1 /*
2 * Copyright (c) 1999, 2016, 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 "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "prims/nativeLookup.hpp"
52 #include "prims/unsafe.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #ifdef TRACE_HAVE_INTRINSICS
55 #include "trace/traceMacros.hpp"
56 #endif
57
58 class LibraryIntrinsic : public InlineCallGenerator {
59 // Extend the set of intrinsics known to the runtime:
60 public:
61 private:
62 bool _is_virtual;
63 bool _does_virtual_dispatch;
64 int8_t _predicates_count; // Intrinsic is predicated by several conditions
65 int8_t _last_predicate; // Last generated predicate
66 vmIntrinsics::ID _intrinsic_id;
67
68 public:
69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
70 : InlineCallGenerator(m),
71 _is_virtual(is_virtual),
72 _does_virtual_dispatch(does_virtual_dispatch),
73 _predicates_count((int8_t)predicates_count),
74 _last_predicate((int8_t)-1),
75 _intrinsic_id(id)
76 {
77 }
78 virtual bool is_intrinsic() const { return true; }
79 virtual bool is_virtual() const { return _is_virtual; }
80 virtual bool is_predicated() const { return _predicates_count > 0; }
81 virtual int predicates_count() const { return _predicates_count; }
82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
83 virtual JVMState* generate(JVMState* jvms);
84 virtual Node* generate_predicate(JVMState* jvms, int predicate);
85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
86 };
87
88
89 // Local helper class for LibraryIntrinsic:
90 class LibraryCallKit : public GraphKit {
91 private:
92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
93 Node* _result; // the result node, if any
94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
95
96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
97
98 public:
99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
100 : GraphKit(jvms),
101 _intrinsic(intrinsic),
102 _result(NULL)
103 {
104 // Check if this is a root compile. In that case we don't have a caller.
105 if (!jvms->has_method()) {
106 _reexecute_sp = sp();
107 } else {
108 // Find out how many arguments the interpreter needs when deoptimizing
109 // and save the stack pointer value so it can used by uncommon_trap.
110 // We find the argument count by looking at the declared signature.
111 bool ignored_will_link;
112 ciSignature* declared_signature = NULL;
113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
116 }
117 }
118
119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
120
121 ciMethod* caller() const { return jvms()->method(); }
122 int bci() const { return jvms()->bci(); }
123 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
125 ciMethod* callee() const { return _intrinsic->method(); }
126
127 bool try_to_inline(int predicate);
128 Node* try_to_predicate(int predicate);
129
130 void push_result() {
131 // Push the result onto the stack.
132 if (!stopped() && result() != NULL) {
133 BasicType bt = result()->bottom_type()->basic_type();
134 push_node(bt, result());
135 }
136 }
137
138 private:
139 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
141 }
142
143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
144 void set_result(RegionNode* region, PhiNode* value);
145 Node* result() { return _result; }
146
147 virtual int reexecute_sp() { return _reexecute_sp; }
148
149 // Helper functions to inline natives
150 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
151 Node* generate_slow_guard(Node* test, RegionNode* region);
152 Node* generate_fair_guard(Node* test, RegionNode* region);
153 Node* generate_negative_guard(Node* index, RegionNode* region,
154 // resulting CastII of index:
155 Node* *pos_index = NULL);
156 Node* generate_limit_guard(Node* offset, Node* subseq_length,
157 Node* array_length,
158 RegionNode* region);
159 void generate_string_range_check(Node* array, Node* offset,
160 Node* length, bool char_count);
161 Node* generate_current_thread(Node* &tls_output);
162 Node* load_mirror_from_klass(Node* klass);
163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path,
165 int offset);
166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
174 RegionNode* region, int null_path) {
175 int offset = java_lang_Class::array_klass_offset_in_bytes();
176 return load_klass_from_mirror_common(mirror, never_see_null,
177 region, null_path,
178 offset);
179 }
180 Node* generate_access_flags_guard(Node* kls,
181 int modifier_mask, int modifier_bits,
182 RegionNode* region);
183 Node* generate_interface_guard(Node* kls, RegionNode* region);
184 Node* generate_array_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, false, false);
186 }
187 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, false, true);
189 }
190 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, true, false);
192 }
193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, true, true);
195 }
196 Node* generate_array_guard_common(Node* kls, RegionNode* region,
197 bool obj_array, bool not_array);
198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
200 bool is_virtual = false, bool is_static = false);
201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
202 return generate_method_call(method_id, false, true);
203 }
204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
205 return generate_method_call(method_id, true, false);
206 }
207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209
210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_indexOfChar();
217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_toBytesU();
219 bool inline_string_getCharsU();
220 bool inline_string_copy(bool compress);
221 bool inline_string_char_access(bool is_store);
222 Node* round_double_node(Node* n);
223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_math_native(vmIntrinsics::ID id);
225 bool inline_math(vmIntrinsics::ID id);
226 template <typename OverflowOp>
227 bool inline_math_overflow(Node* arg1, Node* arg2);
228 void inline_math_mathExact(Node* math, Node* test);
229 bool inline_math_addExactI(bool is_increment);
230 bool inline_math_addExactL(bool is_increment);
231 bool inline_math_multiplyExactI();
232 bool inline_math_multiplyExactL();
233 bool inline_math_negateExactI();
234 bool inline_math_negateExactL();
235 bool inline_math_subtractExactI(bool is_decrement);
236 bool inline_math_subtractExactL(bool is_decrement);
237 bool inline_min_max(vmIntrinsics::ID id);
238 bool inline_notify(vmIntrinsics::ID id);
239 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
240 // This returns Type::AnyPtr, RawPtr, or OopPtr.
241 int classify_unsafe_addr(Node* &base, Node* &offset);
242 Node* make_unsafe_address(Node* base, Node* offset);
243 // Helper for inline_unsafe_access.
244 // Generates the guards that check whether the result of
245 // Unsafe.getObject should be recorded in an SATB log buffer.
246 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
247
248 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
249 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
250 static bool klass_needs_init_guard(Node* kls);
251 bool inline_unsafe_allocate();
252 bool inline_unsafe_newArray(bool uninitialized);
253 bool inline_unsafe_copyMemory();
254 bool inline_native_currentThread();
255
256 bool inline_native_time_funcs(address method, const char* funcName);
257 #ifdef TRACE_HAVE_INTRINSICS
258 bool inline_native_classID();
259 bool inline_native_getBufferWriter();
260 #endif
261 bool inline_native_isInterrupted();
262 bool inline_native_Class_query(vmIntrinsics::ID id);
263 bool inline_native_subtype_check();
264 bool inline_native_getLength();
265 bool inline_array_copyOf(bool is_copyOfRange);
266 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
267 bool inline_preconditions_checkIndex();
268 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
269 bool inline_native_clone(bool is_virtual);
270 bool inline_native_Reflection_getCallerClass();
271 // Helper function for inlining native object hash method
272 bool inline_native_hashcode(bool is_virtual, bool is_static);
273 bool inline_native_getClass();
274
275 // Helper functions for inlining arraycopy
276 bool inline_arraycopy();
277 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
278 RegionNode* slow_region);
279 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
280 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
281
282 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
283 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
284 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
285 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
286 bool inline_unsafe_fence(vmIntrinsics::ID id);
287 bool inline_onspinwait();
288 bool inline_fp_conversions(vmIntrinsics::ID id);
289 bool inline_number_methods(vmIntrinsics::ID id);
290 bool inline_reference_get();
291 bool inline_Class_cast();
292 bool inline_aescrypt_Block(vmIntrinsics::ID id);
293 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
294 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
295 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
296 Node* inline_counterMode_AESCrypt_predicate();
297 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
298 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
299 bool inline_ghash_processBlocks();
300 bool inline_sha_implCompress(vmIntrinsics::ID id);
301 bool inline_digestBase_implCompressMB(int predicate);
302 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
303 bool long_state, address stubAddr, const char *stubName,
304 Node* src_start, Node* ofs, Node* limit);
305 Node* get_state_from_sha_object(Node *sha_object);
306 Node* get_state_from_sha5_object(Node *sha_object);
307 Node* inline_digestBase_implCompressMB_predicate(int predicate);
308 bool inline_encodeISOArray();
309 bool inline_updateCRC32();
310 bool inline_updateBytesCRC32();
311 bool inline_updateByteBufferCRC32();
312 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
313 bool inline_updateBytesCRC32C();
314 bool inline_updateDirectByteBufferCRC32C();
315 bool inline_updateBytesAdler32();
316 bool inline_updateByteBufferAdler32();
317 bool inline_multiplyToLen();
318 bool inline_hasNegatives();
319 bool inline_squareToLen();
320 bool inline_mulAdd();
321 bool inline_montgomeryMultiply();
322 bool inline_montgomerySquare();
323 bool inline_vectorizedMismatch();
324 bool inline_fma(vmIntrinsics::ID id);
325
326 bool inline_profileBoolean();
327 bool inline_isCompileConstant();
328 };
329
330 //---------------------------make_vm_intrinsic----------------------------
331 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
332 vmIntrinsics::ID id = m->intrinsic_id();
333 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
334
335 if (!m->is_loaded()) {
336 // Do not attempt to inline unloaded methods.
337 return NULL;
338 }
339
340 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
341 bool is_available = false;
342
343 {
344 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
345 // the compiler must transition to '_thread_in_vm' state because both
346 // methods access VM-internal data.
347 VM_ENTRY_MARK;
348 methodHandle mh(THREAD, m->get_Method());
349 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
350 !C->directive()->is_intrinsic_disabled(mh) &&
351 !vmIntrinsics::is_disabled_by_flags(mh);
352
353 }
354
355 if (is_available) {
356 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
357 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
358 return new LibraryIntrinsic(m, is_virtual,
359 vmIntrinsics::predicates_needed(id),
360 vmIntrinsics::does_virtual_dispatch(id),
361 (vmIntrinsics::ID) id);
362 } else {
363 return NULL;
364 }
365 }
366
367 //----------------------register_library_intrinsics-----------------------
368 // Initialize this file's data structures, for each Compile instance.
369 void Compile::register_library_intrinsics() {
370 // Nothing to do here.
371 }
372
373 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
374 LibraryCallKit kit(jvms, this);
375 Compile* C = kit.C;
376 int nodes = C->unique();
377 #ifndef PRODUCT
378 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
379 char buf[1000];
380 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
381 tty->print_cr("Intrinsic %s", str);
382 }
383 #endif
384 ciMethod* callee = kit.callee();
385 const int bci = kit.bci();
386
387 // Try to inline the intrinsic.
388 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
389 kit.try_to_inline(_last_predicate)) {
390 if (C->print_intrinsics() || C->print_inlining()) {
391 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
392 }
393 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
394 if (C->log()) {
395 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
396 vmIntrinsics::name_at(intrinsic_id()),
397 (is_virtual() ? " virtual='1'" : ""),
398 C->unique() - nodes);
399 }
400 // Push the result from the inlined method onto the stack.
401 kit.push_result();
402 C->print_inlining_update(this);
403 return kit.transfer_exceptions_into_jvms();
404 }
405
406 // The intrinsic bailed out
407 if (C->print_intrinsics() || C->print_inlining()) {
408 if (jvms->has_method()) {
409 // Not a root compile.
410 const char* msg;
411 if (callee->intrinsic_candidate()) {
412 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
413 } else {
414 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
415 : "failed to inline (intrinsic), method not annotated";
416 }
417 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
418 } else {
419 // Root compile
420 tty->print("Did not generate intrinsic %s%s at bci:%d in",
421 vmIntrinsics::name_at(intrinsic_id()),
422 (is_virtual() ? " (virtual)" : ""), bci);
423 }
424 }
425 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
426 C->print_inlining_update(this);
427 return NULL;
428 }
429
430 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
431 LibraryCallKit kit(jvms, this);
432 Compile* C = kit.C;
433 int nodes = C->unique();
434 _last_predicate = predicate;
435 #ifndef PRODUCT
436 assert(is_predicated() && predicate < predicates_count(), "sanity");
437 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
438 char buf[1000];
439 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
440 tty->print_cr("Predicate for intrinsic %s", str);
441 }
442 #endif
443 ciMethod* callee = kit.callee();
444 const int bci = kit.bci();
445
446 Node* slow_ctl = kit.try_to_predicate(predicate);
447 if (!kit.failing()) {
448 if (C->print_intrinsics() || C->print_inlining()) {
449 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
450 }
451 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
452 if (C->log()) {
453 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
454 vmIntrinsics::name_at(intrinsic_id()),
455 (is_virtual() ? " virtual='1'" : ""),
456 C->unique() - nodes);
457 }
458 return slow_ctl; // Could be NULL if the check folds.
459 }
460
461 // The intrinsic bailed out
462 if (C->print_intrinsics() || C->print_inlining()) {
463 if (jvms->has_method()) {
464 // Not a root compile.
465 const char* msg = "failed to generate predicate for intrinsic";
466 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
467 } else {
468 // Root compile
469 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
470 vmIntrinsics::name_at(intrinsic_id()),
471 (is_virtual() ? " (virtual)" : ""), bci);
472 }
473 }
474 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
475 return NULL;
476 }
477
478 bool LibraryCallKit::try_to_inline(int predicate) {
479 // Handle symbolic names for otherwise undistinguished boolean switches:
480 const bool is_store = true;
481 const bool is_compress = true;
482 const bool is_static = true;
483 const bool is_volatile = true;
484
485 if (!jvms()->has_method()) {
486 // Root JVMState has a null method.
487 assert(map()->memory()->Opcode() == Op_Parm, "");
488 // Insert the memory aliasing node
489 set_all_memory(reset_memory());
490 }
491 assert(merged_memory(), "");
492
493
494 switch (intrinsic_id()) {
495 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
496 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
497 case vmIntrinsics::_getClass: return inline_native_getClass();
498
499 case vmIntrinsics::_dsin:
500 case vmIntrinsics::_dcos:
501 case vmIntrinsics::_dtan:
502 case vmIntrinsics::_dabs:
503 case vmIntrinsics::_datan2:
504 case vmIntrinsics::_dsqrt:
505 case vmIntrinsics::_dexp:
506 case vmIntrinsics::_dlog:
507 case vmIntrinsics::_dlog10:
508 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
509
510 case vmIntrinsics::_min:
511 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
512
513 case vmIntrinsics::_notify:
514 case vmIntrinsics::_notifyAll:
515 if (InlineNotify) {
516 return inline_notify(intrinsic_id());
517 }
518 return false;
519
520 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
521 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
522 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
523 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
524 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
525 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
526 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
527 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
528 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
529 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
530 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
531 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
532
533 case vmIntrinsics::_arraycopy: return inline_arraycopy();
534
535 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
536 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
537 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
538 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
539
540 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
541 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
542 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
543 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
544 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
545 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
546 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
547
548 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
549 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
550
551 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
552 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
553 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
554 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
555
556 case vmIntrinsics::_compressStringC:
557 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
558 case vmIntrinsics::_inflateStringC:
559 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
560
561 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
562 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
563 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
564 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
565 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
566 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
567 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
568 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
569 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
570
571 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
572 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
573 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
574 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
575 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
576 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
577 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
578 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
579 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
580
581 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
582 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
583 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
584 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
585 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
586 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
587 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
588 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
589 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
590
591 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
592 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
593 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
594 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
595 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
596 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
597 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
598 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
599 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
600
601 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
602 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
603 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
604 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
605
606 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
607 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
608 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
609 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
610
611 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
612 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
613 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
614 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
615 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
616 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
617 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
618 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
619 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
620
621 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
622 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
623 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
624 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
625 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
626 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
627 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
628 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
629 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
630
631 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
632 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
633 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
634 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
635 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
636 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
637 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
638 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
639 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
640
641 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
642 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
643 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
644 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
645 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
646 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
647 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
648 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
649 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
650
651 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
652 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
653 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
654 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
655 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
656
657 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
658 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
659 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
660 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
661 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
662 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
663 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
664 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
665 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
666 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
667 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
668 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
669 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
670 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
671 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
672 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
673 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
674 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
675 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
676 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
677
678 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
679 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
680 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
681 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
682 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
683 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
684 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
685 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
686 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
687 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
688 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
689 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
690 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
691 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
692 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
693
694 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
695 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
696 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
697 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
698
699 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
700 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
701 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
702 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
703 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
704
705 case vmIntrinsics::_loadFence:
706 case vmIntrinsics::_storeFence:
707 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
708
709 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
710
711 case vmIntrinsics::_currentThread: return inline_native_currentThread();
712 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
713
714 #ifdef TRACE_HAVE_INTRINSICS
715 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
716 case vmIntrinsics::_getClassId: return inline_native_classID();
717 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
718 #endif
719 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
720 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
721 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
722 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
723 case vmIntrinsics::_getLength: return inline_native_getLength();
724 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
725 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
726 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
727 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
728 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
729 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
730
731 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
732 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
733
734 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
735
736 case vmIntrinsics::_isInstance:
737 case vmIntrinsics::_getModifiers:
738 case vmIntrinsics::_isInterface:
739 case vmIntrinsics::_isArray:
740 case vmIntrinsics::_isPrimitive:
741 case vmIntrinsics::_getSuperclass:
742 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
743
744 case vmIntrinsics::_floatToRawIntBits:
745 case vmIntrinsics::_floatToIntBits:
746 case vmIntrinsics::_intBitsToFloat:
747 case vmIntrinsics::_doubleToRawLongBits:
748 case vmIntrinsics::_doubleToLongBits:
749 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
750
751 case vmIntrinsics::_numberOfLeadingZeros_i:
752 case vmIntrinsics::_numberOfLeadingZeros_l:
753 case vmIntrinsics::_numberOfTrailingZeros_i:
754 case vmIntrinsics::_numberOfTrailingZeros_l:
755 case vmIntrinsics::_bitCount_i:
756 case vmIntrinsics::_bitCount_l:
757 case vmIntrinsics::_reverseBytes_i:
758 case vmIntrinsics::_reverseBytes_l:
759 case vmIntrinsics::_reverseBytes_s:
760 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
761
762 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
763
764 case vmIntrinsics::_Reference_get: return inline_reference_get();
765
766 case vmIntrinsics::_Class_cast: return inline_Class_cast();
767
768 case vmIntrinsics::_aescrypt_encryptBlock:
769 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
770
771 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
772 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
773 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
774
775 case vmIntrinsics::_counterMode_AESCrypt:
776 return inline_counterMode_AESCrypt(intrinsic_id());
777
778 case vmIntrinsics::_sha_implCompress:
779 case vmIntrinsics::_sha2_implCompress:
780 case vmIntrinsics::_sha5_implCompress:
781 return inline_sha_implCompress(intrinsic_id());
782
783 case vmIntrinsics::_digestBase_implCompressMB:
784 return inline_digestBase_implCompressMB(predicate);
785
786 case vmIntrinsics::_multiplyToLen:
787 return inline_multiplyToLen();
788
789 case vmIntrinsics::_squareToLen:
790 return inline_squareToLen();
791
792 case vmIntrinsics::_mulAdd:
793 return inline_mulAdd();
794
795 case vmIntrinsics::_montgomeryMultiply:
796 return inline_montgomeryMultiply();
797 case vmIntrinsics::_montgomerySquare:
798 return inline_montgomerySquare();
799
800 case vmIntrinsics::_vectorizedMismatch:
801 return inline_vectorizedMismatch();
802
803 case vmIntrinsics::_ghash_processBlocks:
804 return inline_ghash_processBlocks();
805
806 case vmIntrinsics::_encodeISOArray:
807 case vmIntrinsics::_encodeByteISOArray:
808 return inline_encodeISOArray();
809
810 case vmIntrinsics::_updateCRC32:
811 return inline_updateCRC32();
812 case vmIntrinsics::_updateBytesCRC32:
813 return inline_updateBytesCRC32();
814 case vmIntrinsics::_updateByteBufferCRC32:
815 return inline_updateByteBufferCRC32();
816
817 case vmIntrinsics::_updateBytesCRC32C:
818 return inline_updateBytesCRC32C();
819 case vmIntrinsics::_updateDirectByteBufferCRC32C:
820 return inline_updateDirectByteBufferCRC32C();
821
822 case vmIntrinsics::_updateBytesAdler32:
823 return inline_updateBytesAdler32();
824 case vmIntrinsics::_updateByteBufferAdler32:
825 return inline_updateByteBufferAdler32();
826
827 case vmIntrinsics::_profileBoolean:
828 return inline_profileBoolean();
829 case vmIntrinsics::_isCompileConstant:
830 return inline_isCompileConstant();
831
832 case vmIntrinsics::_hasNegatives:
833 return inline_hasNegatives();
834
835 case vmIntrinsics::_fmaD:
836 case vmIntrinsics::_fmaF:
837 return inline_fma(intrinsic_id());
838
839 default:
840 // If you get here, it may be that someone has added a new intrinsic
841 // to the list in vmSymbols.hpp without implementing it here.
842 #ifndef PRODUCT
843 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
844 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
845 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
846 }
847 #endif
848 return false;
849 }
850 }
851
852 Node* LibraryCallKit::try_to_predicate(int predicate) {
853 if (!jvms()->has_method()) {
854 // Root JVMState has a null method.
855 assert(map()->memory()->Opcode() == Op_Parm, "");
856 // Insert the memory aliasing node
857 set_all_memory(reset_memory());
858 }
859 assert(merged_memory(), "");
860
861 switch (intrinsic_id()) {
862 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
863 return inline_cipherBlockChaining_AESCrypt_predicate(false);
864 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
865 return inline_cipherBlockChaining_AESCrypt_predicate(true);
866 case vmIntrinsics::_counterMode_AESCrypt:
867 return inline_counterMode_AESCrypt_predicate();
868 case vmIntrinsics::_digestBase_implCompressMB:
869 return inline_digestBase_implCompressMB_predicate(predicate);
870
871 default:
872 // If you get here, it may be that someone has added a new intrinsic
873 // to the list in vmSymbols.hpp without implementing it here.
874 #ifndef PRODUCT
875 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
876 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
877 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
878 }
879 #endif
880 Node* slow_ctl = control();
881 set_control(top()); // No fast path instrinsic
882 return slow_ctl;
883 }
884 }
885
886 //------------------------------set_result-------------------------------
887 // Helper function for finishing intrinsics.
888 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
889 record_for_igvn(region);
890 set_control(_gvn.transform(region));
891 set_result( _gvn.transform(value));
892 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
893 }
894
895 //------------------------------generate_guard---------------------------
896 // Helper function for generating guarded fast-slow graph structures.
897 // The given 'test', if true, guards a slow path. If the test fails
898 // then a fast path can be taken. (We generally hope it fails.)
899 // In all cases, GraphKit::control() is updated to the fast path.
900 // The returned value represents the control for the slow path.
901 // The return value is never 'top'; it is either a valid control
902 // or NULL if it is obvious that the slow path can never be taken.
903 // Also, if region and the slow control are not NULL, the slow edge
904 // is appended to the region.
905 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
906 if (stopped()) {
907 // Already short circuited.
908 return NULL;
909 }
910
911 // Build an if node and its projections.
912 // If test is true we take the slow path, which we assume is uncommon.
913 if (_gvn.type(test) == TypeInt::ZERO) {
914 // The slow branch is never taken. No need to build this guard.
915 return NULL;
916 }
917
918 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
919
920 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
921 if (if_slow == top()) {
922 // The slow branch is never taken. No need to build this guard.
923 return NULL;
924 }
925
926 if (region != NULL)
927 region->add_req(if_slow);
928
929 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
930 set_control(if_fast);
931
932 return if_slow;
933 }
934
935 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
936 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
937 }
938 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
939 return generate_guard(test, region, PROB_FAIR);
940 }
941
942 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
943 Node* *pos_index) {
944 if (stopped())
945 return NULL; // already stopped
946 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
947 return NULL; // index is already adequately typed
948 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
949 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
950 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
951 if (is_neg != NULL && pos_index != NULL) {
952 // Emulate effect of Parse::adjust_map_after_if.
953 Node* ccast = new CastIINode(index, TypeInt::POS);
954 ccast->set_req(0, control());
955 (*pos_index) = _gvn.transform(ccast);
956 }
957 return is_neg;
958 }
959
960 // Make sure that 'position' is a valid limit index, in [0..length].
961 // There are two equivalent plans for checking this:
962 // A. (offset + copyLength) unsigned<= arrayLength
963 // B. offset <= (arrayLength - copyLength)
964 // We require that all of the values above, except for the sum and
965 // difference, are already known to be non-negative.
966 // Plan A is robust in the face of overflow, if offset and copyLength
967 // are both hugely positive.
968 //
969 // Plan B is less direct and intuitive, but it does not overflow at
970 // all, since the difference of two non-negatives is always
971 // representable. Whenever Java methods must perform the equivalent
972 // check they generally use Plan B instead of Plan A.
973 // For the moment we use Plan A.
974 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
975 Node* subseq_length,
976 Node* array_length,
977 RegionNode* region) {
978 if (stopped())
979 return NULL; // already stopped
980 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
981 if (zero_offset && subseq_length->eqv_uncast(array_length))
982 return NULL; // common case of whole-array copy
983 Node* last = subseq_length;
984 if (!zero_offset) // last += offset
985 last = _gvn.transform(new AddINode(last, offset));
986 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
987 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
988 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
989 return is_over;
990 }
991
992 // Emit range checks for the given String.value byte array
993 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
994 if (stopped()) {
995 return; // already stopped
996 }
997 RegionNode* bailout = new RegionNode(1);
998 record_for_igvn(bailout);
999 if (char_count) {
1000 // Convert char count to byte count
1001 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1002 }
1003
1004 // Offset and count must not be negative
1005 generate_negative_guard(offset, bailout);
1006 generate_negative_guard(count, bailout);
1007 // Offset + count must not exceed length of array
1008 generate_limit_guard(offset, count, load_array_length(array), bailout);
1009
1010 if (bailout->req() > 1) {
1011 PreserveJVMState pjvms(this);
1012 set_control(_gvn.transform(bailout));
1013 uncommon_trap(Deoptimization::Reason_intrinsic,
1014 Deoptimization::Action_maybe_recompile);
1015 }
1016 }
1017
1018 //--------------------------generate_current_thread--------------------
1019 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1020 ciKlass* thread_klass = env()->Thread_klass();
1021 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1022 Node* thread = _gvn.transform(new ThreadLocalNode());
1023 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1024 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1025 tls_output = thread;
1026 return threadObj;
1027 }
1028
1029
1030 //------------------------------make_string_method_node------------------------
1031 // Helper method for String intrinsic functions. This version is called with
1032 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1033 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1034 // containing the lengths of str1 and str2.
1035 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1036 Node* result = NULL;
1037 switch (opcode) {
1038 case Op_StrIndexOf:
1039 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1040 str1_start, cnt1, str2_start, cnt2, ae);
1041 break;
1042 case Op_StrComp:
1043 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1044 str1_start, cnt1, str2_start, cnt2, ae);
1045 break;
1046 case Op_StrEquals:
1047 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1048 // Use the constant length if there is one because optimized match rule may exist.
1049 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1050 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1051 break;
1052 default:
1053 ShouldNotReachHere();
1054 return NULL;
1055 }
1056
1057 // All these intrinsics have checks.
1058 C->set_has_split_ifs(true); // Has chance for split-if optimization
1059
1060 return _gvn.transform(result);
1061 }
1062
1063 //------------------------------inline_string_compareTo------------------------
1064 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1065 Node* arg1 = argument(0);
1066 Node* arg2 = argument(1);
1067
1068 // Get start addr and length of first argument
1069 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1070 Node* arg1_cnt = load_array_length(arg1);
1071
1072 // Get start addr and length of second argument
1073 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1074 Node* arg2_cnt = load_array_length(arg2);
1075
1076 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1077 set_result(result);
1078 return true;
1079 }
1080
1081 //------------------------------inline_string_equals------------------------
1082 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1083 Node* arg1 = argument(0);
1084 Node* arg2 = argument(1);
1085
1086 // paths (plus control) merge
1087 RegionNode* region = new RegionNode(3);
1088 Node* phi = new PhiNode(region, TypeInt::BOOL);
1089
1090 if (!stopped()) {
1091 // Get start addr and length of first argument
1092 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1093 Node* arg1_cnt = load_array_length(arg1);
1094
1095 // Get start addr and length of second argument
1096 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1097 Node* arg2_cnt = load_array_length(arg2);
1098
1099 // Check for arg1_cnt != arg2_cnt
1100 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1101 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1102 Node* if_ne = generate_slow_guard(bol, NULL);
1103 if (if_ne != NULL) {
1104 phi->init_req(2, intcon(0));
1105 region->init_req(2, if_ne);
1106 }
1107
1108 // Check for count == 0 is done by assembler code for StrEquals.
1109
1110 if (!stopped()) {
1111 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1112 phi->init_req(1, equals);
1113 region->init_req(1, control());
1114 }
1115 }
1116
1117 // post merge
1118 set_control(_gvn.transform(region));
1119 record_for_igvn(region);
1120
1121 set_result(_gvn.transform(phi));
1122 return true;
1123 }
1124
1125 //------------------------------inline_array_equals----------------------------
1126 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1127 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1128 Node* arg1 = argument(0);
1129 Node* arg2 = argument(1);
1130
1131 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1132 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1133 return true;
1134 }
1135
1136 //------------------------------inline_hasNegatives------------------------------
1137 bool LibraryCallKit::inline_hasNegatives() {
1138 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1139 return false;
1140 }
1141
1142 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1143 // no receiver since it is static method
1144 Node* ba = argument(0);
1145 Node* offset = argument(1);
1146 Node* len = argument(2);
1147
1148 // Range checks
1149 generate_string_range_check(ba, offset, len, false);
1150 if (stopped()) {
1151 return true;
1152 }
1153 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1154 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1155 set_result(_gvn.transform(result));
1156 return true;
1157 }
1158
1159 bool LibraryCallKit::inline_preconditions_checkIndex() {
1160 Node* index = argument(0);
1161 Node* length = argument(1);
1162 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1163 return false;
1164 }
1165
1166 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1167 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1168
1169 {
1170 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1171 uncommon_trap(Deoptimization::Reason_intrinsic,
1172 Deoptimization::Action_make_not_entrant);
1173 }
1174
1175 if (stopped()) {
1176 return false;
1177 }
1178
1179 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1180 BoolTest::mask btest = BoolTest::lt;
1181 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1182 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1183 _gvn.set_type(rc, rc->Value(&_gvn));
1184 if (!rc_bool->is_Con()) {
1185 record_for_igvn(rc);
1186 }
1187 set_control(_gvn.transform(new IfTrueNode(rc)));
1188 {
1189 PreserveJVMState pjvms(this);
1190 set_control(_gvn.transform(new IfFalseNode(rc)));
1191 uncommon_trap(Deoptimization::Reason_range_check,
1192 Deoptimization::Action_make_not_entrant);
1193 }
1194
1195 if (stopped()) {
1196 return false;
1197 }
1198
1199 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1200 result->set_req(0, control());
1201 result = _gvn.transform(result);
1202 set_result(result);
1203 replace_in_map(index, result);
1204 return true;
1205 }
1206
1207 //------------------------------inline_string_indexOf------------------------
1208 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1209 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1210 return false;
1211 }
1212 Node* src = argument(0);
1213 Node* tgt = argument(1);
1214
1215 // Make the merge point
1216 RegionNode* result_rgn = new RegionNode(4);
1217 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1218
1219 // Get start addr and length of source string
1220 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1221 Node* src_count = load_array_length(src);
1222
1223 // Get start addr and length of substring
1224 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1225 Node* tgt_count = load_array_length(tgt);
1226
1227 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1228 // Divide src size by 2 if String is UTF16 encoded
1229 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1230 }
1231 if (ae == StrIntrinsicNode::UU) {
1232 // Divide substring size by 2 if String is UTF16 encoded
1233 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1234 }
1235
1236 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1237 if (result != NULL) {
1238 result_phi->init_req(3, result);
1239 result_rgn->init_req(3, control());
1240 }
1241 set_control(_gvn.transform(result_rgn));
1242 record_for_igvn(result_rgn);
1243 set_result(_gvn.transform(result_phi));
1244
1245 return true;
1246 }
1247
1248 //-----------------------------inline_string_indexOf-----------------------
1249 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1250 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1251 return false;
1252 }
1253 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1254 return false;
1255 }
1256 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1257 Node* src = argument(0); // byte[]
1258 Node* src_count = argument(1); // char count
1259 Node* tgt = argument(2); // byte[]
1260 Node* tgt_count = argument(3); // char count
1261 Node* from_index = argument(4); // char index
1262
1263 // Multiply byte array index by 2 if String is UTF16 encoded
1264 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1265 src_count = _gvn.transform(new SubINode(src_count, from_index));
1266 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1267 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1268
1269 // Range checks
1270 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1271 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1272 if (stopped()) {
1273 return true;
1274 }
1275
1276 RegionNode* region = new RegionNode(5);
1277 Node* phi = new PhiNode(region, TypeInt::INT);
1278
1279 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1280 if (result != NULL) {
1281 // The result is index relative to from_index if substring was found, -1 otherwise.
1282 // Generate code which will fold into cmove.
1283 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1284 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1285
1286 Node* if_lt = generate_slow_guard(bol, NULL);
1287 if (if_lt != NULL) {
1288 // result == -1
1289 phi->init_req(3, result);
1290 region->init_req(3, if_lt);
1291 }
1292 if (!stopped()) {
1293 result = _gvn.transform(new AddINode(result, from_index));
1294 phi->init_req(4, result);
1295 region->init_req(4, control());
1296 }
1297 }
1298
1299 set_control(_gvn.transform(region));
1300 record_for_igvn(region);
1301 set_result(_gvn.transform(phi));
1302
1303 return true;
1304 }
1305
1306 // Create StrIndexOfNode with fast path checks
1307 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1308 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1309 // Check for substr count > string count
1310 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1311 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1312 Node* if_gt = generate_slow_guard(bol, NULL);
1313 if (if_gt != NULL) {
1314 phi->init_req(1, intcon(-1));
1315 region->init_req(1, if_gt);
1316 }
1317 if (!stopped()) {
1318 // Check for substr count == 0
1319 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1320 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1321 Node* if_zero = generate_slow_guard(bol, NULL);
1322 if (if_zero != NULL) {
1323 phi->init_req(2, intcon(0));
1324 region->init_req(2, if_zero);
1325 }
1326 }
1327 if (!stopped()) {
1328 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1329 }
1330 return NULL;
1331 }
1332
1333 //-----------------------------inline_string_indexOfChar-----------------------
1334 bool LibraryCallKit::inline_string_indexOfChar() {
1335 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1336 return false;
1337 }
1338 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1339 return false;
1340 }
1341 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1342 Node* src = argument(0); // byte[]
1343 Node* tgt = argument(1); // tgt is int ch
1344 Node* from_index = argument(2);
1345 Node* max = argument(3);
1346
1347 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1348 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1349 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1350
1351 // Range checks
1352 generate_string_range_check(src, src_offset, src_count, true);
1353 if (stopped()) {
1354 return true;
1355 }
1356
1357 RegionNode* region = new RegionNode(3);
1358 Node* phi = new PhiNode(region, TypeInt::INT);
1359
1360 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1361 C->set_has_split_ifs(true); // Has chance for split-if optimization
1362 _gvn.transform(result);
1363
1364 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1365 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1366
1367 Node* if_lt = generate_slow_guard(bol, NULL);
1368 if (if_lt != NULL) {
1369 // result == -1
1370 phi->init_req(2, result);
1371 region->init_req(2, if_lt);
1372 }
1373 if (!stopped()) {
1374 result = _gvn.transform(new AddINode(result, from_index));
1375 phi->init_req(1, result);
1376 region->init_req(1, control());
1377 }
1378 set_control(_gvn.transform(region));
1379 record_for_igvn(region);
1380 set_result(_gvn.transform(phi));
1381
1382 return true;
1383 }
1384 //---------------------------inline_string_copy---------------------
1385 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1386 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1387 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1388 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1389 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1390 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1391 bool LibraryCallKit::inline_string_copy(bool compress) {
1392 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1393 return false;
1394 }
1395 int nargs = 5; // 2 oops, 3 ints
1396 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1397
1398 Node* src = argument(0);
1399 Node* src_offset = argument(1);
1400 Node* dst = argument(2);
1401 Node* dst_offset = argument(3);
1402 Node* length = argument(4);
1403
1404 // Check for allocation before we add nodes that would confuse
1405 // tightly_coupled_allocation()
1406 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1407
1408 // Figure out the size and type of the elements we will be copying.
1409 const Type* src_type = src->Value(&_gvn);
1410 const Type* dst_type = dst->Value(&_gvn);
1411 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1412 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1413 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1414 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1415 "Unsupported array types for inline_string_copy");
1416
1417 // Convert char[] offsets to byte[] offsets
1418 bool convert_src = (compress && src_elem == T_BYTE);
1419 bool convert_dst = (!compress && dst_elem == T_BYTE);
1420 if (convert_src) {
1421 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1422 } else if (convert_dst) {
1423 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1424 }
1425
1426 // Range checks
1427 generate_string_range_check(src, src_offset, length, convert_src);
1428 generate_string_range_check(dst, dst_offset, length, convert_dst);
1429 if (stopped()) {
1430 return true;
1431 }
1432
1433 Node* src_start = array_element_address(src, src_offset, src_elem);
1434 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1435 // 'src_start' points to src array + scaled offset
1436 // 'dst_start' points to dst array + scaled offset
1437 Node* count = NULL;
1438 if (compress) {
1439 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1440 } else {
1441 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1442 }
1443
1444 if (alloc != NULL) {
1445 if (alloc->maybe_set_complete(&_gvn)) {
1446 // "You break it, you buy it."
1447 InitializeNode* init = alloc->initialization();
1448 assert(init->is_complete(), "we just did this");
1449 init->set_complete_with_arraycopy();
1450 assert(dst->is_CheckCastPP(), "sanity");
1451 assert(dst->in(0)->in(0) == init, "dest pinned");
1452 }
1453 // Do not let stores that initialize this object be reordered with
1454 // a subsequent store that would make this object accessible by
1455 // other threads.
1456 // Record what AllocateNode this StoreStore protects so that
1457 // escape analysis can go from the MemBarStoreStoreNode to the
1458 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1459 // based on the escape status of the AllocateNode.
1460 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1461 }
1462 if (compress) {
1463 set_result(_gvn.transform(count));
1464 }
1465 return true;
1466 }
1467
1468 #ifdef _LP64
1469 #define XTOP ,top() /*additional argument*/
1470 #else //_LP64
1471 #define XTOP /*no additional argument*/
1472 #endif //_LP64
1473
1474 //------------------------inline_string_toBytesU--------------------------
1475 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1476 bool LibraryCallKit::inline_string_toBytesU() {
1477 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1478 return false;
1479 }
1480 // Get the arguments.
1481 Node* value = argument(0);
1482 Node* offset = argument(1);
1483 Node* length = argument(2);
1484
1485 Node* newcopy = NULL;
1486
1487 // Set the original stack and the reexecute bit for the interpreter to reexecute
1488 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1489 { PreserveReexecuteState preexecs(this);
1490 jvms()->set_should_reexecute(true);
1491
1492 // Check if a null path was taken unconditionally.
1493 value = null_check(value);
1494
1495 RegionNode* bailout = new RegionNode(1);
1496 record_for_igvn(bailout);
1497
1498 // Range checks
1499 generate_negative_guard(offset, bailout);
1500 generate_negative_guard(length, bailout);
1501 generate_limit_guard(offset, length, load_array_length(value), bailout);
1502 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1503 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1504
1505 if (bailout->req() > 1) {
1506 PreserveJVMState pjvms(this);
1507 set_control(_gvn.transform(bailout));
1508 uncommon_trap(Deoptimization::Reason_intrinsic,
1509 Deoptimization::Action_maybe_recompile);
1510 }
1511 if (stopped()) {
1512 return true;
1513 }
1514
1515 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1516 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1517 newcopy = new_array(klass_node, size, 0); // no arguments to push
1518 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1519
1520 // Calculate starting addresses.
1521 Node* src_start = array_element_address(value, offset, T_CHAR);
1522 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1523
1524 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1525 const TypeInt* toffset = gvn().type(offset)->is_int();
1526 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1527
1528 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1529 const char* copyfunc_name = "arraycopy";
1530 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1531 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1532 OptoRuntime::fast_arraycopy_Type(),
1533 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1534 src_start, dst_start, ConvI2X(length) XTOP);
1535 // Do not let reads from the cloned object float above the arraycopy.
1536 if (alloc != NULL) {
1537 if (alloc->maybe_set_complete(&_gvn)) {
1538 // "You break it, you buy it."
1539 InitializeNode* init = alloc->initialization();
1540 assert(init->is_complete(), "we just did this");
1541 init->set_complete_with_arraycopy();
1542 assert(newcopy->is_CheckCastPP(), "sanity");
1543 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1544 }
1545 // Do not let stores that initialize this object be reordered with
1546 // a subsequent store that would make this object accessible by
1547 // other threads.
1548 // Record what AllocateNode this StoreStore protects so that
1549 // escape analysis can go from the MemBarStoreStoreNode to the
1550 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1551 // based on the escape status of the AllocateNode.
1552 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1553 } else {
1554 insert_mem_bar(Op_MemBarCPUOrder);
1555 }
1556 } // original reexecute is set back here
1557
1558 C->set_has_split_ifs(true); // Has chance for split-if optimization
1559 if (!stopped()) {
1560 set_result(newcopy);
1561 }
1562 return true;
1563 }
1564
1565 //------------------------inline_string_getCharsU--------------------------
1566 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1567 bool LibraryCallKit::inline_string_getCharsU() {
1568 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1569 return false;
1570 }
1571
1572 // Get the arguments.
1573 Node* src = argument(0);
1574 Node* src_begin = argument(1);
1575 Node* src_end = argument(2); // exclusive offset (i < src_end)
1576 Node* dst = argument(3);
1577 Node* dst_begin = argument(4);
1578
1579 // Check for allocation before we add nodes that would confuse
1580 // tightly_coupled_allocation()
1581 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1582
1583 // Check if a null path was taken unconditionally.
1584 src = null_check(src);
1585 dst = null_check(dst);
1586 if (stopped()) {
1587 return true;
1588 }
1589
1590 // Get length and convert char[] offset to byte[] offset
1591 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1592 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1593
1594 // Range checks
1595 generate_string_range_check(src, src_begin, length, true);
1596 generate_string_range_check(dst, dst_begin, length, false);
1597 if (stopped()) {
1598 return true;
1599 }
1600
1601 if (!stopped()) {
1602 // Calculate starting addresses.
1603 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1604 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1605
1606 // Check if array addresses are aligned to HeapWordSize
1607 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1608 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1609 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1610 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1611
1612 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1613 const char* copyfunc_name = "arraycopy";
1614 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1615 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1616 OptoRuntime::fast_arraycopy_Type(),
1617 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1618 src_start, dst_start, ConvI2X(length) XTOP);
1619 // Do not let reads from the cloned object float above the arraycopy.
1620 if (alloc != NULL) {
1621 if (alloc->maybe_set_complete(&_gvn)) {
1622 // "You break it, you buy it."
1623 InitializeNode* init = alloc->initialization();
1624 assert(init->is_complete(), "we just did this");
1625 init->set_complete_with_arraycopy();
1626 assert(dst->is_CheckCastPP(), "sanity");
1627 assert(dst->in(0)->in(0) == init, "dest pinned");
1628 }
1629 // Do not let stores that initialize this object be reordered with
1630 // a subsequent store that would make this object accessible by
1631 // other threads.
1632 // Record what AllocateNode this StoreStore protects so that
1633 // escape analysis can go from the MemBarStoreStoreNode to the
1634 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1635 // based on the escape status of the AllocateNode.
1636 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1637 } else {
1638 insert_mem_bar(Op_MemBarCPUOrder);
1639 }
1640 }
1641
1642 C->set_has_split_ifs(true); // Has chance for split-if optimization
1643 return true;
1644 }
1645
1646 //----------------------inline_string_char_access----------------------------
1647 // Store/Load char to/from byte[] array.
1648 // static void StringUTF16.putChar(byte[] val, int index, int c)
1649 // static char StringUTF16.getChar(byte[] val, int index)
1650 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1651 Node* value = argument(0);
1652 Node* index = argument(1);
1653 Node* ch = is_store ? argument(2) : NULL;
1654
1655 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1656 // correctly requires matched array shapes.
1657 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1658 "sanity: byte[] and char[] bases agree");
1659 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1660 "sanity: byte[] and char[] scales agree");
1661
1662 // Bail when getChar over constants is requested: constant folding would
1663 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1664 // Java method would constant fold nicely instead.
1665 if (!is_store && value->is_Con() && index->is_Con()) {
1666 return false;
1667 }
1668
1669 Node* adr = array_element_address(value, index, T_CHAR);
1670 if (is_store) {
1671 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1672 false, false, true /* mismatched */);
1673 } else {
1674 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1675 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1676 set_result(ch);
1677 }
1678 return true;
1679 }
1680
1681 //--------------------------round_double_node--------------------------------
1682 // Round a double node if necessary.
1683 Node* LibraryCallKit::round_double_node(Node* n) {
1684 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1685 n = _gvn.transform(new RoundDoubleNode(0, n));
1686 return n;
1687 }
1688
1689 //------------------------------inline_math-----------------------------------
1690 // public static double Math.abs(double)
1691 // public static double Math.sqrt(double)
1692 // public static double Math.log(double)
1693 // public static double Math.log10(double)
1694 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1695 Node* arg = round_double_node(argument(0));
1696 Node* n = NULL;
1697 switch (id) {
1698 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1699 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1700 default: fatal_unexpected_iid(id); break;
1701 }
1702 set_result(_gvn.transform(n));
1703 return true;
1704 }
1705
1706 //------------------------------runtime_math-----------------------------
1707 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1708 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1709 "must be (DD)D or (D)D type");
1710
1711 // Inputs
1712 Node* a = round_double_node(argument(0));
1713 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1714
1715 const TypePtr* no_memory_effects = NULL;
1716 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1717 no_memory_effects,
1718 a, top(), b, b ? top() : NULL);
1719 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1720 #ifdef ASSERT
1721 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1722 assert(value_top == top(), "second value must be top");
1723 #endif
1724
1725 set_result(value);
1726 return true;
1727 }
1728
1729 //------------------------------inline_math_native-----------------------------
1730 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1731 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1732 switch (id) {
1733 // These intrinsics are not properly supported on all hardware
1734 case vmIntrinsics::_dsin:
1735 return StubRoutines::dsin() != NULL ?
1736 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1737 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1738 case vmIntrinsics::_dcos:
1739 return StubRoutines::dcos() != NULL ?
1740 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1741 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1742 case vmIntrinsics::_dtan:
1743 return StubRoutines::dtan() != NULL ?
1744 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1745 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1746 case vmIntrinsics::_dlog:
1747 return StubRoutines::dlog() != NULL ?
1748 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1749 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1750 case vmIntrinsics::_dlog10:
1751 return StubRoutines::dlog10() != NULL ?
1752 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1753 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1754
1755 // These intrinsics are supported on all hardware
1756 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1757 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1758
1759 case vmIntrinsics::_dexp:
1760 return StubRoutines::dexp() != NULL ?
1761 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1762 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1763 case vmIntrinsics::_dpow:
1764 return StubRoutines::dpow() != NULL ?
1765 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1766 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1767 #undef FN_PTR
1768
1769 // These intrinsics are not yet correctly implemented
1770 case vmIntrinsics::_datan2:
1771 return false;
1772
1773 default:
1774 fatal_unexpected_iid(id);
1775 return false;
1776 }
1777 }
1778
1779 static bool is_simple_name(Node* n) {
1780 return (n->req() == 1 // constant
1781 || (n->is_Type() && n->as_Type()->type()->singleton())
1782 || n->is_Proj() // parameter or return value
1783 || n->is_Phi() // local of some sort
1784 );
1785 }
1786
1787 //----------------------------inline_notify-----------------------------------*
1788 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1789 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1790 address func;
1791 if (id == vmIntrinsics::_notify) {
1792 func = OptoRuntime::monitor_notify_Java();
1793 } else {
1794 func = OptoRuntime::monitor_notifyAll_Java();
1795 }
1796 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1797 make_slow_call_ex(call, env()->Throwable_klass(), false);
1798 return true;
1799 }
1800
1801
1802 //----------------------------inline_min_max-----------------------------------
1803 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1804 set_result(generate_min_max(id, argument(0), argument(1)));
1805 return true;
1806 }
1807
1808 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1809 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1810 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1811 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1812 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1813
1814 {
1815 PreserveJVMState pjvms(this);
1816 PreserveReexecuteState preexecs(this);
1817 jvms()->set_should_reexecute(true);
1818
1819 set_control(slow_path);
1820 set_i_o(i_o());
1821
1822 uncommon_trap(Deoptimization::Reason_intrinsic,
1823 Deoptimization::Action_none);
1824 }
1825
1826 set_control(fast_path);
1827 set_result(math);
1828 }
1829
1830 template <typename OverflowOp>
1831 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1832 typedef typename OverflowOp::MathOp MathOp;
1833
1834 MathOp* mathOp = new MathOp(arg1, arg2);
1835 Node* operation = _gvn.transform( mathOp );
1836 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1837 inline_math_mathExact(operation, ofcheck);
1838 return true;
1839 }
1840
1841 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1842 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1843 }
1844
1845 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1846 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1847 }
1848
1849 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1850 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1851 }
1852
1853 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1854 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1855 }
1856
1857 bool LibraryCallKit::inline_math_negateExactI() {
1858 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1859 }
1860
1861 bool LibraryCallKit::inline_math_negateExactL() {
1862 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1863 }
1864
1865 bool LibraryCallKit::inline_math_multiplyExactI() {
1866 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1867 }
1868
1869 bool LibraryCallKit::inline_math_multiplyExactL() {
1870 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1871 }
1872
1873 Node*
1874 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1875 // These are the candidate return value:
1876 Node* xvalue = x0;
1877 Node* yvalue = y0;
1878
1879 if (xvalue == yvalue) {
1880 return xvalue;
1881 }
1882
1883 bool want_max = (id == vmIntrinsics::_max);
1884
1885 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1886 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1887 if (txvalue == NULL || tyvalue == NULL) return top();
1888 // This is not really necessary, but it is consistent with a
1889 // hypothetical MaxINode::Value method:
1890 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1891
1892 // %%% This folding logic should (ideally) be in a different place.
1893 // Some should be inside IfNode, and there to be a more reliable
1894 // transformation of ?: style patterns into cmoves. We also want
1895 // more powerful optimizations around cmove and min/max.
1896
1897 // Try to find a dominating comparison of these guys.
1898 // It can simplify the index computation for Arrays.copyOf
1899 // and similar uses of System.arraycopy.
1900 // First, compute the normalized version of CmpI(x, y).
1901 int cmp_op = Op_CmpI;
1902 Node* xkey = xvalue;
1903 Node* ykey = yvalue;
1904 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1905 if (ideal_cmpxy->is_Cmp()) {
1906 // E.g., if we have CmpI(length - offset, count),
1907 // it might idealize to CmpI(length, count + offset)
1908 cmp_op = ideal_cmpxy->Opcode();
1909 xkey = ideal_cmpxy->in(1);
1910 ykey = ideal_cmpxy->in(2);
1911 }
1912
1913 // Start by locating any relevant comparisons.
1914 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1915 Node* cmpxy = NULL;
1916 Node* cmpyx = NULL;
1917 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1918 Node* cmp = start_from->fast_out(k);
1919 if (cmp->outcnt() > 0 && // must have prior uses
1920 cmp->in(0) == NULL && // must be context-independent
1921 cmp->Opcode() == cmp_op) { // right kind of compare
1922 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1923 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1924 }
1925 }
1926
1927 const int NCMPS = 2;
1928 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1929 int cmpn;
1930 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1931 if (cmps[cmpn] != NULL) break; // find a result
1932 }
1933 if (cmpn < NCMPS) {
1934 // Look for a dominating test that tells us the min and max.
1935 int depth = 0; // Limit search depth for speed
1936 Node* dom = control();
1937 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1938 if (++depth >= 100) break;
1939 Node* ifproj = dom;
1940 if (!ifproj->is_Proj()) continue;
1941 Node* iff = ifproj->in(0);
1942 if (!iff->is_If()) continue;
1943 Node* bol = iff->in(1);
1944 if (!bol->is_Bool()) continue;
1945 Node* cmp = bol->in(1);
1946 if (cmp == NULL) continue;
1947 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1948 if (cmps[cmpn] == cmp) break;
1949 if (cmpn == NCMPS) continue;
1950 BoolTest::mask btest = bol->as_Bool()->_test._test;
1951 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1952 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1953 // At this point, we know that 'x btest y' is true.
1954 switch (btest) {
1955 case BoolTest::eq:
1956 // They are proven equal, so we can collapse the min/max.
1957 // Either value is the answer. Choose the simpler.
1958 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1959 return yvalue;
1960 return xvalue;
1961 case BoolTest::lt: // x < y
1962 case BoolTest::le: // x <= y
1963 return (want_max ? yvalue : xvalue);
1964 case BoolTest::gt: // x > y
1965 case BoolTest::ge: // x >= y
1966 return (want_max ? xvalue : yvalue);
1967 }
1968 }
1969 }
1970
1971 // We failed to find a dominating test.
1972 // Let's pick a test that might GVN with prior tests.
1973 Node* best_bol = NULL;
1974 BoolTest::mask best_btest = BoolTest::illegal;
1975 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1976 Node* cmp = cmps[cmpn];
1977 if (cmp == NULL) continue;
1978 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1979 Node* bol = cmp->fast_out(j);
1980 if (!bol->is_Bool()) continue;
1981 BoolTest::mask btest = bol->as_Bool()->_test._test;
1982 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1983 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1984 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1985 best_bol = bol->as_Bool();
1986 best_btest = btest;
1987 }
1988 }
1989 }
1990
1991 Node* answer_if_true = NULL;
1992 Node* answer_if_false = NULL;
1993 switch (best_btest) {
1994 default:
1995 if (cmpxy == NULL)
1996 cmpxy = ideal_cmpxy;
1997 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
1998 // and fall through:
1999 case BoolTest::lt: // x < y
2000 case BoolTest::le: // x <= y
2001 answer_if_true = (want_max ? yvalue : xvalue);
2002 answer_if_false = (want_max ? xvalue : yvalue);
2003 break;
2004 case BoolTest::gt: // x > y
2005 case BoolTest::ge: // x >= y
2006 answer_if_true = (want_max ? xvalue : yvalue);
2007 answer_if_false = (want_max ? yvalue : xvalue);
2008 break;
2009 }
2010
2011 jint hi, lo;
2012 if (want_max) {
2013 // We can sharpen the minimum.
2014 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2015 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2016 } else {
2017 // We can sharpen the maximum.
2018 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2019 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2020 }
2021
2022 // Use a flow-free graph structure, to avoid creating excess control edges
2023 // which could hinder other optimizations.
2024 // Since Math.min/max is often used with arraycopy, we want
2025 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2026 Node* cmov = CMoveNode::make(NULL, best_bol,
2027 answer_if_false, answer_if_true,
2028 TypeInt::make(lo, hi, widen));
2029
2030 return _gvn.transform(cmov);
2031
2032 /*
2033 // This is not as desirable as it may seem, since Min and Max
2034 // nodes do not have a full set of optimizations.
2035 // And they would interfere, anyway, with 'if' optimizations
2036 // and with CMoveI canonical forms.
2037 switch (id) {
2038 case vmIntrinsics::_min:
2039 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2040 case vmIntrinsics::_max:
2041 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2042 default:
2043 ShouldNotReachHere();
2044 }
2045 */
2046 }
2047
2048 inline int
2049 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2050 const TypePtr* base_type = TypePtr::NULL_PTR;
2051 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2052 if (base_type == NULL) {
2053 // Unknown type.
2054 return Type::AnyPtr;
2055 } else if (base_type == TypePtr::NULL_PTR) {
2056 // Since this is a NULL+long form, we have to switch to a rawptr.
2057 base = _gvn.transform(new CastX2PNode(offset));
2058 offset = MakeConX(0);
2059 return Type::RawPtr;
2060 } else if (base_type->base() == Type::RawPtr) {
2061 return Type::RawPtr;
2062 } else if (base_type->isa_oopptr()) {
2063 // Base is never null => always a heap address.
2064 if (base_type->ptr() == TypePtr::NotNull) {
2065 return Type::OopPtr;
2066 }
2067 // Offset is small => always a heap address.
2068 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2069 if (offset_type != NULL &&
2070 base_type->offset() == 0 && // (should always be?)
2071 offset_type->_lo >= 0 &&
2072 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2073 return Type::OopPtr;
2074 }
2075 // Otherwise, it might either be oop+off or NULL+addr.
2076 return Type::AnyPtr;
2077 } else {
2078 // No information:
2079 return Type::AnyPtr;
2080 }
2081 }
2082
2083 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2084 int kind = classify_unsafe_addr(base, offset);
2085 if (kind == Type::RawPtr) {
2086 return basic_plus_adr(top(), base, offset);
2087 } else {
2088 return basic_plus_adr(base, offset);
2089 }
2090 }
2091
2092 //--------------------------inline_number_methods-----------------------------
2093 // inline int Integer.numberOfLeadingZeros(int)
2094 // inline int Long.numberOfLeadingZeros(long)
2095 //
2096 // inline int Integer.numberOfTrailingZeros(int)
2097 // inline int Long.numberOfTrailingZeros(long)
2098 //
2099 // inline int Integer.bitCount(int)
2100 // inline int Long.bitCount(long)
2101 //
2102 // inline char Character.reverseBytes(char)
2103 // inline short Short.reverseBytes(short)
2104 // inline int Integer.reverseBytes(int)
2105 // inline long Long.reverseBytes(long)
2106 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2107 Node* arg = argument(0);
2108 Node* n = NULL;
2109 switch (id) {
2110 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2111 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2112 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2113 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2114 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2115 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2116 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2117 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2118 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2119 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2120 default: fatal_unexpected_iid(id); break;
2121 }
2122 set_result(_gvn.transform(n));
2123 return true;
2124 }
2125
2126 //----------------------------inline_unsafe_access----------------------------
2127
2128 // Helper that guards and inserts a pre-barrier.
2129 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2130 Node* pre_val, bool need_mem_bar) {
2131 // We could be accessing the referent field of a reference object. If so, when G1
2132 // is enabled, we need to log the value in the referent field in an SATB buffer.
2133 // This routine performs some compile time filters and generates suitable
2134 // runtime filters that guard the pre-barrier code.
2135 // Also add memory barrier for non volatile load from the referent field
2136 // to prevent commoning of loads across safepoint.
2137 if (!UseG1GC && !need_mem_bar)
2138 return;
2139
2140 // Some compile time checks.
2141
2142 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2143 const TypeX* otype = offset->find_intptr_t_type();
2144 if (otype != NULL && otype->is_con() &&
2145 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2146 // Constant offset but not the reference_offset so just return
2147 return;
2148 }
2149
2150 // We only need to generate the runtime guards for instances.
2151 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2152 if (btype != NULL) {
2153 if (btype->isa_aryptr()) {
2154 // Array type so nothing to do
2155 return;
2156 }
2157
2158 const TypeInstPtr* itype = btype->isa_instptr();
2159 if (itype != NULL) {
2160 // Can the klass of base_oop be statically determined to be
2161 // _not_ a sub-class of Reference and _not_ Object?
2162 ciKlass* klass = itype->klass();
2163 if ( klass->is_loaded() &&
2164 !klass->is_subtype_of(env()->Reference_klass()) &&
2165 !env()->Object_klass()->is_subtype_of(klass)) {
2166 return;
2167 }
2168 }
2169 }
2170
2171 // The compile time filters did not reject base_oop/offset so
2172 // we need to generate the following runtime filters
2173 //
2174 // if (offset == java_lang_ref_Reference::_reference_offset) {
2175 // if (instance_of(base, java.lang.ref.Reference)) {
2176 // pre_barrier(_, pre_val, ...);
2177 // }
2178 // }
2179
2180 float likely = PROB_LIKELY( 0.999);
2181 float unlikely = PROB_UNLIKELY(0.999);
2182
2183 IdealKit ideal(this);
2184 #define __ ideal.
2185
2186 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2187
2188 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2189 // Update graphKit memory and control from IdealKit.
2190 sync_kit(ideal);
2191
2192 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2193 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2194
2195 // Update IdealKit memory and control from graphKit.
2196 __ sync_kit(this);
2197
2198 Node* one = __ ConI(1);
2199 // is_instof == 0 if base_oop == NULL
2200 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2201
2202 // Update graphKit from IdeakKit.
2203 sync_kit(ideal);
2204
2205 // Use the pre-barrier to record the value in the referent field
2206 pre_barrier(false /* do_load */,
2207 __ ctrl(),
2208 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2209 pre_val /* pre_val */,
2210 T_OBJECT);
2211 if (need_mem_bar) {
2212 // Add memory barrier to prevent commoning reads from this field
2213 // across safepoint since GC can change its value.
2214 insert_mem_bar(Op_MemBarCPUOrder);
2215 }
2216 // Update IdealKit from graphKit.
2217 __ sync_kit(this);
2218
2219 } __ end_if(); // _ref_type != ref_none
2220 } __ end_if(); // offset == referent_offset
2221
2222 // Final sync IdealKit and GraphKit.
2223 final_sync(ideal);
2224 #undef __
2225 }
2226
2227
2228 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2229 // Attempt to infer a sharper value type from the offset and base type.
2230 ciKlass* sharpened_klass = NULL;
2231
2232 // See if it is an instance field, with an object type.
2233 if (alias_type->field() != NULL) {
2234 if (alias_type->field()->type()->is_klass()) {
2235 sharpened_klass = alias_type->field()->type()->as_klass();
2236 }
2237 }
2238
2239 // See if it is a narrow oop array.
2240 if (adr_type->isa_aryptr()) {
2241 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2242 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2243 if (elem_type != NULL) {
2244 sharpened_klass = elem_type->klass();
2245 }
2246 }
2247 }
2248
2249 // The sharpened class might be unloaded if there is no class loader
2250 // contraint in place.
2251 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2252 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2253
2254 #ifndef PRODUCT
2255 if (C->print_intrinsics() || C->print_inlining()) {
2256 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2257 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2258 }
2259 #endif
2260 // Sharpen the value type.
2261 return tjp;
2262 }
2263 return NULL;
2264 }
2265
2266 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2267 if (callee()->is_static()) return false; // caller must have the capability!
2268 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2269 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2270 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2271
2272 #ifndef PRODUCT
2273 {
2274 ResourceMark rm;
2275 // Check the signatures.
2276 ciSignature* sig = callee()->signature();
2277 #ifdef ASSERT
2278 if (!is_store) {
2279 // Object getObject(Object base, int/long offset), etc.
2280 BasicType rtype = sig->return_type()->basic_type();
2281 assert(rtype == type, "getter must return the expected value");
2282 assert(sig->count() == 2, "oop getter has 2 arguments");
2283 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2284 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2285 } else {
2286 // void putObject(Object base, int/long offset, Object x), etc.
2287 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2288 assert(sig->count() == 3, "oop putter has 3 arguments");
2289 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2290 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2291 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2292 assert(vtype == type, "putter must accept the expected value");
2293 }
2294 #endif // ASSERT
2295 }
2296 #endif //PRODUCT
2297
2298 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2299
2300 Node* receiver = argument(0); // type: oop
2301
2302 // Build address expression.
2303 Node* adr;
2304 Node* heap_base_oop = top();
2305 Node* offset = top();
2306 Node* val;
2307
2308 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2309 Node* base = argument(1); // type: oop
2310 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2311 offset = argument(2); // type: long
2312 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2313 // to be plain byte offsets, which are also the same as those accepted
2314 // by oopDesc::field_base.
2315 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2316 "fieldOffset must be byte-scaled");
2317 // 32-bit machines ignore the high half!
2318 offset = ConvL2X(offset);
2319 adr = make_unsafe_address(base, offset);
2320 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2321 heap_base_oop = base;
2322 } else if (type == T_OBJECT) {
2323 return false; // off-heap oop accesses are not supported
2324 }
2325
2326 // Can base be NULL? Otherwise, always on-heap access.
2327 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2328
2329 val = is_store ? argument(4) : NULL;
2330
2331 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2332
2333 // Try to categorize the address.
2334 Compile::AliasType* alias_type = C->alias_type(adr_type);
2335 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2336
2337 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2338 alias_type->adr_type() == TypeAryPtr::RANGE) {
2339 return false; // not supported
2340 }
2341
2342 bool mismatched = false;
2343 BasicType bt = alias_type->basic_type();
2344 if (bt != T_ILLEGAL) {
2345 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2346 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2347 // Alias type doesn't differentiate between byte[] and boolean[]).
2348 // Use address type to get the element type.
2349 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2350 }
2351 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2352 // accessing an array field with getObject is not a mismatch
2353 bt = T_OBJECT;
2354 }
2355 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2356 // Don't intrinsify mismatched object accesses
2357 return false;
2358 }
2359 mismatched = (bt != type);
2360 } else if (alias_type->adr_type()->isa_oopptr()) {
2361 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2362 }
2363
2364 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2365
2366 // First guess at the value type.
2367 const Type *value_type = Type::get_const_basic_type(type);
2368
2369 // We will need memory barriers unless we can determine a unique
2370 // alias category for this reference. (Note: If for some reason
2371 // the barriers get omitted and the unsafe reference begins to "pollute"
2372 // the alias analysis of the rest of the graph, either Compile::can_alias
2373 // or Compile::must_alias will throw a diagnostic assert.)
2374 bool need_mem_bar;
2375 switch (kind) {
2376 case Relaxed:
2377 need_mem_bar = mismatched || can_access_non_heap;
2378 break;
2379 case Opaque:
2380 // Opaque uses CPUOrder membars for protection against code movement.
2381 case Acquire:
2382 case Release:
2383 case Volatile:
2384 need_mem_bar = true;
2385 break;
2386 default:
2387 ShouldNotReachHere();
2388 }
2389
2390 // Some accesses require access atomicity for all types, notably longs and doubles.
2391 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2392 bool requires_atomic_access = false;
2393 switch (kind) {
2394 case Relaxed:
2395 requires_atomic_access = AlwaysAtomicAccesses;
2396 break;
2397 case Opaque:
2398 // Opaque accesses are atomic.
2399 case Acquire:
2400 case Release:
2401 case Volatile:
2402 requires_atomic_access = true;
2403 break;
2404 default:
2405 ShouldNotReachHere();
2406 }
2407
2408 // Figure out the memory ordering.
2409 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2410 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2411
2412 // If we are reading the value of the referent field of a Reference
2413 // object (either by using Unsafe directly or through reflection)
2414 // then, if G1 is enabled, we need to record the referent in an
2415 // SATB log buffer using the pre-barrier mechanism.
2416 // Also we need to add memory barrier to prevent commoning reads
2417 // from this field across safepoint since GC can change its value.
2418 bool need_read_barrier = !is_store &&
2419 offset != top() && heap_base_oop != top();
2420
2421 if (!is_store && type == T_OBJECT) {
2422 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2423 if (tjp != NULL) {
2424 value_type = tjp;
2425 }
2426 }
2427
2428 receiver = null_check(receiver);
2429 if (stopped()) {
2430 return true;
2431 }
2432 // Heap pointers get a null-check from the interpreter,
2433 // as a courtesy. However, this is not guaranteed by Unsafe,
2434 // and it is not possible to fully distinguish unintended nulls
2435 // from intended ones in this API.
2436
2437 // We need to emit leading and trailing CPU membars (see below) in
2438 // addition to memory membars for special access modes. This is a little
2439 // too strong, but avoids the need to insert per-alias-type
2440 // volatile membars (for stores; compare Parse::do_put_xxx), which
2441 // we cannot do effectively here because we probably only have a
2442 // rough approximation of type.
2443
2444 switch(kind) {
2445 case Relaxed:
2446 case Opaque:
2447 case Acquire:
2448 break;
2449 case Release:
2450 case Volatile:
2451 if (is_store) {
2452 insert_mem_bar(Op_MemBarRelease);
2453 } else {
2454 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2455 insert_mem_bar(Op_MemBarVolatile);
2456 }
2457 }
2458 break;
2459 default:
2460 ShouldNotReachHere();
2461 }
2462
2463 // Memory barrier to prevent normal and 'unsafe' accesses from
2464 // bypassing each other. Happens after null checks, so the
2465 // exception paths do not take memory state from the memory barrier,
2466 // so there's no problems making a strong assert about mixing users
2467 // of safe & unsafe memory.
2468 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2469
2470 if (!is_store) {
2471 Node* p = NULL;
2472 // Try to constant fold a load from a constant field
2473 ciField* field = alias_type->field();
2474 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2475 // final or stable field
2476 p = make_constant_from_field(field, heap_base_oop);
2477 }
2478 if (p == NULL) {
2479 // To be valid, unsafe loads may depend on other conditions than
2480 // the one that guards them: pin the Load node
2481 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched);
2482 // load value
2483 switch (type) {
2484 case T_BOOLEAN:
2485 {
2486 // Normalize the value returned by getBoolean in the following cases
2487 if (mismatched ||
2488 heap_base_oop == top() || // - heap_base_oop is NULL or
2489 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2490 // and the unsafe access is made to large offset
2491 // (i.e., larger than the maximum offset necessary for any
2492 // field access)
2493 ) {
2494 IdealKit ideal = IdealKit(this);
2495 #define __ ideal.
2496 IdealVariable normalized_result(ideal);
2497 __ declarations_done();
2498 __ set(normalized_result, p);
2499 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2500 __ set(normalized_result, ideal.ConI(1));
2501 ideal.end_if();
2502 final_sync(ideal);
2503 p = __ value(normalized_result);
2504 #undef __
2505 }
2506 }
2507 case T_CHAR:
2508 case T_BYTE:
2509 case T_SHORT:
2510 case T_INT:
2511 case T_LONG:
2512 case T_FLOAT:
2513 case T_DOUBLE:
2514 break;
2515 case T_OBJECT:
2516 if (need_read_barrier) {
2517 // We do not require a mem bar inside pre_barrier if need_mem_bar
2518 // is set: the barriers would be emitted by us.
2519 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2520 }
2521 break;
2522 case T_ADDRESS:
2523 // Cast to an int type.
2524 p = _gvn.transform(new CastP2XNode(NULL, p));
2525 p = ConvX2UL(p);
2526 break;
2527 default:
2528 fatal("unexpected type %d: %s", type, type2name(type));
2529 break;
2530 }
2531 }
2532 // The load node has the control of the preceding MemBarCPUOrder. All
2533 // following nodes will have the control of the MemBarCPUOrder inserted at
2534 // the end of this method. So, pushing the load onto the stack at a later
2535 // point is fine.
2536 set_result(p);
2537 } else {
2538 // place effect of store into memory
2539 switch (type) {
2540 case T_DOUBLE:
2541 val = dstore_rounding(val);
2542 break;
2543 case T_ADDRESS:
2544 // Repackage the long as a pointer.
2545 val = ConvL2X(val);
2546 val = _gvn.transform(new CastX2PNode(val));
2547 break;
2548 }
2549
2550 if (type == T_OBJECT) {
2551 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2552 } else {
2553 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2554 }
2555 }
2556
2557 switch(kind) {
2558 case Relaxed:
2559 case Opaque:
2560 case Release:
2561 break;
2562 case Acquire:
2563 case Volatile:
2564 if (!is_store) {
2565 insert_mem_bar(Op_MemBarAcquire);
2566 } else {
2567 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2568 insert_mem_bar(Op_MemBarVolatile);
2569 }
2570 }
2571 break;
2572 default:
2573 ShouldNotReachHere();
2574 }
2575
2576 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2577
2578 return true;
2579 }
2580
2581 //----------------------------inline_unsafe_load_store----------------------------
2582 // This method serves a couple of different customers (depending on LoadStoreKind):
2583 //
2584 // LS_cmp_swap:
2585 //
2586 // boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2587 // boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2588 // boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2589 //
2590 // LS_cmp_swap_weak:
2591 //
2592 // boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x);
2593 // boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x);
2594 // boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x);
2595 //
2596 // boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x);
2597 // boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x);
2598 // boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x);
2599 //
2600 // boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x);
2601 // boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x);
2602 // boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x);
2603 //
2604 // LS_cmp_exchange:
2605 //
2606 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2607 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2608 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2609 //
2610 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2611 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2612 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2613 //
2614 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2615 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2616 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2617 //
2618 // LS_get_add:
2619 //
2620 // int getAndAddInt( Object o, long offset, int delta)
2621 // long getAndAddLong(Object o, long offset, long delta)
2622 //
2623 // LS_get_set:
2624 //
2625 // int getAndSet(Object o, long offset, int newValue)
2626 // long getAndSet(Object o, long offset, long newValue)
2627 // Object getAndSet(Object o, long offset, Object newValue)
2628 //
2629 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2630 // This basic scheme here is the same as inline_unsafe_access, but
2631 // differs in enough details that combining them would make the code
2632 // overly confusing. (This is a true fact! I originally combined
2633 // them, but even I was confused by it!) As much code/comments as
2634 // possible are retained from inline_unsafe_access though to make
2635 // the correspondences clearer. - dl
2636
2637 if (callee()->is_static()) return false; // caller must have the capability!
2638
2639 #ifndef PRODUCT
2640 BasicType rtype;
2641 {
2642 ResourceMark rm;
2643 // Check the signatures.
2644 ciSignature* sig = callee()->signature();
2645 rtype = sig->return_type()->basic_type();
2646 switch(kind) {
2647 case LS_get_add:
2648 case LS_get_set: {
2649 // Check the signatures.
2650 #ifdef ASSERT
2651 assert(rtype == type, "get and set must return the expected type");
2652 assert(sig->count() == 3, "get and set has 3 arguments");
2653 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2654 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2655 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2656 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2657 #endif // ASSERT
2658 break;
2659 }
2660 case LS_cmp_swap:
2661 case LS_cmp_swap_weak: {
2662 // Check the signatures.
2663 #ifdef ASSERT
2664 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2665 assert(sig->count() == 4, "CAS has 4 arguments");
2666 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2667 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2668 #endif // ASSERT
2669 break;
2670 }
2671 case LS_cmp_exchange: {
2672 // Check the signatures.
2673 #ifdef ASSERT
2674 assert(rtype == type, "CAS must return the expected type");
2675 assert(sig->count() == 4, "CAS has 4 arguments");
2676 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2677 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2678 #endif // ASSERT
2679 break;
2680 }
2681 default:
2682 ShouldNotReachHere();
2683 }
2684 }
2685 #endif //PRODUCT
2686
2687 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2688
2689 // Get arguments:
2690 Node* receiver = NULL;
2691 Node* base = NULL;
2692 Node* offset = NULL;
2693 Node* oldval = NULL;
2694 Node* newval = NULL;
2695 switch(kind) {
2696 case LS_cmp_swap:
2697 case LS_cmp_swap_weak:
2698 case LS_cmp_exchange: {
2699 const bool two_slot_type = type2size[type] == 2;
2700 receiver = argument(0); // type: oop
2701 base = argument(1); // type: oop
2702 offset = argument(2); // type: long
2703 oldval = argument(4); // type: oop, int, or long
2704 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2705 break;
2706 }
2707 case LS_get_add:
2708 case LS_get_set: {
2709 receiver = argument(0); // type: oop
2710 base = argument(1); // type: oop
2711 offset = argument(2); // type: long
2712 oldval = NULL;
2713 newval = argument(4); // type: oop, int, or long
2714 break;
2715 }
2716 default:
2717 ShouldNotReachHere();
2718 }
2719
2720 // Build field offset expression.
2721 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2722 // to be plain byte offsets, which are also the same as those accepted
2723 // by oopDesc::field_base.
2724 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2725 // 32-bit machines ignore the high half of long offsets
2726 offset = ConvL2X(offset);
2727 Node* adr = make_unsafe_address(base, offset);
2728 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2729
2730 Compile::AliasType* alias_type = C->alias_type(adr_type);
2731 BasicType bt = alias_type->basic_type();
2732 if (bt != T_ILLEGAL &&
2733 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2734 // Don't intrinsify mismatched object accesses.
2735 return false;
2736 }
2737
2738 // For CAS, unlike inline_unsafe_access, there seems no point in
2739 // trying to refine types. Just use the coarse types here.
2740 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2741 const Type *value_type = Type::get_const_basic_type(type);
2742
2743 switch (kind) {
2744 case LS_get_set:
2745 case LS_cmp_exchange: {
2746 if (type == T_OBJECT) {
2747 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2748 if (tjp != NULL) {
2749 value_type = tjp;
2750 }
2751 }
2752 break;
2753 }
2754 case LS_cmp_swap:
2755 case LS_cmp_swap_weak:
2756 case LS_get_add:
2757 break;
2758 default:
2759 ShouldNotReachHere();
2760 }
2761
2762 // Null check receiver.
2763 receiver = null_check(receiver);
2764 if (stopped()) {
2765 return true;
2766 }
2767
2768 int alias_idx = C->get_alias_index(adr_type);
2769
2770 // Memory-model-wise, a LoadStore acts like a little synchronized
2771 // block, so needs barriers on each side. These don't translate
2772 // into actual barriers on most machines, but we still need rest of
2773 // compiler to respect ordering.
2774
2775 switch (access_kind) {
2776 case Relaxed:
2777 case Acquire:
2778 break;
2779 case Release:
2780 insert_mem_bar(Op_MemBarRelease);
2781 break;
2782 case Volatile:
2783 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2784 insert_mem_bar(Op_MemBarVolatile);
2785 } else {
2786 insert_mem_bar(Op_MemBarRelease);
2787 }
2788 break;
2789 default:
2790 ShouldNotReachHere();
2791 }
2792 insert_mem_bar(Op_MemBarCPUOrder);
2793
2794 // Figure out the memory ordering.
2795 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2796
2797 // 4984716: MemBars must be inserted before this
2798 // memory node in order to avoid a false
2799 // dependency which will confuse the scheduler.
2800 Node *mem = memory(alias_idx);
2801
2802 // For now, we handle only those cases that actually exist: ints,
2803 // longs, and Object. Adding others should be straightforward.
2804 Node* load_store = NULL;
2805 switch(type) {
2806 case T_BYTE:
2807 switch(kind) {
2808 case LS_get_add:
2809 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2810 break;
2811 case LS_get_set:
2812 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2813 break;
2814 case LS_cmp_swap_weak:
2815 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2816 break;
2817 case LS_cmp_swap:
2818 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2819 break;
2820 case LS_cmp_exchange:
2821 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2822 break;
2823 default:
2824 ShouldNotReachHere();
2825 }
2826 break;
2827 case T_SHORT:
2828 switch(kind) {
2829 case LS_get_add:
2830 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2831 break;
2832 case LS_get_set:
2833 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2834 break;
2835 case LS_cmp_swap_weak:
2836 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2837 break;
2838 case LS_cmp_swap:
2839 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2840 break;
2841 case LS_cmp_exchange:
2842 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2843 break;
2844 default:
2845 ShouldNotReachHere();
2846 }
2847 break;
2848 case T_INT:
2849 switch(kind) {
2850 case LS_get_add:
2851 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2852 break;
2853 case LS_get_set:
2854 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2855 break;
2856 case LS_cmp_swap_weak:
2857 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2858 break;
2859 case LS_cmp_swap:
2860 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2861 break;
2862 case LS_cmp_exchange:
2863 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2864 break;
2865 default:
2866 ShouldNotReachHere();
2867 }
2868 break;
2869 case T_LONG:
2870 switch(kind) {
2871 case LS_get_add:
2872 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2873 break;
2874 case LS_get_set:
2875 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2876 break;
2877 case LS_cmp_swap_weak:
2878 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2879 break;
2880 case LS_cmp_swap:
2881 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2882 break;
2883 case LS_cmp_exchange:
2884 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2885 break;
2886 default:
2887 ShouldNotReachHere();
2888 }
2889 break;
2890 case T_OBJECT:
2891 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2892 // could be delayed during Parse (for example, in adjust_map_after_if()).
2893 // Execute transformation here to avoid barrier generation in such case.
2894 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2895 newval = _gvn.makecon(TypePtr::NULL_PTR);
2896
2897 // Reference stores need a store barrier.
2898 switch(kind) {
2899 case LS_get_set: {
2900 // If pre-barrier must execute before the oop store, old value will require do_load here.
2901 if (!can_move_pre_barrier()) {
2902 pre_barrier(true /* do_load*/,
2903 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2904 NULL /* pre_val*/,
2905 T_OBJECT);
2906 } // Else move pre_barrier to use load_store value, see below.
2907 break;
2908 }
2909 case LS_cmp_swap_weak:
2910 case LS_cmp_swap:
2911 case LS_cmp_exchange: {
2912 // Same as for newval above:
2913 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2914 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2915 }
2916 // The only known value which might get overwritten is oldval.
2917 pre_barrier(false /* do_load */,
2918 control(), NULL, NULL, max_juint, NULL, NULL,
2919 oldval /* pre_val */,
2920 T_OBJECT);
2921 break;
2922 }
2923 default:
2924 ShouldNotReachHere();
2925 }
2926
2927 #ifdef _LP64
2928 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2929 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2930
2931 switch(kind) {
2932 case LS_get_set:
2933 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
2934 break;
2935 case LS_cmp_swap_weak: {
2936 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2937 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2938 break;
2939 }
2940 case LS_cmp_swap: {
2941 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2942 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2943 break;
2944 }
2945 case LS_cmp_exchange: {
2946 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2947 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
2948 break;
2949 }
2950 default:
2951 ShouldNotReachHere();
2952 }
2953 } else
2954 #endif
2955 switch (kind) {
2956 case LS_get_set:
2957 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2958 break;
2959 case LS_cmp_swap_weak:
2960 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2961 break;
2962 case LS_cmp_swap:
2963 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2964 break;
2965 case LS_cmp_exchange:
2966 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
2967 break;
2968 default:
2969 ShouldNotReachHere();
2970 }
2971
2972 // Emit the post barrier only when the actual store happened. This makes sense
2973 // to check only for LS_cmp_* that can fail to set the value.
2974 // LS_cmp_exchange does not produce any branches by default, so there is no
2975 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
2976 // CompareAndExchange and move branches here, it would make sense to conditionalize
2977 // post_barriers for LS_cmp_exchange as well.
2978 //
2979 // CAS success path is marked more likely since we anticipate this is a performance
2980 // critical path, while CAS failure path can use the penalty for going through unlikely
2981 // path as backoff. Which is still better than doing a store barrier there.
2982 switch (kind) {
2983 case LS_get_set:
2984 case LS_cmp_exchange: {
2985 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2986 break;
2987 }
2988 case LS_cmp_swap_weak:
2989 case LS_cmp_swap: {
2990 IdealKit ideal(this);
2991 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2992 sync_kit(ideal);
2993 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2994 ideal.sync_kit(this);
2995 } ideal.end_if();
2996 final_sync(ideal);
2997 break;
2998 }
2999 default:
3000 ShouldNotReachHere();
3001 }
3002 break;
3003 default:
3004 fatal("unexpected type %d: %s", type, type2name(type));
3005 break;
3006 }
3007
3008 // SCMemProjNodes represent the memory state of a LoadStore. Their
3009 // main role is to prevent LoadStore nodes from being optimized away
3010 // when their results aren't used.
3011 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3012 set_memory(proj, alias_idx);
3013
3014 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3015 #ifdef _LP64
3016 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3017 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3018 }
3019 #endif
3020 if (can_move_pre_barrier()) {
3021 // Don't need to load pre_val. The old value is returned by load_store.
3022 // The pre_barrier can execute after the xchg as long as no safepoint
3023 // gets inserted between them.
3024 pre_barrier(false /* do_load */,
3025 control(), NULL, NULL, max_juint, NULL, NULL,
3026 load_store /* pre_val */,
3027 T_OBJECT);
3028 }
3029 }
3030
3031 // Add the trailing membar surrounding the access
3032 insert_mem_bar(Op_MemBarCPUOrder);
3033
3034 switch (access_kind) {
3035 case Relaxed:
3036 case Release:
3037 break; // do nothing
3038 case Acquire:
3039 case Volatile:
3040 insert_mem_bar(Op_MemBarAcquire);
3041 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3042 break;
3043 default:
3044 ShouldNotReachHere();
3045 }
3046
3047 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3048 set_result(load_store);
3049 return true;
3050 }
3051
3052 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3053 MemNode::MemOrd mo = MemNode::unset;
3054 switch(kind) {
3055 case Opaque:
3056 case Relaxed: mo = MemNode::unordered; break;
3057 case Acquire: mo = MemNode::acquire; break;
3058 case Release: mo = MemNode::release; break;
3059 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3060 default:
3061 ShouldNotReachHere();
3062 }
3063 guarantee(mo != MemNode::unset, "Should select memory ordering");
3064 return mo;
3065 }
3066
3067 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3068 MemNode::MemOrd mo = MemNode::unset;
3069 switch(kind) {
3070 case Opaque:
3071 case Relaxed: mo = MemNode::unordered; break;
3072 case Acquire: mo = MemNode::acquire; break;
3073 case Release: mo = MemNode::release; break;
3074 case Volatile: mo = MemNode::seqcst; break;
3075 default:
3076 ShouldNotReachHere();
3077 }
3078 guarantee(mo != MemNode::unset, "Should select memory ordering");
3079 return mo;
3080 }
3081
3082 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3083 // Regardless of form, don't allow previous ld/st to move down,
3084 // then issue acquire, release, or volatile mem_bar.
3085 insert_mem_bar(Op_MemBarCPUOrder);
3086 switch(id) {
3087 case vmIntrinsics::_loadFence:
3088 insert_mem_bar(Op_LoadFence);
3089 return true;
3090 case vmIntrinsics::_storeFence:
3091 insert_mem_bar(Op_StoreFence);
3092 return true;
3093 case vmIntrinsics::_fullFence:
3094 insert_mem_bar(Op_MemBarVolatile);
3095 return true;
3096 default:
3097 fatal_unexpected_iid(id);
3098 return false;
3099 }
3100 }
3101
3102 bool LibraryCallKit::inline_onspinwait() {
3103 insert_mem_bar(Op_OnSpinWait);
3104 return true;
3105 }
3106
3107 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3108 if (!kls->is_Con()) {
3109 return true;
3110 }
3111 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3112 if (klsptr == NULL) {
3113 return true;
3114 }
3115 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3116 // don't need a guard for a klass that is already initialized
3117 return !ik->is_initialized();
3118 }
3119
3120 //----------------------------inline_unsafe_allocate---------------------------
3121 // public native Object Unsafe.allocateInstance(Class<?> cls);
3122 bool LibraryCallKit::inline_unsafe_allocate() {
3123 if (callee()->is_static()) return false; // caller must have the capability!
3124
3125 null_check_receiver(); // null-check, then ignore
3126 Node* cls = null_check(argument(1));
3127 if (stopped()) return true;
3128
3129 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3130 kls = null_check(kls);
3131 if (stopped()) return true; // argument was like int.class
3132
3133 Node* test = NULL;
3134 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3135 // Note: The argument might still be an illegal value like
3136 // Serializable.class or Object[].class. The runtime will handle it.
3137 // But we must make an explicit check for initialization.
3138 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3139 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3140 // can generate code to load it as unsigned byte.
3141 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3142 Node* bits = intcon(InstanceKlass::fully_initialized);
3143 test = _gvn.transform(new SubINode(inst, bits));
3144 // The 'test' is non-zero if we need to take a slow path.
3145 }
3146
3147 Node* obj = new_instance(kls, test);
3148 set_result(obj);
3149 return true;
3150 }
3151
3152 //------------------------inline_native_time_funcs--------------
3153 // inline code for System.currentTimeMillis() and System.nanoTime()
3154 // these have the same type and signature
3155 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3156 const TypeFunc* tf = OptoRuntime::void_long_Type();
3157 const TypePtr* no_memory_effects = NULL;
3158 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3159 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3160 #ifdef ASSERT
3161 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3162 assert(value_top == top(), "second value must be top");
3163 #endif
3164 set_result(value);
3165 return true;
3166 }
3167
3168 #ifdef TRACE_HAVE_INTRINSICS
3169
3170 /*
3171 * oop -> myklass
3172 * myklass->trace_id |= USED
3173 * return myklass->trace_id & ~0x3
3174 */
3175 bool LibraryCallKit::inline_native_classID() {
3176 Node* cls = null_check(argument(0), T_OBJECT);
3177 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3178 kls = null_check(kls, T_OBJECT);
3179
3180 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3181 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3182 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3183
3184 Node* clsused = longcon(0x01l); // set the class bit
3185 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3186 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3187 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3188
3189 #ifdef TRACE_ID_META_BITS
3190 Node* mbits = longcon(~TRACE_ID_META_BITS);
3191 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3192 #endif
3193 #ifdef TRACE_ID_CLASS_SHIFT
3194 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3195 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3196 #endif
3197
3198 set_result(tvalue);
3199 return true;
3200
3201 }
3202
3203 bool LibraryCallKit::inline_native_getBufferWriter() {
3204 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3205
3206 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3207 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3208 );
3209
3210 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3211
3212 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3213 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3214
3215 IfNode* iff_jobj_null =
3216 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3217
3218 enum { _normal_path = 1,
3219 _null_path = 2,
3220 PATH_LIMIT };
3221
3222 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3223 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3224
3225 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3226 result_rgn->init_req(_null_path, jobj_is_null);
3227 result_val->init_req(_null_path, null());
3228
3229 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3230 result_rgn->init_req(_normal_path, jobj_is_not_null);
3231
3232 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3233 result_val->init_req(_normal_path, res);
3234
3235 set_result(result_rgn, result_val);
3236
3237 return true;
3238 }
3239
3240 #endif
3241
3242 //------------------------inline_native_currentThread------------------
3243 bool LibraryCallKit::inline_native_currentThread() {
3244 Node* junk = NULL;
3245 set_result(generate_current_thread(junk));
3246 return true;
3247 }
3248
3249 //------------------------inline_native_isInterrupted------------------
3250 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3251 bool LibraryCallKit::inline_native_isInterrupted() {
3252 // Add a fast path to t.isInterrupted(clear_int):
3253 // (t == Thread.current() &&
3254 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3255 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3256 // So, in the common case that the interrupt bit is false,
3257 // we avoid making a call into the VM. Even if the interrupt bit
3258 // is true, if the clear_int argument is false, we avoid the VM call.
3259 // However, if the receiver is not currentThread, we must call the VM,
3260 // because there must be some locking done around the operation.
3261
3262 // We only go to the fast case code if we pass two guards.
3263 // Paths which do not pass are accumulated in the slow_region.
3264
3265 enum {
3266 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3267 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3268 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3269 PATH_LIMIT
3270 };
3271
3272 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3273 // out of the function.
3274 insert_mem_bar(Op_MemBarCPUOrder);
3275
3276 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3277 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3278
3279 RegionNode* slow_region = new RegionNode(1);
3280 record_for_igvn(slow_region);
3281
3282 // (a) Receiving thread must be the current thread.
3283 Node* rec_thr = argument(0);
3284 Node* tls_ptr = NULL;
3285 Node* cur_thr = generate_current_thread(tls_ptr);
3286 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3287 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3288
3289 generate_slow_guard(bol_thr, slow_region);
3290
3291 // (b) Interrupt bit on TLS must be false.
3292 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3293 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3294 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3295
3296 // Set the control input on the field _interrupted read to prevent it floating up.
3297 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3298 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3299 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3300
3301 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3302
3303 // First fast path: if (!TLS._interrupted) return false;
3304 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3305 result_rgn->init_req(no_int_result_path, false_bit);
3306 result_val->init_req(no_int_result_path, intcon(0));
3307
3308 // drop through to next case
3309 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3310
3311 #ifndef _WINDOWS
3312 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3313 Node* clr_arg = argument(1);
3314 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3315 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3316 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3317
3318 // Second fast path: ... else if (!clear_int) return true;
3319 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3320 result_rgn->init_req(no_clear_result_path, false_arg);
3321 result_val->init_req(no_clear_result_path, intcon(1));
3322
3323 // drop through to next case
3324 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3325 #else
3326 // To return true on Windows you must read the _interrupted field
3327 // and check the event state i.e. take the slow path.
3328 #endif // _WINDOWS
3329
3330 // (d) Otherwise, go to the slow path.
3331 slow_region->add_req(control());
3332 set_control( _gvn.transform(slow_region));
3333
3334 if (stopped()) {
3335 // There is no slow path.
3336 result_rgn->init_req(slow_result_path, top());
3337 result_val->init_req(slow_result_path, top());
3338 } else {
3339 // non-virtual because it is a private non-static
3340 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3341
3342 Node* slow_val = set_results_for_java_call(slow_call);
3343 // this->control() comes from set_results_for_java_call
3344
3345 Node* fast_io = slow_call->in(TypeFunc::I_O);
3346 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3347
3348 // These two phis are pre-filled with copies of of the fast IO and Memory
3349 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3350 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3351
3352 result_rgn->init_req(slow_result_path, control());
3353 result_io ->init_req(slow_result_path, i_o());
3354 result_mem->init_req(slow_result_path, reset_memory());
3355 result_val->init_req(slow_result_path, slow_val);
3356
3357 set_all_memory(_gvn.transform(result_mem));
3358 set_i_o( _gvn.transform(result_io));
3359 }
3360
3361 C->set_has_split_ifs(true); // Has chance for split-if optimization
3362 set_result(result_rgn, result_val);
3363 return true;
3364 }
3365
3366 //---------------------------load_mirror_from_klass----------------------------
3367 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3368 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3369 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3370 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3371 }
3372
3373 //-----------------------load_klass_from_mirror_common-------------------------
3374 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3375 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3376 // and branch to the given path on the region.
3377 // If never_see_null, take an uncommon trap on null, so we can optimistically
3378 // compile for the non-null case.
3379 // If the region is NULL, force never_see_null = true.
3380 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3381 bool never_see_null,
3382 RegionNode* region,
3383 int null_path,
3384 int offset) {
3385 if (region == NULL) never_see_null = true;
3386 Node* p = basic_plus_adr(mirror, offset);
3387 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3388 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3389 Node* null_ctl = top();
3390 kls = null_check_oop(kls, &null_ctl, never_see_null);
3391 if (region != NULL) {
3392 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3393 region->init_req(null_path, null_ctl);
3394 } else {
3395 assert(null_ctl == top(), "no loose ends");
3396 }
3397 return kls;
3398 }
3399
3400 //--------------------(inline_native_Class_query helpers)---------------------
3401 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3402 // Fall through if (mods & mask) == bits, take the guard otherwise.
3403 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3404 // Branch around if the given klass has the given modifier bit set.
3405 // Like generate_guard, adds a new path onto the region.
3406 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3407 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3408 Node* mask = intcon(modifier_mask);
3409 Node* bits = intcon(modifier_bits);
3410 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3411 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3412 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3413 return generate_fair_guard(bol, region);
3414 }
3415 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3416 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3417 }
3418
3419 //-------------------------inline_native_Class_query-------------------
3420 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3421 const Type* return_type = TypeInt::BOOL;
3422 Node* prim_return_value = top(); // what happens if it's a primitive class?
3423 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3424 bool expect_prim = false; // most of these guys expect to work on refs
3425
3426 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3427
3428 Node* mirror = argument(0);
3429 Node* obj = top();
3430
3431 switch (id) {
3432 case vmIntrinsics::_isInstance:
3433 // nothing is an instance of a primitive type
3434 prim_return_value = intcon(0);
3435 obj = argument(1);
3436 break;
3437 case vmIntrinsics::_getModifiers:
3438 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3439 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3440 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3441 break;
3442 case vmIntrinsics::_isInterface:
3443 prim_return_value = intcon(0);
3444 break;
3445 case vmIntrinsics::_isArray:
3446 prim_return_value = intcon(0);
3447 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3448 break;
3449 case vmIntrinsics::_isPrimitive:
3450 prim_return_value = intcon(1);
3451 expect_prim = true; // obviously
3452 break;
3453 case vmIntrinsics::_getSuperclass:
3454 prim_return_value = null();
3455 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3456 break;
3457 case vmIntrinsics::_getClassAccessFlags:
3458 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3459 return_type = TypeInt::INT; // not bool! 6297094
3460 break;
3461 default:
3462 fatal_unexpected_iid(id);
3463 break;
3464 }
3465
3466 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3467 if (mirror_con == NULL) return false; // cannot happen?
3468
3469 #ifndef PRODUCT
3470 if (C->print_intrinsics() || C->print_inlining()) {
3471 ciType* k = mirror_con->java_mirror_type();
3472 if (k) {
3473 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3474 k->print_name();
3475 tty->cr();
3476 }
3477 }
3478 #endif
3479
3480 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3481 RegionNode* region = new RegionNode(PATH_LIMIT);
3482 record_for_igvn(region);
3483 PhiNode* phi = new PhiNode(region, return_type);
3484
3485 // The mirror will never be null of Reflection.getClassAccessFlags, however
3486 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3487 // if it is. See bug 4774291.
3488
3489 // For Reflection.getClassAccessFlags(), the null check occurs in
3490 // the wrong place; see inline_unsafe_access(), above, for a similar
3491 // situation.
3492 mirror = null_check(mirror);
3493 // If mirror or obj is dead, only null-path is taken.
3494 if (stopped()) return true;
3495
3496 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3497
3498 // Now load the mirror's klass metaobject, and null-check it.
3499 // Side-effects region with the control path if the klass is null.
3500 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3501 // If kls is null, we have a primitive mirror.
3502 phi->init_req(_prim_path, prim_return_value);
3503 if (stopped()) { set_result(region, phi); return true; }
3504 bool safe_for_replace = (region->in(_prim_path) == top());
3505
3506 Node* p; // handy temp
3507 Node* null_ctl;
3508
3509 // Now that we have the non-null klass, we can perform the real query.
3510 // For constant classes, the query will constant-fold in LoadNode::Value.
3511 Node* query_value = top();
3512 switch (id) {
3513 case vmIntrinsics::_isInstance:
3514 // nothing is an instance of a primitive type
3515 query_value = gen_instanceof(obj, kls, safe_for_replace);
3516 break;
3517
3518 case vmIntrinsics::_getModifiers:
3519 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3520 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3521 break;
3522
3523 case vmIntrinsics::_isInterface:
3524 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3525 if (generate_interface_guard(kls, region) != NULL)
3526 // A guard was added. If the guard is taken, it was an interface.
3527 phi->add_req(intcon(1));
3528 // If we fall through, it's a plain class.
3529 query_value = intcon(0);
3530 break;
3531
3532 case vmIntrinsics::_isArray:
3533 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3534 if (generate_array_guard(kls, region) != NULL)
3535 // A guard was added. If the guard is taken, it was an array.
3536 phi->add_req(intcon(1));
3537 // If we fall through, it's a plain class.
3538 query_value = intcon(0);
3539 break;
3540
3541 case vmIntrinsics::_isPrimitive:
3542 query_value = intcon(0); // "normal" path produces false
3543 break;
3544
3545 case vmIntrinsics::_getSuperclass:
3546 // The rules here are somewhat unfortunate, but we can still do better
3547 // with random logic than with a JNI call.
3548 // Interfaces store null or Object as _super, but must report null.
3549 // Arrays store an intermediate super as _super, but must report Object.
3550 // Other types can report the actual _super.
3551 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3552 if (generate_interface_guard(kls, region) != NULL)
3553 // A guard was added. If the guard is taken, it was an interface.
3554 phi->add_req(null());
3555 if (generate_array_guard(kls, region) != NULL)
3556 // A guard was added. If the guard is taken, it was an array.
3557 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3558 // If we fall through, it's a plain class. Get its _super.
3559 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3560 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3561 null_ctl = top();
3562 kls = null_check_oop(kls, &null_ctl);
3563 if (null_ctl != top()) {
3564 // If the guard is taken, Object.superClass is null (both klass and mirror).
3565 region->add_req(null_ctl);
3566 phi ->add_req(null());
3567 }
3568 if (!stopped()) {
3569 query_value = load_mirror_from_klass(kls);
3570 }
3571 break;
3572
3573 case vmIntrinsics::_getClassAccessFlags:
3574 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3575 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3576 break;
3577
3578 default:
3579 fatal_unexpected_iid(id);
3580 break;
3581 }
3582
3583 // Fall-through is the normal case of a query to a real class.
3584 phi->init_req(1, query_value);
3585 region->init_req(1, control());
3586
3587 C->set_has_split_ifs(true); // Has chance for split-if optimization
3588 set_result(region, phi);
3589 return true;
3590 }
3591
3592 //-------------------------inline_Class_cast-------------------
3593 bool LibraryCallKit::inline_Class_cast() {
3594 Node* mirror = argument(0); // Class
3595 Node* obj = argument(1);
3596 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3597 if (mirror_con == NULL) {
3598 return false; // dead path (mirror->is_top()).
3599 }
3600 if (obj == NULL || obj->is_top()) {
3601 return false; // dead path
3602 }
3603 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3604
3605 // First, see if Class.cast() can be folded statically.
3606 // java_mirror_type() returns non-null for compile-time Class constants.
3607 ciType* tm = mirror_con->java_mirror_type();
3608 if (tm != NULL && tm->is_klass() &&
3609 tp != NULL && tp->klass() != NULL) {
3610 if (!tp->klass()->is_loaded()) {
3611 // Don't use intrinsic when class is not loaded.
3612 return false;
3613 } else {
3614 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3615 if (static_res == Compile::SSC_always_true) {
3616 // isInstance() is true - fold the code.
3617 set_result(obj);
3618 return true;
3619 } else if (static_res == Compile::SSC_always_false) {
3620 // Don't use intrinsic, have to throw ClassCastException.
3621 // If the reference is null, the non-intrinsic bytecode will
3622 // be optimized appropriately.
3623 return false;
3624 }
3625 }
3626 }
3627
3628 // Bailout intrinsic and do normal inlining if exception path is frequent.
3629 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3630 return false;
3631 }
3632
3633 // Generate dynamic checks.
3634 // Class.cast() is java implementation of _checkcast bytecode.
3635 // Do checkcast (Parse::do_checkcast()) optimizations here.
3636
3637 mirror = null_check(mirror);
3638 // If mirror is dead, only null-path is taken.
3639 if (stopped()) {
3640 return true;
3641 }
3642
3643 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3644 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3645 RegionNode* region = new RegionNode(PATH_LIMIT);
3646 record_for_igvn(region);
3647
3648 // Now load the mirror's klass metaobject, and null-check it.
3649 // If kls is null, we have a primitive mirror and
3650 // nothing is an instance of a primitive type.
3651 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3652
3653 Node* res = top();
3654 if (!stopped()) {
3655 Node* bad_type_ctrl = top();
3656 // Do checkcast optimizations.
3657 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3658 region->init_req(_bad_type_path, bad_type_ctrl);
3659 }
3660 if (region->in(_prim_path) != top() ||
3661 region->in(_bad_type_path) != top()) {
3662 // Let Interpreter throw ClassCastException.
3663 PreserveJVMState pjvms(this);
3664 set_control(_gvn.transform(region));
3665 uncommon_trap(Deoptimization::Reason_intrinsic,
3666 Deoptimization::Action_maybe_recompile);
3667 }
3668 if (!stopped()) {
3669 set_result(res);
3670 }
3671 return true;
3672 }
3673
3674
3675 //--------------------------inline_native_subtype_check------------------------
3676 // This intrinsic takes the JNI calls out of the heart of
3677 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3678 bool LibraryCallKit::inline_native_subtype_check() {
3679 // Pull both arguments off the stack.
3680 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3681 args[0] = argument(0);
3682 args[1] = argument(1);
3683 Node* klasses[2]; // corresponding Klasses: superk, subk
3684 klasses[0] = klasses[1] = top();
3685
3686 enum {
3687 // A full decision tree on {superc is prim, subc is prim}:
3688 _prim_0_path = 1, // {P,N} => false
3689 // {P,P} & superc!=subc => false
3690 _prim_same_path, // {P,P} & superc==subc => true
3691 _prim_1_path, // {N,P} => false
3692 _ref_subtype_path, // {N,N} & subtype check wins => true
3693 _both_ref_path, // {N,N} & subtype check loses => false
3694 PATH_LIMIT
3695 };
3696
3697 RegionNode* region = new RegionNode(PATH_LIMIT);
3698 Node* phi = new PhiNode(region, TypeInt::BOOL);
3699 record_for_igvn(region);
3700
3701 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3702 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3703 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3704
3705 // First null-check both mirrors and load each mirror's klass metaobject.
3706 int which_arg;
3707 for (which_arg = 0; which_arg <= 1; which_arg++) {
3708 Node* arg = args[which_arg];
3709 arg = null_check(arg);
3710 if (stopped()) break;
3711 args[which_arg] = arg;
3712
3713 Node* p = basic_plus_adr(arg, class_klass_offset);
3714 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3715 klasses[which_arg] = _gvn.transform(kls);
3716 }
3717
3718 // Having loaded both klasses, test each for null.
3719 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3720 for (which_arg = 0; which_arg <= 1; which_arg++) {
3721 Node* kls = klasses[which_arg];
3722 Node* null_ctl = top();
3723 kls = null_check_oop(kls, &null_ctl, never_see_null);
3724 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3725 region->init_req(prim_path, null_ctl);
3726 if (stopped()) break;
3727 klasses[which_arg] = kls;
3728 }
3729
3730 if (!stopped()) {
3731 // now we have two reference types, in klasses[0..1]
3732 Node* subk = klasses[1]; // the argument to isAssignableFrom
3733 Node* superk = klasses[0]; // the receiver
3734 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3735 // now we have a successful reference subtype check
3736 region->set_req(_ref_subtype_path, control());
3737 }
3738
3739 // If both operands are primitive (both klasses null), then
3740 // we must return true when they are identical primitives.
3741 // It is convenient to test this after the first null klass check.
3742 set_control(region->in(_prim_0_path)); // go back to first null check
3743 if (!stopped()) {
3744 // Since superc is primitive, make a guard for the superc==subc case.
3745 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3746 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3747 generate_guard(bol_eq, region, PROB_FAIR);
3748 if (region->req() == PATH_LIMIT+1) {
3749 // A guard was added. If the added guard is taken, superc==subc.
3750 region->swap_edges(PATH_LIMIT, _prim_same_path);
3751 region->del_req(PATH_LIMIT);
3752 }
3753 region->set_req(_prim_0_path, control()); // Not equal after all.
3754 }
3755
3756 // these are the only paths that produce 'true':
3757 phi->set_req(_prim_same_path, intcon(1));
3758 phi->set_req(_ref_subtype_path, intcon(1));
3759
3760 // pull together the cases:
3761 assert(region->req() == PATH_LIMIT, "sane region");
3762 for (uint i = 1; i < region->req(); i++) {
3763 Node* ctl = region->in(i);
3764 if (ctl == NULL || ctl == top()) {
3765 region->set_req(i, top());
3766 phi ->set_req(i, top());
3767 } else if (phi->in(i) == NULL) {
3768 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3769 }
3770 }
3771
3772 set_control(_gvn.transform(region));
3773 set_result(_gvn.transform(phi));
3774 return true;
3775 }
3776
3777 //---------------------generate_array_guard_common------------------------
3778 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3779 bool obj_array, bool not_array) {
3780
3781 if (stopped()) {
3782 return NULL;
3783 }
3784
3785 // If obj_array/non_array==false/false:
3786 // Branch around if the given klass is in fact an array (either obj or prim).
3787 // If obj_array/non_array==false/true:
3788 // Branch around if the given klass is not an array klass of any kind.
3789 // If obj_array/non_array==true/true:
3790 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3791 // If obj_array/non_array==true/false:
3792 // Branch around if the kls is an oop array (Object[] or subtype)
3793 //
3794 // Like generate_guard, adds a new path onto the region.
3795 jint layout_con = 0;
3796 Node* layout_val = get_layout_helper(kls, layout_con);
3797 if (layout_val == NULL) {
3798 bool query = (obj_array
3799 ? Klass::layout_helper_is_objArray(layout_con)
3800 : Klass::layout_helper_is_array(layout_con));
3801 if (query == not_array) {
3802 return NULL; // never a branch
3803 } else { // always a branch
3804 Node* always_branch = control();
3805 if (region != NULL)
3806 region->add_req(always_branch);
3807 set_control(top());
3808 return always_branch;
3809 }
3810 }
3811 // Now test the correct condition.
3812 jint nval = (obj_array
3813 ? (jint)(Klass::_lh_array_tag_type_value
3814 << Klass::_lh_array_tag_shift)
3815 : Klass::_lh_neutral_value);
3816 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3817 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3818 // invert the test if we are looking for a non-array
3819 if (not_array) btest = BoolTest(btest).negate();
3820 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3821 return generate_fair_guard(bol, region);
3822 }
3823
3824
3825 //-----------------------inline_native_newArray--------------------------
3826 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3827 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3828 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3829 Node* mirror;
3830 Node* count_val;
3831 if (uninitialized) {
3832 mirror = argument(1);
3833 count_val = argument(2);
3834 } else {
3835 mirror = argument(0);
3836 count_val = argument(1);
3837 }
3838
3839 mirror = null_check(mirror);
3840 // If mirror or obj is dead, only null-path is taken.
3841 if (stopped()) return true;
3842
3843 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3844 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3845 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3846 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3847 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3848
3849 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3850 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3851 result_reg, _slow_path);
3852 Node* normal_ctl = control();
3853 Node* no_array_ctl = result_reg->in(_slow_path);
3854
3855 // Generate code for the slow case. We make a call to newArray().
3856 set_control(no_array_ctl);
3857 if (!stopped()) {
3858 // Either the input type is void.class, or else the
3859 // array klass has not yet been cached. Either the
3860 // ensuing call will throw an exception, or else it
3861 // will cache the array klass for next time.
3862 PreserveJVMState pjvms(this);
3863 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3864 Node* slow_result = set_results_for_java_call(slow_call);
3865 // this->control() comes from set_results_for_java_call
3866 result_reg->set_req(_slow_path, control());
3867 result_val->set_req(_slow_path, slow_result);
3868 result_io ->set_req(_slow_path, i_o());
3869 result_mem->set_req(_slow_path, reset_memory());
3870 }
3871
3872 set_control(normal_ctl);
3873 if (!stopped()) {
3874 // Normal case: The array type has been cached in the java.lang.Class.
3875 // The following call works fine even if the array type is polymorphic.
3876 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3877 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3878 result_reg->init_req(_normal_path, control());
3879 result_val->init_req(_normal_path, obj);
3880 result_io ->init_req(_normal_path, i_o());
3881 result_mem->init_req(_normal_path, reset_memory());
3882
3883 if (uninitialized) {
3884 // Mark the allocation so that zeroing is skipped
3885 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3886 alloc->maybe_set_complete(&_gvn);
3887 }
3888 }
3889
3890 // Return the combined state.
3891 set_i_o( _gvn.transform(result_io) );
3892 set_all_memory( _gvn.transform(result_mem));
3893
3894 C->set_has_split_ifs(true); // Has chance for split-if optimization
3895 set_result(result_reg, result_val);
3896 return true;
3897 }
3898
3899 //----------------------inline_native_getLength--------------------------
3900 // public static native int java.lang.reflect.Array.getLength(Object array);
3901 bool LibraryCallKit::inline_native_getLength() {
3902 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3903
3904 Node* array = null_check(argument(0));
3905 // If array is dead, only null-path is taken.
3906 if (stopped()) return true;
3907
3908 // Deoptimize if it is a non-array.
3909 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3910
3911 if (non_array != NULL) {
3912 PreserveJVMState pjvms(this);
3913 set_control(non_array);
3914 uncommon_trap(Deoptimization::Reason_intrinsic,
3915 Deoptimization::Action_maybe_recompile);
3916 }
3917
3918 // If control is dead, only non-array-path is taken.
3919 if (stopped()) return true;
3920
3921 // The works fine even if the array type is polymorphic.
3922 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3923 Node* result = load_array_length(array);
3924
3925 C->set_has_split_ifs(true); // Has chance for split-if optimization
3926 set_result(result);
3927 return true;
3928 }
3929
3930 //------------------------inline_array_copyOf----------------------------
3931 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3932 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3933 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3934 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3935
3936 // Get the arguments.
3937 Node* original = argument(0);
3938 Node* start = is_copyOfRange? argument(1): intcon(0);
3939 Node* end = is_copyOfRange? argument(2): argument(1);
3940 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3941
3942 Node* newcopy = NULL;
3943
3944 // Set the original stack and the reexecute bit for the interpreter to reexecute
3945 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3946 { PreserveReexecuteState preexecs(this);
3947 jvms()->set_should_reexecute(true);
3948
3949 array_type_mirror = null_check(array_type_mirror);
3950 original = null_check(original);
3951
3952 // Check if a null path was taken unconditionally.
3953 if (stopped()) return true;
3954
3955 Node* orig_length = load_array_length(original);
3956
3957 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3958 klass_node = null_check(klass_node);
3959
3960 RegionNode* bailout = new RegionNode(1);
3961 record_for_igvn(bailout);
3962
3963 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3964 // Bail out if that is so.
3965 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3966 if (not_objArray != NULL) {
3967 // Improve the klass node's type from the new optimistic assumption:
3968 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3969 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3970 Node* cast = new CastPPNode(klass_node, akls);
3971 cast->init_req(0, control());
3972 klass_node = _gvn.transform(cast);
3973 }
3974
3975 // Bail out if either start or end is negative.
3976 generate_negative_guard(start, bailout, &start);
3977 generate_negative_guard(end, bailout, &end);
3978
3979 Node* length = end;
3980 if (_gvn.type(start) != TypeInt::ZERO) {
3981 length = _gvn.transform(new SubINode(end, start));
3982 }
3983
3984 // Bail out if length is negative.
3985 // Without this the new_array would throw
3986 // NegativeArraySizeException but IllegalArgumentException is what
3987 // should be thrown
3988 generate_negative_guard(length, bailout, &length);
3989
3990 if (bailout->req() > 1) {
3991 PreserveJVMState pjvms(this);
3992 set_control(_gvn.transform(bailout));
3993 uncommon_trap(Deoptimization::Reason_intrinsic,
3994 Deoptimization::Action_maybe_recompile);
3995 }
3996
3997 if (!stopped()) {
3998 // How many elements will we copy from the original?
3999 // The answer is MinI(orig_length - start, length).
4000 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4001 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4002
4003 // Generate a direct call to the right arraycopy function(s).
4004 // We know the copy is disjoint but we might not know if the
4005 // oop stores need checking.
4006 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4007 // This will fail a store-check if x contains any non-nulls.
4008
4009 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4010 // loads/stores but it is legal only if we're sure the
4011 // Arrays.copyOf would succeed. So we need all input arguments
4012 // to the copyOf to be validated, including that the copy to the
4013 // new array won't trigger an ArrayStoreException. That subtype
4014 // check can be optimized if we know something on the type of
4015 // the input array from type speculation.
4016 if (_gvn.type(klass_node)->singleton()) {
4017 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4018 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4019
4020 int test = C->static_subtype_check(superk, subk);
4021 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4022 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4023 if (t_original->speculative_type() != NULL) {
4024 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4025 }
4026 }
4027 }
4028
4029 bool validated = false;
4030 // Reason_class_check rather than Reason_intrinsic because we
4031 // want to intrinsify even if this traps.
4032 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4033 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4034 klass_node);
4035
4036 if (not_subtype_ctrl != top()) {
4037 PreserveJVMState pjvms(this);
4038 set_control(not_subtype_ctrl);
4039 uncommon_trap(Deoptimization::Reason_class_check,
4040 Deoptimization::Action_make_not_entrant);
4041 assert(stopped(), "Should be stopped");
4042 }
4043 validated = true;
4044 }
4045
4046 if (!stopped()) {
4047 newcopy = new_array(klass_node, length, 0); // no arguments to push
4048
4049 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4050 load_object_klass(original), klass_node);
4051 if (!is_copyOfRange) {
4052 ac->set_copyof(validated);
4053 } else {
4054 ac->set_copyofrange(validated);
4055 }
4056 Node* n = _gvn.transform(ac);
4057 if (n == ac) {
4058 ac->connect_outputs(this);
4059 } else {
4060 assert(validated, "shouldn't transform if all arguments not validated");
4061 set_all_memory(n);
4062 }
4063 }
4064 }
4065 } // original reexecute is set back here
4066
4067 C->set_has_split_ifs(true); // Has chance for split-if optimization
4068 if (!stopped()) {
4069 set_result(newcopy);
4070 }
4071 return true;
4072 }
4073
4074
4075 //----------------------generate_virtual_guard---------------------------
4076 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4077 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4078 RegionNode* slow_region) {
4079 ciMethod* method = callee();
4080 int vtable_index = method->vtable_index();
4081 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4082 "bad index %d", vtable_index);
4083 // Get the Method* out of the appropriate vtable entry.
4084 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4085 vtable_index*vtableEntry::size_in_bytes() +
4086 vtableEntry::method_offset_in_bytes();
4087 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4088 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4089
4090 // Compare the target method with the expected method (e.g., Object.hashCode).
4091 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4092
4093 Node* native_call = makecon(native_call_addr);
4094 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4095 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4096
4097 return generate_slow_guard(test_native, slow_region);
4098 }
4099
4100 //-----------------------generate_method_call----------------------------
4101 // Use generate_method_call to make a slow-call to the real
4102 // method if the fast path fails. An alternative would be to
4103 // use a stub like OptoRuntime::slow_arraycopy_Java.
4104 // This only works for expanding the current library call,
4105 // not another intrinsic. (E.g., don't use this for making an
4106 // arraycopy call inside of the copyOf intrinsic.)
4107 CallJavaNode*
4108 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4109 // When compiling the intrinsic method itself, do not use this technique.
4110 guarantee(callee() != C->method(), "cannot make slow-call to self");
4111
4112 ciMethod* method = callee();
4113 // ensure the JVMS we have will be correct for this call
4114 guarantee(method_id == method->intrinsic_id(), "must match");
4115
4116 const TypeFunc* tf = TypeFunc::make(method);
4117 CallJavaNode* slow_call;
4118 if (is_static) {
4119 assert(!is_virtual, "");
4120 slow_call = new CallStaticJavaNode(C, tf,
4121 SharedRuntime::get_resolve_static_call_stub(),
4122 method, bci());
4123 } else if (is_virtual) {
4124 null_check_receiver();
4125 int vtable_index = Method::invalid_vtable_index;
4126 if (UseInlineCaches) {
4127 // Suppress the vtable call
4128 } else {
4129 // hashCode and clone are not a miranda methods,
4130 // so the vtable index is fixed.
4131 // No need to use the linkResolver to get it.
4132 vtable_index = method->vtable_index();
4133 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4134 "bad index %d", vtable_index);
4135 }
4136 slow_call = new CallDynamicJavaNode(tf,
4137 SharedRuntime::get_resolve_virtual_call_stub(),
4138 method, vtable_index, bci());
4139 } else { // neither virtual nor static: opt_virtual
4140 null_check_receiver();
4141 slow_call = new CallStaticJavaNode(C, tf,
4142 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4143 method, bci());
4144 slow_call->set_optimized_virtual(true);
4145 }
4146 set_arguments_for_java_call(slow_call);
4147 set_edges_for_java_call(slow_call);
4148 return slow_call;
4149 }
4150
4151
4152 /**
4153 * Build special case code for calls to hashCode on an object. This call may
4154 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4155 * slightly different code.
4156 */
4157 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4158 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4159 assert(!(is_virtual && is_static), "either virtual, special, or static");
4160
4161 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4162
4163 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4164 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4165 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4166 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4167 Node* obj = NULL;
4168 if (!is_static) {
4169 // Check for hashing null object
4170 obj = null_check_receiver();
4171 if (stopped()) return true; // unconditionally null
4172 result_reg->init_req(_null_path, top());
4173 result_val->init_req(_null_path, top());
4174 } else {
4175 // Do a null check, and return zero if null.
4176 // System.identityHashCode(null) == 0
4177 obj = argument(0);
4178 Node* null_ctl = top();
4179 obj = null_check_oop(obj, &null_ctl);
4180 result_reg->init_req(_null_path, null_ctl);
4181 result_val->init_req(_null_path, _gvn.intcon(0));
4182 }
4183
4184 // Unconditionally null? Then return right away.
4185 if (stopped()) {
4186 set_control( result_reg->in(_null_path));
4187 if (!stopped())
4188 set_result(result_val->in(_null_path));
4189 return true;
4190 }
4191
4192 // We only go to the fast case code if we pass a number of guards. The
4193 // paths which do not pass are accumulated in the slow_region.
4194 RegionNode* slow_region = new RegionNode(1);
4195 record_for_igvn(slow_region);
4196
4197 // If this is a virtual call, we generate a funny guard. We pull out
4198 // the vtable entry corresponding to hashCode() from the target object.
4199 // If the target method which we are calling happens to be the native
4200 // Object hashCode() method, we pass the guard. We do not need this
4201 // guard for non-virtual calls -- the caller is known to be the native
4202 // Object hashCode().
4203 if (is_virtual) {
4204 // After null check, get the object's klass.
4205 Node* obj_klass = load_object_klass(obj);
4206 generate_virtual_guard(obj_klass, slow_region);
4207 }
4208
4209 // Get the header out of the object, use LoadMarkNode when available
4210 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4211 // The control of the load must be NULL. Otherwise, the load can move before
4212 // the null check after castPP removal.
4213 Node* no_ctrl = NULL;
4214 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4215
4216 // Test the header to see if it is unlocked.
4217 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4218 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4219 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4220 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4221 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4222
4223 generate_slow_guard(test_unlocked, slow_region);
4224
4225 // Get the hash value and check to see that it has been properly assigned.
4226 // We depend on hash_mask being at most 32 bits and avoid the use of
4227 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4228 // vm: see markOop.hpp.
4229 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4230 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4231 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4232 // This hack lets the hash bits live anywhere in the mark object now, as long
4233 // as the shift drops the relevant bits into the low 32 bits. Note that
4234 // Java spec says that HashCode is an int so there's no point in capturing
4235 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4236 hshifted_header = ConvX2I(hshifted_header);
4237 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4238
4239 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4240 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4241 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4242
4243 generate_slow_guard(test_assigned, slow_region);
4244
4245 Node* init_mem = reset_memory();
4246 // fill in the rest of the null path:
4247 result_io ->init_req(_null_path, i_o());
4248 result_mem->init_req(_null_path, init_mem);
4249
4250 result_val->init_req(_fast_path, hash_val);
4251 result_reg->init_req(_fast_path, control());
4252 result_io ->init_req(_fast_path, i_o());
4253 result_mem->init_req(_fast_path, init_mem);
4254
4255 // Generate code for the slow case. We make a call to hashCode().
4256 set_control(_gvn.transform(slow_region));
4257 if (!stopped()) {
4258 // No need for PreserveJVMState, because we're using up the present state.
4259 set_all_memory(init_mem);
4260 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4261 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4262 Node* slow_result = set_results_for_java_call(slow_call);
4263 // this->control() comes from set_results_for_java_call
4264 result_reg->init_req(_slow_path, control());
4265 result_val->init_req(_slow_path, slow_result);
4266 result_io ->set_req(_slow_path, i_o());
4267 result_mem ->set_req(_slow_path, reset_memory());
4268 }
4269
4270 // Return the combined state.
4271 set_i_o( _gvn.transform(result_io) );
4272 set_all_memory( _gvn.transform(result_mem));
4273
4274 set_result(result_reg, result_val);
4275 return true;
4276 }
4277
4278 //---------------------------inline_native_getClass----------------------------
4279 // public final native Class<?> java.lang.Object.getClass();
4280 //
4281 // Build special case code for calls to getClass on an object.
4282 bool LibraryCallKit::inline_native_getClass() {
4283 Node* obj = null_check_receiver();
4284 if (stopped()) return true;
4285 set_result(load_mirror_from_klass(load_object_klass(obj)));
4286 return true;
4287 }
4288
4289 //-----------------inline_native_Reflection_getCallerClass---------------------
4290 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4291 //
4292 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4293 //
4294 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4295 // in that it must skip particular security frames and checks for
4296 // caller sensitive methods.
4297 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4298 #ifndef PRODUCT
4299 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4300 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4301 }
4302 #endif
4303
4304 if (!jvms()->has_method()) {
4305 #ifndef PRODUCT
4306 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4307 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4308 }
4309 #endif
4310 return false;
4311 }
4312
4313 // Walk back up the JVM state to find the caller at the required
4314 // depth.
4315 JVMState* caller_jvms = jvms();
4316
4317 // Cf. JVM_GetCallerClass
4318 // NOTE: Start the loop at depth 1 because the current JVM state does
4319 // not include the Reflection.getCallerClass() frame.
4320 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4321 ciMethod* m = caller_jvms->method();
4322 switch (n) {
4323 case 0:
4324 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4325 break;
4326 case 1:
4327 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4328 if (!m->caller_sensitive()) {
4329 #ifndef PRODUCT
4330 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4331 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4332 }
4333 #endif
4334 return false; // bail-out; let JVM_GetCallerClass do the work
4335 }
4336 break;
4337 default:
4338 if (!m->is_ignored_by_security_stack_walk()) {
4339 // We have reached the desired frame; return the holder class.
4340 // Acquire method holder as java.lang.Class and push as constant.
4341 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4342 ciInstance* caller_mirror = caller_klass->java_mirror();
4343 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4344
4345 #ifndef PRODUCT
4346 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4347 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());
4348 tty->print_cr(" JVM state at this point:");
4349 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4350 ciMethod* m = jvms()->of_depth(i)->method();
4351 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4352 }
4353 }
4354 #endif
4355 return true;
4356 }
4357 break;
4358 }
4359 }
4360
4361 #ifndef PRODUCT
4362 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4363 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4364 tty->print_cr(" JVM state at this point:");
4365 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4366 ciMethod* m = jvms()->of_depth(i)->method();
4367 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4368 }
4369 }
4370 #endif
4371
4372 return false; // bail-out; let JVM_GetCallerClass do the work
4373 }
4374
4375 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4376 Node* arg = argument(0);
4377 Node* result = NULL;
4378
4379 switch (id) {
4380 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4381 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4382 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4383 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4384
4385 case vmIntrinsics::_doubleToLongBits: {
4386 // two paths (plus control) merge in a wood
4387 RegionNode *r = new RegionNode(3);
4388 Node *phi = new PhiNode(r, TypeLong::LONG);
4389
4390 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4391 // Build the boolean node
4392 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4393
4394 // Branch either way.
4395 // NaN case is less traveled, which makes all the difference.
4396 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4397 Node *opt_isnan = _gvn.transform(ifisnan);
4398 assert( opt_isnan->is_If(), "Expect an IfNode");
4399 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4400 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4401
4402 set_control(iftrue);
4403
4404 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4405 Node *slow_result = longcon(nan_bits); // return NaN
4406 phi->init_req(1, _gvn.transform( slow_result ));
4407 r->init_req(1, iftrue);
4408
4409 // Else fall through
4410 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4411 set_control(iffalse);
4412
4413 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4414 r->init_req(2, iffalse);
4415
4416 // Post merge
4417 set_control(_gvn.transform(r));
4418 record_for_igvn(r);
4419
4420 C->set_has_split_ifs(true); // Has chance for split-if optimization
4421 result = phi;
4422 assert(result->bottom_type()->isa_long(), "must be");
4423 break;
4424 }
4425
4426 case vmIntrinsics::_floatToIntBits: {
4427 // two paths (plus control) merge in a wood
4428 RegionNode *r = new RegionNode(3);
4429 Node *phi = new PhiNode(r, TypeInt::INT);
4430
4431 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4432 // Build the boolean node
4433 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4434
4435 // Branch either way.
4436 // NaN case is less traveled, which makes all the difference.
4437 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4438 Node *opt_isnan = _gvn.transform(ifisnan);
4439 assert( opt_isnan->is_If(), "Expect an IfNode");
4440 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4441 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4442
4443 set_control(iftrue);
4444
4445 static const jint nan_bits = 0x7fc00000;
4446 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4447 phi->init_req(1, _gvn.transform( slow_result ));
4448 r->init_req(1, iftrue);
4449
4450 // Else fall through
4451 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4452 set_control(iffalse);
4453
4454 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4455 r->init_req(2, iffalse);
4456
4457 // Post merge
4458 set_control(_gvn.transform(r));
4459 record_for_igvn(r);
4460
4461 C->set_has_split_ifs(true); // Has chance for split-if optimization
4462 result = phi;
4463 assert(result->bottom_type()->isa_int(), "must be");
4464 break;
4465 }
4466
4467 default:
4468 fatal_unexpected_iid(id);
4469 break;
4470 }
4471 set_result(_gvn.transform(result));
4472 return true;
4473 }
4474
4475 //----------------------inline_unsafe_copyMemory-------------------------
4476 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4477 bool LibraryCallKit::inline_unsafe_copyMemory() {
4478 if (callee()->is_static()) return false; // caller must have the capability!
4479 null_check_receiver(); // null-check receiver
4480 if (stopped()) return true;
4481
4482 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4483
4484 Node* src_ptr = argument(1); // type: oop
4485 Node* src_off = ConvL2X(argument(2)); // type: long
4486 Node* dst_ptr = argument(4); // type: oop
4487 Node* dst_off = ConvL2X(argument(5)); // type: long
4488 Node* size = ConvL2X(argument(7)); // type: long
4489
4490 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4491 "fieldOffset must be byte-scaled");
4492
4493 Node* src = make_unsafe_address(src_ptr, src_off);
4494 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4495
4496 // Conservatively insert a memory barrier on all memory slices.
4497 // Do not let writes of the copy source or destination float below the copy.
4498 insert_mem_bar(Op_MemBarCPUOrder);
4499
4500 // Call it. Note that the length argument is not scaled.
4501 make_runtime_call(RC_LEAF|RC_NO_FP,
4502 OptoRuntime::fast_arraycopy_Type(),
4503 StubRoutines::unsafe_arraycopy(),
4504 "unsafe_arraycopy",
4505 TypeRawPtr::BOTTOM,
4506 src, dst, size XTOP);
4507
4508 // Do not let reads of the copy destination float above the copy.
4509 insert_mem_bar(Op_MemBarCPUOrder);
4510
4511 return true;
4512 }
4513
4514 //------------------------clone_coping-----------------------------------
4515 // Helper function for inline_native_clone.
4516 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4517 assert(obj_size != NULL, "");
4518 Node* raw_obj = alloc_obj->in(1);
4519 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4520
4521 AllocateNode* alloc = NULL;
4522 if (ReduceBulkZeroing) {
4523 // We will be completely responsible for initializing this object -
4524 // mark Initialize node as complete.
4525 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4526 // The object was just allocated - there should be no any stores!
4527 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4528 // Mark as complete_with_arraycopy so that on AllocateNode
4529 // expansion, we know this AllocateNode is initialized by an array
4530 // copy and a StoreStore barrier exists after the array copy.
4531 alloc->initialization()->set_complete_with_arraycopy();
4532 }
4533
4534 // Copy the fastest available way.
4535 // TODO: generate fields copies for small objects instead.
4536 Node* src = obj;
4537 Node* dest = alloc_obj;
4538 Node* size = _gvn.transform(obj_size);
4539
4540 // Exclude the header but include array length to copy by 8 bytes words.
4541 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4542 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4543 instanceOopDesc::base_offset_in_bytes();
4544 // base_off:
4545 // 8 - 32-bit VM
4546 // 12 - 64-bit VM, compressed klass
4547 // 16 - 64-bit VM, normal klass
4548 if (base_off % BytesPerLong != 0) {
4549 assert(UseCompressedClassPointers, "");
4550 if (is_array) {
4551 // Exclude length to copy by 8 bytes words.
4552 base_off += sizeof(int);
4553 } else {
4554 // Include klass to copy by 8 bytes words.
4555 base_off = instanceOopDesc::klass_offset_in_bytes();
4556 }
4557 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4558 }
4559 src = basic_plus_adr(src, base_off);
4560 dest = basic_plus_adr(dest, base_off);
4561
4562 // Compute the length also, if needed:
4563 Node* countx = size;
4564 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4565 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4566
4567 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4568
4569 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4570 ac->set_clonebasic();
4571 Node* n = _gvn.transform(ac);
4572 if (n == ac) {
4573 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4574 } else {
4575 set_all_memory(n);
4576 }
4577
4578 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4579 if (card_mark) {
4580 assert(!is_array, "");
4581 // Put in store barrier for any and all oops we are sticking
4582 // into this object. (We could avoid this if we could prove
4583 // that the object type contains no oop fields at all.)
4584 Node* no_particular_value = NULL;
4585 Node* no_particular_field = NULL;
4586 int raw_adr_idx = Compile::AliasIdxRaw;
4587 post_barrier(control(),
4588 memory(raw_adr_type),
4589 alloc_obj,
4590 no_particular_field,
4591 raw_adr_idx,
4592 no_particular_value,
4593 T_OBJECT,
4594 false);
4595 }
4596
4597 // Do not let reads from the cloned object float above the arraycopy.
4598 if (alloc != NULL) {
4599 // Do not let stores that initialize this object be reordered with
4600 // a subsequent store that would make this object accessible by
4601 // other threads.
4602 // Record what AllocateNode this StoreStore protects so that
4603 // escape analysis can go from the MemBarStoreStoreNode to the
4604 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4605 // based on the escape status of the AllocateNode.
4606 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4607 } else {
4608 insert_mem_bar(Op_MemBarCPUOrder);
4609 }
4610 }
4611
4612 //------------------------inline_native_clone----------------------------
4613 // protected native Object java.lang.Object.clone();
4614 //
4615 // Here are the simple edge cases:
4616 // null receiver => normal trap
4617 // virtual and clone was overridden => slow path to out-of-line clone
4618 // not cloneable or finalizer => slow path to out-of-line Object.clone
4619 //
4620 // The general case has two steps, allocation and copying.
4621 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4622 //
4623 // Copying also has two cases, oop arrays and everything else.
4624 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4625 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4626 //
4627 // These steps fold up nicely if and when the cloned object's klass
4628 // can be sharply typed as an object array, a type array, or an instance.
4629 //
4630 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4631 PhiNode* result_val;
4632
4633 // Set the reexecute bit for the interpreter to reexecute
4634 // the bytecode that invokes Object.clone if deoptimization happens.
4635 { PreserveReexecuteState preexecs(this);
4636 jvms()->set_should_reexecute(true);
4637
4638 Node* obj = null_check_receiver();
4639 if (stopped()) return true;
4640
4641 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4642
4643 // If we are going to clone an instance, we need its exact type to
4644 // know the number and types of fields to convert the clone to
4645 // loads/stores. Maybe a speculative type can help us.
4646 if (!obj_type->klass_is_exact() &&
4647 obj_type->speculative_type() != NULL &&
4648 obj_type->speculative_type()->is_instance_klass()) {
4649 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4650 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4651 !spec_ik->has_injected_fields()) {
4652 ciKlass* k = obj_type->klass();
4653 if (!k->is_instance_klass() ||
4654 k->as_instance_klass()->is_interface() ||
4655 k->as_instance_klass()->has_subklass()) {
4656 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4657 }
4658 }
4659 }
4660
4661 Node* obj_klass = load_object_klass(obj);
4662 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4663 const TypeOopPtr* toop = ((tklass != NULL)
4664 ? tklass->as_instance_type()
4665 : TypeInstPtr::NOTNULL);
4666
4667 // Conservatively insert a memory barrier on all memory slices.
4668 // Do not let writes into the original float below the clone.
4669 insert_mem_bar(Op_MemBarCPUOrder);
4670
4671 // paths into result_reg:
4672 enum {
4673 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4674 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4675 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4676 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4677 PATH_LIMIT
4678 };
4679 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4680 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4681 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4682 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4683 record_for_igvn(result_reg);
4684
4685 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4686 int raw_adr_idx = Compile::AliasIdxRaw;
4687
4688 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4689 if (array_ctl != NULL) {
4690 // It's an array.
4691 PreserveJVMState pjvms(this);
4692 set_control(array_ctl);
4693 Node* obj_length = load_array_length(obj);
4694 Node* obj_size = NULL;
4695 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4696
4697 if (!use_ReduceInitialCardMarks()) {
4698 // If it is an oop array, it requires very special treatment,
4699 // because card marking is required on each card of the array.
4700 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4701 if (is_obja != NULL) {
4702 PreserveJVMState pjvms2(this);
4703 set_control(is_obja);
4704 // Generate a direct call to the right arraycopy function(s).
4705 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4706 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4707 ac->set_cloneoop();
4708 Node* n = _gvn.transform(ac);
4709 assert(n == ac, "cannot disappear");
4710 ac->connect_outputs(this);
4711
4712 result_reg->init_req(_objArray_path, control());
4713 result_val->init_req(_objArray_path, alloc_obj);
4714 result_i_o ->set_req(_objArray_path, i_o());
4715 result_mem ->set_req(_objArray_path, reset_memory());
4716 }
4717 }
4718 // Otherwise, there are no card marks to worry about.
4719 // (We can dispense with card marks if we know the allocation
4720 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4721 // causes the non-eden paths to take compensating steps to
4722 // simulate a fresh allocation, so that no further
4723 // card marks are required in compiled code to initialize
4724 // the object.)
4725
4726 if (!stopped()) {
4727 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4728
4729 // Present the results of the copy.
4730 result_reg->init_req(_array_path, control());
4731 result_val->init_req(_array_path, alloc_obj);
4732 result_i_o ->set_req(_array_path, i_o());
4733 result_mem ->set_req(_array_path, reset_memory());
4734 }
4735 }
4736
4737 // We only go to the instance fast case code if we pass a number of guards.
4738 // The paths which do not pass are accumulated in the slow_region.
4739 RegionNode* slow_region = new RegionNode(1);
4740 record_for_igvn(slow_region);
4741 if (!stopped()) {
4742 // It's an instance (we did array above). Make the slow-path tests.
4743 // If this is a virtual call, we generate a funny guard. We grab
4744 // the vtable entry corresponding to clone() from the target object.
4745 // If the target method which we are calling happens to be the
4746 // Object clone() method, we pass the guard. We do not need this
4747 // guard for non-virtual calls; the caller is known to be the native
4748 // Object clone().
4749 if (is_virtual) {
4750 generate_virtual_guard(obj_klass, slow_region);
4751 }
4752
4753 // The object must be easily cloneable and must not have a finalizer.
4754 // Both of these conditions may be checked in a single test.
4755 // We could optimize the test further, but we don't care.
4756 generate_access_flags_guard(obj_klass,
4757 // Test both conditions:
4758 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4759 // Must be cloneable but not finalizer:
4760 JVM_ACC_IS_CLONEABLE_FAST,
4761 slow_region);
4762 }
4763
4764 if (!stopped()) {
4765 // It's an instance, and it passed the slow-path tests.
4766 PreserveJVMState pjvms(this);
4767 Node* obj_size = NULL;
4768 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4769 // is reexecuted if deoptimization occurs and there could be problems when merging
4770 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4771 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4772
4773 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4774
4775 // Present the results of the slow call.
4776 result_reg->init_req(_instance_path, control());
4777 result_val->init_req(_instance_path, alloc_obj);
4778 result_i_o ->set_req(_instance_path, i_o());
4779 result_mem ->set_req(_instance_path, reset_memory());
4780 }
4781
4782 // Generate code for the slow case. We make a call to clone().
4783 set_control(_gvn.transform(slow_region));
4784 if (!stopped()) {
4785 PreserveJVMState pjvms(this);
4786 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4787 Node* slow_result = set_results_for_java_call(slow_call);
4788 // this->control() comes from set_results_for_java_call
4789 result_reg->init_req(_slow_path, control());
4790 result_val->init_req(_slow_path, slow_result);
4791 result_i_o ->set_req(_slow_path, i_o());
4792 result_mem ->set_req(_slow_path, reset_memory());
4793 }
4794
4795 // Return the combined state.
4796 set_control( _gvn.transform(result_reg));
4797 set_i_o( _gvn.transform(result_i_o));
4798 set_all_memory( _gvn.transform(result_mem));
4799 } // original reexecute is set back here
4800
4801 set_result(_gvn.transform(result_val));
4802 return true;
4803 }
4804
4805 // If we have a tighly coupled allocation, the arraycopy may take care
4806 // of the array initialization. If one of the guards we insert between
4807 // the allocation and the arraycopy causes a deoptimization, an
4808 // unitialized array will escape the compiled method. To prevent that
4809 // we set the JVM state for uncommon traps between the allocation and
4810 // the arraycopy to the state before the allocation so, in case of
4811 // deoptimization, we'll reexecute the allocation and the
4812 // initialization.
4813 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4814 if (alloc != NULL) {
4815 ciMethod* trap_method = alloc->jvms()->method();
4816 int trap_bci = alloc->jvms()->bci();
4817
4818 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4819 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4820 // Make sure there's no store between the allocation and the
4821 // arraycopy otherwise visible side effects could be rexecuted
4822 // in case of deoptimization and cause incorrect execution.
4823 bool no_interfering_store = true;
4824 Node* mem = alloc->in(TypeFunc::Memory);
4825 if (mem->is_MergeMem()) {
4826 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4827 Node* n = mms.memory();
4828 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4829 assert(n->is_Store(), "what else?");
4830 no_interfering_store = false;
4831 break;
4832 }
4833 }
4834 } else {
4835 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4836 Node* n = mms.memory();
4837 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4838 assert(n->is_Store(), "what else?");
4839 no_interfering_store = false;
4840 break;
4841 }
4842 }
4843 }
4844
4845 if (no_interfering_store) {
4846 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4847 uint size = alloc->req();
4848 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4849 old_jvms->set_map(sfpt);
4850 for (uint i = 0; i < size; i++) {
4851 sfpt->init_req(i, alloc->in(i));
4852 }
4853 // re-push array length for deoptimization
4854 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4855 old_jvms->set_sp(old_jvms->sp()+1);
4856 old_jvms->set_monoff(old_jvms->monoff()+1);
4857 old_jvms->set_scloff(old_jvms->scloff()+1);
4858 old_jvms->set_endoff(old_jvms->endoff()+1);
4859 old_jvms->set_should_reexecute(true);
4860
4861 sfpt->set_i_o(map()->i_o());
4862 sfpt->set_memory(map()->memory());
4863 sfpt->set_control(map()->control());
4864
4865 JVMState* saved_jvms = jvms();
4866 saved_reexecute_sp = _reexecute_sp;
4867
4868 set_jvms(sfpt->jvms());
4869 _reexecute_sp = jvms()->sp();
4870
4871 return saved_jvms;
4872 }
4873 }
4874 }
4875 return NULL;
4876 }
4877
4878 // In case of a deoptimization, we restart execution at the
4879 // allocation, allocating a new array. We would leave an uninitialized
4880 // array in the heap that GCs wouldn't expect. Move the allocation
4881 // after the traps so we don't allocate the array if we
4882 // deoptimize. This is possible because tightly_coupled_allocation()
4883 // guarantees there's no observer of the allocated array at this point
4884 // and the control flow is simple enough.
4885 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4886 if (saved_jvms != NULL && !stopped()) {
4887 assert(alloc != NULL, "only with a tightly coupled allocation");
4888 // restore JVM state to the state at the arraycopy
4889 saved_jvms->map()->set_control(map()->control());
4890 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4891 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4892 // If we've improved the types of some nodes (null check) while
4893 // emitting the guards, propagate them to the current state
4894 map()->replaced_nodes().apply(saved_jvms->map());
4895 set_jvms(saved_jvms);
4896 _reexecute_sp = saved_reexecute_sp;
4897
4898 // Remove the allocation from above the guards
4899 CallProjections callprojs;
4900 alloc->extract_projections(&callprojs, true);
4901 InitializeNode* init = alloc->initialization();
4902 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4903 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4904 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4905 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4906
4907 // move the allocation here (after the guards)
4908 _gvn.hash_delete(alloc);
4909 alloc->set_req(TypeFunc::Control, control());
4910 alloc->set_req(TypeFunc::I_O, i_o());
4911 Node *mem = reset_memory();
4912 set_all_memory(mem);
4913 alloc->set_req(TypeFunc::Memory, mem);
4914 set_control(init->proj_out(TypeFunc::Control));
4915 set_i_o(callprojs.fallthrough_ioproj);
4916
4917 // Update memory as done in GraphKit::set_output_for_allocation()
4918 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4919 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4920 if (ary_type->isa_aryptr() && length_type != NULL) {
4921 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4922 }
4923 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4924 int elemidx = C->get_alias_index(telemref);
4925 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4926 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4927
4928 Node* allocx = _gvn.transform(alloc);
4929 assert(allocx == alloc, "where has the allocation gone?");
4930 assert(dest->is_CheckCastPP(), "not an allocation result?");
4931
4932 _gvn.hash_delete(dest);
4933 dest->set_req(0, control());
4934 Node* destx = _gvn.transform(dest);
4935 assert(destx == dest, "where has the allocation result gone?");
4936 }
4937 }
4938
4939
4940 //------------------------------inline_arraycopy-----------------------
4941 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4942 // Object dest, int destPos,
4943 // int length);
4944 bool LibraryCallKit::inline_arraycopy() {
4945 // Get the arguments.
4946 Node* src = argument(0); // type: oop
4947 Node* src_offset = argument(1); // type: int
4948 Node* dest = argument(2); // type: oop
4949 Node* dest_offset = argument(3); // type: int
4950 Node* length = argument(4); // type: int
4951
4952
4953 // Check for allocation before we add nodes that would confuse
4954 // tightly_coupled_allocation()
4955 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4956
4957 int saved_reexecute_sp = -1;
4958 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4959 // See arraycopy_restore_alloc_state() comment
4960 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4961 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4962 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4963 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4964
4965 // The following tests must be performed
4966 // (1) src and dest are arrays.
4967 // (2) src and dest arrays must have elements of the same BasicType
4968 // (3) src and dest must not be null.
4969 // (4) src_offset must not be negative.
4970 // (5) dest_offset must not be negative.
4971 // (6) length must not be negative.
4972 // (7) src_offset + length must not exceed length of src.
4973 // (8) dest_offset + length must not exceed length of dest.
4974 // (9) each element of an oop array must be assignable
4975
4976 // (3) src and dest must not be null.
4977 // always do this here because we need the JVM state for uncommon traps
4978 Node* null_ctl = top();
4979 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4980 assert(null_ctl->is_top(), "no null control here");
4981 dest = null_check(dest, T_ARRAY);
4982
4983 if (!can_emit_guards) {
4984 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4985 // guards but the arraycopy node could still take advantage of a
4986 // tightly allocated allocation. tightly_coupled_allocation() is
4987 // called again to make sure it takes the null check above into
4988 // account: the null check is mandatory and if it caused an
4989 // uncommon trap to be emitted then the allocation can't be
4990 // considered tightly coupled in this context.
4991 alloc = tightly_coupled_allocation(dest, NULL);
4992 }
4993
4994 bool validated = false;
4995
4996 const Type* src_type = _gvn.type(src);
4997 const Type* dest_type = _gvn.type(dest);
4998 const TypeAryPtr* top_src = src_type->isa_aryptr();
4999 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5000
5001 // Do we have the type of src?
5002 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5003 // Do we have the type of dest?
5004 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5005 // Is the type for src from speculation?
5006 bool src_spec = false;
5007 // Is the type for dest from speculation?
5008 bool dest_spec = false;
5009
5010 if ((!has_src || !has_dest) && can_emit_guards) {
5011 // We don't have sufficient type information, let's see if
5012 // speculative types can help. We need to have types for both src
5013 // and dest so that it pays off.
5014
5015 // Do we already have or could we have type information for src
5016 bool could_have_src = has_src;
5017 // Do we already have or could we have type information for dest
5018 bool could_have_dest = has_dest;
5019
5020 ciKlass* src_k = NULL;
5021 if (!has_src) {
5022 src_k = src_type->speculative_type_not_null();
5023 if (src_k != NULL && src_k->is_array_klass()) {
5024 could_have_src = true;
5025 }
5026 }
5027
5028 ciKlass* dest_k = NULL;
5029 if (!has_dest) {
5030 dest_k = dest_type->speculative_type_not_null();
5031 if (dest_k != NULL && dest_k->is_array_klass()) {
5032 could_have_dest = true;
5033 }
5034 }
5035
5036 if (could_have_src && could_have_dest) {
5037 // This is going to pay off so emit the required guards
5038 if (!has_src) {
5039 src = maybe_cast_profiled_obj(src, src_k, true);
5040 src_type = _gvn.type(src);
5041 top_src = src_type->isa_aryptr();
5042 has_src = (top_src != NULL && top_src->klass() != NULL);
5043 src_spec = true;
5044 }
5045 if (!has_dest) {
5046 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5047 dest_type = _gvn.type(dest);
5048 top_dest = dest_type->isa_aryptr();
5049 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5050 dest_spec = true;
5051 }
5052 }
5053 }
5054
5055 if (has_src && has_dest && can_emit_guards) {
5056 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5057 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5058 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5059 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5060
5061 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5062 // If both arrays are object arrays then having the exact types
5063 // for both will remove the need for a subtype check at runtime
5064 // before the call and may make it possible to pick a faster copy
5065 // routine (without a subtype check on every element)
5066 // Do we have the exact type of src?
5067 bool could_have_src = src_spec;
5068 // Do we have the exact type of dest?
5069 bool could_have_dest = dest_spec;
5070 ciKlass* src_k = top_src->klass();
5071 ciKlass* dest_k = top_dest->klass();
5072 if (!src_spec) {
5073 src_k = src_type->speculative_type_not_null();
5074 if (src_k != NULL && src_k->is_array_klass()) {
5075 could_have_src = true;
5076 }
5077 }
5078 if (!dest_spec) {
5079 dest_k = dest_type->speculative_type_not_null();
5080 if (dest_k != NULL && dest_k->is_array_klass()) {
5081 could_have_dest = true;
5082 }
5083 }
5084 if (could_have_src && could_have_dest) {
5085 // If we can have both exact types, emit the missing guards
5086 if (could_have_src && !src_spec) {
5087 src = maybe_cast_profiled_obj(src, src_k, true);
5088 }
5089 if (could_have_dest && !dest_spec) {
5090 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5091 }
5092 }
5093 }
5094 }
5095
5096 ciMethod* trap_method = method();
5097 int trap_bci = bci();
5098 if (saved_jvms != NULL) {
5099 trap_method = alloc->jvms()->method();
5100 trap_bci = alloc->jvms()->bci();
5101 }
5102
5103 bool negative_length_guard_generated = false;
5104
5105 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5106 can_emit_guards &&
5107 !src->is_top() && !dest->is_top()) {
5108 // validate arguments: enables transformation the ArrayCopyNode
5109 validated = true;
5110
5111 RegionNode* slow_region = new RegionNode(1);
5112 record_for_igvn(slow_region);
5113
5114 // (1) src and dest are arrays.
5115 generate_non_array_guard(load_object_klass(src), slow_region);
5116 generate_non_array_guard(load_object_klass(dest), slow_region);
5117
5118 // (2) src and dest arrays must have elements of the same BasicType
5119 // done at macro expansion or at Ideal transformation time
5120
5121 // (4) src_offset must not be negative.
5122 generate_negative_guard(src_offset, slow_region);
5123
5124 // (5) dest_offset must not be negative.
5125 generate_negative_guard(dest_offset, slow_region);
5126
5127 // (7) src_offset + length must not exceed length of src.
5128 generate_limit_guard(src_offset, length,
5129 load_array_length(src),
5130 slow_region);
5131
5132 // (8) dest_offset + length must not exceed length of dest.
5133 generate_limit_guard(dest_offset, length,
5134 load_array_length(dest),
5135 slow_region);
5136
5137 // (6) length must not be negative.
5138 // This is also checked in generate_arraycopy() during macro expansion, but
5139 // we also have to check it here for the case where the ArrayCopyNode will
5140 // be eliminated by Escape Analysis.
5141 if (EliminateAllocations) {
5142 generate_negative_guard(length, slow_region);
5143 negative_length_guard_generated = true;
5144 }
5145
5146 // (9) each element of an oop array must be assignable
5147 Node* src_klass = load_object_klass(src);
5148 Node* dest_klass = load_object_klass(dest);
5149 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5150
5151 if (not_subtype_ctrl != top()) {
5152 PreserveJVMState pjvms(this);
5153 set_control(not_subtype_ctrl);
5154 uncommon_trap(Deoptimization::Reason_intrinsic,
5155 Deoptimization::Action_make_not_entrant);
5156 assert(stopped(), "Should be stopped");
5157 }
5158 {
5159 PreserveJVMState pjvms(this);
5160 set_control(_gvn.transform(slow_region));
5161 uncommon_trap(Deoptimization::Reason_intrinsic,
5162 Deoptimization::Action_make_not_entrant);
5163 assert(stopped(), "Should be stopped");
5164 }
5165 }
5166
5167 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5168
5169 if (stopped()) {
5170 return true;
5171 }
5172
5173 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5174 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5175 // so the compiler has a chance to eliminate them: during macro expansion,
5176 // we have to set their control (CastPP nodes are eliminated).
5177 load_object_klass(src), load_object_klass(dest),
5178 load_array_length(src), load_array_length(dest));
5179
5180 ac->set_arraycopy(validated);
5181
5182 Node* n = _gvn.transform(ac);
5183 if (n == ac) {
5184 ac->connect_outputs(this);
5185 } else {
5186 assert(validated, "shouldn't transform if all arguments not validated");
5187 set_all_memory(n);
5188 }
5189
5190 return true;
5191 }
5192
5193
5194 // Helper function which determines if an arraycopy immediately follows
5195 // an allocation, with no intervening tests or other escapes for the object.
5196 AllocateArrayNode*
5197 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5198 RegionNode* slow_region) {
5199 if (stopped()) return NULL; // no fast path
5200 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5201
5202 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5203 if (alloc == NULL) return NULL;
5204
5205 Node* rawmem = memory(Compile::AliasIdxRaw);
5206 // Is the allocation's memory state untouched?
5207 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5208 // Bail out if there have been raw-memory effects since the allocation.
5209 // (Example: There might have been a call or safepoint.)
5210 return NULL;
5211 }
5212 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5213 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5214 return NULL;
5215 }
5216
5217 // There must be no unexpected observers of this allocation.
5218 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5219 Node* obs = ptr->fast_out(i);
5220 if (obs != this->map()) {
5221 return NULL;
5222 }
5223 }
5224
5225 // This arraycopy must unconditionally follow the allocation of the ptr.
5226 Node* alloc_ctl = ptr->in(0);
5227 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5228
5229 Node* ctl = control();
5230 while (ctl != alloc_ctl) {
5231 // There may be guards which feed into the slow_region.
5232 // Any other control flow means that we might not get a chance
5233 // to finish initializing the allocated object.
5234 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5235 IfNode* iff = ctl->in(0)->as_If();
5236 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5237 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5238 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5239 ctl = iff->in(0); // This test feeds the known slow_region.
5240 continue;
5241 }
5242 // One more try: Various low-level checks bottom out in
5243 // uncommon traps. If the debug-info of the trap omits
5244 // any reference to the allocation, as we've already
5245 // observed, then there can be no objection to the trap.
5246 bool found_trap = false;
5247 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5248 Node* obs = not_ctl->fast_out(j);
5249 if (obs->in(0) == not_ctl && obs->is_Call() &&
5250 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5251 found_trap = true; break;
5252 }
5253 }
5254 if (found_trap) {
5255 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5256 continue;
5257 }
5258 }
5259 return NULL;
5260 }
5261
5262 // If we get this far, we have an allocation which immediately
5263 // precedes the arraycopy, and we can take over zeroing the new object.
5264 // The arraycopy will finish the initialization, and provide
5265 // a new control state to which we will anchor the destination pointer.
5266
5267 return alloc;
5268 }
5269
5270 //-------------inline_encodeISOArray-----------------------------------
5271 // encode char[] to byte[] in ISO_8859_1
5272 bool LibraryCallKit::inline_encodeISOArray() {
5273 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5274 // no receiver since it is static method
5275 Node *src = argument(0);
5276 Node *src_offset = argument(1);
5277 Node *dst = argument(2);
5278 Node *dst_offset = argument(3);
5279 Node *length = argument(4);
5280
5281 const Type* src_type = src->Value(&_gvn);
5282 const Type* dst_type = dst->Value(&_gvn);
5283 const TypeAryPtr* top_src = src_type->isa_aryptr();
5284 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5285 if (top_src == NULL || top_src->klass() == NULL ||
5286 top_dest == NULL || top_dest->klass() == NULL) {
5287 // failed array check
5288 return false;
5289 }
5290
5291 // Figure out the size and type of the elements we will be copying.
5292 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5293 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5294 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5295 return false;
5296 }
5297
5298 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5299 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5300 // 'src_start' points to src array + scaled offset
5301 // 'dst_start' points to dst array + scaled offset
5302
5303 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5304 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5305 enc = _gvn.transform(enc);
5306 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5307 set_memory(res_mem, mtype);
5308 set_result(enc);
5309 return true;
5310 }
5311
5312 //-------------inline_multiplyToLen-----------------------------------
5313 bool LibraryCallKit::inline_multiplyToLen() {
5314 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5315
5316 address stubAddr = StubRoutines::multiplyToLen();
5317 if (stubAddr == NULL) {
5318 return false; // Intrinsic's stub is not implemented on this platform
5319 }
5320 const char* stubName = "multiplyToLen";
5321
5322 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5323
5324 // no receiver because it is a static method
5325 Node* x = argument(0);
5326 Node* xlen = argument(1);
5327 Node* y = argument(2);
5328 Node* ylen = argument(3);
5329 Node* z = argument(4);
5330
5331 const Type* x_type = x->Value(&_gvn);
5332 const Type* y_type = y->Value(&_gvn);
5333 const TypeAryPtr* top_x = x_type->isa_aryptr();
5334 const TypeAryPtr* top_y = y_type->isa_aryptr();
5335 if (top_x == NULL || top_x->klass() == NULL ||
5336 top_y == NULL || top_y->klass() == NULL) {
5337 // failed array check
5338 return false;
5339 }
5340
5341 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5342 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5343 if (x_elem != T_INT || y_elem != T_INT) {
5344 return false;
5345 }
5346
5347 // Set the original stack and the reexecute bit for the interpreter to reexecute
5348 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5349 // on the return from z array allocation in runtime.
5350 { PreserveReexecuteState preexecs(this);
5351 jvms()->set_should_reexecute(true);
5352
5353 Node* x_start = array_element_address(x, intcon(0), x_elem);
5354 Node* y_start = array_element_address(y, intcon(0), y_elem);
5355 // 'x_start' points to x array + scaled xlen
5356 // 'y_start' points to y array + scaled ylen
5357
5358 // Allocate the result array
5359 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5360 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5361 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5362
5363 IdealKit ideal(this);
5364
5365 #define __ ideal.
5366 Node* one = __ ConI(1);
5367 Node* zero = __ ConI(0);
5368 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5369 __ set(need_alloc, zero);
5370 __ set(z_alloc, z);
5371 __ if_then(z, BoolTest::eq, null()); {
5372 __ increment (need_alloc, one);
5373 } __ else_(); {
5374 // Update graphKit memory and control from IdealKit.
5375 sync_kit(ideal);
5376 Node* zlen_arg = load_array_length(z);
5377 // Update IdealKit memory and control from graphKit.
5378 __ sync_kit(this);
5379 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5380 __ increment (need_alloc, one);
5381 } __ end_if();
5382 } __ end_if();
5383
5384 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5385 // Update graphKit memory and control from IdealKit.
5386 sync_kit(ideal);
5387 Node * narr = new_array(klass_node, zlen, 1);
5388 // Update IdealKit memory and control from graphKit.
5389 __ sync_kit(this);
5390 __ set(z_alloc, narr);
5391 } __ end_if();
5392
5393 sync_kit(ideal);
5394 z = __ value(z_alloc);
5395 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5396 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5397 // Final sync IdealKit and GraphKit.
5398 final_sync(ideal);
5399 #undef __
5400
5401 Node* z_start = array_element_address(z, intcon(0), T_INT);
5402
5403 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5404 OptoRuntime::multiplyToLen_Type(),
5405 stubAddr, stubName, TypePtr::BOTTOM,
5406 x_start, xlen, y_start, ylen, z_start, zlen);
5407 } // original reexecute is set back here
5408
5409 C->set_has_split_ifs(true); // Has chance for split-if optimization
5410 set_result(z);
5411 return true;
5412 }
5413
5414 //-------------inline_squareToLen------------------------------------
5415 bool LibraryCallKit::inline_squareToLen() {
5416 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5417
5418 address stubAddr = StubRoutines::squareToLen();
5419 if (stubAddr == NULL) {
5420 return false; // Intrinsic's stub is not implemented on this platform
5421 }
5422 const char* stubName = "squareToLen";
5423
5424 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5425
5426 Node* x = argument(0);
5427 Node* len = argument(1);
5428 Node* z = argument(2);
5429 Node* zlen = argument(3);
5430
5431 const Type* x_type = x->Value(&_gvn);
5432 const Type* z_type = z->Value(&_gvn);
5433 const TypeAryPtr* top_x = x_type->isa_aryptr();
5434 const TypeAryPtr* top_z = z_type->isa_aryptr();
5435 if (top_x == NULL || top_x->klass() == NULL ||
5436 top_z == NULL || top_z->klass() == NULL) {
5437 // failed array check
5438 return false;
5439 }
5440
5441 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5442 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5443 if (x_elem != T_INT || z_elem != T_INT) {
5444 return false;
5445 }
5446
5447
5448 Node* x_start = array_element_address(x, intcon(0), x_elem);
5449 Node* z_start = array_element_address(z, intcon(0), z_elem);
5450
5451 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5452 OptoRuntime::squareToLen_Type(),
5453 stubAddr, stubName, TypePtr::BOTTOM,
5454 x_start, len, z_start, zlen);
5455
5456 set_result(z);
5457 return true;
5458 }
5459
5460 //-------------inline_mulAdd------------------------------------------
5461 bool LibraryCallKit::inline_mulAdd() {
5462 assert(UseMulAddIntrinsic, "not implemented on this platform");
5463
5464 address stubAddr = StubRoutines::mulAdd();
5465 if (stubAddr == NULL) {
5466 return false; // Intrinsic's stub is not implemented on this platform
5467 }
5468 const char* stubName = "mulAdd";
5469
5470 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5471
5472 Node* out = argument(0);
5473 Node* in = argument(1);
5474 Node* offset = argument(2);
5475 Node* len = argument(3);
5476 Node* k = argument(4);
5477
5478 const Type* out_type = out->Value(&_gvn);
5479 const Type* in_type = in->Value(&_gvn);
5480 const TypeAryPtr* top_out = out_type->isa_aryptr();
5481 const TypeAryPtr* top_in = in_type->isa_aryptr();
5482 if (top_out == NULL || top_out->klass() == NULL ||
5483 top_in == NULL || top_in->klass() == NULL) {
5484 // failed array check
5485 return false;
5486 }
5487
5488 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5489 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5490 if (out_elem != T_INT || in_elem != T_INT) {
5491 return false;
5492 }
5493
5494 Node* outlen = load_array_length(out);
5495 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5496 Node* out_start = array_element_address(out, intcon(0), out_elem);
5497 Node* in_start = array_element_address(in, intcon(0), in_elem);
5498
5499 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5500 OptoRuntime::mulAdd_Type(),
5501 stubAddr, stubName, TypePtr::BOTTOM,
5502 out_start,in_start, new_offset, len, k);
5503 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5504 set_result(result);
5505 return true;
5506 }
5507
5508 //-------------inline_montgomeryMultiply-----------------------------------
5509 bool LibraryCallKit::inline_montgomeryMultiply() {
5510 address stubAddr = StubRoutines::montgomeryMultiply();
5511 if (stubAddr == NULL) {
5512 return false; // Intrinsic's stub is not implemented on this platform
5513 }
5514
5515 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5516 const char* stubName = "montgomery_multiply";
5517
5518 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5519
5520 Node* a = argument(0);
5521 Node* b = argument(1);
5522 Node* n = argument(2);
5523 Node* len = argument(3);
5524 Node* inv = argument(4);
5525 Node* m = argument(6);
5526
5527 const Type* a_type = a->Value(&_gvn);
5528 const TypeAryPtr* top_a = a_type->isa_aryptr();
5529 const Type* b_type = b->Value(&_gvn);
5530 const TypeAryPtr* top_b = b_type->isa_aryptr();
5531 const Type* n_type = a->Value(&_gvn);
5532 const TypeAryPtr* top_n = n_type->isa_aryptr();
5533 const Type* m_type = a->Value(&_gvn);
5534 const TypeAryPtr* top_m = m_type->isa_aryptr();
5535 if (top_a == NULL || top_a->klass() == NULL ||
5536 top_b == NULL || top_b->klass() == NULL ||
5537 top_n == NULL || top_n->klass() == NULL ||
5538 top_m == NULL || top_m->klass() == NULL) {
5539 // failed array check
5540 return false;
5541 }
5542
5543 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5544 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5545 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5546 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5547 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5548 return false;
5549 }
5550
5551 // Make the call
5552 {
5553 Node* a_start = array_element_address(a, intcon(0), a_elem);
5554 Node* b_start = array_element_address(b, intcon(0), b_elem);
5555 Node* n_start = array_element_address(n, intcon(0), n_elem);
5556 Node* m_start = array_element_address(m, intcon(0), m_elem);
5557
5558 Node* call = make_runtime_call(RC_LEAF,
5559 OptoRuntime::montgomeryMultiply_Type(),
5560 stubAddr, stubName, TypePtr::BOTTOM,
5561 a_start, b_start, n_start, len, inv, top(),
5562 m_start);
5563 set_result(m);
5564 }
5565
5566 return true;
5567 }
5568
5569 bool LibraryCallKit::inline_montgomerySquare() {
5570 address stubAddr = StubRoutines::montgomerySquare();
5571 if (stubAddr == NULL) {
5572 return false; // Intrinsic's stub is not implemented on this platform
5573 }
5574
5575 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5576 const char* stubName = "montgomery_square";
5577
5578 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5579
5580 Node* a = argument(0);
5581 Node* n = argument(1);
5582 Node* len = argument(2);
5583 Node* inv = argument(3);
5584 Node* m = argument(5);
5585
5586 const Type* a_type = a->Value(&_gvn);
5587 const TypeAryPtr* top_a = a_type->isa_aryptr();
5588 const Type* n_type = a->Value(&_gvn);
5589 const TypeAryPtr* top_n = n_type->isa_aryptr();
5590 const Type* m_type = a->Value(&_gvn);
5591 const TypeAryPtr* top_m = m_type->isa_aryptr();
5592 if (top_a == NULL || top_a->klass() == NULL ||
5593 top_n == NULL || top_n->klass() == NULL ||
5594 top_m == NULL || top_m->klass() == NULL) {
5595 // failed array check
5596 return false;
5597 }
5598
5599 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5600 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5601 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5602 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5603 return false;
5604 }
5605
5606 // Make the call
5607 {
5608 Node* a_start = array_element_address(a, intcon(0), a_elem);
5609 Node* n_start = array_element_address(n, intcon(0), n_elem);
5610 Node* m_start = array_element_address(m, intcon(0), m_elem);
5611
5612 Node* call = make_runtime_call(RC_LEAF,
5613 OptoRuntime::montgomerySquare_Type(),
5614 stubAddr, stubName, TypePtr::BOTTOM,
5615 a_start, n_start, len, inv, top(),
5616 m_start);
5617 set_result(m);
5618 }
5619
5620 return true;
5621 }
5622
5623 //-------------inline_vectorizedMismatch------------------------------
5624 bool LibraryCallKit::inline_vectorizedMismatch() {
5625 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5626
5627 address stubAddr = StubRoutines::vectorizedMismatch();
5628 if (stubAddr == NULL) {
5629 return false; // Intrinsic's stub is not implemented on this platform
5630 }
5631 const char* stubName = "vectorizedMismatch";
5632 int size_l = callee()->signature()->size();
5633 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5634
5635 Node* obja = argument(0);
5636 Node* aoffset = argument(1);
5637 Node* objb = argument(3);
5638 Node* boffset = argument(4);
5639 Node* length = argument(6);
5640 Node* scale = argument(7);
5641
5642 const Type* a_type = obja->Value(&_gvn);
5643 const Type* b_type = objb->Value(&_gvn);
5644 const TypeAryPtr* top_a = a_type->isa_aryptr();
5645 const TypeAryPtr* top_b = b_type->isa_aryptr();
5646 if (top_a == NULL || top_a->klass() == NULL ||
5647 top_b == NULL || top_b->klass() == NULL) {
5648 // failed array check
5649 return false;
5650 }
5651
5652 Node* call;
5653 jvms()->set_should_reexecute(true);
5654
5655 Node* obja_adr = make_unsafe_address(obja, aoffset);
5656 Node* objb_adr = make_unsafe_address(objb, boffset);
5657
5658 call = make_runtime_call(RC_LEAF,
5659 OptoRuntime::vectorizedMismatch_Type(),
5660 stubAddr, stubName, TypePtr::BOTTOM,
5661 obja_adr, objb_adr, length, scale);
5662
5663 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5664 set_result(result);
5665 return true;
5666 }
5667
5668 /**
5669 * Calculate CRC32 for byte.
5670 * int java.util.zip.CRC32.update(int crc, int b)
5671 */
5672 bool LibraryCallKit::inline_updateCRC32() {
5673 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5674 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5675 // no receiver since it is static method
5676 Node* crc = argument(0); // type: int
5677 Node* b = argument(1); // type: int
5678
5679 /*
5680 * int c = ~ crc;
5681 * b = timesXtoThe32[(b ^ c) & 0xFF];
5682 * b = b ^ (c >>> 8);
5683 * crc = ~b;
5684 */
5685
5686 Node* M1 = intcon(-1);
5687 crc = _gvn.transform(new XorINode(crc, M1));
5688 Node* result = _gvn.transform(new XorINode(crc, b));
5689 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5690
5691 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5692 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5693 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5694 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5695
5696 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5697 result = _gvn.transform(new XorINode(crc, result));
5698 result = _gvn.transform(new XorINode(result, M1));
5699 set_result(result);
5700 return true;
5701 }
5702
5703 /**
5704 * Calculate CRC32 for byte[] array.
5705 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5706 */
5707 bool LibraryCallKit::inline_updateBytesCRC32() {
5708 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5709 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5710 // no receiver since it is static method
5711 Node* crc = argument(0); // type: int
5712 Node* src = argument(1); // type: oop
5713 Node* offset = argument(2); // type: int
5714 Node* length = argument(3); // type: int
5715
5716 const Type* src_type = src->Value(&_gvn);
5717 const TypeAryPtr* top_src = src_type->isa_aryptr();
5718 if (top_src == NULL || top_src->klass() == NULL) {
5719 // failed array check
5720 return false;
5721 }
5722
5723 // Figure out the size and type of the elements we will be copying.
5724 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5725 if (src_elem != T_BYTE) {
5726 return false;
5727 }
5728
5729 // 'src_start' points to src array + scaled offset
5730 Node* src_start = array_element_address(src, offset, src_elem);
5731
5732 // We assume that range check is done by caller.
5733 // TODO: generate range check (offset+length < src.length) in debug VM.
5734
5735 // Call the stub.
5736 address stubAddr = StubRoutines::updateBytesCRC32();
5737 const char *stubName = "updateBytesCRC32";
5738
5739 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5740 stubAddr, stubName, TypePtr::BOTTOM,
5741 crc, src_start, length);
5742 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5743 set_result(result);
5744 return true;
5745 }
5746
5747 /**
5748 * Calculate CRC32 for ByteBuffer.
5749 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5750 */
5751 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5752 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5753 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5754 // no receiver since it is static method
5755 Node* crc = argument(0); // type: int
5756 Node* src = argument(1); // type: long
5757 Node* offset = argument(3); // type: int
5758 Node* length = argument(4); // type: int
5759
5760 src = ConvL2X(src); // adjust Java long to machine word
5761 Node* base = _gvn.transform(new CastX2PNode(src));
5762 offset = ConvI2X(offset);
5763
5764 // 'src_start' points to src array + scaled offset
5765 Node* src_start = basic_plus_adr(top(), base, offset);
5766
5767 // Call the stub.
5768 address stubAddr = StubRoutines::updateBytesCRC32();
5769 const char *stubName = "updateBytesCRC32";
5770
5771 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5772 stubAddr, stubName, TypePtr::BOTTOM,
5773 crc, src_start, length);
5774 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5775 set_result(result);
5776 return true;
5777 }
5778
5779 //------------------------------get_table_from_crc32c_class-----------------------
5780 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5781 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5782 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5783
5784 return table;
5785 }
5786
5787 //------------------------------inline_updateBytesCRC32C-----------------------
5788 //
5789 // Calculate CRC32C for byte[] array.
5790 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5791 //
5792 bool LibraryCallKit::inline_updateBytesCRC32C() {
5793 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5794 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5795 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5796 // no receiver since it is a static method
5797 Node* crc = argument(0); // type: int
5798 Node* src = argument(1); // type: oop
5799 Node* offset = argument(2); // type: int
5800 Node* end = argument(3); // type: int
5801
5802 Node* length = _gvn.transform(new SubINode(end, offset));
5803
5804 const Type* src_type = src->Value(&_gvn);
5805 const TypeAryPtr* top_src = src_type->isa_aryptr();
5806 if (top_src == NULL || top_src->klass() == NULL) {
5807 // failed array check
5808 return false;
5809 }
5810
5811 // Figure out the size and type of the elements we will be copying.
5812 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5813 if (src_elem != T_BYTE) {
5814 return false;
5815 }
5816
5817 // 'src_start' points to src array + scaled offset
5818 Node* src_start = array_element_address(src, offset, src_elem);
5819
5820 // static final int[] byteTable in class CRC32C
5821 Node* table = get_table_from_crc32c_class(callee()->holder());
5822 Node* table_start = array_element_address(table, intcon(0), T_INT);
5823
5824 // We assume that range check is done by caller.
5825 // TODO: generate range check (offset+length < src.length) in debug VM.
5826
5827 // Call the stub.
5828 address stubAddr = StubRoutines::updateBytesCRC32C();
5829 const char *stubName = "updateBytesCRC32C";
5830
5831 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5832 stubAddr, stubName, TypePtr::BOTTOM,
5833 crc, src_start, length, table_start);
5834 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5835 set_result(result);
5836 return true;
5837 }
5838
5839 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5840 //
5841 // Calculate CRC32C for DirectByteBuffer.
5842 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5843 //
5844 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5845 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5846 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5847 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5848 // no receiver since it is a static method
5849 Node* crc = argument(0); // type: int
5850 Node* src = argument(1); // type: long
5851 Node* offset = argument(3); // type: int
5852 Node* end = argument(4); // type: int
5853
5854 Node* length = _gvn.transform(new SubINode(end, offset));
5855
5856 src = ConvL2X(src); // adjust Java long to machine word
5857 Node* base = _gvn.transform(new CastX2PNode(src));
5858 offset = ConvI2X(offset);
5859
5860 // 'src_start' points to src array + scaled offset
5861 Node* src_start = basic_plus_adr(top(), base, offset);
5862
5863 // static final int[] byteTable in class CRC32C
5864 Node* table = get_table_from_crc32c_class(callee()->holder());
5865 Node* table_start = array_element_address(table, intcon(0), T_INT);
5866
5867 // Call the stub.
5868 address stubAddr = StubRoutines::updateBytesCRC32C();
5869 const char *stubName = "updateBytesCRC32C";
5870
5871 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5872 stubAddr, stubName, TypePtr::BOTTOM,
5873 crc, src_start, length, table_start);
5874 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5875 set_result(result);
5876 return true;
5877 }
5878
5879 //------------------------------inline_updateBytesAdler32----------------------
5880 //
5881 // Calculate Adler32 checksum for byte[] array.
5882 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5883 //
5884 bool LibraryCallKit::inline_updateBytesAdler32() {
5885 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5886 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5887 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5888 // no receiver since it is static method
5889 Node* crc = argument(0); // type: int
5890 Node* src = argument(1); // type: oop
5891 Node* offset = argument(2); // type: int
5892 Node* length = argument(3); // type: int
5893
5894 const Type* src_type = src->Value(&_gvn);
5895 const TypeAryPtr* top_src = src_type->isa_aryptr();
5896 if (top_src == NULL || top_src->klass() == NULL) {
5897 // failed array check
5898 return false;
5899 }
5900
5901 // Figure out the size and type of the elements we will be copying.
5902 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5903 if (src_elem != T_BYTE) {
5904 return false;
5905 }
5906
5907 // 'src_start' points to src array + scaled offset
5908 Node* src_start = array_element_address(src, offset, src_elem);
5909
5910 // We assume that range check is done by caller.
5911 // TODO: generate range check (offset+length < src.length) in debug VM.
5912
5913 // Call the stub.
5914 address stubAddr = StubRoutines::updateBytesAdler32();
5915 const char *stubName = "updateBytesAdler32";
5916
5917 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5918 stubAddr, stubName, TypePtr::BOTTOM,
5919 crc, src_start, length);
5920 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5921 set_result(result);
5922 return true;
5923 }
5924
5925 //------------------------------inline_updateByteBufferAdler32---------------
5926 //
5927 // Calculate Adler32 checksum for DirectByteBuffer.
5928 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5929 //
5930 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5931 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5932 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5933 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5934 // no receiver since it is static method
5935 Node* crc = argument(0); // type: int
5936 Node* src = argument(1); // type: long
5937 Node* offset = argument(3); // type: int
5938 Node* length = argument(4); // type: int
5939
5940 src = ConvL2X(src); // adjust Java long to machine word
5941 Node* base = _gvn.transform(new CastX2PNode(src));
5942 offset = ConvI2X(offset);
5943
5944 // 'src_start' points to src array + scaled offset
5945 Node* src_start = basic_plus_adr(top(), base, offset);
5946
5947 // Call the stub.
5948 address stubAddr = StubRoutines::updateBytesAdler32();
5949 const char *stubName = "updateBytesAdler32";
5950
5951 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5952 stubAddr, stubName, TypePtr::BOTTOM,
5953 crc, src_start, length);
5954
5955 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5956 set_result(result);
5957 return true;
5958 }
5959
5960 //----------------------------inline_reference_get----------------------------
5961 // public T java.lang.ref.Reference.get();
5962 bool LibraryCallKit::inline_reference_get() {
5963 const int referent_offset = java_lang_ref_Reference::referent_offset;
5964 guarantee(referent_offset > 0, "should have already been set");
5965
5966 // Get the argument:
5967 Node* reference_obj = null_check_receiver();
5968 if (stopped()) return true;
5969
5970 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5971
5972 ciInstanceKlass* klass = env()->Object_klass();
5973 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5974
5975 Node* no_ctrl = NULL;
5976 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5977
5978 // Use the pre-barrier to record the value in the referent field
5979 pre_barrier(false /* do_load */,
5980 control(),
5981 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5982 result /* pre_val */,
5983 T_OBJECT);
5984
5985 // Add memory barrier to prevent commoning reads from this field
5986 // across safepoint since GC can change its value.
5987 insert_mem_bar(Op_MemBarCPUOrder);
5988
5989 set_result(result);
5990 return true;
5991 }
5992
5993
5994 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5995 bool is_exact=true, bool is_static=false,
5996 ciInstanceKlass * fromKls=NULL) {
5997 if (fromKls == NULL) {
5998 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5999 assert(tinst != NULL, "obj is null");
6000 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6001 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6002 fromKls = tinst->klass()->as_instance_klass();
6003 } else {
6004 assert(is_static, "only for static field access");
6005 }
6006 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6007 ciSymbol::make(fieldTypeString),
6008 is_static);
6009
6010 assert (field != NULL, "undefined field");
6011 if (field == NULL) return (Node *) NULL;
6012
6013 if (is_static) {
6014 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6015 fromObj = makecon(tip);
6016 }
6017
6018 // Next code copied from Parse::do_get_xxx():
6019
6020 // Compute address and memory type.
6021 int offset = field->offset_in_bytes();
6022 bool is_vol = field->is_volatile();
6023 ciType* field_klass = field->type();
6024 assert(field_klass->is_loaded(), "should be loaded");
6025 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6026 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6027 BasicType bt = field->layout_type();
6028
6029 // Build the resultant type of the load
6030 const Type *type;
6031 if (bt == T_OBJECT) {
6032 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6033 } else {
6034 type = Type::get_const_basic_type(bt);
6035 }
6036
6037 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6038 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6039 }
6040 // Build the load.
6041 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6042 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6043 // If reference is volatile, prevent following memory ops from
6044 // floating up past the volatile read. Also prevents commoning
6045 // another volatile read.
6046 if (is_vol) {
6047 // Memory barrier includes bogus read of value to force load BEFORE membar
6048 insert_mem_bar(Op_MemBarAcquire, loadedField);
6049 }
6050 return loadedField;
6051 }
6052
6053 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6054 bool is_exact = true, bool is_static = false,
6055 ciInstanceKlass * fromKls = NULL) {
6056 if (fromKls == NULL) {
6057 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6058 assert(tinst != NULL, "obj is null");
6059 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6060 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6061 fromKls = tinst->klass()->as_instance_klass();
6062 }
6063 else {
6064 assert(is_static, "only for static field access");
6065 }
6066 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6067 ciSymbol::make(fieldTypeString),
6068 is_static);
6069
6070 assert(field != NULL, "undefined field");
6071 assert(!field->is_volatile(), "not defined for volatile fields");
6072
6073 if (is_static) {
6074 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6075 fromObj = makecon(tip);
6076 }
6077
6078 // Next code copied from Parse::do_get_xxx():
6079
6080 // Compute address and memory type.
6081 int offset = field->offset_in_bytes();
6082 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6083
6084 return adr;
6085 }
6086
6087 //------------------------------inline_aescrypt_Block-----------------------
6088 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6089 address stubAddr = NULL;
6090 const char *stubName;
6091 assert(UseAES, "need AES instruction support");
6092
6093 switch(id) {
6094 case vmIntrinsics::_aescrypt_encryptBlock:
6095 stubAddr = StubRoutines::aescrypt_encryptBlock();
6096 stubName = "aescrypt_encryptBlock";
6097 break;
6098 case vmIntrinsics::_aescrypt_decryptBlock:
6099 stubAddr = StubRoutines::aescrypt_decryptBlock();
6100 stubName = "aescrypt_decryptBlock";
6101 break;
6102 }
6103 if (stubAddr == NULL) return false;
6104
6105 Node* aescrypt_object = argument(0);
6106 Node* src = argument(1);
6107 Node* src_offset = argument(2);
6108 Node* dest = argument(3);
6109 Node* dest_offset = argument(4);
6110
6111 // (1) src and dest are arrays.
6112 const Type* src_type = src->Value(&_gvn);
6113 const Type* dest_type = dest->Value(&_gvn);
6114 const TypeAryPtr* top_src = src_type->isa_aryptr();
6115 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6116 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6117
6118 // for the quick and dirty code we will skip all the checks.
6119 // we are just trying to get the call to be generated.
6120 Node* src_start = src;
6121 Node* dest_start = dest;
6122 if (src_offset != NULL || dest_offset != NULL) {
6123 assert(src_offset != NULL && dest_offset != NULL, "");
6124 src_start = array_element_address(src, src_offset, T_BYTE);
6125 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6126 }
6127
6128 // now need to get the start of its expanded key array
6129 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6130 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6131 if (k_start == NULL) return false;
6132
6133 if (Matcher::pass_original_key_for_aes()) {
6134 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6135 // compatibility issues between Java key expansion and SPARC crypto instructions
6136 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6137 if (original_k_start == NULL) return false;
6138
6139 // Call the stub.
6140 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6141 stubAddr, stubName, TypePtr::BOTTOM,
6142 src_start, dest_start, k_start, original_k_start);
6143 } else {
6144 // Call the stub.
6145 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6146 stubAddr, stubName, TypePtr::BOTTOM,
6147 src_start, dest_start, k_start);
6148 }
6149
6150 return true;
6151 }
6152
6153 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6154 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6155 address stubAddr = NULL;
6156 const char *stubName = NULL;
6157
6158 assert(UseAES, "need AES instruction support");
6159
6160 switch(id) {
6161 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6162 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6163 stubName = "cipherBlockChaining_encryptAESCrypt";
6164 break;
6165 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6166 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6167 stubName = "cipherBlockChaining_decryptAESCrypt";
6168 break;
6169 }
6170 if (stubAddr == NULL) return false;
6171
6172 Node* cipherBlockChaining_object = argument(0);
6173 Node* src = argument(1);
6174 Node* src_offset = argument(2);
6175 Node* len = argument(3);
6176 Node* dest = argument(4);
6177 Node* dest_offset = argument(5);
6178
6179 // (1) src and dest are arrays.
6180 const Type* src_type = src->Value(&_gvn);
6181 const Type* dest_type = dest->Value(&_gvn);
6182 const TypeAryPtr* top_src = src_type->isa_aryptr();
6183 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6184 assert (top_src != NULL && top_src->klass() != NULL
6185 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6186
6187 // checks are the responsibility of the caller
6188 Node* src_start = src;
6189 Node* dest_start = dest;
6190 if (src_offset != NULL || dest_offset != NULL) {
6191 assert(src_offset != NULL && dest_offset != NULL, "");
6192 src_start = array_element_address(src, src_offset, T_BYTE);
6193 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6194 }
6195
6196 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6197 // (because of the predicated logic executed earlier).
6198 // so we cast it here safely.
6199 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6200
6201 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6202 if (embeddedCipherObj == NULL) return false;
6203
6204 // cast it to what we know it will be at runtime
6205 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6206 assert(tinst != NULL, "CBC obj is null");
6207 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6208 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6209 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6210
6211 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6212 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6213 const TypeOopPtr* xtype = aklass->as_instance_type();
6214 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6215 aescrypt_object = _gvn.transform(aescrypt_object);
6216
6217 // we need to get the start of the aescrypt_object's expanded key array
6218 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6219 if (k_start == NULL) return false;
6220
6221 // similarly, get the start address of the r vector
6222 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6223 if (objRvec == NULL) return false;
6224 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6225
6226 Node* cbcCrypt;
6227 if (Matcher::pass_original_key_for_aes()) {
6228 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6229 // compatibility issues between Java key expansion and SPARC crypto instructions
6230 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6231 if (original_k_start == NULL) return false;
6232
6233 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6234 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6235 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6236 stubAddr, stubName, TypePtr::BOTTOM,
6237 src_start, dest_start, k_start, r_start, len, original_k_start);
6238 } else {
6239 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6240 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6241 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6242 stubAddr, stubName, TypePtr::BOTTOM,
6243 src_start, dest_start, k_start, r_start, len);
6244 }
6245
6246 // return cipher length (int)
6247 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6248 set_result(retvalue);
6249 return true;
6250 }
6251
6252 //------------------------------inline_counterMode_AESCrypt-----------------------
6253 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6254 assert(UseAES, "need AES instruction support");
6255 if (!UseAESCTRIntrinsics) return false;
6256
6257 address stubAddr = NULL;
6258 const char *stubName = NULL;
6259 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6260 stubAddr = StubRoutines::counterMode_AESCrypt();
6261 stubName = "counterMode_AESCrypt";
6262 }
6263 if (stubAddr == NULL) return false;
6264
6265 Node* counterMode_object = argument(0);
6266 Node* src = argument(1);
6267 Node* src_offset = argument(2);
6268 Node* len = argument(3);
6269 Node* dest = argument(4);
6270 Node* dest_offset = argument(5);
6271
6272 // (1) src and dest are arrays.
6273 const Type* src_type = src->Value(&_gvn);
6274 const Type* dest_type = dest->Value(&_gvn);
6275 const TypeAryPtr* top_src = src_type->isa_aryptr();
6276 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6277 assert(top_src != NULL && top_src->klass() != NULL &&
6278 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6279
6280 // checks are the responsibility of the caller
6281 Node* src_start = src;
6282 Node* dest_start = dest;
6283 if (src_offset != NULL || dest_offset != NULL) {
6284 assert(src_offset != NULL && dest_offset != NULL, "");
6285 src_start = array_element_address(src, src_offset, T_BYTE);
6286 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6287 }
6288
6289 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6290 // (because of the predicated logic executed earlier).
6291 // so we cast it here safely.
6292 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6293 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6294 if (embeddedCipherObj == NULL) return false;
6295 // cast it to what we know it will be at runtime
6296 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6297 assert(tinst != NULL, "CTR obj is null");
6298 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6299 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6300 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6301 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6302 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6303 const TypeOopPtr* xtype = aklass->as_instance_type();
6304 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6305 aescrypt_object = _gvn.transform(aescrypt_object);
6306 // we need to get the start of the aescrypt_object's expanded key array
6307 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6308 if (k_start == NULL) return false;
6309 // similarly, get the start address of the r vector
6310 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6311 if (obj_counter == NULL) return false;
6312 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6313
6314 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6315 if (saved_encCounter == NULL) return false;
6316 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6317 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6318
6319 Node* ctrCrypt;
6320 if (Matcher::pass_original_key_for_aes()) {
6321 // no SPARC version for AES/CTR intrinsics now.
6322 return false;
6323 }
6324 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6325 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6326 OptoRuntime::counterMode_aescrypt_Type(),
6327 stubAddr, stubName, TypePtr::BOTTOM,
6328 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6329
6330 // return cipher length (int)
6331 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6332 set_result(retvalue);
6333 return true;
6334 }
6335
6336 //------------------------------get_key_start_from_aescrypt_object-----------------------
6337 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6338 #if defined(PPC64) || defined(S390)
6339 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6340 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6341 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6342 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6343 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6344 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6345 if (objSessionK == NULL) {
6346 return (Node *) NULL;
6347 }
6348 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6349 #else
6350 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6351 #endif // PPC64
6352 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6353 if (objAESCryptKey == NULL) return (Node *) NULL;
6354
6355 // now have the array, need to get the start address of the K array
6356 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6357 return k_start;
6358 }
6359
6360 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6361 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6362 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6363 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6364 if (objAESCryptKey == NULL) return (Node *) NULL;
6365
6366 // now have the array, need to get the start address of the lastKey array
6367 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6368 return original_k_start;
6369 }
6370
6371 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6372 // Return node representing slow path of predicate check.
6373 // the pseudo code we want to emulate with this predicate is:
6374 // for encryption:
6375 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6376 // for decryption:
6377 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6378 // note cipher==plain is more conservative than the original java code but that's OK
6379 //
6380 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6381 // The receiver was checked for NULL already.
6382 Node* objCBC = argument(0);
6383
6384 // Load embeddedCipher field of CipherBlockChaining object.
6385 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6386
6387 // get AESCrypt klass for instanceOf check
6388 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6389 // will have same classloader as CipherBlockChaining object
6390 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6391 assert(tinst != NULL, "CBCobj is null");
6392 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6393
6394 // we want to do an instanceof comparison against the AESCrypt class
6395 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6396 if (!klass_AESCrypt->is_loaded()) {
6397 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6398 Node* ctrl = control();
6399 set_control(top()); // no regular fast path
6400 return ctrl;
6401 }
6402 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6403
6404 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6405 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6406 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6407
6408 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6409
6410 // for encryption, we are done
6411 if (!decrypting)
6412 return instof_false; // even if it is NULL
6413
6414 // for decryption, we need to add a further check to avoid
6415 // taking the intrinsic path when cipher and plain are the same
6416 // see the original java code for why.
6417 RegionNode* region = new RegionNode(3);
6418 region->init_req(1, instof_false);
6419 Node* src = argument(1);
6420 Node* dest = argument(4);
6421 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6422 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6423 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6424 region->init_req(2, src_dest_conjoint);
6425
6426 record_for_igvn(region);
6427 return _gvn.transform(region);
6428 }
6429
6430 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6431 // Return node representing slow path of predicate check.
6432 // the pseudo code we want to emulate with this predicate is:
6433 // for encryption:
6434 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6435 // for decryption:
6436 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6437 // note cipher==plain is more conservative than the original java code but that's OK
6438 //
6439
6440 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6441 // The receiver was checked for NULL already.
6442 Node* objCTR = argument(0);
6443
6444 // Load embeddedCipher field of CipherBlockChaining object.
6445 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6446
6447 // get AESCrypt klass for instanceOf check
6448 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6449 // will have same classloader as CipherBlockChaining object
6450 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6451 assert(tinst != NULL, "CTRobj is null");
6452 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6453
6454 // we want to do an instanceof comparison against the AESCrypt class
6455 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6456 if (!klass_AESCrypt->is_loaded()) {
6457 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6458 Node* ctrl = control();
6459 set_control(top()); // no regular fast path
6460 return ctrl;
6461 }
6462
6463 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6464 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6465 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6466 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6467 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6468
6469 return instof_false; // even if it is NULL
6470 }
6471
6472 //------------------------------inline_ghash_processBlocks
6473 bool LibraryCallKit::inline_ghash_processBlocks() {
6474 address stubAddr;
6475 const char *stubName;
6476 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6477
6478 stubAddr = StubRoutines::ghash_processBlocks();
6479 stubName = "ghash_processBlocks";
6480
6481 Node* data = argument(0);
6482 Node* offset = argument(1);
6483 Node* len = argument(2);
6484 Node* state = argument(3);
6485 Node* subkeyH = argument(4);
6486
6487 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6488 assert(state_start, "state is NULL");
6489 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6490 assert(subkeyH_start, "subkeyH is NULL");
6491 Node* data_start = array_element_address(data, offset, T_BYTE);
6492 assert(data_start, "data is NULL");
6493
6494 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6495 OptoRuntime::ghash_processBlocks_Type(),
6496 stubAddr, stubName, TypePtr::BOTTOM,
6497 state_start, subkeyH_start, data_start, len);
6498 return true;
6499 }
6500
6501 //------------------------------inline_sha_implCompress-----------------------
6502 //
6503 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6504 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6505 //
6506 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6507 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6508 //
6509 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6510 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6511 //
6512 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6513 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6514
6515 Node* sha_obj = argument(0);
6516 Node* src = argument(1); // type oop
6517 Node* ofs = argument(2); // type int
6518
6519 const Type* src_type = src->Value(&_gvn);
6520 const TypeAryPtr* top_src = src_type->isa_aryptr();
6521 if (top_src == NULL || top_src->klass() == NULL) {
6522 // failed array check
6523 return false;
6524 }
6525 // Figure out the size and type of the elements we will be copying.
6526 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6527 if (src_elem != T_BYTE) {
6528 return false;
6529 }
6530 // 'src_start' points to src array + offset
6531 Node* src_start = array_element_address(src, ofs, src_elem);
6532 Node* state = NULL;
6533 address stubAddr;
6534 const char *stubName;
6535
6536 switch(id) {
6537 case vmIntrinsics::_sha_implCompress:
6538 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6539 state = get_state_from_sha_object(sha_obj);
6540 stubAddr = StubRoutines::sha1_implCompress();
6541 stubName = "sha1_implCompress";
6542 break;
6543 case vmIntrinsics::_sha2_implCompress:
6544 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6545 state = get_state_from_sha_object(sha_obj);
6546 stubAddr = StubRoutines::sha256_implCompress();
6547 stubName = "sha256_implCompress";
6548 break;
6549 case vmIntrinsics::_sha5_implCompress:
6550 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6551 state = get_state_from_sha5_object(sha_obj);
6552 stubAddr = StubRoutines::sha512_implCompress();
6553 stubName = "sha512_implCompress";
6554 break;
6555 default:
6556 fatal_unexpected_iid(id);
6557 return false;
6558 }
6559 if (state == NULL) return false;
6560
6561 // Call the stub.
6562 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6563 stubAddr, stubName, TypePtr::BOTTOM,
6564 src_start, state);
6565
6566 return true;
6567 }
6568
6569 //------------------------------inline_digestBase_implCompressMB-----------------------
6570 //
6571 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6572 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6573 //
6574 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6575 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6576 "need SHA1/SHA256/SHA512 instruction support");
6577 assert((uint)predicate < 3, "sanity");
6578 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6579
6580 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6581 Node* src = argument(1); // byte[] array
6582 Node* ofs = argument(2); // type int
6583 Node* limit = argument(3); // type int
6584
6585 const Type* src_type = src->Value(&_gvn);
6586 const TypeAryPtr* top_src = src_type->isa_aryptr();
6587 if (top_src == NULL || top_src->klass() == NULL) {
6588 // failed array check
6589 return false;
6590 }
6591 // Figure out the size and type of the elements we will be copying.
6592 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6593 if (src_elem != T_BYTE) {
6594 return false;
6595 }
6596 // 'src_start' points to src array + offset
6597 Node* src_start = array_element_address(src, ofs, src_elem);
6598
6599 const char* klass_SHA_name = NULL;
6600 const char* stub_name = NULL;
6601 address stub_addr = NULL;
6602 bool long_state = false;
6603
6604 switch (predicate) {
6605 case 0:
6606 if (UseSHA1Intrinsics) {
6607 klass_SHA_name = "sun/security/provider/SHA";
6608 stub_name = "sha1_implCompressMB";
6609 stub_addr = StubRoutines::sha1_implCompressMB();
6610 }
6611 break;
6612 case 1:
6613 if (UseSHA256Intrinsics) {
6614 klass_SHA_name = "sun/security/provider/SHA2";
6615 stub_name = "sha256_implCompressMB";
6616 stub_addr = StubRoutines::sha256_implCompressMB();
6617 }
6618 break;
6619 case 2:
6620 if (UseSHA512Intrinsics) {
6621 klass_SHA_name = "sun/security/provider/SHA5";
6622 stub_name = "sha512_implCompressMB";
6623 stub_addr = StubRoutines::sha512_implCompressMB();
6624 long_state = true;
6625 }
6626 break;
6627 default:
6628 fatal("unknown SHA intrinsic predicate: %d", predicate);
6629 }
6630 if (klass_SHA_name != NULL) {
6631 // get DigestBase klass to lookup for SHA klass
6632 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6633 assert(tinst != NULL, "digestBase_obj is not instance???");
6634 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6635
6636 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6637 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6638 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6639 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6640 }
6641 return false;
6642 }
6643 //------------------------------inline_sha_implCompressMB-----------------------
6644 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6645 bool long_state, address stubAddr, const char *stubName,
6646 Node* src_start, Node* ofs, Node* limit) {
6647 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6648 const TypeOopPtr* xtype = aklass->as_instance_type();
6649 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6650 sha_obj = _gvn.transform(sha_obj);
6651
6652 Node* state;
6653 if (long_state) {
6654 state = get_state_from_sha5_object(sha_obj);
6655 } else {
6656 state = get_state_from_sha_object(sha_obj);
6657 }
6658 if (state == NULL) return false;
6659
6660 // Call the stub.
6661 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6662 OptoRuntime::digestBase_implCompressMB_Type(),
6663 stubAddr, stubName, TypePtr::BOTTOM,
6664 src_start, state, ofs, limit);
6665 // return ofs (int)
6666 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6667 set_result(result);
6668
6669 return true;
6670 }
6671
6672 //------------------------------get_state_from_sha_object-----------------------
6673 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6674 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6675 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6676 if (sha_state == NULL) return (Node *) NULL;
6677
6678 // now have the array, need to get the start address of the state array
6679 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6680 return state;
6681 }
6682
6683 //------------------------------get_state_from_sha5_object-----------------------
6684 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6685 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6686 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6687 if (sha_state == NULL) return (Node *) NULL;
6688
6689 // now have the array, need to get the start address of the state array
6690 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6691 return state;
6692 }
6693
6694 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6695 // Return node representing slow path of predicate check.
6696 // the pseudo code we want to emulate with this predicate is:
6697 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6698 //
6699 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6700 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6701 "need SHA1/SHA256/SHA512 instruction support");
6702 assert((uint)predicate < 3, "sanity");
6703
6704 // The receiver was checked for NULL already.
6705 Node* digestBaseObj = argument(0);
6706
6707 // get DigestBase klass for instanceOf check
6708 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6709 assert(tinst != NULL, "digestBaseObj is null");
6710 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6711
6712 const char* klass_SHA_name = NULL;
6713 switch (predicate) {
6714 case 0:
6715 if (UseSHA1Intrinsics) {
6716 // we want to do an instanceof comparison against the SHA class
6717 klass_SHA_name = "sun/security/provider/SHA";
6718 }
6719 break;
6720 case 1:
6721 if (UseSHA256Intrinsics) {
6722 // we want to do an instanceof comparison against the SHA2 class
6723 klass_SHA_name = "sun/security/provider/SHA2";
6724 }
6725 break;
6726 case 2:
6727 if (UseSHA512Intrinsics) {
6728 // we want to do an instanceof comparison against the SHA5 class
6729 klass_SHA_name = "sun/security/provider/SHA5";
6730 }
6731 break;
6732 default:
6733 fatal("unknown SHA intrinsic predicate: %d", predicate);
6734 }
6735
6736 ciKlass* klass_SHA = NULL;
6737 if (klass_SHA_name != NULL) {
6738 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6739 }
6740 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6741 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6742 Node* ctrl = control();
6743 set_control(top()); // no intrinsic path
6744 return ctrl;
6745 }
6746 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6747
6748 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6749 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6750 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6751 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6752
6753 return instof_false; // even if it is NULL
6754 }
6755
6756 //-------------inline_fma-----------------------------------
6757 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6758 Node *a = NULL;
6759 Node *b = NULL;
6760 Node *c = NULL;
6761 Node* result = NULL;
6762 switch (id) {
6763 case vmIntrinsics::_fmaD:
6764 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6765 // no receiver since it is static method
6766 a = round_double_node(argument(0));
6767 b = round_double_node(argument(2));
6768 c = round_double_node(argument(4));
6769 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6770 break;
6771 case vmIntrinsics::_fmaF:
6772 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6773 a = argument(0);
6774 b = argument(1);
6775 c = argument(2);
6776 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6777 break;
6778 default:
6779 fatal_unexpected_iid(id); break;
6780 }
6781 set_result(result);
6782 return true;
6783 }
6784
6785 bool LibraryCallKit::inline_profileBoolean() {
6786 Node* counts = argument(1);
6787 const TypeAryPtr* ary = NULL;
6788 ciArray* aobj = NULL;
6789 if (counts->is_Con()
6790 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6791 && (aobj = ary->const_oop()->as_array()) != NULL
6792 && (aobj->length() == 2)) {
6793 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6794 jint false_cnt = aobj->element_value(0).as_int();
6795 jint true_cnt = aobj->element_value(1).as_int();
6796
6797 if (C->log() != NULL) {
6798 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6799 false_cnt, true_cnt);
6800 }
6801
6802 if (false_cnt + true_cnt == 0) {
6803 // According to profile, never executed.
6804 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6805 Deoptimization::Action_reinterpret);
6806 return true;
6807 }
6808
6809 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6810 // is a number of each value occurrences.
6811 Node* result = argument(0);
6812 if (false_cnt == 0 || true_cnt == 0) {
6813 // According to profile, one value has been never seen.
6814 int expected_val = (false_cnt == 0) ? 1 : 0;
6815
6816 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6817 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6818
6819 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6820 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6821 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6822
6823 { // Slow path: uncommon trap for never seen value and then reexecute
6824 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6825 // the value has been seen at least once.
6826 PreserveJVMState pjvms(this);
6827 PreserveReexecuteState preexecs(this);
6828 jvms()->set_should_reexecute(true);
6829
6830 set_control(slow_path);
6831 set_i_o(i_o());
6832
6833 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6834 Deoptimization::Action_reinterpret);
6835 }
6836 // The guard for never seen value enables sharpening of the result and
6837 // returning a constant. It allows to eliminate branches on the same value
6838 // later on.
6839 set_control(fast_path);
6840 result = intcon(expected_val);
6841 }
6842 // Stop profiling.
6843 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6844 // By replacing method body with profile data (represented as ProfileBooleanNode
6845 // on IR level) we effectively disable profiling.
6846 // It enables full speed execution once optimized code is generated.
6847 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6848 C->record_for_igvn(profile);
6849 set_result(profile);
6850 return true;
6851 } else {
6852 // Continue profiling.
6853 // Profile data isn't available at the moment. So, execute method's bytecode version.
6854 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6855 // is compiled and counters aren't available since corresponding MethodHandle
6856 // isn't a compile-time constant.
6857 return false;
6858 }
6859 }
6860
6861 bool LibraryCallKit::inline_isCompileConstant() {
6862 Node* n = argument(0);
6863 set_result(n->is_Con() ? intcon(1) : intcon(0));
6864 return true;
6865 }