
117
118 void push_result() {
119 // Push the result onto the stack.
120 if (!stopped() && result() != NULL) {
121 BasicType bt = result()->bottom_type()->basic_type();
122 push_node(bt, result());
123 }
124 }
125
126 private:
127 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
128 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
129 }
130
131 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
132 void set_result(RegionNode* region, PhiNode* value);
133 Node* result() { return _result; }
134
135 virtual int reexecute_sp() { return _reexecute_sp; }
136
137 // Helper functions to inline natives
138 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
139 Node* generate_slow_guard(Node* test, RegionNode* region);
140 Node* generate_fair_guard(Node* test, RegionNode* region);
141 Node* generate_negative_guard(Node* index, RegionNode* region,
142 // resulting CastII of index:
143 Node* *pos_index = NULL);
144 Node* generate_nonpositive_guard(Node* index, bool never_negative,
145 // resulting CastII of index:
146 Node* *pos_index = NULL);
147 Node* generate_limit_guard(Node* offset, Node* subseq_length,
148 Node* array_length,
149 RegionNode* region);
150 Node* generate_current_thread(Node* &tls_output);
151 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
152 bool disjoint_bases, const char* &name, bool dest_uninitialized);
153 Node* load_mirror_from_klass(Node* klass);
154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path,
156 int offset);
157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path) {
159 int offset = java_lang_Class::klass_offset_in_bytes();
160 return load_klass_from_mirror_common(mirror, never_see_null,
161 region, null_path,
162 offset);
163 }
164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path) {
166 int offset = java_lang_Class::array_klass_offset_in_bytes();
167 return load_klass_from_mirror_common(mirror, never_see_null,
168 region, null_path,
169 offset);
170 }
171 Node* generate_access_flags_guard(Node* kls,
172 int modifier_mask, int modifier_bits,
173 RegionNode* region);
174 Node* generate_interface_guard(Node* kls, RegionNode* region);
175 Node* generate_array_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, false, false);
177 }
178 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, true);
180 }
181 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, true, false);
183 }
184 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, true);
186 }
187 Node* generate_array_guard_common(Node* kls, RegionNode* region,
188 bool obj_array, bool not_array);
189 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
190 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
191 bool is_virtual = false, bool is_static = false);
192 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
193 return generate_method_call(method_id, false, true);
194 }
195 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, true, false);
197 }
198 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
199
200 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
201 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
202 bool inline_string_compareTo();
203 bool inline_string_indexOf();
204 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
205 bool inline_string_equals();
206 Node* round_double_node(Node* n);
207 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
208 bool inline_math_native(vmIntrinsics::ID id);
209 bool inline_trig(vmIntrinsics::ID id);
210 bool inline_math(vmIntrinsics::ID id);
211 template <typename OverflowOp>
212 bool inline_math_overflow(Node* arg1, Node* arg2);
213 void inline_math_mathExact(Node* math, Node* test);
214 bool inline_math_addExactI(bool is_increment);
215 bool inline_math_addExactL(bool is_increment);
216 bool inline_math_multiplyExactI();
217 bool inline_math_multiplyExactL();
218 bool inline_math_negateExactI();
219 bool inline_math_negateExactL();
220 bool inline_math_subtractExactI(bool is_decrement);
221 bool inline_math_subtractExactL(bool is_decrement);
222 bool inline_exp();
223 bool inline_pow();
224 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_min_max(vmIntrinsics::ID id);
226 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
227 // This returns Type::AnyPtr, RawPtr, or OopPtr.
228 int classify_unsafe_addr(Node* &base, Node* &offset);
229 Node* make_unsafe_address(Node* base, Node* offset);
230 // Helper for inline_unsafe_access.
231 // Generates the guards that check whether the result of
232 // Unsafe.getObject should be recorded in an SATB log buffer.
233 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
234 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
235 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
236 static bool klass_needs_init_guard(Node* kls);
237 bool inline_unsafe_allocate();
238 bool inline_unsafe_copyMemory();
239 bool inline_native_currentThread();
240 #ifdef JFR_HAVE_INTRINSICS
241 bool inline_native_classID();
242 bool inline_native_getEventWriter();
243 #endif
244 bool inline_native_time_funcs(address method, const char* funcName);
245 bool inline_native_isInterrupted();
246 bool inline_native_Class_query(vmIntrinsics::ID id);
247 bool inline_native_subtype_check();
248
249 bool inline_native_newArray();
250 bool inline_native_getLength();
251 bool inline_array_copyOf(bool is_copyOfRange);
252 bool inline_array_equals();
253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
254 bool inline_native_clone(bool is_virtual);
255 bool inline_native_Reflection_getCallerClass();
256 // Helper function for inlining native object hash method
257 bool inline_native_hashcode(bool is_virtual, bool is_static);
258 bool inline_native_getClass();
259
260 // Helper functions for inlining arraycopy
261 bool inline_arraycopy();
262 void generate_arraycopy(const TypePtr* adr_type,
263 BasicType basic_elem_type,
264 Node* src, Node* src_offset,
265 Node* dest, Node* dest_offset,
266 Node* copy_length,
267 bool disjoint_bases = false,
268 bool length_never_negative = false,
269 RegionNode* slow_region = NULL);
270 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
271 RegionNode* slow_region);
272 void generate_clear_array(const TypePtr* adr_type,
273 Node* dest,
274 BasicType basic_elem_type,
275 Node* slice_off,
276 Node* slice_len,
277 Node* slice_end);
278 bool generate_block_arraycopy(const TypePtr* adr_type,
279 BasicType basic_elem_type,
280 AllocateNode* alloc,
281 Node* src, Node* src_offset,
282 Node* dest, Node* dest_offset,
283 Node* dest_size, bool dest_uninitialized);
284 void generate_slow_arraycopy(const TypePtr* adr_type,
285 Node* src, Node* src_offset,
286 Node* dest, Node* dest_offset,
287 Node* copy_length, bool dest_uninitialized);
288 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
289 Node* dest_elem_klass,
290 Node* src, Node* src_offset,
291 Node* dest, Node* dest_offset,
292 Node* copy_length, bool dest_uninitialized);
293 Node* generate_generic_arraycopy(const TypePtr* adr_type,
294 Node* src, Node* src_offset,
295 Node* dest, Node* dest_offset,
296 Node* copy_length, bool dest_uninitialized);
297 void generate_unchecked_arraycopy(const TypePtr* adr_type,
298 BasicType basic_elem_type,
299 bool disjoint_bases,
300 Node* src, Node* src_offset,
301 Node* dest, Node* dest_offset,
302 Node* copy_length, bool dest_uninitialized);
303 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
304 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
305 bool inline_unsafe_ordered_store(BasicType type);
306 bool inline_unsafe_fence(vmIntrinsics::ID id);
307 bool inline_fp_conversions(vmIntrinsics::ID id);
308 bool inline_number_methods(vmIntrinsics::ID id);
309 bool inline_reference_get();
310 bool inline_aescrypt_Block(vmIntrinsics::ID id);
311 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
312 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
313 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
314 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
315 bool inline_ghash_processBlocks();
316 bool inline_sha_implCompress(vmIntrinsics::ID id);
317 bool inline_digestBase_implCompressMB(int predicate);
318 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
319 bool long_state, address stubAddr, const char *stubName,
320 Node* src_start, Node* ofs, Node* limit);
321 Node* get_state_from_sha_object(Node *sha_object);
322 Node* get_state_from_sha5_object(Node *sha_object);
323 Node* inline_digestBase_implCompressMB_predicate(int predicate);
324 bool inline_encodeISOArray();
325 bool inline_updateCRC32();
326 bool inline_updateBytesCRC32();
327 bool inline_updateByteBufferCRC32();
328 bool inline_multiplyToLen();
329 bool inline_squareToLen();
330 bool inline_mulAdd();
331 bool inline_montgomeryMultiply();
332 bool inline_montgomerySquare();
333
334 bool inline_profileBoolean();
335 };
336
337
338 //---------------------------make_vm_intrinsic----------------------------
339 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
340 vmIntrinsics::ID id = m->intrinsic_id();
341 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
342
343 ccstr disable_intr = NULL;
344
345 if ((DisableIntrinsic[0] != '\0'
346 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
347 (method_has_option_value("DisableIntrinsic", disable_intr)
348 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
349 // disabled by a user request on the command line:
350 // example: -XX:DisableIntrinsic=_hashCode,_getClass
351 return NULL;
352 }
353
354 if (!m->is_loaded()) {
355 // do not attempt to inline unloaded methods
356 return NULL;
357 }
358
359 // Only a few intrinsics implement a virtual dispatch.
360 // They are expensive calls which are also frequently overridden.
361 if (is_virtual) {
362 switch (id) {
363 case vmIntrinsics::_hashCode:
364 case vmIntrinsics::_clone:
365 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
366 break;
367 default:
368 return NULL;
369 }
370 }
371
372 // -XX:-InlineNatives disables nearly all intrinsics:
373 if (!InlineNatives) {
374 switch (id) {
375 case vmIntrinsics::_indexOf:
376 case vmIntrinsics::_compareTo:
377 case vmIntrinsics::_equals:
378 case vmIntrinsics::_equalsC:
379 case vmIntrinsics::_getAndAddInt:
380 case vmIntrinsics::_getAndAddLong:
381 case vmIntrinsics::_getAndSetInt:
382 case vmIntrinsics::_getAndSetLong:
383 case vmIntrinsics::_getAndSetObject:
384 case vmIntrinsics::_loadFence:
385 case vmIntrinsics::_storeFence:
386 case vmIntrinsics::_fullFence:
387 break; // InlineNatives does not control String.compareTo
388 case vmIntrinsics::_Reference_get:
389 break; // InlineNatives does not control Reference.get
390 default:
391 return NULL;
392 }
393 }
394
395 int predicates = 0;
396 bool does_virtual_dispatch = false;
397
398 switch (id) {
399 case vmIntrinsics::_compareTo:
400 if (!SpecialStringCompareTo) return NULL;
401 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
402 break;
403 case vmIntrinsics::_indexOf:
404 if (!SpecialStringIndexOf) return NULL;
405 break;
406 case vmIntrinsics::_equals:
407 if (!SpecialStringEquals) return NULL;
408 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
409 break;
410 case vmIntrinsics::_equalsC:
411 if (!SpecialArraysEquals) return NULL;
412 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
413 break;
414 case vmIntrinsics::_arraycopy:
415 if (!InlineArrayCopy) return NULL;
416 break;
417 case vmIntrinsics::_copyMemory:
418 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
419 if (!InlineArrayCopy) return NULL;
420 break;
421 case vmIntrinsics::_hashCode:
422 if (!InlineObjectHash) return NULL;
423 does_virtual_dispatch = true;
424 break;
425 case vmIntrinsics::_clone:
426 does_virtual_dispatch = true;
427 case vmIntrinsics::_copyOf:
428 case vmIntrinsics::_copyOfRange:
429 if (!InlineObjectCopy) return NULL;
430 // These also use the arraycopy intrinsic mechanism:
431 if (!InlineArrayCopy) return NULL;
432 break;
433 case vmIntrinsics::_encodeISOArray:
434 if (!SpecialEncodeISOArray) return NULL;
435 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
436 break;
437 case vmIntrinsics::_checkIndex:
438 // We do not intrinsify this. The optimizer does fine with it.
439 return NULL;
440
441 case vmIntrinsics::_getCallerClass:
442 if (!UseNewReflection) return NULL;
443 if (!InlineReflectionGetCallerClass) return NULL;
444 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
445 break;
446
447 case vmIntrinsics::_bitCount_i:
448 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
449 break;
450
451 case vmIntrinsics::_bitCount_l:
452 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
453 break;
454
455 case vmIntrinsics::_numberOfLeadingZeros_i:
456 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
457 break;
458
459 case vmIntrinsics::_numberOfLeadingZeros_l:
460 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
461 break;
462
463 case vmIntrinsics::_numberOfTrailingZeros_i:
464 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
465 break;
466
467 case vmIntrinsics::_numberOfTrailingZeros_l:
468 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
469 break;
470
471 case vmIntrinsics::_reverseBytes_c:
472 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
473 break;
474 case vmIntrinsics::_reverseBytes_s:
475 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
476 break;
477 case vmIntrinsics::_reverseBytes_i:
478 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
479 break;
480 case vmIntrinsics::_reverseBytes_l:
481 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
482 break;
483
484 case vmIntrinsics::_Reference_get:
485 // Use the intrinsic version of Reference.get() so that the value in
486 // the referent field can be registered by the G1 pre-barrier code.
487 // Also add memory barrier to prevent commoning reads from this field
488 // across safepoint since GC can change it value.
489 break;
490
491 case vmIntrinsics::_compareAndSwapObject:
492 #ifdef _LP64
493 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
494 #endif
495 break;
496
497 case vmIntrinsics::_compareAndSwapLong:
498 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
499 break;
500
501 case vmIntrinsics::_getAndAddInt:
502 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
503 break;
504
505 case vmIntrinsics::_getAndAddLong:
506 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
507 break;
508
509 case vmIntrinsics::_getAndSetInt:
510 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
511 break;
512
513 case vmIntrinsics::_getAndSetLong:
514 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
515 break;
516
517 case vmIntrinsics::_getAndSetObject:
518 #ifdef _LP64
519 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
520 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
521 break;
522 #else
523 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
524 break;
525 #endif
526
527 case vmIntrinsics::_aescrypt_encryptBlock:
528 case vmIntrinsics::_aescrypt_decryptBlock:
529 if (!UseAESIntrinsics) return NULL;
530 break;
531
532 case vmIntrinsics::_multiplyToLen:
533 if (!UseMultiplyToLenIntrinsic) return NULL;
534 break;
535
536 case vmIntrinsics::_squareToLen:
537 if (!UseSquareToLenIntrinsic) return NULL;
538 break;
539
540 case vmIntrinsics::_mulAdd:
541 if (!UseMulAddIntrinsic) return NULL;
542 break;
543
544 case vmIntrinsics::_montgomeryMultiply:
545 if (!UseMontgomeryMultiplyIntrinsic) return NULL;
546 break;
547 case vmIntrinsics::_montgomerySquare:
548 if (!UseMontgomerySquareIntrinsic) return NULL;
549 break;
550
551 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
552 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
553 if (!UseAESIntrinsics) return NULL;
554 // these two require the predicated logic
555 predicates = 1;
556 break;
557
558 case vmIntrinsics::_sha_implCompress:
559 if (!UseSHA1Intrinsics) return NULL;
560 break;
561
562 case vmIntrinsics::_sha2_implCompress:
563 if (!UseSHA256Intrinsics) return NULL;
564 break;
565
566 case vmIntrinsics::_sha5_implCompress:
567 if (!UseSHA512Intrinsics) return NULL;
568 break;
569
570 case vmIntrinsics::_digestBase_implCompressMB:
571 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
572 predicates = 3;
573 break;
574
575 case vmIntrinsics::_ghash_processBlocks:
576 if (!UseGHASHIntrinsics) return NULL;
577 break;
578
579 case vmIntrinsics::_updateCRC32:
580 case vmIntrinsics::_updateBytesCRC32:
581 case vmIntrinsics::_updateByteBufferCRC32:
582 if (!UseCRC32Intrinsics) return NULL;
583 break;
584
585 case vmIntrinsics::_incrementExactI:
586 case vmIntrinsics::_addExactI:
587 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
588 break;
589 case vmIntrinsics::_incrementExactL:
590 case vmIntrinsics::_addExactL:
591 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
592 break;
593 case vmIntrinsics::_decrementExactI:
594 case vmIntrinsics::_subtractExactI:
595 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
596 break;
597 case vmIntrinsics::_decrementExactL:
598 case vmIntrinsics::_subtractExactL:
599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
600 break;
601 case vmIntrinsics::_negateExactI:
602 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
603 break;
604 case vmIntrinsics::_negateExactL:
605 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
606 break;
607 case vmIntrinsics::_multiplyExactI:
608 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
609 break;
610 case vmIntrinsics::_multiplyExactL:
611 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
612 break;
613
614 default:
615 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
616 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
617 break;
618 }
619
620 // -XX:-InlineClassNatives disables natives from the Class class.
621 // The flag applies to all reflective calls, notably Array.newArray
622 // (visible to Java programmers as Array.newInstance).
623 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
624 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
625 if (!InlineClassNatives) return NULL;
626 }
627
628 // -XX:-InlineThreadNatives disables natives from the Thread class.
629 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
630 if (!InlineThreadNatives) return NULL;
631 }
632
633 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
634 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
635 m->holder()->name() == ciSymbol::java_lang_Float() ||
636 m->holder()->name() == ciSymbol::java_lang_Double()) {
637 if (!InlineMathNatives) return NULL;
638 }
639
640 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
641 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
642 if (!InlineUnsafeOps) return NULL;
643 }
644
645 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
646 }
647
648 //----------------------register_library_intrinsics-----------------------
649 // Initialize this file's data structures, for each Compile instance.
650 void Compile::register_library_intrinsics() {
651 // Nothing to do here.
652 }
653
654 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
655 LibraryCallKit kit(jvms, this);
656 Compile* C = kit.C;
657 int nodes = C->unique();
658 #ifndef PRODUCT
659 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
660 char buf[1000];
661 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
662 tty->print_cr("Intrinsic %s", str);
663 }
664 #endif
665 ciMethod* callee = kit.callee();
666 const int bci = kit.bci();
667
668 // Try to inline the intrinsic.
669 if (kit.try_to_inline(_last_predicate)) {
670 if (C->print_intrinsics() || C->print_inlining()) {
671 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
672 }
673 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
674 if (C->log()) {
675 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
676 vmIntrinsics::name_at(intrinsic_id()),
677 (is_virtual() ? " virtual='1'" : ""),
678 C->unique() - nodes);
679 }
680 // Push the result from the inlined method onto the stack.
681 kit.push_result();
682 return kit.transfer_exceptions_into_jvms();
683 }
684
685 // The intrinsic bailed out
686 if (C->print_intrinsics() || C->print_inlining()) {
687 if (jvms->has_method()) {
688 // Not a root compile.
689 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
690 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
691 } else {
692 // Root compile
693 tty->print("Did not generate intrinsic %s%s at bci:%d in",
694 vmIntrinsics::name_at(intrinsic_id()),
695 (is_virtual() ? " (virtual)" : ""), bci);
696 }
697 }
698 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
699 return NULL;
700 }
701
702 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
703 LibraryCallKit kit(jvms, this);
704 Compile* C = kit.C;
705 int nodes = C->unique();
706 _last_predicate = predicate;
707 #ifndef PRODUCT
708 assert(is_predicated() && predicate < predicates_count(), "sanity");
709 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
710 char buf[1000];
711 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
712 tty->print_cr("Predicate for intrinsic %s", str);
713 }
714 #endif
715 ciMethod* callee = kit.callee();
716 const int bci = kit.bci();
717
718 Node* slow_ctl = kit.try_to_predicate(predicate);
719 if (!kit.failing()) {
720 if (C->print_intrinsics() || C->print_inlining()) {
721 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
722 }
723 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
724 if (C->log()) {
725 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
726 vmIntrinsics::name_at(intrinsic_id()),
727 (is_virtual() ? " virtual='1'" : ""),
728 C->unique() - nodes);
729 }
730 return slow_ctl; // Could be NULL if the check folds.
731 }
732
733 // The intrinsic bailed out
734 if (C->print_intrinsics() || C->print_inlining()) {
735 if (jvms->has_method()) {
736 // Not a root compile.
737 const char* msg = "failed to generate predicate for intrinsic";
738 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
739 } else {
740 // Root compile
741 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
742 vmIntrinsics::name_at(intrinsic_id()),
743 (is_virtual() ? " (virtual)" : ""), bci);
744 }
745 }
746 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
747 return NULL;
748 }
749
750 bool LibraryCallKit::try_to_inline(int predicate) {
751 // Handle symbolic names for otherwise undistinguished boolean switches:
752 const bool is_store = true;
753 const bool is_native_ptr = true;
754 const bool is_static = true;
755 const bool is_volatile = true;
756
757 if (!jvms()->has_method()) {
758 // Root JVMState has a null method.
759 assert(map()->memory()->Opcode() == Op_Parm, "");
760 // Insert the memory aliasing node
761 set_all_memory(reset_memory());
762 }
763 assert(merged_memory(), "");
764
765
766 switch (intrinsic_id()) {
767 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
768 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
769 case vmIntrinsics::_getClass: return inline_native_getClass();
770
771 case vmIntrinsics::_dsin:
772 case vmIntrinsics::_dcos:
773 case vmIntrinsics::_dtan:
774 case vmIntrinsics::_dabs:
775 case vmIntrinsics::_datan2:
776 case vmIntrinsics::_dsqrt:
777 case vmIntrinsics::_dexp:
778 case vmIntrinsics::_dlog:
779 case vmIntrinsics::_dlog10:
780 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
781
782 case vmIntrinsics::_min:
783 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
784
785 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
786 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
787 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
788 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
789 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
790 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
791 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
792 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
793 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
794 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
795 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
796 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
797
798 case vmIntrinsics::_arraycopy: return inline_arraycopy();
799
800 case vmIntrinsics::_compareTo: return inline_string_compareTo();
801 case vmIntrinsics::_indexOf: return inline_string_indexOf();
802 case vmIntrinsics::_equals: return inline_string_equals();
803
804 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
805 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
806 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
807 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
808 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
809 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
810 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
811 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
812 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
813
814 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
815 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
816 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
817 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
818 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
819 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
820 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
821 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
822 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
823
824 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
825 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
826 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
827 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
828 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
829 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
830 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
831 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
832
833 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
834 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
835 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
836 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
837 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
838 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
839 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
840 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
841
842 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
843 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
844 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
845 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
846 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
847 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
848 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
849 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
850 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
851
852 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
853 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
854 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
855 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
856 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
857 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
858 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
859 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
860 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
861
862 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
863 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
864 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
865 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
866
867 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
868 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
869 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
870
871 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
872 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
873 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
874
875 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
876 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
877 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
878 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
879 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
880
881 case vmIntrinsics::_loadFence:
882 case vmIntrinsics::_storeFence:
883 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
884
885 case vmIntrinsics::_currentThread: return inline_native_currentThread();
886 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
887
888 #ifdef JFR_HAVE_INTRINSICS
889 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
890 case vmIntrinsics::_getClassId: return inline_native_classID();
891 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
892 #endif
893 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
894 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
895 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
896 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
897 case vmIntrinsics::_newArray: return inline_native_newArray();
898 case vmIntrinsics::_getLength: return inline_native_getLength();
899 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
900 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
901 case vmIntrinsics::_equalsC: return inline_array_equals();
902 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
903
904 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
905
906 case vmIntrinsics::_isInstance:
907 case vmIntrinsics::_getModifiers:
908 case vmIntrinsics::_isInterface:
909 case vmIntrinsics::_isArray:
910 case vmIntrinsics::_isPrimitive:
911 case vmIntrinsics::_getSuperclass:
912 case vmIntrinsics::_getComponentType:
913 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
914
915 case vmIntrinsics::_floatToRawIntBits:
916 case vmIntrinsics::_floatToIntBits:
917 case vmIntrinsics::_intBitsToFloat:
918 case vmIntrinsics::_doubleToRawLongBits:
919 case vmIntrinsics::_doubleToLongBits:
920 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
921
922 case vmIntrinsics::_numberOfLeadingZeros_i:
923 case vmIntrinsics::_numberOfLeadingZeros_l:
924 case vmIntrinsics::_numberOfTrailingZeros_i:
925 case vmIntrinsics::_numberOfTrailingZeros_l:
926 case vmIntrinsics::_bitCount_i:
927 case vmIntrinsics::_bitCount_l:
928 case vmIntrinsics::_reverseBytes_i:
929 case vmIntrinsics::_reverseBytes_l:
930 case vmIntrinsics::_reverseBytes_s:
931 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
932
933 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
934
935 case vmIntrinsics::_Reference_get: return inline_reference_get();
936
937 case vmIntrinsics::_aescrypt_encryptBlock:
938 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
939
940 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
941 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
942 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
943
944 case vmIntrinsics::_sha_implCompress:
945 case vmIntrinsics::_sha2_implCompress:
946 case vmIntrinsics::_sha5_implCompress:
947 return inline_sha_implCompress(intrinsic_id());
948
949 case vmIntrinsics::_digestBase_implCompressMB:
950 return inline_digestBase_implCompressMB(predicate);
951
952 case vmIntrinsics::_multiplyToLen:
953 return inline_multiplyToLen();
954
955 case vmIntrinsics::_squareToLen:
956 return inline_squareToLen();
957
958 case vmIntrinsics::_mulAdd:
959 return inline_mulAdd();
960
961 case vmIntrinsics::_montgomeryMultiply:
962 return inline_montgomeryMultiply();
963 case vmIntrinsics::_montgomerySquare:
964 return inline_montgomerySquare();
965
966 case vmIntrinsics::_ghash_processBlocks:
967 return inline_ghash_processBlocks();
968
969 case vmIntrinsics::_encodeISOArray:
970 return inline_encodeISOArray();
971
972 case vmIntrinsics::_updateCRC32:
973 return inline_updateCRC32();
974 case vmIntrinsics::_updateBytesCRC32:
975 return inline_updateBytesCRC32();
976 case vmIntrinsics::_updateByteBufferCRC32:
977 return inline_updateByteBufferCRC32();
978
979 case vmIntrinsics::_profileBoolean:
980 return inline_profileBoolean();
981
982 default:
983 // If you get here, it may be that someone has added a new intrinsic
984 // to the list in vmSymbols.hpp without implementing it here.
985 #ifndef PRODUCT
986 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
987 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
988 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
989 }
990 #endif
991 return false;
992 }
993 }
994
995 Node* LibraryCallKit::try_to_predicate(int predicate) {
996 if (!jvms()->has_method()) {
997 // Root JVMState has a null method.
998 assert(map()->memory()->Opcode() == Op_Parm, "");
999 // Insert the memory aliasing node
1000 set_all_memory(reset_memory());
1001 }
1002 assert(merged_memory(), "");
1003
1004 switch (intrinsic_id()) {
1005 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1006 return inline_cipherBlockChaining_AESCrypt_predicate(false);
1007 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1008 return inline_cipherBlockChaining_AESCrypt_predicate(true);
1009 case vmIntrinsics::_digestBase_implCompressMB:
1010 return inline_digestBase_implCompressMB_predicate(predicate);
1011
1012 default:
1013 // If you get here, it may be that someone has added a new intrinsic
1014 // to the list in vmSymbols.hpp without implementing it here.
1015 #ifndef PRODUCT
1016 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1017 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1018 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1019 }
1020 #endif
1021 Node* slow_ctl = control();
1022 set_control(top()); // No fast path instrinsic
1023 return slow_ctl;
1024 }
1025 }
1026
1027 //------------------------------set_result-------------------------------
1028 // Helper function for finishing intrinsics.
1029 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1030 record_for_igvn(region);
1031 set_control(_gvn.transform(region));
1032 set_result( _gvn.transform(value));
1033 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1034 }
1035
1036 //------------------------------generate_guard---------------------------
1037 // Helper function for generating guarded fast-slow graph structures.
1038 // The given 'test', if true, guards a slow path. If the test fails
1039 // then a fast path can be taken. (We generally hope it fails.)
1040 // In all cases, GraphKit::control() is updated to the fast path.
1041 // The returned value represents the control for the slow path.
1042 // The return value is never 'top'; it is either a valid control
1043 // or NULL if it is obvious that the slow path can never be taken.
1044 // Also, if region and the slow control are not NULL, the slow edge
1045 // is appended to the region.
1046 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1047 if (stopped()) {
1048 // Already short circuited.
1049 return NULL;
1050 }
1051
1052 // Build an if node and its projections.
1053 // If test is true we take the slow path, which we assume is uncommon.
1054 if (_gvn.type(test) == TypeInt::ZERO) {
1055 // The slow branch is never taken. No need to build this guard.
1056 return NULL;
1057 }
1058
1059 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1060
1061 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1062 if (if_slow == top()) {
1063 // The slow branch is never taken. No need to build this guard.
1064 return NULL;
1065 }
1066
1067 if (region != NULL)
1068 region->add_req(if_slow);
1069
1070 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1071 set_control(if_fast);
1072
1073 return if_slow;
1074 }
1075
1076 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1077 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1078 }
1079 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1080 return generate_guard(test, region, PROB_FAIR);
1081 }
1082
1083 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1084 Node* *pos_index) {
1085 if (stopped())
1086 return NULL; // already stopped
1087 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1088 return NULL; // index is already adequately typed
1089 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1090 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1091 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1092 if (is_neg != NULL && pos_index != NULL) {
1093 // Emulate effect of Parse::adjust_map_after_if.
1094 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1095 ccast->set_req(0, control());
1096 (*pos_index) = _gvn.transform(ccast);
1097 }
1098 return is_neg;
1099 }
1100
1101 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1102 Node* *pos_index) {
1103 if (stopped())
1104 return NULL; // already stopped
1105 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1106 return NULL; // index is already adequately typed
1107 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1108 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1109 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1110 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1111 if (is_notp != NULL && pos_index != NULL) {
1112 // Emulate effect of Parse::adjust_map_after_if.
1113 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1114 ccast->set_req(0, control());
1115 (*pos_index) = _gvn.transform(ccast);
1116 }
1117 return is_notp;
1118 }
1119
1120 // Make sure that 'position' is a valid limit index, in [0..length].
1121 // There are two equivalent plans for checking this:
1122 // A. (offset + copyLength) unsigned<= arrayLength
1123 // B. offset <= (arrayLength - copyLength)
1124 // We require that all of the values above, except for the sum and
1125 // difference, are already known to be non-negative.
1126 // Plan A is robust in the face of overflow, if offset and copyLength
1127 // are both hugely positive.
1128 //
1129 // Plan B is less direct and intuitive, but it does not overflow at
1130 // all, since the difference of two non-negatives is always
1131 // representable. Whenever Java methods must perform the equivalent
1132 // check they generally use Plan B instead of Plan A.
1133 // For the moment we use Plan A.
1134 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1135 Node* subseq_length,
1136 Node* array_length,
1137 RegionNode* region) {
1138 if (stopped())
1139 return NULL; // already stopped
1140 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1141 if (zero_offset && subseq_length->eqv_uncast(array_length))
1142 return NULL; // common case of whole-array copy
1143 Node* last = subseq_length;
1144 if (!zero_offset) // last += offset
1145 last = _gvn.transform(new (C) AddINode(last, offset));
1146 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1147 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1148 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1149 return is_over;
1150 }
1151
1152
1153 //--------------------------generate_current_thread--------------------
1154 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1155 ciKlass* thread_klass = env()->Thread_klass();
1156 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1157 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1158 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1159 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1160 tls_output = thread;
1161 return threadObj;
1162 }
1163
1164
1165 //------------------------------make_string_method_node------------------------
1166 // Helper method for String intrinsic functions. This version is called
1167 // with str1 and str2 pointing to String object nodes.
1168 //
1169 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1170 Node* no_ctrl = NULL;
1171
1172 // Get start addr of string
1173 Node* str1_value = load_String_value(no_ctrl, str1);
1174 Node* str1_offset = load_String_offset(no_ctrl, str1);
1175 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1176
1177 // Get length of string 1
1178 Node* str1_len = load_String_length(no_ctrl, str1);
1179
1180 Node* str2_value = load_String_value(no_ctrl, str2);
1181 Node* str2_offset = load_String_offset(no_ctrl, str2);
1182 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1183
1184 Node* str2_len = NULL;
1185 Node* result = NULL;
1186
1187 switch (opcode) {
1188 case Op_StrIndexOf:
1189 // Get length of string 2
1190 str2_len = load_String_length(no_ctrl, str2);
1191
1192 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1193 str1_start, str1_len, str2_start, str2_len);
1194 break;
1195 case Op_StrComp:
1196 // Get length of string 2
1197 str2_len = load_String_length(no_ctrl, str2);
1198
1199 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1200 str1_start, str1_len, str2_start, str2_len);
1201 break;
1202 case Op_StrEquals:
1203 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1204 str1_start, str2_start, str1_len);
1205 break;
1206 default:
1207 ShouldNotReachHere();
1208 return NULL;
1209 }
1210
1211 // All these intrinsics have checks.
1212 C->set_has_split_ifs(true); // Has chance for split-if optimization
1213
1214 return _gvn.transform(result);
1215 }
1216
1217 // Helper method for String intrinsic functions. This version is called
1218 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1219 // to Int nodes containing the lenghts of str1 and str2.
1220 //
1221 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1222 Node* result = NULL;
1223 switch (opcode) {
1224 case Op_StrIndexOf:
1225 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1226 str1_start, cnt1, str2_start, cnt2);
1227 break;
1228 case Op_StrComp:
1229 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1230 str1_start, cnt1, str2_start, cnt2);
1231 break;
1232 case Op_StrEquals:
1233 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1234 str1_start, str2_start, cnt1);
1235 break;
1236 default:
1237 ShouldNotReachHere();
1238 return NULL;
1239 }
1240
1241 // All these intrinsics have checks.
1242 C->set_has_split_ifs(true); // Has chance for split-if optimization
1243
1244 return _gvn.transform(result);
1245 }
1246
1247 //------------------------------inline_string_compareTo------------------------
1248 // public int java.lang.String.compareTo(String anotherString);
1249 bool LibraryCallKit::inline_string_compareTo() {
1250 Node* receiver = null_check(argument(0));
1251 Node* arg = null_check(argument(1));
1252 if (stopped()) {
1253 return true;
1254 }
1255 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1256 return true;
1257 }
1258
1259 //------------------------------inline_string_equals------------------------
1260 bool LibraryCallKit::inline_string_equals() {
1261 Node* receiver = null_check_receiver();
1262 // NOTE: Do not null check argument for String.equals() because spec
1263 // allows to specify NULL as argument.
1264 Node* argument = this->argument(1);
1265 if (stopped()) {
1266 return true;
1267 }
1268
1269 // paths (plus control) merge
1270 RegionNode* region = new (C) RegionNode(5);
1271 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1272
1273 // does source == target string?
1274 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1275 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1276
1277 Node* if_eq = generate_slow_guard(bol, NULL);
1278 if (if_eq != NULL) {
1279 // receiver == argument
1280 phi->init_req(2, intcon(1));
1281 region->init_req(2, if_eq);
1282 }
1283
1284 // get String klass for instanceOf
1285 ciInstanceKlass* klass = env()->String_klass();
1286
1287 if (!stopped()) {
1288 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1289 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1290 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1291
1292 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1293 //instanceOf == true, fallthrough
1294
1295 if (inst_false != NULL) {
1296 phi->init_req(3, intcon(0));
1297 region->init_req(3, inst_false);
1298 }
1299 }
1300
1301 if (!stopped()) {
1302 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1303
1304 // Properly cast the argument to String
1305 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1306 // This path is taken only when argument's type is String:NotNull.
1307 argument = cast_not_null(argument, false);
1308
1309 Node* no_ctrl = NULL;
1310
1311 // Get start addr of receiver
1312 Node* receiver_val = load_String_value(no_ctrl, receiver);
1313 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1314 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1315
1316 // Get length of receiver
1317 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1318
1319 // Get start addr of argument
1320 Node* argument_val = load_String_value(no_ctrl, argument);
1321 Node* argument_offset = load_String_offset(no_ctrl, argument);
1322 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1323
1324 // Get length of argument
1325 Node* argument_cnt = load_String_length(no_ctrl, argument);
1326
1327 // Check for receiver count != argument count
1328 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1329 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1330 Node* if_ne = generate_slow_guard(bol, NULL);
1331 if (if_ne != NULL) {
1332 phi->init_req(4, intcon(0));
1333 region->init_req(4, if_ne);
1334 }
1335
1336 // Check for count == 0 is done by assembler code for StrEquals.
1337
1338 if (!stopped()) {
1339 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1340 phi->init_req(1, equals);
1341 region->init_req(1, control());
1342 }
1343 }
1344
1345 // post merge
1346 set_control(_gvn.transform(region));
1347 record_for_igvn(region);
1348
1349 set_result(_gvn.transform(phi));
1350 return true;
1351 }
1352
1353 //------------------------------inline_array_equals----------------------------
1354 bool LibraryCallKit::inline_array_equals() {
1355 Node* arg1 = argument(0);
1356 Node* arg2 = argument(1);
1357 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1358 return true;
1359 }
1360
1361 // Java version of String.indexOf(constant string)
1362 // class StringDecl {
1363 // StringDecl(char[] ca) {
1364 // offset = 0;
1365 // count = ca.length;
1366 // value = ca;
1367 // }
1368 // int offset;
1369 // int count;
1370 // char[] value;
1371 // }
1372 //
1373 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1374 // int targetOffset, int cache_i, int md2) {
1375 // int cache = cache_i;
1376 // int sourceOffset = string_object.offset;
1377 // int sourceCount = string_object.count;
1378 // int targetCount = target_object.length;
1379 //
1380 // int targetCountLess1 = targetCount - 1;
1381 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1382 //
1383 // char[] source = string_object.value;
1384 // char[] target = target_object;
1385 // int lastChar = target[targetCountLess1];
1386 //
1387 // outer_loop:
1388 // for (int i = sourceOffset; i < sourceEnd; ) {
1389 // int src = source[i + targetCountLess1];
1390 // if (src == lastChar) {
1391 // // With random strings and a 4-character alphabet,
1392 // // reverse matching at this point sets up 0.8% fewer
1393 // // frames, but (paradoxically) makes 0.3% more probes.
1394 // // Since those probes are nearer the lastChar probe,
1395 // // there is may be a net D$ win with reverse matching.
1396 // // But, reversing loop inhibits unroll of inner loop
1397 // // for unknown reason. So, does running outer loop from
1398 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1399 // for (int j = 0; j < targetCountLess1; j++) {
1400 // if (target[targetOffset + j] != source[i+j]) {
1401 // if ((cache & (1 << source[i+j])) == 0) {
1402 // if (md2 < j+1) {
1403 // i += j+1;
1404 // continue outer_loop;
1405 // }
1406 // }
1407 // i += md2;
1408 // continue outer_loop;
1409 // }
1410 // }
1411 // return i - sourceOffset;
1412 // }
1413 // if ((cache & (1 << src)) == 0) {
1414 // i += targetCountLess1;
1415 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1416 // i++;
1417 // }
1418 // return -1;
1419 // }
1420
1421 //------------------------------string_indexOf------------------------
1422 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1423 jint cache_i, jint md2_i) {
1424
1425 Node* no_ctrl = NULL;
1426 float likely = PROB_LIKELY(0.9);
1427 float unlikely = PROB_UNLIKELY(0.9);
1428
1429 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1430
1431 Node* source = load_String_value(no_ctrl, string_object);
1432 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1433 Node* sourceCount = load_String_length(no_ctrl, string_object);
1434
1435 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1436 jint target_length = target_array->length();
1437 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1438 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1439
1440 // String.value field is known to be @Stable.
1441 if (UseImplicitStableValues) {
1442 target = cast_array_to_stable(target, target_type);
1443 }
1444
1445 IdealKit kit(this, false, true);
1446 #define __ kit.
1447 Node* zero = __ ConI(0);
1448 Node* one = __ ConI(1);
1449 Node* cache = __ ConI(cache_i);
1450 Node* md2 = __ ConI(md2_i);
1451 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1452 Node* targetCount = __ ConI(target_length);
1453 Node* targetCountLess1 = __ ConI(target_length - 1);
1454 Node* targetOffset = __ ConI(targetOffset_i);
1455 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1456
1457 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1458 Node* outer_loop = __ make_label(2 /* goto */);
1459 Node* return_ = __ make_label(1);
1460
1461 __ set(rtn,__ ConI(-1));
1462 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1463 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1464 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1465 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1466 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1467 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1468 Node* tpj = __ AddI(targetOffset, __ value(j));
1469 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1470 Node* ipj = __ AddI(__ value(i), __ value(j));
1471 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1472 __ if_then(targ, BoolTest::ne, src2); {
1473 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1474 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1475 __ increment(i, __ AddI(__ value(j), one));
1476 __ goto_(outer_loop);
1477 } __ end_if(); __ dead(j);
1478 }__ end_if(); __ dead(j);
1479 __ increment(i, md2);
1480 __ goto_(outer_loop);
1481 }__ end_if();
1482 __ increment(j, one);
1483 }__ end_loop(); __ dead(j);
1484 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1485 __ goto_(return_);
1486 }__ end_if();
1487 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1488 __ increment(i, targetCountLess1);
1489 }__ end_if();
1490 __ increment(i, one);
1491 __ bind(outer_loop);
1492 }__ end_loop(); __ dead(i);
1493 __ bind(return_);
1494
1495 // Final sync IdealKit and GraphKit.
1496 final_sync(kit);
1497 Node* result = __ value(rtn);
1498 #undef __
1499 C->set_has_loops(true);
1500 return result;
1501 }
1502
1503 //------------------------------inline_string_indexOf------------------------
1504 bool LibraryCallKit::inline_string_indexOf() {
1505 Node* receiver = argument(0);
1506 Node* arg = argument(1);
1507
1508 Node* result;
1509 // Disable the use of pcmpestri until it can be guaranteed that
1510 // the load doesn't cross into the uncommited space.
1511 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1512 UseSSE42Intrinsics) {
1513 // Generate SSE4.2 version of indexOf
1514 // We currently only have match rules that use SSE4.2
1515
1516 receiver = null_check(receiver);
1517 arg = null_check(arg);
1518 if (stopped()) {
1519 return true;
1520 }
1521
1522 ciInstanceKlass* str_klass = env()->String_klass();
1523 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1524
1525 // Make the merge point
1526 RegionNode* result_rgn = new (C) RegionNode(4);
1527 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1528 Node* no_ctrl = NULL;
1529
1530 // Get start addr of source string
1531 Node* source = load_String_value(no_ctrl, receiver);
1532 Node* source_offset = load_String_offset(no_ctrl, receiver);
1533 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1534
1535 // Get length of source string
1536 Node* source_cnt = load_String_length(no_ctrl, receiver);
1537
1538 // Get start addr of substring
1539 Node* substr = load_String_value(no_ctrl, arg);
1540 Node* substr_offset = load_String_offset(no_ctrl, arg);
1541 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1542
1543 // Get length of source string
1544 Node* substr_cnt = load_String_length(no_ctrl, arg);
1545
1546 // Check for substr count > string count
1547 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1548 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1549 Node* if_gt = generate_slow_guard(bol, NULL);
1550 if (if_gt != NULL) {
1551 result_phi->init_req(2, intcon(-1));
1552 result_rgn->init_req(2, if_gt);
1553 }
1554
1555 if (!stopped()) {
1556 // Check for substr count == 0
1557 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1558 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1559 Node* if_zero = generate_slow_guard(bol, NULL);
1560 if (if_zero != NULL) {
1561 result_phi->init_req(3, intcon(0));
1562 result_rgn->init_req(3, if_zero);
1563 }
1564 }
1565
1566 if (!stopped()) {
1567 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1568 result_phi->init_req(1, result);
1569 result_rgn->init_req(1, control());
1570 }
1571 set_control(_gvn.transform(result_rgn));
1572 record_for_igvn(result_rgn);
1573 result = _gvn.transform(result_phi);
1574
1575 } else { // Use LibraryCallKit::string_indexOf
1576 // don't intrinsify if argument isn't a constant string.
1577 if (!arg->is_Con()) {
1578 return false;
1579 }
1580 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1581 if (str_type == NULL) {
1582 return false;
1583 }
1584 ciInstanceKlass* klass = env()->String_klass();
1585 ciObject* str_const = str_type->const_oop();
1586 if (str_const == NULL || str_const->klass() != klass) {
1587 return false;
1588 }
1589 ciInstance* str = str_const->as_instance();
1590 assert(str != NULL, "must be instance");
1591
1592 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1593 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1594
1595 int o;
1596 int c;
1597 if (java_lang_String::has_offset_field()) {
1598 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1599 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1600 } else {
1601 o = 0;
1602 c = pat->length();
1603 }
1604
1605 // constant strings have no offset and count == length which
1606 // simplifies the resulting code somewhat so lets optimize for that.
1607 if (o != 0 || c != pat->length()) {
1608 return false;
1609 }
1610
1611 receiver = null_check(receiver, T_OBJECT);
1612 // NOTE: No null check on the argument is needed since it's a constant String oop.
1613 if (stopped()) {
1614 return true;
1615 }
1616
1617 // The null string as a pattern always returns 0 (match at beginning of string)
1618 if (c == 0) {
1619 set_result(intcon(0));
1620 return true;
1621 }
1622
1623 // Generate default indexOf
1624 jchar lastChar = pat->char_at(o + (c - 1));
1625 int cache = 0;
1626 int i;
1627 for (i = 0; i < c - 1; i++) {
1628 assert(i < pat->length(), "out of range");
1629 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1630 }
1631
1632 int md2 = c;
1633 for (i = 0; i < c - 1; i++) {
1634 assert(i < pat->length(), "out of range");
1635 if (pat->char_at(o + i) == lastChar) {
1636 md2 = (c - 1) - i;
1637 }
1638 }
1639
1640 result = string_indexOf(receiver, pat, o, cache, md2);
1641 }
1642 set_result(result);
1643 return true;
1644 }
1645
1646 //--------------------------round_double_node--------------------------------
1647 // Round a double node if necessary.
1648 Node* LibraryCallKit::round_double_node(Node* n) {
1649 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1650 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1651 return n;
1652 }
1653
1654 //------------------------------inline_math-----------------------------------
1655 // public static double Math.abs(double)
1656 // public static double Math.sqrt(double)
1657 // public static double Math.log(double)
1658 // public static double Math.log10(double)
1659 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1660 Node* arg = round_double_node(argument(0));
1661 Node* n = NULL;
1662 switch (id) {
1663 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1664 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1665 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1666 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1667 default: fatal_unexpected_iid(id); break;
1668 }
1669 set_result(_gvn.transform(n));
1670 return true;
1671 }
1672
1673 //------------------------------inline_trig----------------------------------
1674 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1675 // argument reduction which will turn into a fast/slow diamond.
1676 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1677 Node* arg = round_double_node(argument(0));
1678 Node* n = NULL;
1679
1680 switch (id) {
1681 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1682 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1683 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1684 default: fatal_unexpected_iid(id); break;
1685 }
1686 n = _gvn.transform(n);
1687
1688 // Rounding required? Check for argument reduction!
1689 if (Matcher::strict_fp_requires_explicit_rounding) {
1690 static const double pi_4 = 0.7853981633974483;
1691 static const double neg_pi_4 = -0.7853981633974483;
1692 // pi/2 in 80-bit extended precision
1693 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1694 // -pi/2 in 80-bit extended precision
1695 // 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};
1696 // Cutoff value for using this argument reduction technique
1697 //static const double pi_2_minus_epsilon = 1.564660403643354;
1698 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1699
1700 // Pseudocode for sin:
1701 // if (x <= Math.PI / 4.0) {
1702 // if (x >= -Math.PI / 4.0) return fsin(x);
1703 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1704 // } else {
1705 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1706 // }
1707 // return StrictMath.sin(x);
1708
1709 // Pseudocode for cos:
1710 // if (x <= Math.PI / 4.0) {
1711 // if (x >= -Math.PI / 4.0) return fcos(x);
1712 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1713 // } else {
1714 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1715 // }
1716 // return StrictMath.cos(x);
1717
1718 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1719 // requires a special machine instruction to load it. Instead we'll try
1720 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1721 // probably do the math inside the SIN encoding.
1722
1723 // Make the merge point
1724 RegionNode* r = new (C) RegionNode(3);
1725 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1726
1727 // Flatten arg so we need only 1 test
1728 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1729 // Node for PI/4 constant
1730 Node *pi4 = makecon(TypeD::make(pi_4));
1731 // Check PI/4 : abs(arg)
1732 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1733 // Check: If PI/4 < abs(arg) then go slow
1734 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1735 // Branch either way
1736 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1737 set_control(opt_iff(r,iff));
1738
1739 // Set fast path result
1740 phi->init_req(2, n);
1741
1742 // Slow path - non-blocking leaf call
1743 Node* call = NULL;
1744 switch (id) {
1745 case vmIntrinsics::_dsin:
1746 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1747 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1748 "Sin", NULL, arg, top());
1749 break;
1750 case vmIntrinsics::_dcos:
1751 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1752 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1753 "Cos", NULL, arg, top());
1754 break;
1755 case vmIntrinsics::_dtan:
1756 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1757 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1758 "Tan", NULL, arg, top());
1759 break;
1760 }
1761 assert(control()->in(0) == call, "");
1762 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1763 r->init_req(1, control());
1764 phi->init_req(1, slow_result);
1765
1766 // Post-merge
1767 set_control(_gvn.transform(r));
1768 record_for_igvn(r);
1769 n = _gvn.transform(phi);
1770
1771 C->set_has_split_ifs(true); // Has chance for split-if optimization
1772 }
1773 set_result(n);
1774 return true;
1775 }
1776
1777 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1778 //-------------------
1779 //result=(result.isNaN())? funcAddr():result;
1780 // Check: If isNaN() by checking result!=result? then either trap
1781 // or go to runtime
1782 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1783 // Build the boolean node
1784 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1785
1786 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1787 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1788 // The pow or exp intrinsic returned a NaN, which requires a call
1789 // to the runtime. Recompile with the runtime call.
1790 uncommon_trap(Deoptimization::Reason_intrinsic,
1791 Deoptimization::Action_make_not_entrant);
1792 }
1793 return result;
1794 } else {
1795 // If this inlining ever returned NaN in the past, we compile a call
1796 // to the runtime to properly handle corner cases
1797
1798 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1799 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1800 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1801
1802 if (!if_slow->is_top()) {
1803 RegionNode* result_region = new (C) RegionNode(3);
1804 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1805
1806 result_region->init_req(1, if_fast);
1807 result_val->init_req(1, result);
1808
1809 set_control(if_slow);
1810
1811 const TypePtr* no_memory_effects = NULL;
1812 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1813 no_memory_effects,
1814 x, top(), y, y ? top() : NULL);
1815 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1816 #ifdef ASSERT
1817 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1818 assert(value_top == top(), "second value must be top");
1819 #endif
1820
1821 result_region->init_req(2, control());
1822 result_val->init_req(2, value);
1823 set_control(_gvn.transform(result_region));
1824 return _gvn.transform(result_val);
1825 } else {
1826 return result;
1827 }
1828 }
1829 }
1830
1831 //------------------------------inline_exp-------------------------------------
1832 // Inline exp instructions, if possible. The Intel hardware only misses
1833 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1834 bool LibraryCallKit::inline_exp() {
1835 Node* arg = round_double_node(argument(0));
1836 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1837
1838 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1839 set_result(n);
1840
1841 C->set_has_split_ifs(true); // Has chance for split-if optimization
1842 return true;
1843 }
1844
1845 //------------------------------inline_pow-------------------------------------
1846 // Inline power instructions, if possible.
1847 bool LibraryCallKit::inline_pow() {
1848 // Pseudocode for pow
1849 // if (y == 2) {
1850 // return x * x;
1851 // } else {
1852 // if (x <= 0.0) {
1853 // long longy = (long)y;
1854 // if ((double)longy == y) { // if y is long
1855 // if (y + 1 == y) longy = 0; // huge number: even
1856 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1857 // } else {
1858 // result = NaN;
1859 // }
1860 // } else {
1861 // result = DPow(x,y);
1862 // }
1863 // if (result != result)? {
1864 // result = uncommon_trap() or runtime_call();
1865 // }
1866 // return result;
1867 // }
1868
1869 Node* x = round_double_node(argument(0));
1870 Node* y = round_double_node(argument(2));
1871
1872 Node* result = NULL;
1873
1874 Node* const_two_node = makecon(TypeD::make(2.0));
1875 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1876 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1877 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1878 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1879 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1880
1881 RegionNode* region_node = new (C) RegionNode(3);
1882 region_node->init_req(1, if_true);
1883
1884 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1885 // special case for x^y where y == 2, we can convert it to x * x
1886 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1887
1888 // set control to if_false since we will now process the false branch
1889 set_control(if_false);
1890
1891 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1892 // Short form: skip the fancy tests and just check for NaN result.
1893 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1894 } else {
1895 // If this inlining ever returned NaN in the past, include all
1896 // checks + call to the runtime.
1897
1898 // Set the merge point for If node with condition of (x <= 0.0)
1899 // There are four possible paths to region node and phi node
1900 RegionNode *r = new (C) RegionNode(4);
1901 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1902
1903 // Build the first if node: if (x <= 0.0)
1904 // Node for 0 constant
1905 Node *zeronode = makecon(TypeD::ZERO);
1906 // Check x:0
1907 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1908 // Check: If (x<=0) then go complex path
1909 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1910 // Branch either way
1911 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1912 // Fast path taken; set region slot 3
1913 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1914 r->init_req(3,fast_taken); // Capture fast-control
1915
1916 // Fast path not-taken, i.e. slow path
1917 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1918
1919 // Set fast path result
1920 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1921 phi->init_req(3, fast_result);
1922
1923 // Complex path
1924 // Build the second if node (if y is long)
1925 // Node for (long)y
1926 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1927 // Node for (double)((long) y)
1928 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1929 // Check (double)((long) y) : y
1930 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1931 // Check if (y isn't long) then go to slow path
1932
1933 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1934 // Branch either way
1935 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1936 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1937
1938 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1939
1940 // Calculate DPow(abs(x), y)*(1 & (long)y)
1941 // Node for constant 1
1942 Node *conone = longcon(1);
1943 // 1& (long)y
1944 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1945
1946 // A huge number is always even. Detect a huge number by checking
1947 // if y + 1 == y and set integer to be tested for parity to 0.
1948 // Required for corner case:
1949 // (long)9.223372036854776E18 = max_jlong
1950 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1951 // max_jlong is odd but 9.223372036854776E18 is even
1952 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1953 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1954 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1955 Node* correctedsign = NULL;
1956 if (ConditionalMoveLimit != 0) {
1957 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1958 } else {
1959 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1960 RegionNode *r = new (C) RegionNode(3);
1961 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1962 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1963 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1964 phi->init_req(1, signnode);
1965 phi->init_req(2, longcon(0));
1966 correctedsign = _gvn.transform(phi);
1967 ylong_path = _gvn.transform(r);
1968 record_for_igvn(r);
1969 }
1970
1971 // zero node
1972 Node *conzero = longcon(0);
1973 // Check (1&(long)y)==0?
1974 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1975 // Check if (1&(long)y)!=0?, if so the result is negative
1976 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1977 // abs(x)
1978 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1979 // abs(x)^y
1980 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1981 // -abs(x)^y
1982 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1983 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1984 Node *signresult = NULL;
1985 if (ConditionalMoveLimit != 0) {
1986 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1987 } else {
1988 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1989 RegionNode *r = new (C) RegionNode(3);
1990 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1991 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1992 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1993 phi->init_req(1, absxpowy);
1994 phi->init_req(2, negabsxpowy);
1995 signresult = _gvn.transform(phi);
1996 ylong_path = _gvn.transform(r);
1997 record_for_igvn(r);
1998 }
1999 // Set complex path fast result
2000 r->init_req(2, ylong_path);
2001 phi->init_req(2, signresult);
2002
2003 static const jlong nan_bits = CONST64(0x7ff8000000000000);
2004 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
2005 r->init_req(1,slow_path);
2006 phi->init_req(1,slow_result);
2007
2008 // Post merge
2009 set_control(_gvn.transform(r));
2010 record_for_igvn(r);
2011 result = _gvn.transform(phi);
2012 }
2013
2014 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2015
2016 // control from finish_pow_exp is now input to the region node
2017 region_node->set_req(2, control());
2018 // the result from finish_pow_exp is now input to the phi node
2019 phi_node->init_req(2, result);
2020 set_control(_gvn.transform(region_node));
2021 record_for_igvn(region_node);
2022 set_result(_gvn.transform(phi_node));
2023
2024 C->set_has_split_ifs(true); // Has chance for split-if optimization
2025 return true;
2026 }
2027
2028 //------------------------------runtime_math-----------------------------
2029 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2030 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2031 "must be (DD)D or (D)D type");
2032
2033 // Inputs
2034 Node* a = round_double_node(argument(0));
2035 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2036
2037 const TypePtr* no_memory_effects = NULL;
2038 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2039 no_memory_effects,
2040 a, top(), b, b ? top() : NULL);
2041 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
2042 #ifdef ASSERT
2043 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
2044 assert(value_top == top(), "second value must be top");
2045 #endif
2046
2047 set_result(value);
2048 return true;
2049 }
2050
2051 //------------------------------inline_math_native-----------------------------
2052 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2053 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2054 switch (id) {
2055 // These intrinsics are not properly supported on all hardware
2056 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
2057 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
2058 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
2059 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
2060 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
2061 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
2062
2063 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
2064 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
2065 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2066 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2067
2068 // These intrinsics are supported on all hardware
2069 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2070 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
2071
2072 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
2073 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
2074 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2075 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2076 #undef FN_PTR
2077
2078 // These intrinsics are not yet correctly implemented
2079 case vmIntrinsics::_datan2:
2080 return false;
2081
2082 default:
2083 fatal_unexpected_iid(id);
2084 return false;
2085 }
2086 }
2087
2088 static bool is_simple_name(Node* n) {
2089 return (n->req() == 1 // constant
2090 || (n->is_Type() && n->as_Type()->type()->singleton())
2091 || n->is_Proj() // parameter or return value
2092 || n->is_Phi() // local of some sort
2093 );
2094 }
2095
2096 //----------------------------inline_min_max-----------------------------------
2097 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2098 set_result(generate_min_max(id, argument(0), argument(1)));
2099 return true;
2100 }
2101
2102 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2103 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
2104 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2105 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
2106 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
2107
2108 {
2109 PreserveJVMState pjvms(this);
2110 PreserveReexecuteState preexecs(this);
2111 jvms()->set_should_reexecute(true);
2112
2113 set_control(slow_path);
2114 set_i_o(i_o());
2115
2116 uncommon_trap(Deoptimization::Reason_intrinsic,
2117 Deoptimization::Action_none);
2118 }
2119
2120 set_control(fast_path);
2121 set_result(math);
2122 }
2123
2124 template <typename OverflowOp>
2125 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2126 typedef typename OverflowOp::MathOp MathOp;
2127
2128 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2129 Node* operation = _gvn.transform( mathOp );
2130 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2131 inline_math_mathExact(operation, ofcheck);
2132 return true;
2133 }
2134
2135 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2136 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2137 }
2138
2139 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2140 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2141 }
2142
2143 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2144 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2145 }
2146
2147 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2148 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2149 }
2150
2151 bool LibraryCallKit::inline_math_negateExactI() {
2152 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2153 }
2154
2155 bool LibraryCallKit::inline_math_negateExactL() {
2156 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2157 }
2158
2159 bool LibraryCallKit::inline_math_multiplyExactI() {
2160 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2161 }
2162
2163 bool LibraryCallKit::inline_math_multiplyExactL() {
2164 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2165 }
2166
2167 Node*
2168 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2169 // These are the candidate return value:
2170 Node* xvalue = x0;
2171 Node* yvalue = y0;
2172
2173 if (xvalue == yvalue) {
2174 return xvalue;
2175 }
2176
2177 bool want_max = (id == vmIntrinsics::_max);
2178
2179 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2180 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2181 if (txvalue == NULL || tyvalue == NULL) return top();
2182 // This is not really necessary, but it is consistent with a
2183 // hypothetical MaxINode::Value method:
2184 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2185
2186 // %%% This folding logic should (ideally) be in a different place.
2187 // Some should be inside IfNode, and there to be a more reliable
2188 // transformation of ?: style patterns into cmoves. We also want
2189 // more powerful optimizations around cmove and min/max.
2190
2191 // Try to find a dominating comparison of these guys.
2192 // It can simplify the index computation for Arrays.copyOf
2193 // and similar uses of System.arraycopy.
2194 // First, compute the normalized version of CmpI(x, y).
2195 int cmp_op = Op_CmpI;
2196 Node* xkey = xvalue;
2197 Node* ykey = yvalue;
2198 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2199 if (ideal_cmpxy->is_Cmp()) {
2200 // E.g., if we have CmpI(length - offset, count),
2201 // it might idealize to CmpI(length, count + offset)
2202 cmp_op = ideal_cmpxy->Opcode();
2203 xkey = ideal_cmpxy->in(1);
2204 ykey = ideal_cmpxy->in(2);
2205 }
2206
2207 // Start by locating any relevant comparisons.
2208 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2209 Node* cmpxy = NULL;
2210 Node* cmpyx = NULL;
2211 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2212 Node* cmp = start_from->fast_out(k);
2213 if (cmp->outcnt() > 0 && // must have prior uses
2214 cmp->in(0) == NULL && // must be context-independent
2215 cmp->Opcode() == cmp_op) { // right kind of compare
2216 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2217 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2218 }
2219 }
2220
2221 const int NCMPS = 2;
2222 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2223 int cmpn;
2224 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2225 if (cmps[cmpn] != NULL) break; // find a result
2226 }
2227 if (cmpn < NCMPS) {
2228 // Look for a dominating test that tells us the min and max.
2229 int depth = 0; // Limit search depth for speed
2230 Node* dom = control();
2231 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2232 if (++depth >= 100) break;
2233 Node* ifproj = dom;
2234 if (!ifproj->is_Proj()) continue;
2235 Node* iff = ifproj->in(0);
2236 if (!iff->is_If()) continue;
2237 Node* bol = iff->in(1);
2238 if (!bol->is_Bool()) continue;
2239 Node* cmp = bol->in(1);
2240 if (cmp == NULL) continue;
2241 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2242 if (cmps[cmpn] == cmp) break;
2243 if (cmpn == NCMPS) continue;
2244 BoolTest::mask btest = bol->as_Bool()->_test._test;
2245 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2246 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2247 // At this point, we know that 'x btest y' is true.
2248 switch (btest) {
2249 case BoolTest::eq:
2250 // They are proven equal, so we can collapse the min/max.
2251 // Either value is the answer. Choose the simpler.
2252 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2253 return yvalue;
2254 return xvalue;
2255 case BoolTest::lt: // x < y
2256 case BoolTest::le: // x <= y
2257 return (want_max ? yvalue : xvalue);
2258 case BoolTest::gt: // x > y
2259 case BoolTest::ge: // x >= y
2260 return (want_max ? xvalue : yvalue);
2261 }
2262 }
2263 }
2264
2265 // We failed to find a dominating test.
2266 // Let's pick a test that might GVN with prior tests.
2267 Node* best_bol = NULL;
2268 BoolTest::mask best_btest = BoolTest::illegal;
2269 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2270 Node* cmp = cmps[cmpn];
2271 if (cmp == NULL) continue;
2272 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2273 Node* bol = cmp->fast_out(j);
2274 if (!bol->is_Bool()) continue;
2275 BoolTest::mask btest = bol->as_Bool()->_test._test;
2276 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2277 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2278 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2279 best_bol = bol->as_Bool();
2280 best_btest = btest;
2281 }
2282 }
2283 }
2284
2285 Node* answer_if_true = NULL;
2286 Node* answer_if_false = NULL;
2287 switch (best_btest) {
2288 default:
2289 if (cmpxy == NULL)
2290 cmpxy = ideal_cmpxy;
2291 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2292 // and fall through:
2293 case BoolTest::lt: // x < y
2294 case BoolTest::le: // x <= y
2295 answer_if_true = (want_max ? yvalue : xvalue);
2296 answer_if_false = (want_max ? xvalue : yvalue);
2297 break;
2298 case BoolTest::gt: // x > y
2299 case BoolTest::ge: // x >= y
2300 answer_if_true = (want_max ? xvalue : yvalue);
2301 answer_if_false = (want_max ? yvalue : xvalue);
2302 break;
2303 }
2304
2305 jint hi, lo;
2306 if (want_max) {
2307 // We can sharpen the minimum.
2308 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2309 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2310 } else {
2311 // We can sharpen the maximum.
2312 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2313 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2314 }
2315
2316 // Use a flow-free graph structure, to avoid creating excess control edges
2317 // which could hinder other optimizations.
2318 // Since Math.min/max is often used with arraycopy, we want
2319 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2320 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2321 answer_if_false, answer_if_true,
2322 TypeInt::make(lo, hi, widen));
2323
2324 return _gvn.transform(cmov);
2325
2326 /*
2327 // This is not as desirable as it may seem, since Min and Max
2328 // nodes do not have a full set of optimizations.
2329 // And they would interfere, anyway, with 'if' optimizations
2330 // and with CMoveI canonical forms.
2331 switch (id) {
2332 case vmIntrinsics::_min:
2333 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2334 case vmIntrinsics::_max:
2335 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2336 default:
2337 ShouldNotReachHere();
2338 }
2339 */
2340 }
2341
2342 inline int
2343 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2344 const TypePtr* base_type = TypePtr::NULL_PTR;
2345 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2346 if (base_type == NULL) {
2347 // Unknown type.
2348 return Type::AnyPtr;
2349 } else if (base_type == TypePtr::NULL_PTR) {
2350 // Since this is a NULL+long form, we have to switch to a rawptr.
2351 base = _gvn.transform(new (C) CastX2PNode(offset));
2352 offset = MakeConX(0);
2353 return Type::RawPtr;
2354 } else if (base_type->base() == Type::RawPtr) {
2355 return Type::RawPtr;
2356 } else if (base_type->isa_oopptr()) {
2357 // Base is never null => always a heap address.
2358 if (base_type->ptr() == TypePtr::NotNull) {
2359 return Type::OopPtr;
2360 }
2361 // Offset is small => always a heap address.
2362 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2363 if (offset_type != NULL &&
2364 base_type->offset() == 0 && // (should always be?)
2365 offset_type->_lo >= 0 &&
2366 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2367 return Type::OopPtr;
2368 }
2369 // Otherwise, it might either be oop+off or NULL+addr.
2370 return Type::AnyPtr;
2371 } else {
2372 // No information:
2373 return Type::AnyPtr;
2374 }
2375 }
2376
2377 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2378 int kind = classify_unsafe_addr(base, offset);
2379 if (kind == Type::RawPtr) {
2380 return basic_plus_adr(top(), base, offset);
2381 } else {
2382 return basic_plus_adr(base, offset);
2383 }
2384 }
2385
2386 //--------------------------inline_number_methods-----------------------------
2387 // inline int Integer.numberOfLeadingZeros(int)
2388 // inline int Long.numberOfLeadingZeros(long)
2389 //
2390 // inline int Integer.numberOfTrailingZeros(int)
2391 // inline int Long.numberOfTrailingZeros(long)
2392 //
2393 // inline int Integer.bitCount(int)
2394 // inline int Long.bitCount(long)
2395 //
2396 // inline char Character.reverseBytes(char)
2397 // inline short Short.reverseBytes(short)
2398 // inline int Integer.reverseBytes(int)
2399 // inline long Long.reverseBytes(long)
2400 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2401 Node* arg = argument(0);
2402 Node* n = NULL;
2403 switch (id) {
2404 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2405 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2406 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2407 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2408 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2409 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2410 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2411 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2412 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2413 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2414 default: fatal_unexpected_iid(id); break;
2415 }
2416 set_result(_gvn.transform(n));
2417 return true;
2418 }
2419
2420 //----------------------------inline_unsafe_access----------------------------
2421
2422 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2423
2424 // Helper that guards and inserts a pre-barrier.
2425 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2426 Node* pre_val, bool need_mem_bar) {
2427 // We could be accessing the referent field of a reference object. If so, when G1
2428 // is enabled, we need to log the value in the referent field in an SATB buffer.
2429 // This routine performs some compile time filters and generates suitable
2430 // runtime filters that guard the pre-barrier code.
2431 // Also add memory barrier for non volatile load from the referent field
2432 // to prevent commoning of loads across safepoint.
2433 if (!UseG1GC && !need_mem_bar)
2434 return;
2435
2436 // Some compile time checks.
2437
2438 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2439 const TypeX* otype = offset->find_intptr_t_type();
2440 if (otype != NULL && otype->is_con() &&
2441 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2442 // Constant offset but not the reference_offset so just return
2443 return;
2444 }
2445
2446 // We only need to generate the runtime guards for instances.
2447 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2448 if (btype != NULL) {
2449 if (btype->isa_aryptr()) {
2450 // Array type so nothing to do
2451 return;
2452 }
2453
2454 const TypeInstPtr* itype = btype->isa_instptr();
2455 if (itype != NULL) {
2456 // Can the klass of base_oop be statically determined to be
2457 // _not_ a sub-class of Reference and _not_ Object?
2458 ciKlass* klass = itype->klass();
2459 if ( klass->is_loaded() &&
2460 !klass->is_subtype_of(env()->Reference_klass()) &&
2461 !env()->Object_klass()->is_subtype_of(klass)) {
2462 return;
2463 }
2464 }
2465 }
2466
2467 // The compile time filters did not reject base_oop/offset so
2468 // we need to generate the following runtime filters
2469 //
2470 // if (offset == java_lang_ref_Reference::_reference_offset) {
2471 // if (instance_of(base, java.lang.ref.Reference)) {
2472 // pre_barrier(_, pre_val, ...);
2473 // }
2474 // }
2475
2476 float likely = PROB_LIKELY( 0.999);
2477 float unlikely = PROB_UNLIKELY(0.999);
2478
2479 IdealKit ideal(this);
2480 #define __ ideal.
2481
2482 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2483
2484 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2485 // Update graphKit memory and control from IdealKit.
2486 sync_kit(ideal);
2487
2488 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2489 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2490
2491 // Update IdealKit memory and control from graphKit.
2492 __ sync_kit(this);
2493
2494 Node* one = __ ConI(1);
2495 // is_instof == 0 if base_oop == NULL
2496 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2497
2498 // Update graphKit from IdeakKit.
2499 sync_kit(ideal);
2500
2501 // Use the pre-barrier to record the value in the referent field
2502 pre_barrier(false /* do_load */,
2503 __ ctrl(),
2504 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2505 pre_val /* pre_val */,
2506 T_OBJECT);
2507 if (need_mem_bar) {
2508 // Add memory barrier to prevent commoning reads from this field
2509 // across safepoint since GC can change its value.
2510 insert_mem_bar(Op_MemBarCPUOrder);
2511 }
2512 // Update IdealKit from graphKit.
2513 __ sync_kit(this);
2514
2515 } __ end_if(); // _ref_type != ref_none
2516 } __ end_if(); // offset == referent_offset
2517
2518 // Final sync IdealKit and GraphKit.
2519 final_sync(ideal);
2520 #undef __
2521 }
2522
2523
2524 // Interpret Unsafe.fieldOffset cookies correctly:
2525 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2526
2527 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2528 // Attempt to infer a sharper value type from the offset and base type.
2529 ciKlass* sharpened_klass = NULL;
2530
2531 // See if it is an instance field, with an object type.
2532 if (alias_type->field() != NULL) {
2533 assert(!is_native_ptr, "native pointer op cannot use a java address");
2534 if (alias_type->field()->type()->is_klass()) {
2535 sharpened_klass = alias_type->field()->type()->as_klass();
2536 }
2537 }
2538
2539 // See if it is a narrow oop array.
2540 if (adr_type->isa_aryptr()) {
2541 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2542 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2543 if (elem_type != NULL) {
2544 sharpened_klass = elem_type->klass();
2545 }
2546 }
2547 }
2548
2549 // The sharpened class might be unloaded if there is no class loader
2550 // contraint in place.
2551 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2552 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2553
2554 #ifndef PRODUCT
2555 if (C->print_intrinsics() || C->print_inlining()) {
2556 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2557 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2558 }
2559 #endif
2560 // Sharpen the value type.
2561 return tjp;
2562 }
2563 return NULL;
2564 }
2565
2566 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2567 if (callee()->is_static()) return false; // caller must have the capability!
2568 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2569
2570 #ifndef PRODUCT
2571 {
2572 ResourceMark rm;
2573 // Check the signatures.
2574 ciSignature* sig = callee()->signature();
2575 #ifdef ASSERT
2576 if (!is_store) {
2577 // Object getObject(Object base, int/long offset), etc.
2578 BasicType rtype = sig->return_type()->basic_type();
2579 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2580 rtype = T_ADDRESS; // it is really a C void*
2581 assert(rtype == type, "getter must return the expected value");
2582 if (!is_native_ptr) {
2583 assert(sig->count() == 2, "oop getter has 2 arguments");
2584 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2585 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2586 } else {
2587 assert(sig->count() == 1, "native getter has 1 argument");
2588 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2589 }
2590 } else {
2591 // void putObject(Object base, int/long offset, Object x), etc.
2592 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2593 if (!is_native_ptr) {
2594 assert(sig->count() == 3, "oop putter has 3 arguments");
2595 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2596 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2597 } else {
2598 assert(sig->count() == 2, "native putter has 2 arguments");
2599 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2600 }
2601 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2602 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2603 vtype = T_ADDRESS; // it is really a C void*
2604 assert(vtype == type, "putter must accept the expected value");
2605 }
2606 #endif // ASSERT
2607 }
2608 #endif //PRODUCT
2609
2610 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2611
2612 Node* receiver = argument(0); // type: oop
2613
2614 // Build address expression. See the code in inline_unsafe_prefetch.
2615 Node* adr;
2616 Node* heap_base_oop = top();
2617 Node* offset = top();
2618 Node* val;
2619
2620 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2621 Node* base = argument(1); // type: oop
2622
2623 if (!is_native_ptr) {
2624 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2625 offset = argument(2); // type: long
2626 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2627 // to be plain byte offsets, which are also the same as those accepted
2628 // by oopDesc::field_base.
2629 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2630 "fieldOffset must be byte-scaled");
2631 // 32-bit machines ignore the high half!
2632 offset = ConvL2X(offset);
2633 adr = make_unsafe_address(base, offset);
2634 heap_base_oop = base;
2635 val = is_store ? argument(4) : NULL;
2636 } else {
2637 Node* ptr = argument(1); // type: long
2638 ptr = ConvL2X(ptr); // adjust Java long to machine word
2639 adr = make_unsafe_address(NULL, ptr);
2640 val = is_store ? argument(3) : NULL;
2641 }
2642
2643 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) {
2644 return false; // off-heap oop accesses are not supported
2645 }
2646
2647 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2648
2649 // Try to categorize the address.
2650 Compile::AliasType* alias_type = C->alias_type(adr_type);
2651 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2652
2653 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2654 alias_type->adr_type() == TypeAryPtr::RANGE) {
2655 return false; // not supported
2656 }
2657
2658 bool mismatched = false;
2659 BasicType bt = alias_type->basic_type();
2660 if (bt != T_ILLEGAL) {
2661 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2662 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2663 // Alias type doesn't differentiate between byte[] and boolean[]).
2664 // Use address type to get the element type.
2665 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2666 }
2667 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2668 // accessing an array field with getObject is not a mismatch
2669 bt = T_OBJECT;
2670 }
2671 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2672 // Don't intrinsify mismatched object accesses
2673 return false;
2674 }
2675 mismatched = (bt != type);
2676 }
2677
2678 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2679
2680 // First guess at the value type.
2681 const Type *value_type = Type::get_const_basic_type(type);
2682
2683 // We will need memory barriers unless we can determine a unique
2684 // alias category for this reference. (Note: If for some reason
2685 // the barriers get omitted and the unsafe reference begins to "pollute"
2686 // the alias analysis of the rest of the graph, either Compile::can_alias
2687 // or Compile::must_alias will throw a diagnostic assert.)
2688 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2689
2690 // If we are reading the value of the referent field of a Reference
2691 // object (either by using Unsafe directly or through reflection)
2692 // then, if G1 is enabled, we need to record the referent in an
2693 // SATB log buffer using the pre-barrier mechanism.
2694 // Also we need to add memory barrier to prevent commoning reads
2695 // from this field across safepoint since GC can change its value.
2696 bool need_read_barrier = !is_native_ptr && !is_store &&
2697 offset != top() && heap_base_oop != top();
2698
2699 if (!is_store && type == T_OBJECT) {
2700 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2701 if (tjp != NULL) {
2702 value_type = tjp;
2703 }
2704 }
2705
2706 receiver = null_check(receiver);
2707 if (stopped()) {
2708 return true;
2709 }
2710 // Heap pointers get a null-check from the interpreter,
2711 // as a courtesy. However, this is not guaranteed by Unsafe,
2712 // and it is not possible to fully distinguish unintended nulls
2713 // from intended ones in this API.
2714
2715 Node* load = NULL;
2716 Node* store = NULL;
2717 Node* leading_membar = NULL;
2718 if (is_volatile) {
2719 // We need to emit leading and trailing CPU membars (see below) in
2720 // addition to memory membars when is_volatile. This is a little
2721 // too strong, but avoids the need to insert per-alias-type
2722 // volatile membars (for stores; compare Parse::do_put_xxx), which
2723 // we cannot do effectively here because we probably only have a
2724 // rough approximation of type.
2725 need_mem_bar = true;
2726 // For Stores, place a memory ordering barrier now.
2727 if (is_store) {
2728 leading_membar = insert_mem_bar(Op_MemBarRelease);
2729 } else {
2730 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2731 leading_membar = insert_mem_bar(Op_MemBarVolatile);
2732 }
2733 }
2734 }
2735
2736 // Memory barrier to prevent normal and 'unsafe' accesses from
2737 // bypassing each other. Happens after null checks, so the
2738 // exception paths do not take memory state from the memory barrier,
2739 // so there's no problems making a strong assert about mixing users
2740 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2741 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2742 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2743
2744 if (!is_store) {
2745 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2746 // To be valid, unsafe loads may depend on other conditions than
2747 // the one that guards them: pin the Load node
2748 load = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2749 // load value
2750 switch (type) {
2751 case T_BOOLEAN:
2752 case T_CHAR:
2753 case T_BYTE:
2754 case T_SHORT:
2755 case T_INT:
2756 case T_LONG:
2757 case T_FLOAT:
2758 case T_DOUBLE:
2759 break;
2760 case T_OBJECT:
2761 if (need_read_barrier) {
2762 insert_pre_barrier(heap_base_oop, offset, load, !(is_volatile || need_mem_bar));
2763 }
2764 break;
2765 case T_ADDRESS:
2766 // Cast to an int type.
2767 load = _gvn.transform(new (C) CastP2XNode(NULL, load));
2768 load = ConvX2UL(load);
2769 break;
2770 default:
2771 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2772 break;
2773 }
2774 // The load node has the control of the preceding MemBarCPUOrder. All
2775 // following nodes will have the control of the MemBarCPUOrder inserted at
2776 // the end of this method. So, pushing the load onto the stack at a later
2777 // point is fine.
2778 set_result(load);
2779 } else {
2780 // place effect of store into memory
2781 switch (type) {
2782 case T_DOUBLE:
2783 val = dstore_rounding(val);
2784 break;
2785 case T_ADDRESS:
2786 // Repackage the long as a pointer.
2787 val = ConvL2X(val);
2788 val = _gvn.transform(new (C) CastX2PNode(val));
2789 break;
2790 }
2791
2792 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2793 if (type == T_OBJECT ) {
2794 store = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2795 } else {
2796 store = store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2797 }
2798 }
2799
2800 if (is_volatile) {
2801 if (!is_store) {
2802 Node* mb = insert_mem_bar(Op_MemBarAcquire, load);
2803 mb->as_MemBar()->set_trailing_load();
2804 } else {
2805 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2806 Node* mb = insert_mem_bar(Op_MemBarVolatile, store);
2807 MemBarNode::set_store_pair(leading_membar->as_MemBar(), mb->as_MemBar());
2808 }
2809 }
2810 }
2811
2812 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2813
2814 return true;
2815 }
2816
2817 //----------------------------inline_unsafe_prefetch----------------------------
2818
2819 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2820 #ifndef PRODUCT
2821 {
2822 ResourceMark rm;
2823 // Check the signatures.
2824 ciSignature* sig = callee()->signature();
2825 #ifdef ASSERT
2826 // Object getObject(Object base, int/long offset), etc.
2827 BasicType rtype = sig->return_type()->basic_type();
2828 if (!is_native_ptr) {
2829 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2830 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2831 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2832 } else {
2833 assert(sig->count() == 1, "native prefetch has 1 argument");
2834 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2835 }
2836 #endif // ASSERT
2837 }
2838 #endif // !PRODUCT
2839
2840 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2841
2842 const int idx = is_static ? 0 : 1;
2843 if (!is_static) {
2844 null_check_receiver();
2845 if (stopped()) {
2846 return true;
2847 }
2848 }
2849
2850 // Build address expression. See the code in inline_unsafe_access.
2851 Node *adr;
2852 if (!is_native_ptr) {
2853 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2854 Node* base = argument(idx + 0); // type: oop
2855 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2856 Node* offset = argument(idx + 1); // type: long
2857 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2858 // to be plain byte offsets, which are also the same as those accepted
2859 // by oopDesc::field_base.
2860 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2861 "fieldOffset must be byte-scaled");
2862 // 32-bit machines ignore the high half!
2863 offset = ConvL2X(offset);
2864 adr = make_unsafe_address(base, offset);
2865 } else {
2866 Node* ptr = argument(idx + 0); // type: long
2867 ptr = ConvL2X(ptr); // adjust Java long to machine word
2868 adr = make_unsafe_address(NULL, ptr);
2869 }
2870
2871 // Generate the read or write prefetch
2872 Node *prefetch;
2873 if (is_store) {
2874 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2875 } else {
2876 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2877 }
2878 prefetch->init_req(0, control());
2879 set_i_o(_gvn.transform(prefetch));
2880
2881 return true;
2882 }
2883
2884 //----------------------------inline_unsafe_load_store----------------------------
2885 // This method serves a couple of different customers (depending on LoadStoreKind):
2886 //
2887 // LS_cmpxchg:
2888 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2889 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2890 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2891 //
2892 // LS_xadd:
2893 // public int getAndAddInt( Object o, long offset, int delta)
2894 // public long getAndAddLong(Object o, long offset, long delta)
2895 //
2896 // LS_xchg:
2897 // int getAndSet(Object o, long offset, int newValue)
2898 // long getAndSet(Object o, long offset, long newValue)
2899 // Object getAndSet(Object o, long offset, Object newValue)
2900 //
2901 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2902 // This basic scheme here is the same as inline_unsafe_access, but
2903 // differs in enough details that combining them would make the code
2904 // overly confusing. (This is a true fact! I originally combined
2905 // them, but even I was confused by it!) As much code/comments as
2906 // possible are retained from inline_unsafe_access though to make
2907 // the correspondences clearer. - dl
2908
2909 if (callee()->is_static()) return false; // caller must have the capability!
2910
2911 #ifndef PRODUCT
2912 BasicType rtype;
2913 {
2914 ResourceMark rm;
2915 // Check the signatures.
2916 ciSignature* sig = callee()->signature();
2917 rtype = sig->return_type()->basic_type();
2918 if (kind == LS_xadd || kind == LS_xchg) {
2919 // Check the signatures.
2920 #ifdef ASSERT
2921 assert(rtype == type, "get and set must return the expected type");
2922 assert(sig->count() == 3, "get and set has 3 arguments");
2923 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2924 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2925 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2926 #endif // ASSERT
2927 } else if (kind == LS_cmpxchg) {
2928 // Check the signatures.
2929 #ifdef ASSERT
2930 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2931 assert(sig->count() == 4, "CAS has 4 arguments");
2932 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2933 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2934 #endif // ASSERT
2935 } else {
2936 ShouldNotReachHere();
2937 }
2938 }
2939 #endif //PRODUCT
2940
2941 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2942
2943 // Get arguments:
2944 Node* receiver = NULL;
2945 Node* base = NULL;
2946 Node* offset = NULL;
2947 Node* oldval = NULL;
2948 Node* newval = NULL;
2949 if (kind == LS_cmpxchg) {
2950 const bool two_slot_type = type2size[type] == 2;
2951 receiver = argument(0); // type: oop
2952 base = argument(1); // type: oop
2953 offset = argument(2); // type: long
2954 oldval = argument(4); // type: oop, int, or long
2955 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2956 } else if (kind == LS_xadd || kind == LS_xchg){
2957 receiver = argument(0); // type: oop
2958 base = argument(1); // type: oop
2959 offset = argument(2); // type: long
2960 oldval = NULL;
2961 newval = argument(4); // type: oop, int, or long
2962 }
2963
2964 // Build field offset expression.
2965 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2966 // to be plain byte offsets, which are also the same as those accepted
2967 // by oopDesc::field_base.
2968 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2969 // 32-bit machines ignore the high half of long offsets
2970 offset = ConvL2X(offset);
2971 Node* adr = make_unsafe_address(base, offset);
2972 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2973
2974 Compile::AliasType* alias_type = C->alias_type(adr_type);
2975 BasicType bt = alias_type->basic_type();
2976 if (bt != T_ILLEGAL &&
2977 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2978 // Don't intrinsify mismatched object accesses.
2979 return false;
2980 }
2981
2982 // For CAS, unlike inline_unsafe_access, there seems no point in
2983 // trying to refine types. Just use the coarse types here.
2984 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2985 const Type *value_type = Type::get_const_basic_type(type);
2986
2987 if (kind == LS_xchg && type == T_OBJECT) {
2988 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2989 if (tjp != NULL) {
2990 value_type = tjp;
2991 }
2992 }
2993
2994 // Null check receiver.
2995 receiver = null_check(receiver);
2996 if (stopped()) {
2997 return true;
2998 }
2999
3000 int alias_idx = C->get_alias_index(adr_type);
3001
3002 // Memory-model-wise, a LoadStore acts like a little synchronized
3003 // block, so needs barriers on each side. These don't translate
3004 // into actual barriers on most machines, but we still need rest of
3005 // compiler to respect ordering.
3006
3007 Node* leading_membar = insert_mem_bar(Op_MemBarRelease);
3008 insert_mem_bar(Op_MemBarCPUOrder);
3009
3010 // 4984716: MemBars must be inserted before this
3011 // memory node in order to avoid a false
3012 // dependency which will confuse the scheduler.
3013 Node *mem = memory(alias_idx);
3014
3015 // For now, we handle only those cases that actually exist: ints,
3016 // longs, and Object. Adding others should be straightforward.
3017 Node* load_store = NULL;
3018 switch(type) {
3019 case T_INT:
3020 if (kind == LS_xadd) {
3021 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
3022 } else if (kind == LS_xchg) {
3023 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
3024 } else if (kind == LS_cmpxchg) {
3025 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
3026 } else {
3027 ShouldNotReachHere();
3028 }
3029 break;
3030 case T_LONG:
3031 if (kind == LS_xadd) {
3032 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
3033 } else if (kind == LS_xchg) {
3034 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
3035 } else if (kind == LS_cmpxchg) {
3036 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
3037 } else {
3038 ShouldNotReachHere();
3039 }
3040 break;
3041 case T_OBJECT:
3042 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3043 // could be delayed during Parse (for example, in adjust_map_after_if()).
3044 // Execute transformation here to avoid barrier generation in such case.
3045 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3046 newval = _gvn.makecon(TypePtr::NULL_PTR);
3047
3048 // Reference stores need a store barrier.
3049 if (kind == LS_xchg) {
3050 // If pre-barrier must execute before the oop store, old value will require do_load here.
3051 if (!can_move_pre_barrier()) {
3052 pre_barrier(true /* do_load*/,
3053 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3054 NULL /* pre_val*/,
3055 T_OBJECT);
3056 } // Else move pre_barrier to use load_store value, see below.
3057 } else if (kind == LS_cmpxchg) {
3058 // Same as for newval above:
3059 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3060 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3061 }
3062 // The only known value which might get overwritten is oldval.
3063 pre_barrier(false /* do_load */,
3064 control(), NULL, NULL, max_juint, NULL, NULL,
3065 oldval /* pre_val */,
3066 T_OBJECT);
3067 } else {
3068 ShouldNotReachHere();
3069 }
3070
3071 #ifdef _LP64
3072 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3073 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3074 if (kind == LS_xchg) {
3075 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
3076 newval_enc, adr_type, value_type->make_narrowoop()));
3077 } else {
3078 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3079 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3080 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
3081 newval_enc, oldval_enc));
3082 }
3083 } else
3084 #endif
3085 {
3086 if (kind == LS_xchg) {
3087 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3088 } else {
3089 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3090 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3091 }
3092 }
3093 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3094 break;
3095 default:
3096 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3097 break;
3098 }
3099
3100 // SCMemProjNodes represent the memory state of a LoadStore. Their
3101 // main role is to prevent LoadStore nodes from being optimized away
3102 // when their results aren't used.
3103 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3104 set_memory(proj, alias_idx);
3105
3106 Node* access = load_store;
3107
3108 if (type == T_OBJECT && kind == LS_xchg) {
3109 #ifdef _LP64
3110 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3111 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3112 }
3113 #endif
3114 if (can_move_pre_barrier()) {
3115 // Don't need to load pre_val. The old value is returned by load_store.
3116 // The pre_barrier can execute after the xchg as long as no safepoint
3117 // gets inserted between them.
3118 pre_barrier(false /* do_load */,
3119 control(), NULL, NULL, max_juint, NULL, NULL,
3120 load_store /* pre_val */,
3121 T_OBJECT);
3122 }
3123 }
3124
3125 // Add the trailing membar surrounding the access
3126 insert_mem_bar(Op_MemBarCPUOrder);
3127 Node* mb = insert_mem_bar(Op_MemBarAcquire, access);
3128 MemBarNode::set_load_store_pair(leading_membar->as_MemBar(), mb->as_MemBar());
3129
3130 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3131 set_result(load_store);
3132 return true;
3133 }
3134
3135 //----------------------------inline_unsafe_ordered_store----------------------
3136 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3137 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3138 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3139 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3140 // This is another variant of inline_unsafe_access, differing in
3141 // that it always issues store-store ("release") barrier and ensures
3142 // store-atomicity (which only matters for "long").
3143
3144 if (callee()->is_static()) return false; // caller must have the capability!
3145
3146 #ifndef PRODUCT
3147 {
3148 ResourceMark rm;
3149 // Check the signatures.
3150 ciSignature* sig = callee()->signature();
3151 #ifdef ASSERT
3152 BasicType rtype = sig->return_type()->basic_type();
3153 assert(rtype == T_VOID, "must return void");
3154 assert(sig->count() == 3, "has 3 arguments");
3155 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3156 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3157 #endif // ASSERT
3158 }
3159 #endif //PRODUCT
3160
3161 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3162
3163 // Get arguments:
3164 Node* receiver = argument(0); // type: oop
3165 Node* base = argument(1); // type: oop
3166 Node* offset = argument(2); // type: long
3167 Node* val = argument(4); // type: oop, int, or long
3168
3169 // Null check receiver.
3170 receiver = null_check(receiver);
3171 if (stopped()) {
3172 return true;
3173 }
3174
3175 // Build field offset expression.
3176 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3177 // 32-bit machines ignore the high half of long offsets
3178 offset = ConvL2X(offset);
3179 Node* adr = make_unsafe_address(base, offset);
3180 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3181 const Type *value_type = Type::get_const_basic_type(type);
3182 Compile::AliasType* alias_type = C->alias_type(adr_type);
3183
3184 insert_mem_bar(Op_MemBarRelease);
3185 insert_mem_bar(Op_MemBarCPUOrder);
3186 // Ensure that the store is atomic for longs:
3187 const bool require_atomic_access = true;
3188 Node* store;
3189 if (type == T_OBJECT) // reference stores need a store barrier.
3190 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3191 else {
3192 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3193 }
3194 insert_mem_bar(Op_MemBarCPUOrder);
3195 return true;
3196 }
3197
3198 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3199 // Regardless of form, don't allow previous ld/st to move down,
3200 // then issue acquire, release, or volatile mem_bar.
3201 insert_mem_bar(Op_MemBarCPUOrder);
3202 switch(id) {
3203 case vmIntrinsics::_loadFence:
3204 insert_mem_bar(Op_LoadFence);
3205 return true;
3206 case vmIntrinsics::_storeFence:
3207 insert_mem_bar(Op_StoreFence);
3208 return true;
3209 case vmIntrinsics::_fullFence:
3210 insert_mem_bar(Op_MemBarVolatile);
3211 return true;
3212 default:
3213 fatal_unexpected_iid(id);
3214 return false;
3215 }
3216 }
3217
3218 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3219 if (!kls->is_Con()) {
3220 return true;
3221 }
3222 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3223 if (klsptr == NULL) {
3224 return true;
3225 }
3226 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3227 // don't need a guard for a klass that is already initialized
3228 return !ik->is_initialized();
3229 }
3230
3231 //----------------------------inline_unsafe_allocate---------------------------
3232 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3233 bool LibraryCallKit::inline_unsafe_allocate() {
3234 if (callee()->is_static()) return false; // caller must have the capability!
3235
3236 null_check_receiver(); // null-check, then ignore
3237 Node* cls = null_check(argument(1));
3238 if (stopped()) return true;
3239
3240 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3241 kls = null_check(kls);
3242 if (stopped()) return true; // argument was like int.class
3243
3244 Node* test = NULL;
3245 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3246 // Note: The argument might still be an illegal value like
3247 // Serializable.class or Object[].class. The runtime will handle it.
3248 // But we must make an explicit check for initialization.
3249 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3250 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3251 // can generate code to load it as unsigned byte.
3252 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3253 Node* bits = intcon(InstanceKlass::fully_initialized);
3254 test = _gvn.transform(new (C) SubINode(inst, bits));
3255 // The 'test' is non-zero if we need to take a slow path.
3256 }
3257
3258 Node* obj = new_instance(kls, test);
3259 set_result(obj);
3260 return true;
3261 }
3262
3263 #ifdef JFR_HAVE_INTRINSICS
3264 /*
3265 * oop -> myklass
3266 * myklass->trace_id |= USED
3267 * return myklass->trace_id & ~0x3
3268 */
3269 bool LibraryCallKit::inline_native_classID() {
3270 Node* cls = null_check(argument(0), T_OBJECT);
3271 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3272 kls = null_check(kls, T_OBJECT);
3273
3274 ByteSize offset = KLASS_TRACE_ID_OFFSET;
3275 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3276 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3277
3278 Node* clsused = longcon(0x01l); // set the class bit
3279 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3280 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3281 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3282
3283 #ifdef TRACE_ID_META_BITS
3284 Node* mbits = longcon(~TRACE_ID_META_BITS);
3285 tvalue = _gvn.transform(new (C) AndLNode(tvalue, mbits));
3286 #endif
3287 #ifdef TRACE_ID_SHIFT
3288 Node* cbits = intcon(TRACE_ID_SHIFT);
3289 tvalue = _gvn.transform(new (C) URShiftLNode(tvalue, cbits));
3290 #endif
3291
3292 set_result(tvalue);
3293 return true;
3294 }
3295
3296 bool LibraryCallKit::inline_native_getEventWriter() {
3297 Node* tls_ptr = _gvn.transform(new (C) ThreadLocalNode());
3298
3299 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3300 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)
3301 );
3302
3303 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3304
3305 Node* jobj_cmp_null = _gvn.transform( new (C) CmpPNode(jobj, null()) );
3306 Node* test_jobj_eq_null = _gvn.transform( new (C) BoolNode(jobj_cmp_null, BoolTest::eq) );
3307
3308 IfNode* iff_jobj_null =
3309 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3310
3311 enum { _normal_path = 1,
3312 _null_path = 2,
3313 PATH_LIMIT };
3314
3315 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3316 PhiNode* result_val = new (C) PhiNode(result_rgn, TypePtr::BOTTOM);
3317
3318 Node* jobj_is_null = _gvn.transform(new (C) IfTrueNode(iff_jobj_null));
3319 result_rgn->init_req(_null_path, jobj_is_null);
3320 result_val->init_req(_null_path, null());
3321
3322 Node* jobj_is_not_null = _gvn.transform(new (C) IfFalseNode(iff_jobj_null));
3323 result_rgn->init_req(_normal_path, jobj_is_not_null);
3324
3325 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3326 result_val->init_req(_normal_path, res);
3327
3328 set_result(result_rgn, result_val);
3329
3330 return true;
3331 }
3332 #endif // JFR_HAVE_INTRINSICS
3333
3334 //------------------------inline_native_time_funcs--------------
3335 // inline code for System.currentTimeMillis() and System.nanoTime()
3336 // these have the same type and signature
3337 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3338 const TypeFunc* tf = OptoRuntime::void_long_Type();
3339 const TypePtr* no_memory_effects = NULL;
3340 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3341 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3342 #ifdef ASSERT
3343 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3344 assert(value_top == top(), "second value must be top");
3345 #endif
3346 set_result(value);
3347 return true;
3348 }
3349
3350 //------------------------inline_native_currentThread------------------
3351 bool LibraryCallKit::inline_native_currentThread() {
3352 Node* junk = NULL;
3353 set_result(generate_current_thread(junk));
3354 return true;
3355 }
3356
3357 //------------------------inline_native_isInterrupted------------------
3358 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3359 bool LibraryCallKit::inline_native_isInterrupted() {
3360 // Add a fast path to t.isInterrupted(clear_int):
3361 // (t == Thread.current() &&
3362 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3363 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3364 // So, in the common case that the interrupt bit is false,
3365 // we avoid making a call into the VM. Even if the interrupt bit
3366 // is true, if the clear_int argument is false, we avoid the VM call.
3367 // However, if the receiver is not currentThread, we must call the VM,
3368 // because there must be some locking done around the operation.
3369
3370 // We only go to the fast case code if we pass two guards.
3371 // Paths which do not pass are accumulated in the slow_region.
3372
3373 enum {
3374 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3375 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3376 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3377 PATH_LIMIT
3378 };
3379
3380 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3381 // out of the function.
3382 insert_mem_bar(Op_MemBarCPUOrder);
3383
3384 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3385 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3386
3387 RegionNode* slow_region = new (C) RegionNode(1);
3388 record_for_igvn(slow_region);
3389
3390 // (a) Receiving thread must be the current thread.
3391 Node* rec_thr = argument(0);
3392 Node* tls_ptr = NULL;
3393 Node* cur_thr = generate_current_thread(tls_ptr);
3394 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3395 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3396
3397 generate_slow_guard(bol_thr, slow_region);
3398
3399 // (b) Interrupt bit on TLS must be false.
3400 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3401 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3402 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3403
3404 // Set the control input on the field _interrupted read to prevent it floating up.
3405 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3406 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3407 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3408
3409 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3410
3411 // First fast path: if (!TLS._interrupted) return false;
3412 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3413 result_rgn->init_req(no_int_result_path, false_bit);
3414 result_val->init_req(no_int_result_path, intcon(0));
3415
3416 // drop through to next case
3417 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3418
3419 #ifndef TARGET_OS_FAMILY_windows
3420 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3421 Node* clr_arg = argument(1);
3422 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3423 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3424 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3425
3426 // Second fast path: ... else if (!clear_int) return true;
3427 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3428 result_rgn->init_req(no_clear_result_path, false_arg);
3429 result_val->init_req(no_clear_result_path, intcon(1));
3430
3431 // drop through to next case
3432 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3433 #else
3434 // To return true on Windows you must read the _interrupted field
3435 // and check the the event state i.e. take the slow path.
3436 #endif // TARGET_OS_FAMILY_windows
3437
3438 // (d) Otherwise, go to the slow path.
3439 slow_region->add_req(control());
3440 set_control( _gvn.transform(slow_region));
3441
3442 if (stopped()) {
3443 // There is no slow path.
3444 result_rgn->init_req(slow_result_path, top());
3445 result_val->init_req(slow_result_path, top());
3446 } else {
3447 // non-virtual because it is a private non-static
3448 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3449
3450 Node* slow_val = set_results_for_java_call(slow_call);
3451 // this->control() comes from set_results_for_java_call
3452
3453 Node* fast_io = slow_call->in(TypeFunc::I_O);
3454 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3455
3456 // These two phis are pre-filled with copies of of the fast IO and Memory
3457 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3458 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3459
3460 result_rgn->init_req(slow_result_path, control());
3461 result_io ->init_req(slow_result_path, i_o());
3462 result_mem->init_req(slow_result_path, reset_memory());
3463 result_val->init_req(slow_result_path, slow_val);
3464
3465 set_all_memory(_gvn.transform(result_mem));
3466 set_i_o( _gvn.transform(result_io));
3467 }
3468
3469 C->set_has_split_ifs(true); // Has chance for split-if optimization
3470 set_result(result_rgn, result_val);
3471 return true;
3472 }
3473
3474 //---------------------------load_mirror_from_klass----------------------------
3475 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3476 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3477 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3478 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3479 }
3480
3481 //-----------------------load_klass_from_mirror_common-------------------------
3482 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3483 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3484 // and branch to the given path on the region.
3485 // If never_see_null, take an uncommon trap on null, so we can optimistically
3486 // compile for the non-null case.
3487 // If the region is NULL, force never_see_null = true.
3488 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3489 bool never_see_null,
3490 RegionNode* region,
3491 int null_path,
3492 int offset) {
3493 if (region == NULL) never_see_null = true;
3494 Node* p = basic_plus_adr(mirror, offset);
3495 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3496 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3497 Node* null_ctl = top();
3498 kls = null_check_oop(kls, &null_ctl, never_see_null);
3499 if (region != NULL) {
3500 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3501 region->init_req(null_path, null_ctl);
3502 } else {
3503 assert(null_ctl == top(), "no loose ends");
3504 }
3505 return kls;
3506 }
3507
3508 //--------------------(inline_native_Class_query helpers)---------------------
3509 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3510 // Fall through if (mods & mask) == bits, take the guard otherwise.
3511 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3512 // Branch around if the given klass has the given modifier bit set.
3513 // Like generate_guard, adds a new path onto the region.
3514 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3515 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3516 Node* mask = intcon(modifier_mask);
3517 Node* bits = intcon(modifier_bits);
3518 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3519 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3520 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3521 return generate_fair_guard(bol, region);
3522 }
3523 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3524 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3525 }
3526
3527 //-------------------------inline_native_Class_query-------------------
3528 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3529 const Type* return_type = TypeInt::BOOL;
3530 Node* prim_return_value = top(); // what happens if it's a primitive class?
3531 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3532 bool expect_prim = false; // most of these guys expect to work on refs
3533
3534 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3535
3536 Node* mirror = argument(0);
3537 Node* obj = top();
3538
3539 switch (id) {
3540 case vmIntrinsics::_isInstance:
3541 // nothing is an instance of a primitive type
3542 prim_return_value = intcon(0);
3543 obj = argument(1);
3544 break;
3545 case vmIntrinsics::_getModifiers:
3546 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3547 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3548 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3549 break;
3550 case vmIntrinsics::_isInterface:
3551 prim_return_value = intcon(0);
3552 break;
3553 case vmIntrinsics::_isArray:
3554 prim_return_value = intcon(0);
3555 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3556 break;
3557 case vmIntrinsics::_isPrimitive:
3558 prim_return_value = intcon(1);
3559 expect_prim = true; // obviously
3560 break;
3561 case vmIntrinsics::_getSuperclass:
3562 prim_return_value = null();
3563 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3564 break;
3565 case vmIntrinsics::_getComponentType:
3566 prim_return_value = null();
3567 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3568 break;
3569 case vmIntrinsics::_getClassAccessFlags:
3570 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3571 return_type = TypeInt::INT; // not bool! 6297094
3572 break;
3573 default:
3574 fatal_unexpected_iid(id);
3575 break;
3576 }
3577
3578 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3579 if (mirror_con == NULL) return false; // cannot happen?
3580
3581 #ifndef PRODUCT
3582 if (C->print_intrinsics() || C->print_inlining()) {
3583 ciType* k = mirror_con->java_mirror_type();
3584 if (k) {
3585 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3586 k->print_name();
3587 tty->cr();
3588 }
3589 }
3590 #endif
3591
3592 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3593 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3594 record_for_igvn(region);
3595 PhiNode* phi = new (C) PhiNode(region, return_type);
3596
3597 // The mirror will never be null of Reflection.getClassAccessFlags, however
3598 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3599 // if it is. See bug 4774291.
3600
3601 // For Reflection.getClassAccessFlags(), the null check occurs in
3602 // the wrong place; see inline_unsafe_access(), above, for a similar
3603 // situation.
3604 mirror = null_check(mirror);
3605 // If mirror or obj is dead, only null-path is taken.
3606 if (stopped()) return true;
3607
3608 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3609
3610 // Now load the mirror's klass metaobject, and null-check it.
3611 // Side-effects region with the control path if the klass is null.
3612 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3613 // If kls is null, we have a primitive mirror.
3614 phi->init_req(_prim_path, prim_return_value);
3615 if (stopped()) { set_result(region, phi); return true; }
3616 bool safe_for_replace = (region->in(_prim_path) == top());
3617
3618 Node* p; // handy temp
3619 Node* null_ctl;
3620
3621 // Now that we have the non-null klass, we can perform the real query.
3622 // For constant classes, the query will constant-fold in LoadNode::Value.
3623 Node* query_value = top();
3624 switch (id) {
3625 case vmIntrinsics::_isInstance:
3626 // nothing is an instance of a primitive type
3627 query_value = gen_instanceof(obj, kls, safe_for_replace);
3628 break;
3629
3630 case vmIntrinsics::_getModifiers:
3631 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3632 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3633 break;
3634
3635 case vmIntrinsics::_isInterface:
3636 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3637 if (generate_interface_guard(kls, region) != NULL)
3638 // A guard was added. If the guard is taken, it was an interface.
3639 phi->add_req(intcon(1));
3640 // If we fall through, it's a plain class.
3641 query_value = intcon(0);
3642 break;
3643
3644 case vmIntrinsics::_isArray:
3645 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3646 if (generate_array_guard(kls, region) != NULL)
3647 // A guard was added. If the guard is taken, it was an array.
3648 phi->add_req(intcon(1));
3649 // If we fall through, it's a plain class.
3650 query_value = intcon(0);
3651 break;
3652
3653 case vmIntrinsics::_isPrimitive:
3654 query_value = intcon(0); // "normal" path produces false
3655 break;
3656
3657 case vmIntrinsics::_getSuperclass:
3658 // The rules here are somewhat unfortunate, but we can still do better
3659 // with random logic than with a JNI call.
3660 // Interfaces store null or Object as _super, but must report null.
3661 // Arrays store an intermediate super as _super, but must report Object.
3662 // Other types can report the actual _super.
3663 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3664 if (generate_interface_guard(kls, region) != NULL)
3665 // A guard was added. If the guard is taken, it was an interface.
3666 phi->add_req(null());
3667 if (generate_array_guard(kls, region) != NULL)
3668 // A guard was added. If the guard is taken, it was an array.
3669 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3670 // If we fall through, it's a plain class. Get its _super.
3671 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3672 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3673 null_ctl = top();
3674 kls = null_check_oop(kls, &null_ctl);
3675 if (null_ctl != top()) {
3676 // If the guard is taken, Object.superClass is null (both klass and mirror).
3677 region->add_req(null_ctl);
3678 phi ->add_req(null());
3679 }
3680 if (!stopped()) {
3681 query_value = load_mirror_from_klass(kls);
3682 }
3683 break;
3684
3685 case vmIntrinsics::_getComponentType:
3686 if (generate_array_guard(kls, region) != NULL) {
3687 // Be sure to pin the oop load to the guard edge just created:
3688 Node* is_array_ctrl = region->in(region->req()-1);
3689 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3690 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3691 phi->add_req(cmo);
3692 }
3693 query_value = null(); // non-array case is null
3694 break;
3695
3696 case vmIntrinsics::_getClassAccessFlags:
3697 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3698 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3699 break;
3700
3701 default:
3702 fatal_unexpected_iid(id);
3703 break;
3704 }
3705
3706 // Fall-through is the normal case of a query to a real class.
3707 phi->init_req(1, query_value);
3708 region->init_req(1, control());
3709
3710 C->set_has_split_ifs(true); // Has chance for split-if optimization
3711 set_result(region, phi);
3712 return true;
3713 }
3714
3715 //--------------------------inline_native_subtype_check------------------------
3716 // This intrinsic takes the JNI calls out of the heart of
3717 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3718 bool LibraryCallKit::inline_native_subtype_check() {
3719 // Pull both arguments off the stack.
3720 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3721 args[0] = argument(0);
3722 args[1] = argument(1);
3723 Node* klasses[2]; // corresponding Klasses: superk, subk
3724 klasses[0] = klasses[1] = top();
3725
3726 enum {
3727 // A full decision tree on {superc is prim, subc is prim}:
3728 _prim_0_path = 1, // {P,N} => false
3729 // {P,P} & superc!=subc => false
3730 _prim_same_path, // {P,P} & superc==subc => true
3731 _prim_1_path, // {N,P} => false
3732 _ref_subtype_path, // {N,N} & subtype check wins => true
3733 _both_ref_path, // {N,N} & subtype check loses => false
3734 PATH_LIMIT
3735 };
3736
3737 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3738 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3739 record_for_igvn(region);
3740
3741 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3742 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3743 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3744
3745 // First null-check both mirrors and load each mirror's klass metaobject.
3746 int which_arg;
3747 for (which_arg = 0; which_arg <= 1; which_arg++) {
3748 Node* arg = args[which_arg];
3749 arg = null_check(arg);
3750 if (stopped()) break;
3751 args[which_arg] = arg;
3752
3753 Node* p = basic_plus_adr(arg, class_klass_offset);
3754 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3755 klasses[which_arg] = _gvn.transform(kls);
3756 }
3757
3758 // Having loaded both klasses, test each for null.
3759 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3760 for (which_arg = 0; which_arg <= 1; which_arg++) {
3761 Node* kls = klasses[which_arg];
3762 Node* null_ctl = top();
3763 kls = null_check_oop(kls, &null_ctl, never_see_null);
3764 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3765 region->init_req(prim_path, null_ctl);
3766 if (stopped()) break;
3767 klasses[which_arg] = kls;
3768 }
3769
3770 if (!stopped()) {
3771 // now we have two reference types, in klasses[0..1]
3772 Node* subk = klasses[1]; // the argument to isAssignableFrom
3773 Node* superk = klasses[0]; // the receiver
3774 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3775 // now we have a successful reference subtype check
3776 region->set_req(_ref_subtype_path, control());
3777 }
3778
3779 // If both operands are primitive (both klasses null), then
3780 // we must return true when they are identical primitives.
3781 // It is convenient to test this after the first null klass check.
3782 set_control(region->in(_prim_0_path)); // go back to first null check
3783 if (!stopped()) {
3784 // Since superc is primitive, make a guard for the superc==subc case.
3785 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3786 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3787 generate_guard(bol_eq, region, PROB_FAIR);
3788 if (region->req() == PATH_LIMIT+1) {
3789 // A guard was added. If the added guard is taken, superc==subc.
3790 region->swap_edges(PATH_LIMIT, _prim_same_path);
3791 region->del_req(PATH_LIMIT);
3792 }
3793 region->set_req(_prim_0_path, control()); // Not equal after all.
3794 }
3795
3796 // these are the only paths that produce 'true':
3797 phi->set_req(_prim_same_path, intcon(1));
3798 phi->set_req(_ref_subtype_path, intcon(1));
3799
3800 // pull together the cases:
3801 assert(region->req() == PATH_LIMIT, "sane region");
3802 for (uint i = 1; i < region->req(); i++) {
3803 Node* ctl = region->in(i);
3804 if (ctl == NULL || ctl == top()) {
3805 region->set_req(i, top());
3806 phi ->set_req(i, top());
3807 } else if (phi->in(i) == NULL) {
3808 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3809 }
3810 }
3811
3812 set_control(_gvn.transform(region));
3813 set_result(_gvn.transform(phi));
3814 return true;
3815 }
3816
3817 //---------------------generate_array_guard_common------------------------
3818 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3819 bool obj_array, bool not_array) {
3820 // If obj_array/non_array==false/false:
3821 // Branch around if the given klass is in fact an array (either obj or prim).
3822 // If obj_array/non_array==false/true:
3823 // Branch around if the given klass is not an array klass of any kind.
3824 // If obj_array/non_array==true/true:
3825 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3826 // If obj_array/non_array==true/false:
3827 // Branch around if the kls is an oop array (Object[] or subtype)
3828 //
3829 // Like generate_guard, adds a new path onto the region.
3830 jint layout_con = 0;
3831 Node* layout_val = get_layout_helper(kls, layout_con);
3832 if (layout_val == NULL) {
3833 bool query = (obj_array
3834 ? Klass::layout_helper_is_objArray(layout_con)
3835 : Klass::layout_helper_is_array(layout_con));
3836 if (query == not_array) {
3837 return NULL; // never a branch
3838 } else { // always a branch
3839 Node* always_branch = control();
3840 if (region != NULL)
3841 region->add_req(always_branch);
3842 set_control(top());
3843 return always_branch;
3844 }
3845 }
3846 // Now test the correct condition.
3847 jint nval = (obj_array
3848 ? (jint)(Klass::_lh_array_tag_type_value
3849 << Klass::_lh_array_tag_shift)
3850 : Klass::_lh_neutral_value);
3851 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3852 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3853 // invert the test if we are looking for a non-array
3854 if (not_array) btest = BoolTest(btest).negate();
3855 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3856 return generate_fair_guard(bol, region);
3857 }
3858
3859
3860 //-----------------------inline_native_newArray--------------------------
3861 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3862 bool LibraryCallKit::inline_native_newArray() {
3863 Node* mirror = argument(0);
3864 Node* count_val = argument(1);
3865
3866 mirror = null_check(mirror);
3867 // If mirror or obj is dead, only null-path is taken.
3868 if (stopped()) return true;
3869
3870 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3871 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3872 PhiNode* result_val = new(C) PhiNode(result_reg,
3873 TypeInstPtr::NOTNULL);
3874 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3875 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3876 TypePtr::BOTTOM);
3877
3878 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3879 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3880 result_reg, _slow_path);
3881 Node* normal_ctl = control();
3882 Node* no_array_ctl = result_reg->in(_slow_path);
3883
3884 // Generate code for the slow case. We make a call to newArray().
3885 set_control(no_array_ctl);
3886 if (!stopped()) {
3887 // Either the input type is void.class, or else the
3888 // array klass has not yet been cached. Either the
3889 // ensuing call will throw an exception, or else it
3890 // will cache the array klass for next time.
3891 PreserveJVMState pjvms(this);
3892 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3893 Node* slow_result = set_results_for_java_call(slow_call);
3894 // this->control() comes from set_results_for_java_call
3895 result_reg->set_req(_slow_path, control());
3896 result_val->set_req(_slow_path, slow_result);
3897 result_io ->set_req(_slow_path, i_o());
3898 result_mem->set_req(_slow_path, reset_memory());
3899 }
3900
3901 set_control(normal_ctl);
3902 if (!stopped()) {
3903 // Normal case: The array type has been cached in the java.lang.Class.
3904 // The following call works fine even if the array type is polymorphic.
3905 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3906 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3907 result_reg->init_req(_normal_path, control());
3908 result_val->init_req(_normal_path, obj);
3909 result_io ->init_req(_normal_path, i_o());
3910 result_mem->init_req(_normal_path, reset_memory());
3911 }
3912
3913 // Return the combined state.
3914 set_i_o( _gvn.transform(result_io) );
3915 set_all_memory( _gvn.transform(result_mem));
3916
3917 C->set_has_split_ifs(true); // Has chance for split-if optimization
3918 set_result(result_reg, result_val);
3919 return true;
3920 }
3921
3922 //----------------------inline_native_getLength--------------------------
3923 // public static native int java.lang.reflect.Array.getLength(Object array);
3924 bool LibraryCallKit::inline_native_getLength() {
3925 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3926
3927 Node* array = null_check(argument(0));
3928 // If array is dead, only null-path is taken.
3929 if (stopped()) return true;
3930
3931 // Deoptimize if it is a non-array.
3932 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3933
3934 if (non_array != NULL) {
3935 PreserveJVMState pjvms(this);
3936 set_control(non_array);
3937 uncommon_trap(Deoptimization::Reason_intrinsic,
3938 Deoptimization::Action_maybe_recompile);
3939 }
3940
3941 // If control is dead, only non-array-path is taken.
3942 if (stopped()) return true;
3943
3944 // The works fine even if the array type is polymorphic.
3945 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3946 Node* result = load_array_length(array);
3947
3948 C->set_has_split_ifs(true); // Has chance for split-if optimization
3949 set_result(result);
3950 return true;
3951 }
3952
3953 //------------------------inline_array_copyOf----------------------------
3954 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3955 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3956 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3957 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3958
3959 // Get the arguments.
3960 Node* original = argument(0);
3961 Node* start = is_copyOfRange? argument(1): intcon(0);
3962 Node* end = is_copyOfRange? argument(2): argument(1);
3963 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3964
3965 Node* newcopy = NULL;
3966
3967 // Set the original stack and the reexecute bit for the interpreter to reexecute
3968 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3969 { PreserveReexecuteState preexecs(this);
3970 jvms()->set_should_reexecute(true);
3971
3972 array_type_mirror = null_check(array_type_mirror);
3973 original = null_check(original);
3974
3975 // Check if a null path was taken unconditionally.
3976 if (stopped()) return true;
3977
3978 Node* orig_length = load_array_length(original);
3979
3980 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3981 klass_node = null_check(klass_node);
3982
3983 RegionNode* bailout = new (C) RegionNode(1);
3984 record_for_igvn(bailout);
3985
3986 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3987 // Bail out if that is so.
3988 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3989 if (not_objArray != NULL) {
3990 // Improve the klass node's type from the new optimistic assumption:
3991 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3992 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3993 Node* cast = new (C) CastPPNode(klass_node, akls);
3994 cast->init_req(0, control());
3995 klass_node = _gvn.transform(cast);
3996 }
3997
3998 // Bail out if either start or end is negative.
3999 generate_negative_guard(start, bailout, &start);
4000 generate_negative_guard(end, bailout, &end);
4001
4002 Node* length = end;
4003 if (_gvn.type(start) != TypeInt::ZERO) {
4004 length = _gvn.transform(new (C) SubINode(end, start));
4005 }
4006
4007 // Bail out if length is negative.
4008 // Without this the new_array would throw
4009 // NegativeArraySizeException but IllegalArgumentException is what
4010 // should be thrown
4011 generate_negative_guard(length, bailout, &length);
4012
4013 if (bailout->req() > 1) {
4014 PreserveJVMState pjvms(this);
4015 set_control(_gvn.transform(bailout));
4016 uncommon_trap(Deoptimization::Reason_intrinsic,
4017 Deoptimization::Action_maybe_recompile);
4018 }
4019
4020 if (!stopped()) {
4021 // How many elements will we copy from the original?
4022 // The answer is MinI(orig_length - start, length).
4023 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
4024 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4025
4026 newcopy = new_array(klass_node, length, 0); // no argments to push
4027
4028 // Generate a direct call to the right arraycopy function(s).
4029 // We know the copy is disjoint but we might not know if the
4030 // oop stores need checking.
4031 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4032 // This will fail a store-check if x contains any non-nulls.
4033 bool disjoint_bases = true;
4034 // if start > orig_length then the length of the copy may be
4035 // negative.
4036 bool length_never_negative = !is_copyOfRange;
4037 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4038 original, start, newcopy, intcon(0), moved,
4039 disjoint_bases, length_never_negative);
4040 }
4041 } // original reexecute is set back here
4042
4043 C->set_has_split_ifs(true); // Has chance for split-if optimization
4044 if (!stopped()) {
4045 set_result(newcopy);
4046 }
4047 return true;
4048 }
4049
4050
4051 //----------------------generate_virtual_guard---------------------------
4052 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4053 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4054 RegionNode* slow_region) {
4055 ciMethod* method = callee();
4056 int vtable_index = method->vtable_index();
4057 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4058 err_msg_res("bad index %d", vtable_index));
4059 // Get the Method* out of the appropriate vtable entry.
4060 int entry_offset = (InstanceKlass::vtable_start_offset() +
4061 vtable_index*vtableEntry::size()) * wordSize +
4062 vtableEntry::method_offset_in_bytes();
4063 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4064 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4065
4066 // Compare the target method with the expected method (e.g., Object.hashCode).
4067 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4068
4069 Node* native_call = makecon(native_call_addr);
4070 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
4071 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
4072
4073 return generate_slow_guard(test_native, slow_region);
4074 }
4075
4076 //-----------------------generate_method_call----------------------------
4077 // Use generate_method_call to make a slow-call to the real
4078 // method if the fast path fails. An alternative would be to
4079 // use a stub like OptoRuntime::slow_arraycopy_Java.
4080 // This only works for expanding the current library call,
4081 // not another intrinsic. (E.g., don't use this for making an
4082 // arraycopy call inside of the copyOf intrinsic.)
4083 CallJavaNode*
4084 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4085 // When compiling the intrinsic method itself, do not use this technique.
4086 guarantee(callee() != C->method(), "cannot make slow-call to self");
4087
4088 ciMethod* method = callee();
4089 // ensure the JVMS we have will be correct for this call
4090 guarantee(method_id == method->intrinsic_id(), "must match");
4091
4092 const TypeFunc* tf = TypeFunc::make(method);
4093 CallJavaNode* slow_call;
4094 if (is_static) {
4095 assert(!is_virtual, "");
4096 slow_call = new(C) CallStaticJavaNode(C, tf,
4097 SharedRuntime::get_resolve_static_call_stub(),
4098 method, bci());
4099 } else if (is_virtual) {
4100 null_check_receiver();
4101 int vtable_index = Method::invalid_vtable_index;
4102 if (UseInlineCaches) {
4103 // Suppress the vtable call
4104 } else {
4105 // hashCode and clone are not a miranda methods,
4106 // so the vtable index is fixed.
4107 // No need to use the linkResolver to get it.
4108 vtable_index = method->vtable_index();
4109 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4110 err_msg_res("bad index %d", vtable_index));
4111 }
4112 slow_call = new(C) CallDynamicJavaNode(tf,
4113 SharedRuntime::get_resolve_virtual_call_stub(),
4114 method, vtable_index, bci());
4115 } else { // neither virtual nor static: opt_virtual
4116 null_check_receiver();
4117 slow_call = new(C) CallStaticJavaNode(C, tf,
4118 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4119 method, bci());
4120 slow_call->set_optimized_virtual(true);
4121 }
4122 set_arguments_for_java_call(slow_call);
4123 set_edges_for_java_call(slow_call);
4124 return slow_call;
4125 }
4126
4127
4128 /**
4129 * Build special case code for calls to hashCode on an object. This call may
4130 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4131 * slightly different code.
4132 */
4133 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4134 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4135 assert(!(is_virtual && is_static), "either virtual, special, or static");
4136
4137 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4138
4139 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4140 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
4141 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4142 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4143 Node* obj = NULL;
4144 if (!is_static) {
4145 // Check for hashing null object
4146 obj = null_check_receiver();
4147 if (stopped()) return true; // unconditionally null
4148 result_reg->init_req(_null_path, top());
4149 result_val->init_req(_null_path, top());
4150 } else {
4151 // Do a null check, and return zero if null.
4152 // System.identityHashCode(null) == 0
4153 obj = argument(0);
4154 Node* null_ctl = top();
4155 obj = null_check_oop(obj, &null_ctl);
4156 result_reg->init_req(_null_path, null_ctl);
4157 result_val->init_req(_null_path, _gvn.intcon(0));
4158 }
4159
4160 // Unconditionally null? Then return right away.
4161 if (stopped()) {
4162 set_control( result_reg->in(_null_path));
4163 if (!stopped())
4164 set_result(result_val->in(_null_path));
4165 return true;
4166 }
4167
4168 // We only go to the fast case code if we pass a number of guards. The
4169 // paths which do not pass are accumulated in the slow_region.
4170 RegionNode* slow_region = new (C) RegionNode(1);
4171 record_for_igvn(slow_region);
4172
4173 // If this is a virtual call, we generate a funny guard. We pull out
4174 // the vtable entry corresponding to hashCode() from the target object.
4175 // If the target method which we are calling happens to be the native
4176 // Object hashCode() method, we pass the guard. We do not need this
4177 // guard for non-virtual calls -- the caller is known to be the native
4178 // Object hashCode().
4179 if (is_virtual) {
4180 // After null check, get the object's klass.
4181 Node* obj_klass = load_object_klass(obj);
4182 generate_virtual_guard(obj_klass, slow_region);
4183 }
4184
4185 // Get the header out of the object, use LoadMarkNode when available
4186 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4187 // The control of the load must be NULL. Otherwise, the load can move before
4188 // the null check after castPP removal.
4189 Node* no_ctrl = NULL;
4190 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4191
4192 // Test the header to see if it is unlocked.
4193 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4194 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4195 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4196 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4197 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4198
4199 generate_slow_guard(test_unlocked, slow_region);
4200
4201 // Get the hash value and check to see that it has been properly assigned.
4202 // We depend on hash_mask being at most 32 bits and avoid the use of
4203 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4204 // vm: see markOop.hpp.
4205 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4206 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4207 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4208 // This hack lets the hash bits live anywhere in the mark object now, as long
4209 // as the shift drops the relevant bits into the low 32 bits. Note that
4210 // Java spec says that HashCode is an int so there's no point in capturing
4211 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4212 hshifted_header = ConvX2I(hshifted_header);
4213 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4214
4215 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4216 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4217 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4218
4219 generate_slow_guard(test_assigned, slow_region);
4220
4221 Node* init_mem = reset_memory();
4222 // fill in the rest of the null path:
4223 result_io ->init_req(_null_path, i_o());
4224 result_mem->init_req(_null_path, init_mem);
4225
4226 result_val->init_req(_fast_path, hash_val);
4227 result_reg->init_req(_fast_path, control());
4228 result_io ->init_req(_fast_path, i_o());
4229 result_mem->init_req(_fast_path, init_mem);
4230
4231 // Generate code for the slow case. We make a call to hashCode().
4232 set_control(_gvn.transform(slow_region));
4233 if (!stopped()) {
4234 // No need for PreserveJVMState, because we're using up the present state.
4235 set_all_memory(init_mem);
4236 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4237 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4238 Node* slow_result = set_results_for_java_call(slow_call);
4239 // this->control() comes from set_results_for_java_call
4240 result_reg->init_req(_slow_path, control());
4241 result_val->init_req(_slow_path, slow_result);
4242 result_io ->set_req(_slow_path, i_o());
4243 result_mem ->set_req(_slow_path, reset_memory());
4244 }
4245
4246 // Return the combined state.
4247 set_i_o( _gvn.transform(result_io) );
4248 set_all_memory( _gvn.transform(result_mem));
4249
4250 set_result(result_reg, result_val);
4251 return true;
4252 }
4253
4254 //---------------------------inline_native_getClass----------------------------
4255 // public final native Class<?> java.lang.Object.getClass();
4256 //
4257 // Build special case code for calls to getClass on an object.
4258 bool LibraryCallKit::inline_native_getClass() {
4259 Node* obj = null_check_receiver();
4260 if (stopped()) return true;
4261 set_result(load_mirror_from_klass(load_object_klass(obj)));
4262 return true;
4263 }
4264
4265 //-----------------inline_native_Reflection_getCallerClass---------------------
4266 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4267 //
4268 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4269 //
4270 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4271 // in that it must skip particular security frames and checks for
4272 // caller sensitive methods.
4273 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4274 #ifndef PRODUCT
4275 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4276 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4277 }
4278 #endif
4279
4280 if (!jvms()->has_method()) {
4281 #ifndef PRODUCT
4282 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4283 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4284 }
4285 #endif
4286 return false;
4287 }
4288
4289 // Walk back up the JVM state to find the caller at the required
4290 // depth.
4291 JVMState* caller_jvms = jvms();
4292
4293 // Cf. JVM_GetCallerClass
4294 // NOTE: Start the loop at depth 1 because the current JVM state does
4295 // not include the Reflection.getCallerClass() frame.
4296 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4297 ciMethod* m = caller_jvms->method();
4298 switch (n) {
4299 case 0:
4300 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4301 break;
4302 case 1:
4303 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4304 if (!m->caller_sensitive()) {
4305 #ifndef PRODUCT
4306 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4307 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4308 }
4309 #endif
4310 return false; // bail-out; let JVM_GetCallerClass do the work
4311 }
4312 break;
4313 default:
4314 if (!m->is_ignored_by_security_stack_walk()) {
4315 // We have reached the desired frame; return the holder class.
4316 // Acquire method holder as java.lang.Class and push as constant.
4317 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4318 ciInstance* caller_mirror = caller_klass->java_mirror();
4319 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4320
4321 #ifndef PRODUCT
4322 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4323 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());
4324 tty->print_cr(" JVM state at this point:");
4325 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4326 ciMethod* m = jvms()->of_depth(i)->method();
4327 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4328 }
4329 }
4330 #endif
4331 return true;
4332 }
4333 break;
4334 }
4335 }
4336
4337 #ifndef PRODUCT
4338 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4339 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4340 tty->print_cr(" JVM state at this point:");
4341 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4342 ciMethod* m = jvms()->of_depth(i)->method();
4343 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4344 }
4345 }
4346 #endif
4347
4348 return false; // bail-out; let JVM_GetCallerClass do the work
4349 }
4350
4351 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4352 Node* arg = argument(0);
4353 Node* result = NULL;
4354
4355 switch (id) {
4356 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4357 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4358 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4359 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4360
4361 case vmIntrinsics::_doubleToLongBits: {
4362 // two paths (plus control) merge in a wood
4363 RegionNode *r = new (C) RegionNode(3);
4364 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4365
4366 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4367 // Build the boolean node
4368 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4369
4370 // Branch either way.
4371 // NaN case is less traveled, which makes all the difference.
4372 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4373 Node *opt_isnan = _gvn.transform(ifisnan);
4374 assert( opt_isnan->is_If(), "Expect an IfNode");
4375 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4376 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4377
4378 set_control(iftrue);
4379
4380 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4381 Node *slow_result = longcon(nan_bits); // return NaN
4382 phi->init_req(1, _gvn.transform( slow_result ));
4383 r->init_req(1, iftrue);
4384
4385 // Else fall through
4386 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4387 set_control(iffalse);
4388
4389 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4390 r->init_req(2, iffalse);
4391
4392 // Post merge
4393 set_control(_gvn.transform(r));
4394 record_for_igvn(r);
4395
4396 C->set_has_split_ifs(true); // Has chance for split-if optimization
4397 result = phi;
4398 assert(result->bottom_type()->isa_long(), "must be");
4399 break;
4400 }
4401
4402 case vmIntrinsics::_floatToIntBits: {
4403 // two paths (plus control) merge in a wood
4404 RegionNode *r = new (C) RegionNode(3);
4405 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4406
4407 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4408 // Build the boolean node
4409 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4410
4411 // Branch either way.
4412 // NaN case is less traveled, which makes all the difference.
4413 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4414 Node *opt_isnan = _gvn.transform(ifisnan);
4415 assert( opt_isnan->is_If(), "Expect an IfNode");
4416 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4417 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4418
4419 set_control(iftrue);
4420
4421 static const jint nan_bits = 0x7fc00000;
4422 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4423 phi->init_req(1, _gvn.transform( slow_result ));
4424 r->init_req(1, iftrue);
4425
4426 // Else fall through
4427 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4428 set_control(iffalse);
4429
4430 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4431 r->init_req(2, iffalse);
4432
4433 // Post merge
4434 set_control(_gvn.transform(r));
4435 record_for_igvn(r);
4436
4437 C->set_has_split_ifs(true); // Has chance for split-if optimization
4438 result = phi;
4439 assert(result->bottom_type()->isa_int(), "must be");
4440 break;
4441 }
4442
4443 default:
4444 fatal_unexpected_iid(id);
4445 break;
4446 }
4447 set_result(_gvn.transform(result));
4448 return true;
4449 }
4450
4451 #ifdef _LP64
4452 #define XTOP ,top() /*additional argument*/
4453 #else //_LP64
4454 #define XTOP /*no additional argument*/
4455 #endif //_LP64
4456
4457 //----------------------inline_unsafe_copyMemory-------------------------
4458 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4459 bool LibraryCallKit::inline_unsafe_copyMemory() {
4460 if (callee()->is_static()) return false; // caller must have the capability!
4461 null_check_receiver(); // null-check receiver
4462 if (stopped()) return true;
4463
4464 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4465
4466 Node* src_ptr = argument(1); // type: oop
4467 Node* src_off = ConvL2X(argument(2)); // type: long
4468 Node* dst_ptr = argument(4); // type: oop
4469 Node* dst_off = ConvL2X(argument(5)); // type: long
4470 Node* size = ConvL2X(argument(7)); // type: long
4471
4472 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4473 "fieldOffset must be byte-scaled");
4474
4475 Node* src = make_unsafe_address(src_ptr, src_off);
4476 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4477
4478 // Conservatively insert a memory barrier on all memory slices.
4479 // Do not let writes of the copy source or destination float below the copy.
4480 insert_mem_bar(Op_MemBarCPUOrder);
4481
4482 // Call it. Note that the length argument is not scaled.
4483 make_runtime_call(RC_LEAF|RC_NO_FP,
4484 OptoRuntime::fast_arraycopy_Type(),
4485 StubRoutines::unsafe_arraycopy(),
4486 "unsafe_arraycopy",
4487 TypeRawPtr::BOTTOM,
4488 src, dst, size XTOP);
4489
4490 // Do not let reads of the copy destination float above the copy.
4491 insert_mem_bar(Op_MemBarCPUOrder);
4492
4493 return true;
4494 }
4495
4496 //------------------------clone_coping-----------------------------------
4497 // Helper function for inline_native_clone.
4498 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4499 assert(obj_size != NULL, "");
4500 Node* raw_obj = alloc_obj->in(1);
4501 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4502
4503 AllocateNode* alloc = NULL;
4504 if (ReduceBulkZeroing) {
4505 // We will be completely responsible for initializing this object -
4506 // mark Initialize node as complete.
4507 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4508 // The object was just allocated - there should be no any stores!
4509 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4510 // Mark as complete_with_arraycopy so that on AllocateNode
4511 // expansion, we know this AllocateNode is initialized by an array
4512 // copy and a StoreStore barrier exists after the array copy.
4513 alloc->initialization()->set_complete_with_arraycopy();
4514 }
4515
4516 // Copy the fastest available way.
4517 // TODO: generate fields copies for small objects instead.
4518 Node* src = obj;
4519 Node* dest = alloc_obj;
4520 Node* size = _gvn.transform(obj_size);
4521
4522 // Exclude the header but include array length to copy by 8 bytes words.
4523 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4524 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4525 instanceOopDesc::base_offset_in_bytes();
4526 // base_off:
4527 // 8 - 32-bit VM
4528 // 12 - 64-bit VM, compressed klass
4529 // 16 - 64-bit VM, normal klass
4530 if (base_off % BytesPerLong != 0) {
4531 assert(UseCompressedClassPointers, "");
4532 if (is_array) {
4533 // Exclude length to copy by 8 bytes words.
4534 base_off += sizeof(int);
4535 } else {
4536 // Include klass to copy by 8 bytes words.
4537 base_off = instanceOopDesc::klass_offset_in_bytes();
4538 }
4539 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4540 }
4541 src = basic_plus_adr(src, base_off);
4542 dest = basic_plus_adr(dest, base_off);
4543
4544 // Compute the length also, if needed:
4545 Node* countx = size;
4546 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4547 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4548
4549 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4550 bool disjoint_bases = true;
4551 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4552 src, NULL, dest, NULL, countx,
4553 /*dest_uninitialized*/true);
4554
4555 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4556 if (card_mark) {
4557 assert(!is_array, "");
4558 // Put in store barrier for any and all oops we are sticking
4559 // into this object. (We could avoid this if we could prove
4560 // that the object type contains no oop fields at all.)
4561 Node* no_particular_value = NULL;
4562 Node* no_particular_field = NULL;
4563 int raw_adr_idx = Compile::AliasIdxRaw;
4564 post_barrier(control(),
4565 memory(raw_adr_type),
4566 alloc_obj,
4567 no_particular_field,
4568 raw_adr_idx,
4569 no_particular_value,
4570 T_OBJECT,
4571 false);
4572 }
4573
4574 // Do not let reads from the cloned object float above the arraycopy.
4575 if (alloc != NULL) {
4576 // Do not let stores that initialize this object be reordered with
4577 // a subsequent store that would make this object accessible by
4578 // other threads.
4579 // Record what AllocateNode this StoreStore protects so that
4580 // escape analysis can go from the MemBarStoreStoreNode to the
4581 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4582 // based on the escape status of the AllocateNode.
4583 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4584 } else {
4585 insert_mem_bar(Op_MemBarCPUOrder);
4586 }
4587 }
4588
4589 //------------------------inline_native_clone----------------------------
4590 // protected native Object java.lang.Object.clone();
4591 //
4592 // Here are the simple edge cases:
4593 // null receiver => normal trap
4594 // virtual and clone was overridden => slow path to out-of-line clone
4595 // not cloneable or finalizer => slow path to out-of-line Object.clone
4596 //
4597 // The general case has two steps, allocation and copying.
4598 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4599 //
4600 // Copying also has two cases, oop arrays and everything else.
4601 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4602 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4603 //
4604 // These steps fold up nicely if and when the cloned object's klass
4605 // can be sharply typed as an object array, a type array, or an instance.
4606 //
4607 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4608 PhiNode* result_val;
4609
4610 // Set the reexecute bit for the interpreter to reexecute
4611 // the bytecode that invokes Object.clone if deoptimization happens.
4612 { PreserveReexecuteState preexecs(this);
4613 jvms()->set_should_reexecute(true);
4614
4615 Node* obj = null_check_receiver();
4616 if (stopped()) return true;
4617
4618 Node* obj_klass = load_object_klass(obj);
4619 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4620 const TypeOopPtr* toop = ((tklass != NULL)
4621 ? tklass->as_instance_type()
4622 : TypeInstPtr::NOTNULL);
4623
4624 // Conservatively insert a memory barrier on all memory slices.
4625 // Do not let writes into the original float below the clone.
4626 insert_mem_bar(Op_MemBarCPUOrder);
4627
4628 // paths into result_reg:
4629 enum {
4630 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4631 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4632 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4633 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4634 PATH_LIMIT
4635 };
4636 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4637 result_val = new(C) PhiNode(result_reg,
4638 TypeInstPtr::NOTNULL);
4639 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4640 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4641 TypePtr::BOTTOM);
4642 record_for_igvn(result_reg);
4643
4644 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4645 int raw_adr_idx = Compile::AliasIdxRaw;
4646
4647 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4648 if (array_ctl != NULL) {
4649 // It's an array.
4650 PreserveJVMState pjvms(this);
4651 set_control(array_ctl);
4652 Node* obj_length = load_array_length(obj);
4653 Node* obj_size = NULL;
4654 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4655
4656 if (!use_ReduceInitialCardMarks()) {
4657 // If it is an oop array, it requires very special treatment,
4658 // because card marking is required on each card of the array.
4659 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4660 if (is_obja != NULL) {
4661 PreserveJVMState pjvms2(this);
4662 set_control(is_obja);
4663 // Generate a direct call to the right arraycopy function(s).
4664 bool disjoint_bases = true;
4665 bool length_never_negative = true;
4666 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4667 obj, intcon(0), alloc_obj, intcon(0),
4668 obj_length,
4669 disjoint_bases, length_never_negative);
4670 result_reg->init_req(_objArray_path, control());
4671 result_val->init_req(_objArray_path, alloc_obj);
4672 result_i_o ->set_req(_objArray_path, i_o());
4673 result_mem ->set_req(_objArray_path, reset_memory());
4674 }
4675 }
4676 // Otherwise, there are no card marks to worry about.
4677 // (We can dispense with card marks if we know the allocation
4678 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4679 // causes the non-eden paths to take compensating steps to
4680 // simulate a fresh allocation, so that no further
4681 // card marks are required in compiled code to initialize
4682 // the object.)
4683
4684 if (!stopped()) {
4685 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4686
4687 // Present the results of the copy.
4688 result_reg->init_req(_array_path, control());
4689 result_val->init_req(_array_path, alloc_obj);
4690 result_i_o ->set_req(_array_path, i_o());
4691 result_mem ->set_req(_array_path, reset_memory());
4692 }
4693 }
4694
4695 // We only go to the instance fast case code if we pass a number of guards.
4696 // The paths which do not pass are accumulated in the slow_region.
4697 RegionNode* slow_region = new (C) RegionNode(1);
4698 record_for_igvn(slow_region);
4699 if (!stopped()) {
4700 // It's an instance (we did array above). Make the slow-path tests.
4701 // If this is a virtual call, we generate a funny guard. We grab
4702 // the vtable entry corresponding to clone() from the target object.
4703 // If the target method which we are calling happens to be the
4704 // Object clone() method, we pass the guard. We do not need this
4705 // guard for non-virtual calls; the caller is known to be the native
4706 // Object clone().
4707 if (is_virtual) {
4708 generate_virtual_guard(obj_klass, slow_region);
4709 }
4710
4711 // The object must be cloneable and must not have a finalizer.
4712 // Both of these conditions may be checked in a single test.
4713 // We could optimize the cloneable test further, but we don't care.
4714 generate_access_flags_guard(obj_klass,
4715 // Test both conditions:
4716 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4717 // Must be cloneable but not finalizer:
4718 JVM_ACC_IS_CLONEABLE,
4719 slow_region);
4720 }
4721
4722 if (!stopped()) {
4723 // It's an instance, and it passed the slow-path tests.
4724 PreserveJVMState pjvms(this);
4725 Node* obj_size = NULL;
4726 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4727 // is reexecuted if deoptimization occurs and there could be problems when merging
4728 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4729 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4730
4731 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4732
4733 // Present the results of the slow call.
4734 result_reg->init_req(_instance_path, control());
4735 result_val->init_req(_instance_path, alloc_obj);
4736 result_i_o ->set_req(_instance_path, i_o());
4737 result_mem ->set_req(_instance_path, reset_memory());
4738 }
4739
4740 // Generate code for the slow case. We make a call to clone().
4741 set_control(_gvn.transform(slow_region));
4742 if (!stopped()) {
4743 PreserveJVMState pjvms(this);
4744 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4745 Node* slow_result = set_results_for_java_call(slow_call);
4746 // this->control() comes from set_results_for_java_call
4747 result_reg->init_req(_slow_path, control());
4748 result_val->init_req(_slow_path, slow_result);
4749 result_i_o ->set_req(_slow_path, i_o());
4750 result_mem ->set_req(_slow_path, reset_memory());
4751 }
4752
4753 // Return the combined state.
4754 set_control( _gvn.transform(result_reg));
4755 set_i_o( _gvn.transform(result_i_o));
4756 set_all_memory( _gvn.transform(result_mem));
4757 } // original reexecute is set back here
4758
4759 set_result(_gvn.transform(result_val));
4760 return true;
4761 }
4762
4763 //------------------------------basictype2arraycopy----------------------------
4764 address LibraryCallKit::basictype2arraycopy(BasicType t,
4765 Node* src_offset,
4766 Node* dest_offset,
4767 bool disjoint_bases,
4768 const char* &name,
4769 bool dest_uninitialized) {
4770 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4771 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4772
4773 bool aligned = false;
4774 bool disjoint = disjoint_bases;
4775
4776 // if the offsets are the same, we can treat the memory regions as
4777 // disjoint, because either the memory regions are in different arrays,
4778 // or they are identical (which we can treat as disjoint.) We can also
4779 // treat a copy with a destination index less that the source index
4780 // as disjoint since a low->high copy will work correctly in this case.
4781 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4782 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4783 // both indices are constants
4784 int s_offs = src_offset_inttype->get_con();
4785 int d_offs = dest_offset_inttype->get_con();
4786 int element_size = type2aelembytes(t);
4787 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4788 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4789 if (s_offs >= d_offs) disjoint = true;
4790 } else if (src_offset == dest_offset && src_offset != NULL) {
4791 // This can occur if the offsets are identical non-constants.
4792 disjoint = true;
4793 }
4794
4795 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4796 }
4797
4798
4799 //------------------------------inline_arraycopy-----------------------
4800 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4801 // Object dest, int destPos,
4802 // int length);
4803 bool LibraryCallKit::inline_arraycopy() {
4804 // Get the arguments.
4805 Node* src = argument(0); // type: oop
4806 Node* src_offset = argument(1); // type: int
4807 Node* dest = argument(2); // type: oop
4808 Node* dest_offset = argument(3); // type: int
4809 Node* length = argument(4); // type: int
4810
4811 // Compile time checks. If any of these checks cannot be verified at compile time,
4812 // we do not make a fast path for this call. Instead, we let the call remain as it
4813 // is. The checks we choose to mandate at compile time are:
4814 //
4815 // (1) src and dest are arrays.
4816 const Type* src_type = src->Value(&_gvn);
4817 const Type* dest_type = dest->Value(&_gvn);
4818 const TypeAryPtr* top_src = src_type->isa_aryptr();
4819 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4820
4821 // Do we have the type of src?
4822 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4823 // Do we have the type of dest?
4824 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4825 // Is the type for src from speculation?
4826 bool src_spec = false;
4827 // Is the type for dest from speculation?
4828 bool dest_spec = false;
4829
4830 if (!has_src || !has_dest) {
4831 // We don't have sufficient type information, let's see if
4832 // speculative types can help. We need to have types for both src
4833 // and dest so that it pays off.
4834
4835 // Do we already have or could we have type information for src
4836 bool could_have_src = has_src;
4837 // Do we already have or could we have type information for dest
4838 bool could_have_dest = has_dest;
4839
4840 ciKlass* src_k = NULL;
4841 if (!has_src) {
4842 src_k = src_type->speculative_type();
4843 if (src_k != NULL && src_k->is_array_klass()) {
4844 could_have_src = true;
4845 }
4846 }
4847
4848 ciKlass* dest_k = NULL;
4849 if (!has_dest) {
4850 dest_k = dest_type->speculative_type();
4851 if (dest_k != NULL && dest_k->is_array_klass()) {
4852 could_have_dest = true;
4853 }
4854 }
4855
4856 if (could_have_src && could_have_dest) {
4857 // This is going to pay off so emit the required guards
4858 if (!has_src) {
4859 src = maybe_cast_profiled_obj(src, src_k);
4860 src_type = _gvn.type(src);
4861 top_src = src_type->isa_aryptr();
4862 has_src = (top_src != NULL && top_src->klass() != NULL);
4863 src_spec = true;
4864 }
4865 if (!has_dest) {
4866 dest = maybe_cast_profiled_obj(dest, dest_k);
4867 dest_type = _gvn.type(dest);
4868 top_dest = dest_type->isa_aryptr();
4869 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4870 dest_spec = true;
4871 }
4872 }
4873 }
4874
4875 if (!has_src || !has_dest) {
4876 // Conservatively insert a memory barrier on all memory slices.
4877 // Do not let writes into the source float below the arraycopy.
4878 insert_mem_bar(Op_MemBarCPUOrder);
4879
4880 // Call StubRoutines::generic_arraycopy stub.
4881 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4882 src, src_offset, dest, dest_offset, length);
4883
4884 // Do not let reads from the destination float above the arraycopy.
4885 // Since we cannot type the arrays, we don't know which slices
4886 // might be affected. We could restrict this barrier only to those
4887 // memory slices which pertain to array elements--but don't bother.
4888 if (!InsertMemBarAfterArraycopy)
4889 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4890 insert_mem_bar(Op_MemBarCPUOrder);
4891 return true;
4892 }
4893
4894 // (2) src and dest arrays must have elements of the same BasicType
4895 // Figure out the size and type of the elements we will be copying.
4896 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4897 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4898 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4899 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4900
4901 if (src_elem != dest_elem || dest_elem == T_VOID) {
4902 // The component types are not the same or are not recognized. Punt.
4903 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4904 generate_slow_arraycopy(TypePtr::BOTTOM,
4905 src, src_offset, dest, dest_offset, length,
4906 /*dest_uninitialized*/false);
4907 return true;
4908 }
4909
4910 if (src_elem == T_OBJECT) {
4911 // If both arrays are object arrays then having the exact types
4912 // for both will remove the need for a subtype check at runtime
4913 // before the call and may make it possible to pick a faster copy
4914 // routine (without a subtype check on every element)
4915 // Do we have the exact type of src?
4916 bool could_have_src = src_spec;
4917 // Do we have the exact type of dest?
4918 bool could_have_dest = dest_spec;
4919 ciKlass* src_k = top_src->klass();
4920 ciKlass* dest_k = top_dest->klass();
4921 if (!src_spec) {
4922 src_k = src_type->speculative_type();
4923 if (src_k != NULL && src_k->is_array_klass()) {
4924 could_have_src = true;
4925 }
4926 }
4927 if (!dest_spec) {
4928 dest_k = dest_type->speculative_type();
4929 if (dest_k != NULL && dest_k->is_array_klass()) {
4930 could_have_dest = true;
4931 }
4932 }
4933 if (could_have_src && could_have_dest) {
4934 // If we can have both exact types, emit the missing guards
4935 if (could_have_src && !src_spec) {
4936 src = maybe_cast_profiled_obj(src, src_k);
4937 }
4938 if (could_have_dest && !dest_spec) {
4939 dest = maybe_cast_profiled_obj(dest, dest_k);
4940 }
4941 }
4942 }
4943
4944 //---------------------------------------------------------------------------
4945 // We will make a fast path for this call to arraycopy.
4946
4947 // We have the following tests left to perform:
4948 //
4949 // (3) src and dest must not be null.
4950 // (4) src_offset must not be negative.
4951 // (5) dest_offset must not be negative.
4952 // (6) length must not be negative.
4953 // (7) src_offset + length must not exceed length of src.
4954 // (8) dest_offset + length must not exceed length of dest.
4955 // (9) each element of an oop array must be assignable
4956
4957 RegionNode* slow_region = new (C) RegionNode(1);
4958 record_for_igvn(slow_region);
4959
4960 // (3) operands must not be null
4961 // We currently perform our null checks with the null_check routine.
4962 // This means that the null exceptions will be reported in the caller
4963 // rather than (correctly) reported inside of the native arraycopy call.
4964 // This should be corrected, given time. We do our null check with the
4965 // stack pointer restored.
4966 src = null_check(src, T_ARRAY);
4967 dest = null_check(dest, T_ARRAY);
4968
4969 // (4) src_offset must not be negative.
4970 generate_negative_guard(src_offset, slow_region);
4971
4972 // (5) dest_offset must not be negative.
4973 generate_negative_guard(dest_offset, slow_region);
4974
4975 // (6) length must not be negative (moved to generate_arraycopy()).
4976 // generate_negative_guard(length, slow_region);
4977
4978 // (7) src_offset + length must not exceed length of src.
4979 generate_limit_guard(src_offset, length,
4980 load_array_length(src),
4981 slow_region);
4982
4983 // (8) dest_offset + length must not exceed length of dest.
4984 generate_limit_guard(dest_offset, length,
4985 load_array_length(dest),
4986 slow_region);
4987
4988 // (9) each element of an oop array must be assignable
4989 // The generate_arraycopy subroutine checks this.
4990
4991 // This is where the memory effects are placed:
4992 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4993 generate_arraycopy(adr_type, dest_elem,
4994 src, src_offset, dest, dest_offset, length,
4995 false, false, slow_region);
4996
4997 return true;
4998 }
4999
5000 //-----------------------------generate_arraycopy----------------------
5001 // Generate an optimized call to arraycopy.
5002 // Caller must guard against non-arrays.
5003 // Caller must determine a common array basic-type for both arrays.
5004 // Caller must validate offsets against array bounds.
5005 // The slow_region has already collected guard failure paths
5006 // (such as out of bounds length or non-conformable array types).
5007 // The generated code has this shape, in general:
5008 //
5009 // if (length == 0) return // via zero_path
5010 // slowval = -1
5011 // if (types unknown) {
5012 // slowval = call generic copy loop
5013 // if (slowval == 0) return // via checked_path
5014 // } else if (indexes in bounds) {
5015 // if ((is object array) && !(array type check)) {
5016 // slowval = call checked copy loop
5017 // if (slowval == 0) return // via checked_path
5018 // } else {
5019 // call bulk copy loop
5020 // return // via fast_path
5021 // }
5022 // }
5023 // // adjust params for remaining work:
5024 // if (slowval != -1) {
5025 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
5026 // }
5027 // slow_region:
5028 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
5029 // return // via slow_call_path
5030 //
5031 // This routine is used from several intrinsics: System.arraycopy,
5032 // Object.clone (the array subcase), and Arrays.copyOf[Range].
5033 //
5034 void
5035 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
5036 BasicType basic_elem_type,
5037 Node* src, Node* src_offset,
5038 Node* dest, Node* dest_offset,
5039 Node* copy_length,
5040 bool disjoint_bases,
5041 bool length_never_negative,
5042 RegionNode* slow_region) {
5043
5044 if (slow_region == NULL) {
5045 slow_region = new(C) RegionNode(1);
5046 record_for_igvn(slow_region);
5047 }
5048
5049 Node* original_dest = dest;
5050 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
5051 bool dest_uninitialized = false;
5052
5053 // See if this is the initialization of a newly-allocated array.
5054 // If so, we will take responsibility here for initializing it to zero.
5055 // (Note: Because tightly_coupled_allocation performs checks on the
5056 // out-edges of the dest, we need to avoid making derived pointers
5057 // from it until we have checked its uses.)
5058 if (ReduceBulkZeroing
5059 && !ZeroTLAB // pointless if already zeroed
5060 && basic_elem_type != T_CONFLICT // avoid corner case
5061 && !src->eqv_uncast(dest)
5062 && ((alloc = tightly_coupled_allocation(dest, slow_region))
5063 != NULL)
5064 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
5065 && alloc->maybe_set_complete(&_gvn)) {
5066 // "You break it, you buy it."
5067 InitializeNode* init = alloc->initialization();
5068 assert(init->is_complete(), "we just did this");
5069 init->set_complete_with_arraycopy();
5070 assert(dest->is_CheckCastPP(), "sanity");
5071 assert(dest->in(0)->in(0) == init, "dest pinned");
5072 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
5073 // From this point on, every exit path is responsible for
5074 // initializing any non-copied parts of the object to zero.
5075 // Also, if this flag is set we make sure that arraycopy interacts properly
5076 // with G1, eliding pre-barriers. See CR 6627983.
5077 dest_uninitialized = true;
5078 } else {
5079 // No zeroing elimination here.
5080 alloc = NULL;
5081 //original_dest = dest;
5082 //dest_uninitialized = false;
5083 }
5084
5085 // Results are placed here:
5086 enum { fast_path = 1, // normal void-returning assembly stub
5087 checked_path = 2, // special assembly stub with cleanup
5088 slow_call_path = 3, // something went wrong; call the VM
5089 zero_path = 4, // bypass when length of copy is zero
5090 bcopy_path = 5, // copy primitive array by 64-bit blocks
5091 PATH_LIMIT = 6
5092 };
5093 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
5094 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
5095 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
5096 record_for_igvn(result_region);
5097 _gvn.set_type_bottom(result_i_o);
5098 _gvn.set_type_bottom(result_memory);
5099 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
5100
5101 // The slow_control path:
5102 Node* slow_control;
5103 Node* slow_i_o = i_o();
5104 Node* slow_mem = memory(adr_type);
5105 debug_only(slow_control = (Node*) badAddress);
5106
5107 // Checked control path:
5108 Node* checked_control = top();
5109 Node* checked_mem = NULL;
5110 Node* checked_i_o = NULL;
5111 Node* checked_value = NULL;
5112
5113 if (basic_elem_type == T_CONFLICT) {
5114 assert(!dest_uninitialized, "");
5115 Node* cv = generate_generic_arraycopy(adr_type,
5116 src, src_offset, dest, dest_offset,
5117 copy_length, dest_uninitialized);
5118 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5119 checked_control = control();
5120 checked_i_o = i_o();
5121 checked_mem = memory(adr_type);
5122 checked_value = cv;
5123 set_control(top()); // no fast path
5124 }
5125
5126 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5127 if (not_pos != NULL) {
5128 PreserveJVMState pjvms(this);
5129 set_control(not_pos);
5130
5131 // (6) length must not be negative.
5132 if (!length_never_negative) {
5133 generate_negative_guard(copy_length, slow_region);
5134 }
5135
5136 // copy_length is 0.
5137 if (!stopped() && dest_uninitialized) {
5138 Node* dest_length = alloc->in(AllocateNode::ALength);
5139 if (copy_length->eqv_uncast(dest_length)
5140 || _gvn.find_int_con(dest_length, 1) <= 0) {
5141 // There is no zeroing to do. No need for a secondary raw memory barrier.
5142 } else {
5143 // Clear the whole thing since there are no source elements to copy.
5144 generate_clear_array(adr_type, dest, basic_elem_type,
5145 intcon(0), NULL,
5146 alloc->in(AllocateNode::AllocSize));
5147 // Use a secondary InitializeNode as raw memory barrier.
5148 // Currently it is needed only on this path since other
5149 // paths have stub or runtime calls as raw memory barriers.
5150 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5151 Compile::AliasIdxRaw,
5152 top())->as_Initialize();
5153 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5154 }
5155 }
5156
5157 // Present the results of the fast call.
5158 result_region->init_req(zero_path, control());
5159 result_i_o ->init_req(zero_path, i_o());
5160 result_memory->init_req(zero_path, memory(adr_type));
5161 }
5162
5163 if (!stopped() && dest_uninitialized) {
5164 // We have to initialize the *uncopied* part of the array to zero.
5165 // The copy destination is the slice dest[off..off+len]. The other slices
5166 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5167 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5168 Node* dest_length = alloc->in(AllocateNode::ALength);
5169 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5170 copy_length));
5171
5172 // If there is a head section that needs zeroing, do it now.
5173 if (find_int_con(dest_offset, -1) != 0) {
5174 generate_clear_array(adr_type, dest, basic_elem_type,
5175 intcon(0), dest_offset,
5176 NULL);
5177 }
5178
5179 // Next, perform a dynamic check on the tail length.
5180 // It is often zero, and we can win big if we prove this.
5181 // There are two wins: Avoid generating the ClearArray
5182 // with its attendant messy index arithmetic, and upgrade
5183 // the copy to a more hardware-friendly word size of 64 bits.
5184 Node* tail_ctl = NULL;
5185 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5186 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5187 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5188 tail_ctl = generate_slow_guard(bol_lt, NULL);
5189 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5190 }
5191
5192 // At this point, let's assume there is no tail.
5193 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5194 // There is no tail. Try an upgrade to a 64-bit copy.
5195 bool didit = false;
5196 { PreserveJVMState pjvms(this);
5197 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5198 src, src_offset, dest, dest_offset,
5199 dest_size, dest_uninitialized);
5200 if (didit) {
5201 // Present the results of the block-copying fast call.
5202 result_region->init_req(bcopy_path, control());
5203 result_i_o ->init_req(bcopy_path, i_o());
5204 result_memory->init_req(bcopy_path, memory(adr_type));
5205 }
5206 }
5207 if (didit)
5208 set_control(top()); // no regular fast path
5209 }
5210
5211 // Clear the tail, if any.
5212 if (tail_ctl != NULL) {
5213 Node* notail_ctl = stopped() ? NULL : control();
5214 set_control(tail_ctl);
5215 if (notail_ctl == NULL) {
5216 generate_clear_array(adr_type, dest, basic_elem_type,
5217 dest_tail, NULL,
5218 dest_size);
5219 } else {
5220 // Make a local merge.
5221 Node* done_ctl = new(C) RegionNode(3);
5222 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5223 done_ctl->init_req(1, notail_ctl);
5224 done_mem->init_req(1, memory(adr_type));
5225 generate_clear_array(adr_type, dest, basic_elem_type,
5226 dest_tail, NULL,
5227 dest_size);
5228 done_ctl->init_req(2, control());
5229 done_mem->init_req(2, memory(adr_type));
5230 set_control( _gvn.transform(done_ctl));
5231 set_memory( _gvn.transform(done_mem), adr_type );
5232 }
5233 }
5234 }
5235
5236 BasicType copy_type = basic_elem_type;
5237 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5238 if (!stopped() && copy_type == T_OBJECT) {
5239 // If src and dest have compatible element types, we can copy bits.
5240 // Types S[] and D[] are compatible if D is a supertype of S.
5241 //
5242 // If they are not, we will use checked_oop_disjoint_arraycopy,
5243 // which performs a fast optimistic per-oop check, and backs off
5244 // further to JVM_ArrayCopy on the first per-oop check that fails.
5245 // (Actually, we don't move raw bits only; the GC requires card marks.)
5246
5247 // Get the Klass* for both src and dest
5248 Node* src_klass = load_object_klass(src);
5249 Node* dest_klass = load_object_klass(dest);
5250
5251 // Generate the subtype check.
5252 // This might fold up statically, or then again it might not.
5253 //
5254 // Non-static example: Copying List<String>.elements to a new String[].
5255 // The backing store for a List<String> is always an Object[],
5256 // but its elements are always type String, if the generic types
5257 // are correct at the source level.
5258 //
5259 // Test S[] against D[], not S against D, because (probably)
5260 // the secondary supertype cache is less busy for S[] than S.
5261 // This usually only matters when D is an interface.
5262 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5263 // Plug failing path into checked_oop_disjoint_arraycopy
5264 if (not_subtype_ctrl != top()) {
5265 PreserveJVMState pjvms(this);
5266 set_control(not_subtype_ctrl);
5267 // (At this point we can assume disjoint_bases, since types differ.)
5268 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5269 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5270 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5271 Node* dest_elem_klass = _gvn.transform(n1);
5272 Node* cv = generate_checkcast_arraycopy(adr_type,
5273 dest_elem_klass,
5274 src, src_offset, dest, dest_offset,
5275 ConvI2X(copy_length), dest_uninitialized);
5276 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5277 checked_control = control();
5278 checked_i_o = i_o();
5279 checked_mem = memory(adr_type);
5280 checked_value = cv;
5281 }
5282 // At this point we know we do not need type checks on oop stores.
5283
5284 // Let's see if we need card marks:
5285 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5286 // If we do not need card marks, copy using the jint or jlong stub.
5287 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5288 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5289 "sizes agree");
5290 }
5291 }
5292
5293 if (!stopped()) {
5294 // Generate the fast path, if possible.
5295 PreserveJVMState pjvms(this);
5296 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5297 src, src_offset, dest, dest_offset,
5298 ConvI2X(copy_length), dest_uninitialized);
5299
5300 // Present the results of the fast call.
5301 result_region->init_req(fast_path, control());
5302 result_i_o ->init_req(fast_path, i_o());
5303 result_memory->init_req(fast_path, memory(adr_type));
5304 }
5305
5306 // Here are all the slow paths up to this point, in one bundle:
5307 slow_control = top();
5308 if (slow_region != NULL)
5309 slow_control = _gvn.transform(slow_region);
5310 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5311
5312 set_control(checked_control);
5313 if (!stopped()) {
5314 // Clean up after the checked call.
5315 // The returned value is either 0 or -1^K,
5316 // where K = number of partially transferred array elements.
5317 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5318 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5319 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5320
5321 // If it is 0, we are done, so transfer to the end.
5322 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5323 result_region->init_req(checked_path, checks_done);
5324 result_i_o ->init_req(checked_path, checked_i_o);
5325 result_memory->init_req(checked_path, checked_mem);
5326
5327 // If it is not zero, merge into the slow call.
5328 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5329 RegionNode* slow_reg2 = new(C) RegionNode(3);
5330 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5331 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5332 record_for_igvn(slow_reg2);
5333 slow_reg2 ->init_req(1, slow_control);
5334 slow_i_o2 ->init_req(1, slow_i_o);
5335 slow_mem2 ->init_req(1, slow_mem);
5336 slow_reg2 ->init_req(2, control());
5337 slow_i_o2 ->init_req(2, checked_i_o);
5338 slow_mem2 ->init_req(2, checked_mem);
5339
5340 slow_control = _gvn.transform(slow_reg2);
5341 slow_i_o = _gvn.transform(slow_i_o2);
5342 slow_mem = _gvn.transform(slow_mem2);
5343
5344 if (alloc != NULL) {
5345 // We'll restart from the very beginning, after zeroing the whole thing.
5346 // This can cause double writes, but that's OK since dest is brand new.
5347 // So we ignore the low 31 bits of the value returned from the stub.
5348 } else {
5349 // We must continue the copy exactly where it failed, or else
5350 // another thread might see the wrong number of writes to dest.
5351 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5352 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5353 slow_offset->init_req(1, intcon(0));
5354 slow_offset->init_req(2, checked_offset);
5355 slow_offset = _gvn.transform(slow_offset);
5356
5357 // Adjust the arguments by the conditionally incoming offset.
5358 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5359 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5360 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5361
5362 // Tweak the node variables to adjust the code produced below:
5363 src_offset = src_off_plus;
5364 dest_offset = dest_off_plus;
5365 copy_length = length_minus;
5366 }
5367 }
5368
5369 set_control(slow_control);
5370 if (!stopped()) {
5371 // Generate the slow path, if needed.
5372 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5373
5374 set_memory(slow_mem, adr_type);
5375 set_i_o(slow_i_o);
5376
5377 if (dest_uninitialized) {
5378 generate_clear_array(adr_type, dest, basic_elem_type,
5379 intcon(0), NULL,
5380 alloc->in(AllocateNode::AllocSize));
5381 }
5382
5383 generate_slow_arraycopy(adr_type,
5384 src, src_offset, dest, dest_offset,
5385 copy_length, /*dest_uninitialized*/false);
5386
5387 result_region->init_req(slow_call_path, control());
5388 result_i_o ->init_req(slow_call_path, i_o());
5389 result_memory->init_req(slow_call_path, memory(adr_type));
5390 }
5391
5392 // Remove unused edges.
5393 for (uint i = 1; i < result_region->req(); i++) {
5394 if (result_region->in(i) == NULL)
5395 result_region->init_req(i, top());
5396 }
5397
5398 // Finished; return the combined state.
5399 set_control( _gvn.transform(result_region));
5400 set_i_o( _gvn.transform(result_i_o) );
5401 set_memory( _gvn.transform(result_memory), adr_type );
5402
5403 // The memory edges above are precise in order to model effects around
5404 // array copies accurately to allow value numbering of field loads around
5405 // arraycopy. Such field loads, both before and after, are common in Java
5406 // collections and similar classes involving header/array data structures.
5407 //
5408 // But with low number of register or when some registers are used or killed
5409 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5410 // The next memory barrier is added to avoid it. If the arraycopy can be
5411 // optimized away (which it can, sometimes) then we can manually remove
5412 // the membar also.
5413 //
5414 // Do not let reads from the cloned object float above the arraycopy.
5415 if (alloc != NULL) {
5416 // Do not let stores that initialize this object be reordered with
5417 // a subsequent store that would make this object accessible by
5418 // other threads.
5419 // Record what AllocateNode this StoreStore protects so that
5420 // escape analysis can go from the MemBarStoreStoreNode to the
5421 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5422 // based on the escape status of the AllocateNode.
5423 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5424 } else if (InsertMemBarAfterArraycopy)
5425 insert_mem_bar(Op_MemBarCPUOrder);
5426 }
5427
5428
5429 // Helper function which determines if an arraycopy immediately follows
5430 // an allocation, with no intervening tests or other escapes for the object.
5431 AllocateArrayNode*
5432 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5433 RegionNode* slow_region) {
5434 if (stopped()) return NULL; // no fast path
5435 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5436
5437 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5438 if (alloc == NULL) return NULL;
5439
5440 Node* rawmem = memory(Compile::AliasIdxRaw);
5441 // Is the allocation's memory state untouched?
5442 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5443 // Bail out if there have been raw-memory effects since the allocation.
5444 // (Example: There might have been a call or safepoint.)
5445 return NULL;
5446 }
5447 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5448 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5449 return NULL;
5450 }
5451
5452 // There must be no unexpected observers of this allocation.
5453 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5454 Node* obs = ptr->fast_out(i);
5455 if (obs != this->map()) {
5456 return NULL;
5457 }
5458 }
5459
5460 // This arraycopy must unconditionally follow the allocation of the ptr.
5461 Node* alloc_ctl = ptr->in(0);
5462 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5463
5464 Node* ctl = control();
5465 while (ctl != alloc_ctl) {
5466 // There may be guards which feed into the slow_region.
5467 // Any other control flow means that we might not get a chance
5468 // to finish initializing the allocated object.
5469 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5470 IfNode* iff = ctl->in(0)->as_If();
5471 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5472 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5473 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5474 ctl = iff->in(0); // This test feeds the known slow_region.
5475 continue;
5476 }
5477 // One more try: Various low-level checks bottom out in
5478 // uncommon traps. If the debug-info of the trap omits
5479 // any reference to the allocation, as we've already
5480 // observed, then there can be no objection to the trap.
5481 bool found_trap = false;
5482 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5483 Node* obs = not_ctl->fast_out(j);
5484 if (obs->in(0) == not_ctl && obs->is_Call() &&
5485 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5486 found_trap = true; break;
5487 }
5488 }
5489 if (found_trap) {
5490 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5491 continue;
5492 }
5493 }
5494 return NULL;
5495 }
5496
5497 // If we get this far, we have an allocation which immediately
5498 // precedes the arraycopy, and we can take over zeroing the new object.
5499 // The arraycopy will finish the initialization, and provide
5500 // a new control state to which we will anchor the destination pointer.
5501
5502 return alloc;
5503 }
5504
5505 // Helper for initialization of arrays, creating a ClearArray.
5506 // It writes zero bits in [start..end), within the body of an array object.
5507 // The memory effects are all chained onto the 'adr_type' alias category.
5508 //
5509 // Since the object is otherwise uninitialized, we are free
5510 // to put a little "slop" around the edges of the cleared area,
5511 // as long as it does not go back into the array's header,
5512 // or beyond the array end within the heap.
5513 //
5514 // The lower edge can be rounded down to the nearest jint and the
5515 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5516 //
5517 // Arguments:
5518 // adr_type memory slice where writes are generated
5519 // dest oop of the destination array
5520 // basic_elem_type element type of the destination
5521 // slice_idx array index of first element to store
5522 // slice_len number of elements to store (or NULL)
5523 // dest_size total size in bytes of the array object
5524 //
5525 // Exactly one of slice_len or dest_size must be non-NULL.
5526 // If dest_size is non-NULL, zeroing extends to the end of the object.
5527 // If slice_len is non-NULL, the slice_idx value must be a constant.
5528 void
5529 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5530 Node* dest,
5531 BasicType basic_elem_type,
5532 Node* slice_idx,
5533 Node* slice_len,
5534 Node* dest_size) {
5535 // one or the other but not both of slice_len and dest_size:
5536 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5537 if (slice_len == NULL) slice_len = top();
5538 if (dest_size == NULL) dest_size = top();
5539
5540 // operate on this memory slice:
5541 Node* mem = memory(adr_type); // memory slice to operate on
5542
5543 // scaling and rounding of indexes:
5544 int scale = exact_log2(type2aelembytes(basic_elem_type));
5545 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5546 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5547 int bump_bit = (-1 << scale) & BytesPerInt;
5548
5549 // determine constant starts and ends
5550 const intptr_t BIG_NEG = -128;
5551 assert(BIG_NEG + 2*abase < 0, "neg enough");
5552 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5553 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5554 if (slice_len_con == 0) {
5555 return; // nothing to do here
5556 }
5557 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5558 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5559 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5560 assert(end_con < 0, "not two cons");
5561 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5562 BytesPerLong);
5563 }
5564
5565 if (start_con >= 0 && end_con >= 0) {
5566 // Constant start and end. Simple.
5567 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5568 start_con, end_con, &_gvn);
5569 } else if (start_con >= 0 && dest_size != top()) {
5570 // Constant start, pre-rounded end after the tail of the array.
5571 Node* end = dest_size;
5572 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5573 start_con, end, &_gvn);
5574 } else if (start_con >= 0 && slice_len != top()) {
5575 // Constant start, non-constant end. End needs rounding up.
5576 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5577 intptr_t end_base = abase + (slice_idx_con << scale);
5578 int end_round = (-1 << scale) & (BytesPerLong - 1);
5579 Node* end = ConvI2X(slice_len);
5580 if (scale != 0)
5581 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5582 end_base += end_round;
5583 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5584 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5585 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5586 start_con, end, &_gvn);
5587 } else if (start_con < 0 && dest_size != top()) {
5588 // Non-constant start, pre-rounded end after the tail of the array.
5589 // This is almost certainly a "round-to-end" operation.
5590 Node* start = slice_idx;
5591 start = ConvI2X(start);
5592 if (scale != 0)
5593 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5594 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5595 if ((bump_bit | clear_low) != 0) {
5596 int to_clear = (bump_bit | clear_low);
5597 // Align up mod 8, then store a jint zero unconditionally
5598 // just before the mod-8 boundary.
5599 if (((abase + bump_bit) & ~to_clear) - bump_bit
5600 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5601 bump_bit = 0;
5602 assert((abase & to_clear) == 0, "array base must be long-aligned");
5603 } else {
5604 // Bump 'start' up to (or past) the next jint boundary:
5605 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5606 assert((abase & clear_low) == 0, "array base must be int-aligned");
5607 }
5608 // Round bumped 'start' down to jlong boundary in body of array.
5609 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5610 if (bump_bit != 0) {
5611 // Store a zero to the immediately preceding jint:
5612 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5613 Node* p1 = basic_plus_adr(dest, x1);
5614 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5615 mem = _gvn.transform(mem);
5616 }
5617 }
5618 Node* end = dest_size; // pre-rounded
5619 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5620 start, end, &_gvn);
5621 } else {
5622 // Non-constant start, unrounded non-constant end.
5623 // (Nobody zeroes a random midsection of an array using this routine.)
5624 ShouldNotReachHere(); // fix caller
5625 }
5626
5627 // Done.
5628 set_memory(mem, adr_type);
5629 }
5630
5631
5632 bool
5633 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5634 BasicType basic_elem_type,
5635 AllocateNode* alloc,
5636 Node* src, Node* src_offset,
5637 Node* dest, Node* dest_offset,
5638 Node* dest_size, bool dest_uninitialized) {
5639 // See if there is an advantage from block transfer.
5640 int scale = exact_log2(type2aelembytes(basic_elem_type));
5641 if (scale >= LogBytesPerLong)
5642 return false; // it is already a block transfer
5643
5644 // Look at the alignment of the starting offsets.
5645 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5646
5647 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5648 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5649 if (src_off_con < 0 || dest_off_con < 0)
5650 // At present, we can only understand constants.
5651 return false;
5652
5653 intptr_t src_off = abase + (src_off_con << scale);
5654 intptr_t dest_off = abase + (dest_off_con << scale);
5655
5656 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5657 // Non-aligned; too bad.
5658 // One more chance: Pick off an initial 32-bit word.
5659 // This is a common case, since abase can be odd mod 8.
5660 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5661 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5662 Node* sptr = basic_plus_adr(src, src_off);
5663 Node* dptr = basic_plus_adr(dest, dest_off);
5664 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5665 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5666 src_off += BytesPerInt;
5667 dest_off += BytesPerInt;
5668 } else {
5669 return false;
5670 }
5671 }
5672 assert(src_off % BytesPerLong == 0, "");
5673 assert(dest_off % BytesPerLong == 0, "");
5674
5675 // Do this copy by giant steps.
5676 Node* sptr = basic_plus_adr(src, src_off);
5677 Node* dptr = basic_plus_adr(dest, dest_off);
5678 Node* countx = dest_size;
5679 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5680 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5681
5682 bool disjoint_bases = true; // since alloc != NULL
5683 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5684 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5685
5686 return true;
5687 }
5688
5689
5690 // Helper function; generates code for the slow case.
5691 // We make a call to a runtime method which emulates the native method,
5692 // but without the native wrapper overhead.
5693 void
5694 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5695 Node* src, Node* src_offset,
5696 Node* dest, Node* dest_offset,
5697 Node* copy_length, bool dest_uninitialized) {
5698 assert(!dest_uninitialized, "Invariant");
5699 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5700 OptoRuntime::slow_arraycopy_Type(),
5701 OptoRuntime::slow_arraycopy_Java(),
5702 "slow_arraycopy", adr_type,
5703 src, src_offset, dest, dest_offset,
5704 copy_length);
5705
5706 // Handle exceptions thrown by this fellow:
5707 make_slow_call_ex(call, env()->Throwable_klass(), false);
5708 }
5709
5710 // Helper function; generates code for cases requiring runtime checks.
5711 Node*
5712 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5713 Node* dest_elem_klass,
5714 Node* src, Node* src_offset,
5715 Node* dest, Node* dest_offset,
5716 Node* copy_length, bool dest_uninitialized) {
5717 if (stopped()) return NULL;
5718
5719 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5720 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5721 return NULL;
5722 }
5723
5724 // Pick out the parameters required to perform a store-check
5725 // for the target array. This is an optimistic check. It will
5726 // look in each non-null element's class, at the desired klass's
5727 // super_check_offset, for the desired klass.
5728 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5729 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5730 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5731 Node* check_offset = ConvI2X(_gvn.transform(n3));
5732 Node* check_value = dest_elem_klass;
5733
5734 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5735 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5736
5737 // (We know the arrays are never conjoint, because their types differ.)
5738 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5739 OptoRuntime::checkcast_arraycopy_Type(),
5740 copyfunc_addr, "checkcast_arraycopy", adr_type,
5741 // five arguments, of which two are
5742 // intptr_t (jlong in LP64)
5743 src_start, dest_start,
5744 copy_length XTOP,
5745 check_offset XTOP,
5746 check_value);
5747
5748 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5749 }
5750
5751
5752 // Helper function; generates code for cases requiring runtime checks.
5753 Node*
5754 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5755 Node* src, Node* src_offset,
5756 Node* dest, Node* dest_offset,
5757 Node* copy_length, bool dest_uninitialized) {
5758 assert(!dest_uninitialized, "Invariant");
5759 if (stopped()) return NULL;
5760 address copyfunc_addr = StubRoutines::generic_arraycopy();
5761 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5762 return NULL;
5763 }
5764
5765 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5766 OptoRuntime::generic_arraycopy_Type(),
5767 copyfunc_addr, "generic_arraycopy", adr_type,
5768 src, src_offset, dest, dest_offset, copy_length);
5769
5770 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5771 }
5772
5773 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5774 void
5775 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5776 BasicType basic_elem_type,
5777 bool disjoint_bases,
5778 Node* src, Node* src_offset,
5779 Node* dest, Node* dest_offset,
5780 Node* copy_length, bool dest_uninitialized) {
5781 if (stopped()) return; // nothing to do
5782
5783 Node* src_start = src;
5784 Node* dest_start = dest;
5785 if (src_offset != NULL || dest_offset != NULL) {
5786 assert(src_offset != NULL && dest_offset != NULL, "");
5787 src_start = array_element_address(src, src_offset, basic_elem_type);
5788 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5789 }
5790
5791 // Figure out which arraycopy runtime method to call.
5792 const char* copyfunc_name = "arraycopy";
5793 address copyfunc_addr =
5794 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5795 disjoint_bases, copyfunc_name, dest_uninitialized);
5796
5797 // Call it. Note that the count_ix value is not scaled to a byte-size.
5798 make_runtime_call(RC_LEAF|RC_NO_FP,
5799 OptoRuntime::fast_arraycopy_Type(),
5800 copyfunc_addr, copyfunc_name, adr_type,
5801 src_start, dest_start, copy_length XTOP);
5802 }
5803
5804 //-------------inline_encodeISOArray-----------------------------------
5805 // encode char[] to byte[] in ISO_8859_1
5806 bool LibraryCallKit::inline_encodeISOArray() {
5807 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5808 // no receiver since it is static method
5809 Node *src = argument(0);
5810 Node *src_offset = argument(1);
5811 Node *dst = argument(2);
5812 Node *dst_offset = argument(3);
5813 Node *length = argument(4);
5814
5815 const Type* src_type = src->Value(&_gvn);
5816 const Type* dst_type = dst->Value(&_gvn);
5817 const TypeAryPtr* top_src = src_type->isa_aryptr();
5818 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5819 if (top_src == NULL || top_src->klass() == NULL ||
5820 top_dest == NULL || top_dest->klass() == NULL) {
5821 // failed array check
5822 return false;
5823 }
5824
5825 // Figure out the size and type of the elements we will be copying.
5826 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5827 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5828 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5829 return false;
5830 }
5831 Node* src_start = array_element_address(src, src_offset, src_elem);
5832 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5833 // 'src_start' points to src array + scaled offset
5834 // 'dst_start' points to dst array + scaled offset
5835
5836 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5837 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5838 enc = _gvn.transform(enc);
5839 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5840 set_memory(res_mem, mtype);
5841 set_result(enc);
5842 return true;
5843 }
5844
5845 //-------------inline_multiplyToLen-----------------------------------
5846 bool LibraryCallKit::inline_multiplyToLen() {
5847 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5848
5849 address stubAddr = StubRoutines::multiplyToLen();
5850 if (stubAddr == NULL) {
5851 return false; // Intrinsic's stub is not implemented on this platform
5852 }
5853 const char* stubName = "multiplyToLen";
5854
5855 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5856
5857 // no receiver because it is a static method
5858 Node* x = argument(0);
5859 Node* xlen = argument(1);
5860 Node* y = argument(2);
5861 Node* ylen = argument(3);
5862 Node* z = argument(4);
5863
5864 const Type* x_type = x->Value(&_gvn);
5865 const Type* y_type = y->Value(&_gvn);
5866 const TypeAryPtr* top_x = x_type->isa_aryptr();
5867 const TypeAryPtr* top_y = y_type->isa_aryptr();
5868 if (top_x == NULL || top_x->klass() == NULL ||
5869 top_y == NULL || top_y->klass() == NULL) {
5870 // failed array check
5871 return false;
5872 }
5873
5874 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5875 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5876 if (x_elem != T_INT || y_elem != T_INT) {
5877 return false;
5878 }
5879
5880 // Set the original stack and the reexecute bit for the interpreter to reexecute
5881 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5882 // on the return from z array allocation in runtime.
5883 { PreserveReexecuteState preexecs(this);
5884 jvms()->set_should_reexecute(true);
5885
5886 Node* x_start = array_element_address(x, intcon(0), x_elem);
5887 Node* y_start = array_element_address(y, intcon(0), y_elem);
5888 // 'x_start' points to x array + scaled xlen
5889 // 'y_start' points to y array + scaled ylen
5890
5891 // Allocate the result array
5892 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
5893 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5894 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5895
5896 IdealKit ideal(this);
5897
5898 #define __ ideal.
5899 Node* one = __ ConI(1);
5900 Node* zero = __ ConI(0);
5901 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5902 __ set(need_alloc, zero);
5903 __ set(z_alloc, z);
5904 __ if_then(z, BoolTest::eq, null()); {
5905 __ increment (need_alloc, one);
5906 } __ else_(); {
5907 // Update graphKit memory and control from IdealKit.
5908 sync_kit(ideal);
5909 Node* zlen_arg = load_array_length(z);
5910 // Update IdealKit memory and control from graphKit.
5911 __ sync_kit(this);
5912 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5913 __ increment (need_alloc, one);
5914 } __ end_if();
5915 } __ end_if();
5916
5917 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5918 // Update graphKit memory and control from IdealKit.
5919 sync_kit(ideal);
5920 Node * narr = new_array(klass_node, zlen, 1);
5921 // Update IdealKit memory and control from graphKit.
5922 __ sync_kit(this);
5923 __ set(z_alloc, narr);
5924 } __ end_if();
5925
5926 sync_kit(ideal);
5927 z = __ value(z_alloc);
5928 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5929 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5930 // Final sync IdealKit and GraphKit.
5931 final_sync(ideal);
5932 #undef __
5933
5934 Node* z_start = array_element_address(z, intcon(0), T_INT);
5935
5936 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5937 OptoRuntime::multiplyToLen_Type(),
5938 stubAddr, stubName, TypePtr::BOTTOM,
5939 x_start, xlen, y_start, ylen, z_start, zlen);
5940 } // original reexecute is set back here
5941
5942 C->set_has_split_ifs(true); // Has chance for split-if optimization
5943 set_result(z);
5944 return true;
5945 }
5946
5947 //-------------inline_squareToLen------------------------------------
5948 bool LibraryCallKit::inline_squareToLen() {
5949 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5950
5951 address stubAddr = StubRoutines::squareToLen();
5952 if (stubAddr == NULL) {
5953 return false; // Intrinsic's stub is not implemented on this platform
5954 }
5955 const char* stubName = "squareToLen";
5956
5957 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5958
5959 Node* x = argument(0);
5960 Node* len = argument(1);
5961 Node* z = argument(2);
5962 Node* zlen = argument(3);
5963
5964 const Type* x_type = x->Value(&_gvn);
5965 const Type* z_type = z->Value(&_gvn);
5966 const TypeAryPtr* top_x = x_type->isa_aryptr();
5967 const TypeAryPtr* top_z = z_type->isa_aryptr();
5968 if (top_x == NULL || top_x->klass() == NULL ||
5969 top_z == NULL || top_z->klass() == NULL) {
5970 // failed array check
5971 return false;
5972 }
5973
5974 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5975 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5976 if (x_elem != T_INT || z_elem != T_INT) {
5977 return false;
5978 }
5979
5980
5981 Node* x_start = array_element_address(x, intcon(0), x_elem);
5982 Node* z_start = array_element_address(z, intcon(0), z_elem);
5983
5984 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5985 OptoRuntime::squareToLen_Type(),
5986 stubAddr, stubName, TypePtr::BOTTOM,
5987 x_start, len, z_start, zlen);
5988
5989 set_result(z);
5990 return true;
5991 }
5992
5993 //-------------inline_mulAdd------------------------------------------
5994 bool LibraryCallKit::inline_mulAdd() {
5995 assert(UseMulAddIntrinsic, "not implementated on this platform");
5996
5997 address stubAddr = StubRoutines::mulAdd();
5998 if (stubAddr == NULL) {
5999 return false; // Intrinsic's stub is not implemented on this platform
6000 }
6001 const char* stubName = "mulAdd";
6002
6003 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
6004
6005 Node* out = argument(0);
6006 Node* in = argument(1);
6007 Node* offset = argument(2);
6008 Node* len = argument(3);
6009 Node* k = argument(4);
6010
6011 const Type* out_type = out->Value(&_gvn);
6012 const Type* in_type = in->Value(&_gvn);
6013 const TypeAryPtr* top_out = out_type->isa_aryptr();
6014 const TypeAryPtr* top_in = in_type->isa_aryptr();
6015 if (top_out == NULL || top_out->klass() == NULL ||
6016 top_in == NULL || top_in->klass() == NULL) {
6017 // failed array check
6018 return false;
6019 }
6020
6021 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6022 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6023 if (out_elem != T_INT || in_elem != T_INT) {
6024 return false;
6025 }
6026
6027 Node* outlen = load_array_length(out);
6028 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
6029 Node* out_start = array_element_address(out, intcon(0), out_elem);
6030 Node* in_start = array_element_address(in, intcon(0), in_elem);
6031
6032 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6033 OptoRuntime::mulAdd_Type(),
6034 stubAddr, stubName, TypePtr::BOTTOM,
6035 out_start,in_start, new_offset, len, k);
6036 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6037 set_result(result);
6038 return true;
6039 }
6040
6041 //-------------inline_montgomeryMultiply-----------------------------------
6042 bool LibraryCallKit::inline_montgomeryMultiply() {
6043 address stubAddr = StubRoutines::montgomeryMultiply();
6044 if (stubAddr == NULL) {
6045 return false; // Intrinsic's stub is not implemented on this platform
6046 }
6047
6048 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
6049 const char* stubName = "montgomery_multiply";
6050
6051 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
6052
6053 Node* a = argument(0);
6054 Node* b = argument(1);
6055 Node* n = argument(2);
6056 Node* len = argument(3);
6057 Node* inv = argument(4);
6058 Node* m = argument(6);
6059
6060 const Type* a_type = a->Value(&_gvn);
6061 const TypeAryPtr* top_a = a_type->isa_aryptr();
6062 const Type* b_type = b->Value(&_gvn);
6063 const TypeAryPtr* top_b = b_type->isa_aryptr();
6064 const Type* n_type = a->Value(&_gvn);
6065 const TypeAryPtr* top_n = n_type->isa_aryptr();
6066 const Type* m_type = a->Value(&_gvn);
6067 const TypeAryPtr* top_m = m_type->isa_aryptr();
6068 if (top_a == NULL || top_a->klass() == NULL ||
6069 top_b == NULL || top_b->klass() == NULL ||
6070 top_n == NULL || top_n->klass() == NULL ||
6071 top_m == NULL || top_m->klass() == NULL) {
6072 // failed array check
6073 return false;
6074 }
6075
6076 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6077 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6078 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6079 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6080 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6081 return false;
6082 }
6083
6084 // Make the call
6085 {
6086 Node* a_start = array_element_address(a, intcon(0), a_elem);
6087 Node* b_start = array_element_address(b, intcon(0), b_elem);
6088 Node* n_start = array_element_address(n, intcon(0), n_elem);
6089 Node* m_start = array_element_address(m, intcon(0), m_elem);
6090
6091 Node* call = NULL;
6092 if (CCallingConventionRequiresIntsAsLongs) {
6093 Node* len_I2L = ConvI2L(len);
6094 call = make_runtime_call(RC_LEAF,
6095 OptoRuntime::montgomeryMultiply_Type(),
6096 stubAddr, stubName, TypePtr::BOTTOM,
6097 a_start, b_start, n_start, len_I2L XTOP, inv,
6098 top(), m_start);
6099 } else {
6100 call = make_runtime_call(RC_LEAF,
6101 OptoRuntime::montgomeryMultiply_Type(),
6102 stubAddr, stubName, TypePtr::BOTTOM,
6103 a_start, b_start, n_start, len, inv, top(),
6104 m_start);
6105 }
6106 set_result(m);
6107 }
6108
6109 return true;
6110 }
6111
6112 bool LibraryCallKit::inline_montgomerySquare() {
6113 address stubAddr = StubRoutines::montgomerySquare();
6114 if (stubAddr == NULL) {
6115 return false; // Intrinsic's stub is not implemented on this platform
6116 }
6117
6118 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
6119 const char* stubName = "montgomery_square";
6120
6121 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
6122
6123 Node* a = argument(0);
6124 Node* n = argument(1);
6125 Node* len = argument(2);
6126 Node* inv = argument(3);
6127 Node* m = argument(5);
6128
6129 const Type* a_type = a->Value(&_gvn);
6130 const TypeAryPtr* top_a = a_type->isa_aryptr();
6131 const Type* n_type = a->Value(&_gvn);
6132 const TypeAryPtr* top_n = n_type->isa_aryptr();
6133 const Type* m_type = a->Value(&_gvn);
6134 const TypeAryPtr* top_m = m_type->isa_aryptr();
6135 if (top_a == NULL || top_a->klass() == NULL ||
6136 top_n == NULL || top_n->klass() == NULL ||
6137 top_m == NULL || top_m->klass() == NULL) {
6138 // failed array check
6139 return false;
6140 }
6141
6142 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6143 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6144 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6145 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
6146 return false;
6147 }
6148
6149 // Make the call
6150 {
6151 Node* a_start = array_element_address(a, intcon(0), a_elem);
6152 Node* n_start = array_element_address(n, intcon(0), n_elem);
6153 Node* m_start = array_element_address(m, intcon(0), m_elem);
6154
6155 Node* call = NULL;
6156 if (CCallingConventionRequiresIntsAsLongs) {
6157 Node* len_I2L = ConvI2L(len);
6158 call = make_runtime_call(RC_LEAF,
6159 OptoRuntime::montgomerySquare_Type(),
6160 stubAddr, stubName, TypePtr::BOTTOM,
6161 a_start, n_start, len_I2L XTOP, inv, top(),
6162 m_start);
6163 } else {
6164 call = make_runtime_call(RC_LEAF,
6165 OptoRuntime::montgomerySquare_Type(),
6166 stubAddr, stubName, TypePtr::BOTTOM,
6167 a_start, n_start, len, inv, top(),
6168 m_start);
6169 }
6170
6171 set_result(m);
6172 }
6173
6174 return true;
6175 }
6176
6177
6178 /**
6179 * Calculate CRC32 for byte.
6180 * int java.util.zip.CRC32.update(int crc, int b)
6181 */
6182 bool LibraryCallKit::inline_updateCRC32() {
6183 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6184 assert(callee()->signature()->size() == 2, "update has 2 parameters");
6185 // no receiver since it is static method
6186 Node* crc = argument(0); // type: int
6187 Node* b = argument(1); // type: int
6188
6189 /*
6190 * int c = ~ crc;
6191 * b = timesXtoThe32[(b ^ c) & 0xFF];
6192 * b = b ^ (c >>> 8);
6193 * crc = ~b;
6194 */
6195
6196 Node* M1 = intcon(-1);
6197 crc = _gvn.transform(new (C) XorINode(crc, M1));
6198 Node* result = _gvn.transform(new (C) XorINode(crc, b));
6199 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
6200
6201 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
6202 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
6203 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
6204 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
6205
6206 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
6207 result = _gvn.transform(new (C) XorINode(crc, result));
6208 result = _gvn.transform(new (C) XorINode(result, M1));
6209 set_result(result);
6210 return true;
6211 }
6212
6213 /**
6214 * Calculate CRC32 for byte[] array.
6215 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
6216 */
6217 bool LibraryCallKit::inline_updateBytesCRC32() {
6218 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6219 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6220 // no receiver since it is static method
6221 Node* crc = argument(0); // type: int
6222 Node* src = argument(1); // type: oop
6223 Node* offset = argument(2); // type: int
6224 Node* length = argument(3); // type: int
6225
6226 const Type* src_type = src->Value(&_gvn);
6227 const TypeAryPtr* top_src = src_type->isa_aryptr();
6228 if (top_src == NULL || top_src->klass() == NULL) {
6229 // failed array check
6230 return false;
6231 }
6232
6233 // Figure out the size and type of the elements we will be copying.
6234 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6235 if (src_elem != T_BYTE) {
6236 return false;
6237 }
6238
6239 // 'src_start' points to src array + scaled offset
6240 Node* src_start = array_element_address(src, offset, src_elem);
6241
6242 // We assume that range check is done by caller.
6243 // TODO: generate range check (offset+length < src.length) in debug VM.
6244
6245 // Call the stub.
6246 address stubAddr = StubRoutines::updateBytesCRC32();
6247 const char *stubName = "updateBytesCRC32";
6248 Node* call;
6249 if (CCallingConventionRequiresIntsAsLongs) {
6250 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6251 stubAddr, stubName, TypePtr::BOTTOM,
6252 crc XTOP, src_start, length XTOP);
6253 } else {
6254 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6255 stubAddr, stubName, TypePtr::BOTTOM,
6256 crc, src_start, length);
6257 }
6258 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6259 set_result(result);
6260 return true;
6261 }
6262
6263 /**
6264 * Calculate CRC32 for ByteBuffer.
6265 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
6266 */
6267 bool LibraryCallKit::inline_updateByteBufferCRC32() {
6268 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
6269 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6270 // no receiver since it is static method
6271 Node* crc = argument(0); // type: int
6272 Node* src = argument(1); // type: long
6273 Node* offset = argument(3); // type: int
6274 Node* length = argument(4); // type: int
6275
6276 src = ConvL2X(src); // adjust Java long to machine word
6277 Node* base = _gvn.transform(new (C) CastX2PNode(src));
6278 offset = ConvI2X(offset);
6279
6280 // 'src_start' points to src array + scaled offset
6281 Node* src_start = basic_plus_adr(top(), base, offset);
6282
6283 // Call the stub.
6284 address stubAddr = StubRoutines::updateBytesCRC32();
6285 const char *stubName = "updateBytesCRC32";
6286 Node* call;
6287 if (CCallingConventionRequiresIntsAsLongs) {
6288 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6289 stubAddr, stubName, TypePtr::BOTTOM,
6290 crc XTOP, src_start, length XTOP);
6291 } else {
6292 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
6293 stubAddr, stubName, TypePtr::BOTTOM,
6294 crc, src_start, length);
6295 }
6296 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6297 set_result(result);
6298 return true;
6299 }
6300
6301 //----------------------------inline_reference_get----------------------------
6302 // public T java.lang.ref.Reference.get();
6303 bool LibraryCallKit::inline_reference_get() {
6304 const int referent_offset = java_lang_ref_Reference::referent_offset;
6305 guarantee(referent_offset > 0, "should have already been set");
6306
6307 // Get the argument:
6308 Node* reference_obj = null_check_receiver();
6309 if (stopped()) return true;
6310
6311 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6312
6313 ciInstanceKlass* klass = env()->Object_klass();
6314 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6315
6316 Node* no_ctrl = NULL;
6317 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6318
6319 // Use the pre-barrier to record the value in the referent field
6320 pre_barrier(false /* do_load */,
6321 control(),
6322 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6323 result /* pre_val */,
6324 T_OBJECT);
6325
6326 // Add memory barrier to prevent commoning reads from this field
6327 // across safepoint since GC can change its value.
6328 insert_mem_bar(Op_MemBarCPUOrder);
6329
6330 set_result(result);
6331 return true;
6332 }
6333
6334
6335 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6336 bool is_exact=true, bool is_static=false) {
6337
6338 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6339 assert(tinst != NULL, "obj is null");
6340 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6341 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6342
6343 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
6344 ciSymbol::make(fieldTypeString),
6345 is_static);
6346 if (field == NULL) return (Node *) NULL;
6347 assert (field != NULL, "undefined field");
6348
6349 // Next code copied from Parse::do_get_xxx():
6350
6351 // Compute address and memory type.
6352 int offset = field->offset_in_bytes();
6353 bool is_vol = field->is_volatile();
6354 ciType* field_klass = field->type();
6355 assert(field_klass->is_loaded(), "should be loaded");
6356 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6357 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6358 BasicType bt = field->layout_type();
6359
6360 // Build the resultant type of the load
6361 const Type *type;
6362 if (bt == T_OBJECT) {
6363 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6364 } else {
6365 type = Type::get_const_basic_type(bt);
6366 }
6367
6368 Node* leading_membar = NULL;
6369 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6370 leading_membar = insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6371 }
6372 // Build the load.
6373 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6374 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6375 // If reference is volatile, prevent following memory ops from
6376 // floating up past the volatile read. Also prevents commoning
6377 // another volatile read.
6378 if (is_vol) {
6379 // Memory barrier includes bogus read of value to force load BEFORE membar
6380 Node* mb = insert_mem_bar(Op_MemBarAcquire, loadedField);
6381 mb->as_MemBar()->set_trailing_load();
6382 }
6383 return loadedField;
6384 }
6385
6386
6387 //------------------------------inline_aescrypt_Block-----------------------
6388 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6389 address stubAddr = NULL;
6390 const char *stubName;
6391 assert(UseAES, "need AES instruction support");
6392
6393 switch(id) {
6394 case vmIntrinsics::_aescrypt_encryptBlock:
6395 stubAddr = StubRoutines::aescrypt_encryptBlock();
6396 stubName = "aescrypt_encryptBlock";
6397 break;
6398 case vmIntrinsics::_aescrypt_decryptBlock:
6399 stubAddr = StubRoutines::aescrypt_decryptBlock();
6400 stubName = "aescrypt_decryptBlock";
6401 break;
6402 }
6403 if (stubAddr == NULL) return false;
6404
6405 Node* aescrypt_object = argument(0);
6406 Node* src = argument(1);
6407 Node* src_offset = argument(2);
6408 Node* dest = argument(3);
6409 Node* dest_offset = argument(4);
6410
6411 // (1) src and dest are arrays.
6412 const Type* src_type = src->Value(&_gvn);
6413 const Type* dest_type = dest->Value(&_gvn);
6414 const TypeAryPtr* top_src = src_type->isa_aryptr();
6415 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6416 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6417
6418 // for the quick and dirty code we will skip all the checks.
6419 // we are just trying to get the call to be generated.
6420 Node* src_start = src;
6421 Node* dest_start = dest;
6422 if (src_offset != NULL || dest_offset != NULL) {
6423 assert(src_offset != NULL && dest_offset != NULL, "");
6424 src_start = array_element_address(src, src_offset, T_BYTE);
6425 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6426 }
6427
6428 // now need to get the start of its expanded key array
6429 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6430 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6431 if (k_start == NULL) return false;
6432
6433 if (Matcher::pass_original_key_for_aes()) {
6434 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6435 // compatibility issues between Java key expansion and SPARC crypto instructions
6436 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6437 if (original_k_start == NULL) return false;
6438
6439 // Call the stub.
6440 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6441 stubAddr, stubName, TypePtr::BOTTOM,
6442 src_start, dest_start, k_start, original_k_start);
6443 } else {
6444 // Call the stub.
6445 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6446 stubAddr, stubName, TypePtr::BOTTOM,
6447 src_start, dest_start, k_start);
6448 }
6449
6450 return true;
6451 }
6452
6453 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6454 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6455 address stubAddr = NULL;
6456 const char *stubName = NULL;
6457
6458 assert(UseAES, "need AES instruction support");
6459
6460 switch(id) {
6461 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6462 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6463 stubName = "cipherBlockChaining_encryptAESCrypt";
6464 break;
6465 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6466 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6467 stubName = "cipherBlockChaining_decryptAESCrypt";
6468 break;
6469 }
6470 if (stubAddr == NULL) return false;
6471
6472 Node* cipherBlockChaining_object = argument(0);
6473 Node* src = argument(1);
6474 Node* src_offset = argument(2);
6475 Node* len = argument(3);
6476 Node* dest = argument(4);
6477 Node* dest_offset = argument(5);
6478
6479 // (1) src and dest are arrays.
6480 const Type* src_type = src->Value(&_gvn);
6481 const Type* dest_type = dest->Value(&_gvn);
6482 const TypeAryPtr* top_src = src_type->isa_aryptr();
6483 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6484 assert (top_src != NULL && top_src->klass() != NULL
6485 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6486
6487 // checks are the responsibility of the caller
6488 Node* src_start = src;
6489 Node* dest_start = dest;
6490 if (src_offset != NULL || dest_offset != NULL) {
6491 assert(src_offset != NULL && dest_offset != NULL, "");
6492 src_start = array_element_address(src, src_offset, T_BYTE);
6493 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6494 }
6495
6496 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6497 // (because of the predicated logic executed earlier).
6498 // so we cast it here safely.
6499 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6500
6501 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6502 if (embeddedCipherObj == NULL) return false;
6503
6504 // cast it to what we know it will be at runtime
6505 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6506 assert(tinst != NULL, "CBC obj is null");
6507 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6508 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6509 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6510
6511 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6512 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6513 const TypeOopPtr* xtype = aklass->as_instance_type();
6514 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6515 aescrypt_object = _gvn.transform(aescrypt_object);
6516
6517 // we need to get the start of the aescrypt_object's expanded key array
6518 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6519 if (k_start == NULL) return false;
6520
6521 // similarly, get the start address of the r vector
6522 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6523 if (objRvec == NULL) return false;
6524 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6525
6526 Node* cbcCrypt;
6527 if (Matcher::pass_original_key_for_aes()) {
6528 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6529 // compatibility issues between Java key expansion and SPARC crypto instructions
6530 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6531 if (original_k_start == NULL) return false;
6532
6533 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6534 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6535 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6536 stubAddr, stubName, TypePtr::BOTTOM,
6537 src_start, dest_start, k_start, r_start, len, original_k_start);
6538 } else {
6539 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6540 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6541 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6542 stubAddr, stubName, TypePtr::BOTTOM,
6543 src_start, dest_start, k_start, r_start, len);
6544 }
6545
6546 // return cipher length (int)
6547 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6548 set_result(retvalue);
6549 return true;
6550 }
6551
6552 //------------------------------get_key_start_from_aescrypt_object-----------------------
6553 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6554 #ifdef PPC64
6555 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6556 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6557 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6558 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6559 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6560 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6561 if (objSessionK == NULL) {
6562 return (Node *) NULL;
6563 }
6564 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6565 #else
6566 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6567 #endif // PPC64
6568 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6569 if (objAESCryptKey == NULL) return (Node *) NULL;
6570
6571 // now have the array, need to get the start address of the K array
6572 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6573 return k_start;
6574 }
6575
6576 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6577 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6578 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6579 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6580 if (objAESCryptKey == NULL) return (Node *) NULL;
6581
6582 // now have the array, need to get the start address of the lastKey array
6583 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6584 return original_k_start;
6585 }
6586
6587 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6588 // Return node representing slow path of predicate check.
6589 // the pseudo code we want to emulate with this predicate is:
6590 // for encryption:
6591 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6592 // for decryption:
6593 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6594 // note cipher==plain is more conservative than the original java code but that's OK
6595 //
6596 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6597 // The receiver was checked for NULL already.
6598 Node* objCBC = argument(0);
6599
6600 // Load embeddedCipher field of CipherBlockChaining object.
6601 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6602
6603 // get AESCrypt klass for instanceOf check
6604 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6605 // will have same classloader as CipherBlockChaining object
6606 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6607 assert(tinst != NULL, "CBCobj is null");
6608 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6609
6610 // we want to do an instanceof comparison against the AESCrypt class
6611 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6612 if (!klass_AESCrypt->is_loaded()) {
6613 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6614 Node* ctrl = control();
6615 set_control(top()); // no regular fast path
6616 return ctrl;
6617 }
6618 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6619
6620 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6621 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6622 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6623
6624 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6625
6626 // for encryption, we are done
6627 if (!decrypting)
6628 return instof_false; // even if it is NULL
6629
6630 // for decryption, we need to add a further check to avoid
6631 // taking the intrinsic path when cipher and plain are the same
6632 // see the original java code for why.
6633 RegionNode* region = new(C) RegionNode(3);
6634 region->init_req(1, instof_false);
6635 Node* src = argument(1);
6636 Node* dest = argument(4);
6637 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6638 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6639 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6640 region->init_req(2, src_dest_conjoint);
6641
6642 record_for_igvn(region);
6643 return _gvn.transform(region);
6644 }
6645
6646 //------------------------------inline_ghash_processBlocks
6647 bool LibraryCallKit::inline_ghash_processBlocks() {
6648 address stubAddr;
6649 const char *stubName;
6650 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6651
6652 stubAddr = StubRoutines::ghash_processBlocks();
6653 stubName = "ghash_processBlocks";
6654
6655 Node* data = argument(0);
6656 Node* offset = argument(1);
6657 Node* len = argument(2);
6658 Node* state = argument(3);
6659 Node* subkeyH = argument(4);
6660
6661 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6662 assert(state_start, "state is NULL");
6663 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6664 assert(subkeyH_start, "subkeyH is NULL");
6665 Node* data_start = array_element_address(data, offset, T_BYTE);
6666 assert(data_start, "data is NULL");
6667
6668 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6669 OptoRuntime::ghash_processBlocks_Type(),
6670 stubAddr, stubName, TypePtr::BOTTOM,
6671 state_start, subkeyH_start, data_start, len);
6672 return true;
6673 }
6674
6675 //------------------------------inline_sha_implCompress-----------------------
6676 //
6677 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6678 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6679 //
6680 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6681 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6682 //
6683 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6684 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6685 //
6686 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6687 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6688
6689 Node* sha_obj = argument(0);
6690 Node* src = argument(1); // type oop
6691 Node* ofs = argument(2); // type int
6692
6693 const Type* src_type = src->Value(&_gvn);
6694 const TypeAryPtr* top_src = src_type->isa_aryptr();
6695 if (top_src == NULL || top_src->klass() == NULL) {
6696 // failed array check
6697 return false;
6698 }
6699 // Figure out the size and type of the elements we will be copying.
6700 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6701 if (src_elem != T_BYTE) {
6702 return false;
6703 }
6704 // 'src_start' points to src array + offset
6705 Node* src_start = array_element_address(src, ofs, src_elem);
6706 Node* state = NULL;
6707 address stubAddr;
6708 const char *stubName;
6709
6710 switch(id) {
6711 case vmIntrinsics::_sha_implCompress:
6712 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6713 state = get_state_from_sha_object(sha_obj);
6714 stubAddr = StubRoutines::sha1_implCompress();
6715 stubName = "sha1_implCompress";
6716 break;
6717 case vmIntrinsics::_sha2_implCompress:
6718 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6719 state = get_state_from_sha_object(sha_obj);
6720 stubAddr = StubRoutines::sha256_implCompress();
6721 stubName = "sha256_implCompress";
6722 break;
6723 case vmIntrinsics::_sha5_implCompress:
6724 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6725 state = get_state_from_sha5_object(sha_obj);
6726 stubAddr = StubRoutines::sha512_implCompress();
6727 stubName = "sha512_implCompress";
6728 break;
6729 default:
6730 fatal_unexpected_iid(id);
6731 return false;
6732 }
6733 if (state == NULL) return false;
6734
6735 // Call the stub.
6736 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6737 stubAddr, stubName, TypePtr::BOTTOM,
6738 src_start, state);
6739
6740 return true;
6741 }
6742
6743 //------------------------------inline_digestBase_implCompressMB-----------------------
6744 //
6745 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6746 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6747 //
6748 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6749 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6750 "need SHA1/SHA256/SHA512 instruction support");
6751 assert((uint)predicate < 3, "sanity");
6752 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6753
6754 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6755 Node* src = argument(1); // byte[] array
6756 Node* ofs = argument(2); // type int
6757 Node* limit = argument(3); // type int
6758
6759 const Type* src_type = src->Value(&_gvn);
6760 const TypeAryPtr* top_src = src_type->isa_aryptr();
6761 if (top_src == NULL || top_src->klass() == NULL) {
6762 // failed array check
6763 return false;
6764 }
6765 // Figure out the size and type of the elements we will be copying.
6766 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6767 if (src_elem != T_BYTE) {
6768 return false;
6769 }
6770 // 'src_start' points to src array + offset
6771 Node* src_start = array_element_address(src, ofs, src_elem);
6772
6773 const char* klass_SHA_name = NULL;
6774 const char* stub_name = NULL;
6775 address stub_addr = NULL;
6776 bool long_state = false;
6777
6778 switch (predicate) {
6779 case 0:
6780 if (UseSHA1Intrinsics) {
6781 klass_SHA_name = "sun/security/provider/SHA";
6782 stub_name = "sha1_implCompressMB";
6783 stub_addr = StubRoutines::sha1_implCompressMB();
6784 }
6785 break;
6786 case 1:
6787 if (UseSHA256Intrinsics) {
6788 klass_SHA_name = "sun/security/provider/SHA2";
6789 stub_name = "sha256_implCompressMB";
6790 stub_addr = StubRoutines::sha256_implCompressMB();
6791 }
6792 break;
6793 case 2:
6794 if (UseSHA512Intrinsics) {
6795 klass_SHA_name = "sun/security/provider/SHA5";
6796 stub_name = "sha512_implCompressMB";
6797 stub_addr = StubRoutines::sha512_implCompressMB();
6798 long_state = true;
6799 }
6800 break;
6801 default:
6802 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6803 }
6804 if (klass_SHA_name != NULL) {
6805 // get DigestBase klass to lookup for SHA klass
6806 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6807 assert(tinst != NULL, "digestBase_obj is not instance???");
6808 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6809
6810 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6811 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6812 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6813 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6814 }
6815 return false;
6816 }
6817 //------------------------------inline_sha_implCompressMB-----------------------
6818 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6819 bool long_state, address stubAddr, const char *stubName,
6820 Node* src_start, Node* ofs, Node* limit) {
6821 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6822 const TypeOopPtr* xtype = aklass->as_instance_type();
6823 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
6824 sha_obj = _gvn.transform(sha_obj);
6825
6826 Node* state;
6827 if (long_state) {
6828 state = get_state_from_sha5_object(sha_obj);
6829 } else {
6830 state = get_state_from_sha_object(sha_obj);
6831 }
6832 if (state == NULL) return false;
6833
6834 // Call the stub.
6835 Node *call;
6836 if (CCallingConventionRequiresIntsAsLongs) {
6837 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6838 OptoRuntime::digestBase_implCompressMB_Type(),
6839 stubAddr, stubName, TypePtr::BOTTOM,
6840 src_start, state, ofs XTOP, limit XTOP);
6841 } else {
6842 call = make_runtime_call(RC_LEAF|RC_NO_FP,
6843 OptoRuntime::digestBase_implCompressMB_Type(),
6844 stubAddr, stubName, TypePtr::BOTTOM,
6845 src_start, state, ofs, limit);
6846 }
6847 // return ofs (int)
6848 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
6849 set_result(result);
6850
6851 return true;
6852 }
6853
6854 //------------------------------get_state_from_sha_object-----------------------
6855 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6856 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6857 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6858 if (sha_state == NULL) return (Node *) NULL;
6859
6860 // now have the array, need to get the start address of the state array
6861 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6862 return state;
6863 }
6864
6865 //------------------------------get_state_from_sha5_object-----------------------
6866 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6867 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6868 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6869 if (sha_state == NULL) return (Node *) NULL;
6870
6871 // now have the array, need to get the start address of the state array
6872 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6873 return state;
6874 }
6875
6876 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6877 // Return node representing slow path of predicate check.
6878 // the pseudo code we want to emulate with this predicate is:
6879 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6880 //
6881 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6882 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6883 "need SHA1/SHA256/SHA512 instruction support");
6884 assert((uint)predicate < 3, "sanity");
6885
6886 // The receiver was checked for NULL already.
6887 Node* digestBaseObj = argument(0);
6888
6889 // get DigestBase klass for instanceOf check
6890 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6891 assert(tinst != NULL, "digestBaseObj is null");
6892 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6893
6894 const char* klass_SHA_name = NULL;
6895 switch (predicate) {
6896 case 0:
6897 if (UseSHA1Intrinsics) {
6898 // we want to do an instanceof comparison against the SHA class
6899 klass_SHA_name = "sun/security/provider/SHA";
6900 }
6901 break;
6902 case 1:
6903 if (UseSHA256Intrinsics) {
6904 // we want to do an instanceof comparison against the SHA2 class
6905 klass_SHA_name = "sun/security/provider/SHA2";
6906 }
6907 break;
6908 case 2:
6909 if (UseSHA512Intrinsics) {
6910 // we want to do an instanceof comparison against the SHA5 class
6911 klass_SHA_name = "sun/security/provider/SHA5";
6912 }
6913 break;
6914 default:
6915 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6916 }
6917
6918 ciKlass* klass_SHA = NULL;
6919 if (klass_SHA_name != NULL) {
6920 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6921 }
6922 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6923 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6924 Node* ctrl = control();
6925 set_control(top()); // no intrinsic path
6926 return ctrl;
6927 }
6928 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6929
6930 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6931 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
6932 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6933 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6934
6935 return instof_false; // even if it is NULL
6936 }
6937
6938 bool LibraryCallKit::inline_profileBoolean() {
6939 Node* counts = argument(1);
6940 const TypeAryPtr* ary = NULL;
6941 ciArray* aobj = NULL;
6942 if (counts->is_Con()
6943 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6944 && (aobj = ary->const_oop()->as_array()) != NULL
6945 && (aobj->length() == 2)) {
6946 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6947 jint false_cnt = aobj->element_value(0).as_int();
6948 jint true_cnt = aobj->element_value(1).as_int();
6949
6950 if (C->log() != NULL) {
6951 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6952 false_cnt, true_cnt);
6953 }
6954
6955 if (false_cnt + true_cnt == 0) {
6956 // According to profile, never executed.
6957 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6958 Deoptimization::Action_reinterpret);
6959 return true;
6960 }
6961
6962 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6963 // is a number of each value occurrences.
6964 Node* result = argument(0);
6965 if (false_cnt == 0 || true_cnt == 0) {
6966 // According to profile, one value has been never seen.
6967 int expected_val = (false_cnt == 0) ? 1 : 0;
6968
6969 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
6970 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
6971
6972 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6973 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
6974 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
6975
6976 { // Slow path: uncommon trap for never seen value and then reexecute
6977 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6978 // the value has been seen at least once.
6979 PreserveJVMState pjvms(this);
6980 PreserveReexecuteState preexecs(this);
6981 jvms()->set_should_reexecute(true);
6982
6983 set_control(slow_path);
6984 set_i_o(i_o());
6985
6986 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6987 Deoptimization::Action_reinterpret);
6988 }
6989 // The guard for never seen value enables sharpening of the result and
6990 // returning a constant. It allows to eliminate branches on the same value
6991 // later on.
6992 set_control(fast_path);
6993 result = intcon(expected_val);
6994 }
6995 // Stop profiling.
6996 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6997 // By replacing method body with profile data (represented as ProfileBooleanNode
6998 // on IR level) we effectively disable profiling.
6999 // It enables full speed execution once optimized code is generated.
7000 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
7001 C->record_for_igvn(profile);
7002 set_result(profile);
7003 return true;
7004 } else {
7005 // Continue profiling.
7006 // Profile data isn't available at the moment. So, execute method's bytecode version.
7007 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7008 // is compiled and counters aren't available since corresponding MethodHandle
7009 // isn't a compile-time constant.
7010 return false;
7011 }
7012 }